TTiti的学习笔记
首页 / 专业知识 / 40-References/Papers/gpt3 - GPT3/01_original.md

Language Models are Few-Shot Learners Tom B.

专业知识 · 40-References/Papers/gpt3 - GPT3/01_original.md

--- title: "Language Models are Few-Shot Learners Tom B." aliases: - "GPT3" - "arXiv:2005.14165" source: "https://arxiv.org/abs/2005.14165" arxiv: "2005.14165" created: 2026-07-16 type: paper-translation status: extraction-complete_translation-pending tags: - paper - ml - deep-learning - nlp


Language Models are Few-Shot Learners Tom B.

原文全文

Page 1

<a id="S0001"></a> Source: p.1 S0001

Language Models are Few-Shot Learners Tom B.

<a id="S0002"></a> Source: p.1 S0002

Brown∗ Benjamin Mann∗ Nick Ryder∗ Melanie Subbiah∗ Jared Kaplan† Prafulla Dhariwal Arvind Neelakantan Pranav Shyam Girish Sastry Amanda Askell Sandhini Agarwal Ariel Herbert-Voss Gretchen Krueger Tom Henighan Rewon Child Aditya Ramesh Daniel M.

<a id="S0003"></a> Source: p.1 S0003

Ziegler Jeffrey Wu Clemens Winter Christopher Hesse Mark Chen Eric Sigler Mateusz Litwin Scott Gray Benjamin Chess Jack Clark Christopher Berner Sam McCandlish Alec Radford Ilya Sutskever Dario Amodei OpenAI Abstract Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.

<a id="S0004"></a> Source: p.1 S0004

While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples.

<a id="S0005"></a> Source: p.1 S0005

By contrast, humans can generally perform a new language task from only a few examples or from simple instructions – something which current NLP systems still largely struggle to do.

<a id="S0006"></a> Source: p.1 S0006

Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art finetuning approaches.

<a id="S0007"></a> Source: p.1 S0007

Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting.

<a id="S0008"></a> Source: p.1 S0008

For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model.

<a id="S0009"></a> Source: p.1 S0009

GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic.

<a id="S0010"></a> Source: p.1 S0010

At the same time, we also identify some datasets where GPT-3’s few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora.

<a id="S0011"></a> Source: p.1 S0011

Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans.

<a id="S0012"></a> Source: p.1 S0012

We discuss broader societal impacts of this finding and of GPT-3 in general. ∗Equal contribution †Johns Hopkins University, OpenAI Author contributions listed at end of paper. 0202 luJ 22 ]LC.sc[ 4v56141.5002:viXra

Page 2

<a id="S0013"></a> Source: p.2 S0013

Contents 1 Introduction 3 2 Approach 6 2.1 Model and Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2 Training Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.3 Training Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 3 Results 10 3.1 Language Modeling, Cloze, and Completion Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2 Closed Book Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.3 Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.4 Winograd-Style Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3.5 Common Sense Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 3.6 Reading Comprehension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.7 SuperGLUE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.8 NLI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 3.9 Synthetic and Qualitative Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4 Measuring and Preventing Memorization Of Benchmarks 29 5 Limitations 33 6 Broader Impacts 34 6.1 Misuse of Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 6.2 Fairness, Bias, and Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 6.3 Energy Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 7 Related Work 39 8 Conclusion 40 A Details of Common Crawl Filtering 43 B Details of Model Training 43 C Details of Test Set Contamination Studies 43 D Total Compute Used to Train Language Models 46 E Human Quality Assessment of Synthetic News Articles 46 F Additional Samples from GPT-3 48 G Details of Task Phrasing and Specifications 50 H Results on All Tasks for All Model Sizes 63 2

Page 3

<a id="S0014"></a> Source: p.3 S0014

1 Introduction Recent years have featured a trend towards pre-trained language representations in NLP systems, applied in increasingly flexible and task-agnostic ways for downstream transfer.

<a id="S0015"></a> Source: p.3 S0015

First, single-layer representations were learned using word vectors [MCCD13, PSM14] and fed to task-specific architectures, then RNNs with multiple layers of representations and contextual state were used to form stronger representations [DL15, MBXS17, PNZtY18] (though still applied to task-specific architectures), and more recently pre-trained recurrent or transformer language models [VSP+17] have been directly fine-tuned, entirely removing the need for task-specific architectures [RNSS18, DCLT18, HR18].

<a id="S0016"></a> Source: p.3 S0016

This last paradigm has led to substantial progress on many challenging NLP tasks such as reading comprehension, question answering, textual entailment, and many others, and has continued to advance based on new architectures and algorithms [RSR+19, LOG+19, YDY+19, LCG+19].

<a id="S0017"></a> Source: p.3 S0017

However, a major limitation to this approach is that while the architecture is task-agnostic, there is still a need for task-specific datasets and task-specific fine-tuning: to achieve strong performance on a desired task typically requires fine-tuning on a dataset of thousands to hundreds of thousands of examples specific to that task.

<a id="S0018"></a> Source: p.3 S0018

Removing this limitation would be desirable, for several reasons.

<a id="S0019"></a> Source: p.3 S0019

First, from a practical perspective, the need for a large dataset of labeled examples for every new task limits the applicability of language models.

<a id="S0020"></a> Source: p.3 S0020

There exists a very wide range of possible useful language tasks, encompassing anything from correcting grammar, to generating examples of an abstract concept, to critiquing a short story.

<a id="S0021"></a> Source: p.3 S0021

For many of these tasks it is difficult to collect a large supervised training dataset, especially when the process must be repeated for every new task.

<a id="S0022"></a> Source: p.3 S0022

Second, the potential to exploit spurious correlations in training data fundamentally grows with the expressiveness of the model and the narrowness of the training distribution.

<a id="S0023"></a> Source: p.3 S0023

This can create problems for the pre-training plus fine-tuning paradigm, where models are designed to be large to absorb information during pre-training, but are then fine-tuned on very narrow task distributions.

<a id="S0024"></a> Source: p.3 S0024

For instance [HLW+20] observe that larger models do not necessarily generalize better out-of-distribution.

<a id="S0025"></a> Source: p.3 S0025

There is evidence that suggests that the generalization achieved under this paradigm can be poor because the model is overly specific to the training distribution and does not generalize well outside it [YdC+19, MPL19].

<a id="S0026"></a> Source: p.3 S0026

Thus, the performance of fine-tuned models on specific benchmarks, even when it is nominally at human-level, may exaggerate actual performance on the underlying task [GSL+18, NK19].

<a id="S0027"></a> Source: p.3 S0027

Third, humans do not require large supervised datasets to learn most language tasks – a brief directive in natural language (e.g. “please tell me if this sentence describes something happy or something sad”) or at most a tiny number of demonstrations (e.g. “here are two examples of people acting brave; please give a third example of bravery”) is often Figure 1.1: Language model meta-learning.

<a id="S0028"></a> Source: p.3 S0028

During unsupervised pre-training, a language model develops a broad set of skills and pattern recognition abilities.

<a id="S0029"></a> Source: p.3 S0029

It then uses these abilities at inference time to rapidly adapt to or recognize the desired task.

<a id="S0030"></a> Source: p.3 S0030

We use the term “in-context learning” to describe the inner loop of this process, which occurs within the forward-pass upon each sequence.

<a id="S0031"></a> Source: p.3 S0031

The sequences in this diagram are not intended to be representative of the data a model would see during pre-training, but are intended to show that there are sometimes repeated sub-tasks embedded within a single sequence. 3

Page 4

<a id="S0032"></a> Source: p.4 S0032

Figure 1.2: Larger models make increasingly efficient use of in-context information.

<a id="S0033"></a> Source: p.4 S0033

We show in-context learning performance on a simple task requiring the model to remove random symbols from a word, both with and without a natural language task description (see Sec. 3.9.2).

<a id="S0034"></a> Source: p.4 S0034

The steeper “in-context learning curves” for large models demonstrate improved ability to learn a task from contextual information.

<a id="S0035"></a> Source: p.4 S0035

We see qualitatively similar behavior across a wide range of tasks. sufficient to enable a human to perform a new task to at least a reasonable degree of competence.

<a id="S0036"></a> Source: p.4 S0036

Aside from pointing to a conceptual limitation in our current NLP techniques, this adaptability has practical advantages – it allows humans to seamlessly mix together or switch between many tasks and skills, for example performing addition during a lengthy dialogue.

<a id="S0037"></a> Source: p.4 S0037

To be broadly useful, we would someday like our NLP systems to have this same fluidity and generality.

<a id="S0038"></a> Source: p.4 S0038

One potential route towards addressing these issues is meta-learning1 – which in the context of language models means the model develops a broad set of skills and pattern recognition abilities at training time, and then uses those abilities at inference time to rapidly adapt to or recognize the desired task (illustrated in Figure 1.1).

<a id="S0039"></a> Source: p.4 S0039

Recent work [RWC+19] attempts to do this via what we call “in-context learning”, using the text input of a pretrained language model as a form of task specification: the model is conditioned on a natural language instruction and/or a few demonstrations of the task and is then expected to complete further instances of the task simply by predicting what comes next.

<a id="S0040"></a> Source: p.4 S0040

While it has shown some initial promise, this approach still achieves results far inferior to fine-tuning – for example [RWC+19] achieves only 4% on Natural Questions, and even its 55 F1 CoQa result is now more than 35 points behind the state of the art.

<a id="S0041"></a> Source: p.4 S0041

Meta-learning clearly requires substantial improvement in order to be viable as a practical method of solving language tasks.

<a id="S0042"></a> Source: p.4 S0042

Another recent trend in language modeling may offer a way forward.

<a id="S0043"></a> Source: p.4 S0043

In recent years the capacity of transformer language models has increased substantially, from 100 million parameters [RNSS18], to 300 million parameters [DCLT18], to 1.5 billion parameters [RWC+19], to 8 billion parameters [SPP+19], 11 billion parameters [RSR+19], and finally 17 billion parameters [Tur20].

<a id="S0044"></a> Source: p.4 S0044

Each increase has brought improvements in text synthesis and/or downstream NLP tasks, and there is evidence suggesting that log loss, which correlates well with many downstream tasks, follows a smooth trend of improvement with scale [KMH+20].

<a id="S0045"></a> Source: p.4 S0045

Since in-context learning involves absorbing many skills and tasks within the parameters of the model, it is plausible that in-context learning abilities might show similarly strong gains with scale. 1In the context of language models this has sometimes been called “zero-shot transfer”, but this term is potentially ambiguous: the method is “zero-shot” in the sense that no gradient updates are performed, but it often involves providing inference-time demonstrations to the model, so is not truly learning from zero examples.

<a id="S0046"></a> Source: p.4 S0046

To avoid this confusion, we use the term “meta-learning” to capture the inner-loop / outer-loop structure of the general method, and the term “in context-learning” to refer to the inner loop of meta-learning.

<a id="S0047"></a> Source: p.4 S0047

We further specialize the description to “zero-shot”, “one-shot”, or “few-shot” depending on how many demonstrations are provided at inference time.

<a id="S0048"></a> Source: p.4 S0048

These terms are intended to remain agnostic on the question of whether the model learns new tasks from scratch at inference time or simply recognizes patterns seen during training – this is an important issue which we discuss later in the paper, but “meta-learning” is intended to encompass both possibilities, and simply describes the inner-outer loop structure. 4

Page 5

<a id="S0049"></a> Source: p.5 S0049

Figure 1.3: Aggregate performance for all 42 accuracy-denominated benchmarks While zero-shot performance improves steadily with model size, few-shot performance increases more rapidly, demonstrating that larger models are more proficient at in-context learning.

<a id="S0050"></a> Source: p.5 S0050

See Figure 3.8 for a more detailed analysis on SuperGLUE, a standard NLP benchmark suite.

<a id="S0051"></a> Source: p.5 S0051

In this paper, we test this hypothesis by training a 175 billion parameter autoregressive language model, which we call GPT-3, and measuring its in-context learning abilities.

<a id="S0052"></a> Source: p.5 S0052

Specifically, we evaluate GPT-3 on over two dozen NLP datasets, as well as several novel tasks designed to test rapid adaptation to tasks unlikely to be directly contained in the training set.

<a id="S0053"></a> Source: p.5 S0053

For each task, we evaluate GPT-3 under 3 conditions: (a) “few-shot learning”, or in-context learning where we allow as many demonstrations as will fit into the model’s context window (typically 10 to 100), (b) “one-shot learning”, where we allow only one demonstration, and (c) “zero-shot” learning, where no demonstrations are allowed and only an instruction in natural language is given to the model.

<a id="S0054"></a> Source: p.5 S0054

GPT-3 could also in principle be evaluated in the traditional fine-tuning setting, but we leave this to future work.

<a id="S0055"></a> Source: p.5 S0055

Figure 1.2 illustrates the conditions we study, and shows few-shot learning of a simple task requiring the model to remove extraneous symbols from a word.

<a id="S0056"></a> Source: p.5 S0056

Model performance improves with the addition of a natural language task description, and with the number of examples in the model’s context, K.

<a id="S0057"></a> Source: p.5 S0057

Few-shot learning also improves dramatically with model size.

<a id="S0058"></a> Source: p.5 S0058

Though the results in this case are particularly striking, the general trends with both model size and number of examples in-context hold for most tasks we study.

<a id="S0059"></a> Source: p.5 S0059

We emphasize that these “learning” curves involve no gradient updates or fine-tuning, just increasing numbers of demonstrations given as conditioning.

<a id="S0060"></a> Source: p.5 S0060

Broadly, on NLP tasks GPT-3 achieves promising results in the zero-shot and one-shot settings, and in the the few-shot setting is sometimes competitive with or even occasionally surpasses state-of-the-art (despite state-of-the-art being held by fine-tuned models).

<a id="S0061"></a> Source: p.5 S0061

For example, GPT-3 achieves 81.5 F1 on CoQA in the zero-shot setting, 84.0 F1 on CoQA in the one-shot setting, 85.0 F1 in the few-shot setting.

<a id="S0062"></a> Source: p.5 S0062

Similarly, GPT-3 achieves 64.3% accuracy on TriviaQA in the zero-shot setting, 68.0% in the one-shot setting, and 71.2% in the few-shot setting, the last of which is state-of-the-art relative to fine-tuned models operating in the same closed-book setting.

<a id="S0063"></a> Source: p.5 S0063

GPT-3 also displays one-shot and few-shot proficiency at tasks designed to test rapid adaption or on-the-fly reasoning, which include unscrambling words, performing arithmetic, and using novel words in a sentence after seeing them defined only once.

<a id="S0064"></a> Source: p.5 S0064

We also show that in the few-shot setting, GPT-3 can generate synthetic news articles which human evaluators have difficulty distinguishing from human-generated articles.

<a id="S0065"></a> Source: p.5 S0065

At the same time, we also find some tasks on which few-shot performance struggles, even at the scale of GPT-3.

<a id="S0066"></a> Source: p.5 S0066

This includes natural language inference tasks like the ANLI dataset, and some reading comprehension datasets like RACE or QuAC.

<a id="S0067"></a> Source: p.5 S0067

By presenting a broad characterization of GPT-3’s strengths and weaknesses, including these limitations, we hope to stimulate study of few-shot learning in language models and draw attention to where progress is most needed. A heuristic sense of the overall results can be seen in Figure 1.3, which aggregates the various tasks (though it should not be seen as a rigorous or meaningful benchmark in itself). 5

Page 6

<a id="S0068"></a> Source: p.6 S0068

We also undertake a systematic study of “data contamination” – a growing problem when training high capacity models on datasets such as Common Crawl, which can potentially include content from test datasets simply because such content often exists on the web.

<a id="S0069"></a> Source: p.6 S0069

In this paper we develop systematic tools to measure data contamination and quantify its distorting effects.

<a id="S0070"></a> Source: p.6 S0070

Although we find that data contamination has a minimal effect on GPT-3’s performance on most datasets, we do identify a few datasets where it could be inflating results, and we either do not report results on these datasets or we note them with an asterisk, depending on the severity.

<a id="S0071"></a> Source: p.6 S0071

In addition to all the above, we also train a series of smaller models (ranging from 125 million parameters to 13 billion parameters) in order to compare their performance to GPT-3 in the zero, one and few-shot settings.

<a id="S0072"></a> Source: p.6 S0072

Broadly, for most tasks we find relatively smooth scaling with model capacity in all three settings; one notable pattern is that the gap between zero-, one-, and few-shot performance often grows with model capacity, perhaps suggesting that larger models are more proficient meta-learners.

<a id="S0073"></a> Source: p.6 S0073

Finally, given the broad spectrum of capabilities displayed by GPT-3, we discuss concerns about bias, fairness, and broader societal impacts, and attempt a preliminary analysis of GPT-3’s characteristics in this regard.

<a id="S0074"></a> Source: p.6 S0074

The remainder of this paper is organized as follows.

<a id="S0075"></a> Source: p.6 S0075

In Section 2, we describe our approach and methods for training GPT-3 and evaluating it.

<a id="S0076"></a> Source: p.6 S0076

Section 3 presents results on the full range of tasks in the zero-, one- and few-shot settings.

<a id="S0077"></a> Source: p.6 S0077

Section 4 addresses questions of data contamination (train-test overlap).

<a id="S0078"></a> Source: p.6 S0078

Section 5 discusses limitations of GPT-3.

<a id="S0079"></a> Source: p.6 S0079

Section 7 reviews related work and Section 8 concludes. 2 Approach Our basic pre-training approach, including model, data, and training, is similar to the process described in [RWC+19], with relatively straightforward scaling up of the model size, dataset size and diversity, and length of training.

<a id="S0080"></a> Source: p.6 S0080

Our use of in-context learning is also similar to [RWC+19], but in this work we systematically explore different settings for learning within the context.

<a id="S0081"></a> Source: p.6 S0081

Therefore, we start this section by explicitly defining and contrasting the different settings that we will be evaluating GPT-3 on or could in principle evaluate GPT-3 on.

<a id="S0082"></a> Source: p.6 S0082

These settings can be seen as lying on a spectrum of how much task-specific data they tend to rely on.

<a id="S0083"></a> Source: p.6 S0083

Specifically, we can identify at least four points on this spectrum (see Figure 2.1 for an illustration): • Fine-Tuning (FT) has been the most common approach in recent years, and involves updating the weights of a pre-trained model by training on a supervised dataset specific to the desired task.

<a id="S0084"></a> Source: p.6 S0084

Typically thousands to hundreds of thousands of labeled examples are used.

<a id="S0085"></a> Source: p.6 S0085

The main advantage of fine-tuning is strong performance on many benchmarks.

<a id="S0086"></a> Source: p.6 S0086

The main disadvantages are the need for a new large dataset for every task, the potential for poor generalization out-of-distribution [MPL19], and the potential to exploit spurious features of the training data [GSL+18, NK19], potentially resulting in an unfair comparison with human performance.

<a id="S0087"></a> Source: p.6 S0087

In this work we do not fine-tune GPT-3 because our focus is on task-agnostic performance, but GPT-3 can be fine-tuned in principle and this is a promising direction for future work. • Few-Shot (FS) is the term we will use in this work to refer to the setting where the model is given a few demonstrations of the task at inference time as conditioning [RWC+19], but no weight updates are allowed.

<a id="S0088"></a> Source: p.6 S0088

As shown in Figure 2.1, for a typical dataset an example has a context and a desired completion (for example an English sentence and the French translation), and few-shot works by giving K examples of context and completion, and then one final example of context, with the model expected to provide the completion.

<a id="S0089"></a> Source: p.6 S0089

We typically set K in the range of 10 to 100 as this is how many examples can fit in the model’s context window (n = 2048).

<a id="S0090"></a> Source: p.6 S0090

The main advantages of few-shot are a major reduction in the need for task-specific data and ctx reduced potential to learn an overly narrow distribution from a large but narrow fine-tuning dataset.

<a id="S0091"></a> Source: p.6 S0091

The main disadvantage is that results from this method have so far been much worse than state-of-the-art fine-tuned models.

<a id="S0092"></a> Source: p.6 S0092

Also, a small amount of task specific data is still required.

<a id="S0093"></a> Source: p.6 S0093

As indicated by the name, few-shot learning as described here for language models is related to few-shot learning as used in other contexts in ML [HYC01, VBL+16] – both involve learning based on a broad distribution of tasks (in this case implicit in the pre-training data) and then rapidly adapting to a new task. • One-Shot (1S) is the same as few-shot except that only one demonstration is allowed, in addition to a natural language description of the task, as shown in Figure 1.

<a id="S0094"></a> Source: p.6 S0094

The reason to distinguish one-shot from few-shot and zero-shot (below) is that it most closely matches the way in which some tasks are communicated to humans.

<a id="S0095"></a> Source: p.6 S0095

For example, when asking humans to generate a dataset on a human worker service (for example Mechanical Turk), it is common to give one demonstration of the task.

<a id="S0096"></a> Source: p.6 S0096

By contrast it is sometimes difficult to communicate the content or format of a task if no examples are given. 6

Page 7

<a id="S0097"></a> Source: p.7 S0097

Figure 2.1: Zero-shot, one-shot and few-shot, contrasted with traditional fine-tuning.

<a id="S0098"></a> Source: p.7 S0098

The panels above show four methods for performing a task with a language model – fine-tuning is the traditional method, whereas zero-, one-, and few-shot, which we study in this work, require the model to perform the task with only forward passes at test time.

<a id="S0099"></a> Source: p.7 S0099

We typically present the model with a few dozen examples in the few shot setting.

<a id="S0100"></a> Source: p.7 S0100

Exact phrasings for all task descriptions, examples and prompts can be found in Appendix G. • Zero-Shot (0S) is the same as one-shot except that no demonstrations are allowed, and the model is only given a natural language instruction describing the task.

<a id="S0101"></a> Source: p.7 S0101

This method provides maximum convenience, potential for robustness, and avoidance of spurious correlations (unless they occur very broadly across the large corpus of pre-training data), but is also the most challenging setting.

<a id="S0102"></a> Source: p.7 S0102

In some cases it may even be difficult for humans to understand the format of the task without prior examples, so this setting is in some cases “unfairly hard”.

<a id="S0103"></a> Source: p.7 S0103

For example, if someone is asked to “make a table of world records for the 200m dash”, this request can be ambiguous, as it may not be clear exactly what format the table should have or what should be included (and even with careful clarification, understanding precisely what is desired can be difficult).

<a id="S0104"></a> Source: p.7 S0104

Nevertheless, for at least some settings zero-shot is closest to how humans perform tasks – for example, in the translation example in Figure 2.1, a human would likely know what to do from just the text instruction.

<a id="S0105"></a> Source: p.7 S0105

Figure 2.1 shows the four methods using the example of translating English to French.

<a id="S0106"></a> Source: p.7 S0106

In this paper we focus on zero-shot, one-shot and few-shot, with the aim of comparing them not as competing alternatives, but as different problem settings which offer a varying trade-off between performance on specific benchmarks and sample efficiency.

<a id="S0107"></a> Source: p.7 S0107

We especially highlight the few-shot results as many of them are only slightly behind state-of-the-art fine-tuned models.

<a id="S0108"></a> Source: p.7 S0108

Ultimately, however, one-shot, or even sometimes zero-shot, seem like the fairest comparisons to human performance, and are important targets for future work.

<a id="S0109"></a> Source: p.7 S0109

Sections 2.1-2.3 below give details on our models, training data, and training process respectively.

<a id="S0110"></a> Source: p.7 S0110

Section 2.4 discusses the details of how we do few-shot, one-shot, and zero-shot evaluations. 7

Page 8

<a id="S0111"></a> Source: p.8 S0111

Model Name n n d n d Batch Size Learning Rate params layers model heads head GPT-3 Small 125M 12 768 12 64 0.5M 6.0 × 10−4 GPT-3 Medium 350M 24 1024 16 64 0.5M 3.0 × 10−4 GPT-3 Large 760M 24 1536 16 96 0.5M 2.5 × 10−4 GPT-3 XL 1.3B 24 2048 24 128 1M 2.0 × 10−4 GPT-3 2.7B 2.7B 32 2560 32 80 1M 1.6 × 10−4 GPT-3 6.7B 6.7B 32 4096 32 128 2M 1.2 × 10−4 GPT-3 13B 13.0B 40 5140 40 128 2M 1.0 × 10−4 GPT-3 175B or “GPT-3” 175.0B 96 12288 96 128 3.2M 0.6 × 10−4 Table 2.1: Sizes, architectures, and learning hyper-parameters (batch size in tokens and learning rate) of the models which we trained.

<a id="S0112"></a> Source: p.8 S0112

All models were trained for a total of 300 billion tokens. 2.1 Model and Architectures We use the same model and architecture as GPT-2 [RWC+19], including the modified initialization, pre-normalization, and reversible tokenization described therein, with the exception that we use alternating dense and locally banded sparse attention patterns in the layers of the transformer, similar to the Sparse Transformer [CGRS19].

<a id="S0113"></a> Source: p.8 S0113

To study the dependence of ML performance on model size, we train 8 different sizes of model, ranging over three orders of magnitude from 125 million parameters to 175 billion parameters, with the last being the model we call GPT-3.

<a id="S0114"></a> Source: p.8 S0114

Previous work [KMH+20] suggests that with enough training data, scaling of validation loss should be approximately a smooth power law as a function of size; training models of many different sizes allows us to test this hypothesis both for validation loss and for downstream language tasks.

<a id="S0115"></a> Source: p.8 S0115

Table 2.1 shows the sizes and architectures of our 8 models.

<a id="S0116"></a> Source: p.8 S0116

Here n is the total number of trainable parameters, params n is the total number of layers, d is the number of units in each bottleneck layer (we always have the layers model feedforward layer four times the size of the bottleneck layer, d = 4 ∗ d ), and d is the dimension of each ff model head attention head.

<a id="S0117"></a> Source: p.8 S0117

All models use a context window of n = 2048 tokens.

<a id="S0118"></a> Source: p.8 S0118

We partition the model across GPUs along ctx both the depth and width dimension in order to minimize data-transfer between nodes.

<a id="S0119"></a> Source: p.8 S0119

The precise architectural parameters for each model are chosen based on computational efficiency and load-balancing in the layout of models across GPU’s.

<a id="S0120"></a> Source: p.8 S0120

Previous work [KMH+20] suggests that validation loss is not strongly sensitive to these parameters within a reasonably broad range. 2.2 Training Dataset Datasets for language models have rapidly expanded, culminating in the Common Crawl dataset2 [RSR+19] constituting nearly a trillion words.

<a id="S0121"></a> Source: p.8 S0121

This size of dataset is sufficient to train our largest models without ever updating on the same sequence twice.

<a id="S0122"></a> Source: p.8 S0122

However, we have found that unfiltered or lightly filtered versions of Common Crawl tend to have lower quality than more curated datasets.

<a id="S0123"></a> Source: p.8 S0123

Therefore, we took 3 steps to improve the average quality of our datasets: (1) we downloaded and filtered a version of CommonCrawl based on similarity to a range of high-quality reference corpora, (2) we performed fuzzy deduplication at the document level, within and across datasets, to prevent redundancy and preserve the integrity of our held-out validation set as an accurate measure of overfitting, and (3) we also added known high-quality reference corpora to the training mix to augment CommonCrawl and increase its diversity.

<a id="S0124"></a> Source: p.8 S0124

Details of the first two points (processing of Common Crawl) are described in Appendix A.

<a id="S0125"></a> Source: p.8 S0125

For the third, we added several curated high-quality datasets, including an expanded version of the WebText dataset [RWC+19], collected by scraping links over a longer period of time, and first described in [KMH+20], two internet-based books corpora (Books1 and Books2) and English-language Wikipedia.

<a id="S0126"></a> Source: p.8 S0126

Table 2.2 shows the final mixture of datasets that we used in training.

<a id="S0127"></a> Source: p.8 S0127

The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl covering 2016 to 2019, constituting 45TB of compressed plaintext before filtering and 570GB after filtering, roughly equivalent to 400 billion byte-pair-encoded tokens.

<a id="S0128"></a> Source: p.8 S0128

Note that during training, datasets are not sampled in proportion to their size, but rather datasets we view as higher-quality are sampled more frequently, such that CommonCrawl and Books2 datasets are sampled less than once during training, but the other datasets are sampled 2-3 times.

<a id="S0129"></a> Source: p.8 S0129

This essentially accepts a small amount of overfitting in exchange for higher quality training data. 2https://commoncrawl.org/the-data/ 8

Page 9

<a id="S0130"></a> Source: p.9 S0130

Figure 2.2: Total compute used during training.

<a id="S0131"></a> Source: p.9 S0131

Based on the analysis in Scaling Laws For Neural Language Models [KMH+20] we train much larger models on many fewer tokens than is typical.

<a id="S0132"></a> Source: p.9 S0132

As a consequence, although GPT-3 3B is almost 10x larger than RoBERTa-Large (355M params), both models took roughly 50 petaflop/s-days of compute during pre-training.

<a id="S0133"></a> Source: p.9 S0133

Methodology for these calculations can be found in Appendix D.

<a id="S0134"></a> Source: p.9 S0134

Quantity Weight in Epochs elapsed when Dataset (tokens) training mix training for 300B tokens Common Crawl (filtered) 410 billion 60% 0.44 WebText2 19 billion 22% 2.9 Books1 12 billion 8% 1.9 Books2 55 billion 8% 0.43 Wikipedia 3 billion 3% 3.4 Table 2.2: Datasets used to train GPT-3. “Weight in training mix” refers to the fraction of examples during training that are drawn from a given dataset, which we intentionally do not make proportional to the size of the dataset.

<a id="S0135"></a> Source: p.9 S0135

As a result, when we train for 300 billion tokens, some datasets are seen up to 3.4 times during training while other datasets are seen less than once. A major methodological concern with language models pretrained on a broad swath of internet data, particularly large models with the capacity to memorize vast amounts of content, is potential contamination of downstream tasks by having their test or development sets inadvertently seen during pre-training.

<a id="S0136"></a> Source: p.9 S0136

To reduce such contamination, we searched for and attempted to remove any overlaps with the development and test sets of all benchmarks studied in this paper.

<a id="S0137"></a> Source: p.9 S0137

Unfortunately, a bug in the filtering caused us to ignore some overlaps, and due to the cost of training it was not feasible to retrain the model.

<a id="S0138"></a> Source: p.9 S0138

In Section 4 we characterize the impact of the remaining overlaps, and in future work we will more aggressively remove data contamination. 2.3 Training Process As found in [KMH+20, MKAT18], larger models can typically use a larger batch size, but require a smaller learning rate.

<a id="S0139"></a> Source: p.9 S0139

We measure the gradient noise scale during training and use it to guide our choice of batch size [MKAT18].

<a id="S0140"></a> Source: p.9 S0140

Table 2.1 shows the parameter settings we used.

<a id="S0141"></a> Source: p.9 S0141

To train the larger models without running out of memory, we use a mixture of model parallelism within each matrix multiply and model parallelism across the layers of the network.

<a id="S0142"></a> Source: p.9 S0142

All models were trained on V100 GPU’s on part of a high-bandwidth cluster provided by Microsoft.

<a id="S0143"></a> Source: p.9 S0143

Details of the training process and hyperparameter settings are described in Appendix B. 9

Page 10

<a id="S0144"></a> Source: p.10 S0144

2.4 Evaluation For few-shot learning, we evaluate each example in the evaluation set by randomly drawing K examples from that task’s training set as conditioning, delimited by 1 or 2 newlines depending on the task.

<a id="S0145"></a> Source: p.10 S0145

For LAMBADA and Storycloze there is no supervised training set available so we draw conditioning examples from the development set and evaluate on the test set.

<a id="S0146"></a> Source: p.10 S0146

For Winograd (the original, not SuperGLUE version) there is only one dataset, so we draw conditioning examples directly from it. K can be any value from 0 to the maximum amount allowed by the model’s context window, which is n = 2048 ctx for all models and typically fits 10 to 100 examples.

<a id="S0147"></a> Source: p.10 S0147

Larger values of K are usually but not always better, so when a separate development and test set are available, we experiment with a few values of K on the development set and then run the best value on the test set.

<a id="S0148"></a> Source: p.10 S0148

For some tasks (see Appendix G) we also use a natural language prompt in addition to (or for K = 0, instead of) demonstrations.

<a id="S0149"></a> Source: p.10 S0149

On tasks that involve choosing one correct completion from several options (multiple choice), we provide K examples of context plus correct completion, followed by one example of context only, and compare the LM likelihood of each completion.

<a id="S0150"></a> Source: p.10 S0150

For most tasks we compare the per-token likelihood (to normalize for length), however on a small number of datasets (ARC, OpenBookQA, and RACE) we gain additional benefit as measured on the development set by normalizing by the unconditional probability of each completion, by computing P (completion|context) , where P (completion|answer context) answer context is the string "Answer: " or "A: " and is used to prompt that the completion should be an answer but is otherwise generic.

<a id="S0151"></a> Source: p.10 S0151

On tasks that involve binary classification, we give the options more semantically meaningful names (e.g. “True” or “False” rather than 0 or 1) and then treat the task like multiple choice; we also sometimes frame the task similar to what is done by [RSR+19] (see Appendix G) for details.

<a id="S0152"></a> Source: p.10 S0152

On tasks with free-form completion, we use beam search with the same parameters as [RSR+19]: a beam width of 4 and a length penalty of α = 0.6.

<a id="S0153"></a> Source: p.10 S0153

We score the model using F1 similarity score, BLEU, or exact match, depending on what is standard for the dataset at hand.

<a id="S0154"></a> Source: p.10 S0154

Final results are reported on the test set when publicly available, for each model size and learning setting (zero-, one-, and few-shot).

<a id="S0155"></a> Source: p.10 S0155

When the test set is private, our model is often too large to fit on the test server, so we report results on the development set.

<a id="S0156"></a> Source: p.10 S0156

We do submit to the test server on a small number of datasets (SuperGLUE, TriviaQA, PiQa) where we were able to make submission work, and we submit only the 200B few-shot results, and report development set results for everything else. 3 Results In Figure 3.1 we display training curves for the 8 models described in Section 2.

<a id="S0157"></a> Source: p.10 S0157

For this graph we also include 6 additional extra-small models with as few as 100,000 parameters.

<a id="S0158"></a> Source: p.10 S0158

As observed in [KMH+20], language modeling performance follows a power-law when making efficient use of training compute.

<a id="S0159"></a> Source: p.10 S0159

After extending this trend by two more orders of magnitude, we observe only a slight (if any) departure from the power-law.

<a id="S0160"></a> Source: p.10 S0160

One might worry that these improvements in cross-entropy loss come only from modeling spurious details of our training corpus.

<a id="S0161"></a> Source: p.10 S0161

However, we will see in the following sections that improvements in cross-entropy loss lead to consistent performance gains across a broad spectrum of natural language tasks.

<a id="S0162"></a> Source: p.10 S0162

Below, we evaluate the 8 models described in Section 2 (the 175 billion parameter parameter GPT-3 and 7 smaller models) on a wide range of datasets.

<a id="S0163"></a> Source: p.10 S0163

We group the datasets into 9 categories representing roughly similar tasks.

<a id="S0164"></a> Source: p.10 S0164

In Section 3.1 we evaluate on traditional language modeling tasks and tasks that are similar to language modeling, such as Cloze tasks and sentence/paragraph completion tasks.

<a id="S0165"></a> Source: p.10 S0165

In Section 3.2 we evaluate on “closed book” question answering tasks: tasks which require using the information stored in the model’s parameters to answer general knowledge questions.

<a id="S0166"></a> Source: p.10 S0166

In Section 3.3 we evaluate the model’s ability to translate between languages (especially one-shot and few-shot).

<a id="S0167"></a> Source: p.10 S0167

In Section 3.4 we evaluate the model’s performance on Winograd Schema-like tasks.

<a id="S0168"></a> Source: p.10 S0168

In Section 3.5 we evaluate on datasets that involve commonsense reasoning or question answering.

<a id="S0169"></a> Source: p.10 S0169

In Section 3.6 we evaluate on reading comprehension tasks, in Section 3.7 we evaluate on the SuperGLUE benchmark suite, and in 3.8 we briefly explore NLI.

<a id="S0170"></a> Source: p.10 S0170

Finally, in Section 3.9, we invent some additional tasks designed especially to probe in-context learning abilities – these tasks focus on on-the-fly reasoning, adaptation skills, or open-ended text synthesis.

<a id="S0171"></a> Source: p.10 S0171

We evaluate all tasks in the few-shot, one-shot, and zero-shot settings. 10

Page 11

<a id="S0172"></a> Source: p.11 S0172

Figure 3.1: Smooth scaling of performance with compute.

<a id="S0173"></a> Source: p.11 S0173

Performance (measured in terms of cross-entropy validation loss) follows a power-law trend with the amount of compute used for training.

<a id="S0174"></a> Source: p.11 S0174

The power-law behavior observed in [KMH+20] continues for an additional two orders of magnitude with only small deviations from the predicted curve.

<a id="S0175"></a> Source: p.11 S0175

For this figure, we exclude embedding parameters from compute and parameter counts.

<a id="S0176"></a> Source: p.11 S0176

Setting PTB SOTA (Zero-Shot) 35.8a GPT-3 Zero-Shot 20.5 Table 3.1: Zero-shot results on PTB language modeling dataset.

<a id="S0177"></a> Source: p.11 S0177

Many other common language modeling datasets are omitted because they are derived from Wikipedia or other sources which are included in GPT-3’s training data. a[RWC+19] 3.1 Language Modeling, Cloze, and Completion Tasks In this section we test GPT-3’s performance on the traditional task of language modeling, as well as related tasks that involve predicting a single word of interest, completing a sentence or paragraph, or choosing between possible completions of a piece of text. 3.1.1 Language Modeling We calculate zero-shot perplexity on the Penn Tree Bank (PTB) [MKM+94] dataset measured in [RWC+19].

<a id="S0178"></a> Source: p.11 S0178

We omit the 4 Wikipedia-related tasks in that work because they are entirely contained in our training data, and we also omit the one-billion word benchmark due to a high fraction of the dataset being contained in our training set.

<a id="S0179"></a> Source: p.11 S0179

PTB escapes these issues due to predating the modern internet.

<a id="S0180"></a> Source: p.11 S0180

Our largest model sets a new SOTA on PTB by a substantial margin of 15 points, achieving a perplexity of 20.50.

<a id="S0181"></a> Source: p.11 S0181

Note that since PTB is a traditional language modeling dataset it does not have a clear separation of examples to define one-shot or few-shot evaluation around, so we measure only zero-shot. 3.1.2 LAMBADA The LAMBADA dataset [PKL+16] tests the modeling of long-range dependencies in text – the model is asked to predict the last word of sentences which require reading a paragraph of context.

<a id="S0182"></a> Source: p.11 S0182

It has recently been suggested that the continued scaling of language models is yielding diminishing returns on this difficult benchmark. [BHT+20] reflect on the small 1.5% improvement achieved by a doubling of model size between two recent state of the art results ([SPP+19] 11

Page 12

<a id="S0183"></a> Source: p.12 S0183

LAMBADA LAMBADA StoryCloze HellaSwag Setting (acc) (ppl) (acc) (acc) SOTA 68.0a 8.63b 91.8c 85.6d GPT-3 Zero-Shot 76.2 3.00 83.2 78.9 GPT-3 One-Shot 72.5 3.35 84.7 78.1 GPT-3 Few-Shot 86.4 1.92 87.7 79.3 Table 3.2: Performance on cloze and completion tasks.

<a id="S0184"></a> Source: p.12 S0184

GPT-3 significantly improves SOTA on LAMBADA while achieving respectable performance on two difficult completion prediction datasets. a[Tur20] b[RWC+19] c[LDL19] d[LCH+20] Figure 3.2: On LAMBADA, the few-shot capability of language models results in a strong boost to accuracy.

<a id="S0185"></a> Source: p.12 S0185

GPT-3 2.7B outperforms the SOTA 17B parameter Turing-NLG [Tur20] in this setting, and GPT-3 175B advances the state of the art by 18%.

<a id="S0186"></a> Source: p.12 S0186

Note zero-shot uses a different format from one-shot and few-shot as described in the text. and [Tur20]) and argue that “continuing to expand hardware and data sizes by orders of magnitude is not the path forward”.

<a id="S0187"></a> Source: p.12 S0187

We find that path is still promising and in a zero-shot setting GPT-3 achieves 76% on LAMBADA, a gain of 8% over the previous state of the art.

<a id="S0188"></a> Source: p.12 S0188

LAMBADA is also a demonstration of the flexibility of few-shot learning as it provides a way to address a problem that classically occurs with this dataset.

<a id="S0189"></a> Source: p.12 S0189

Although the completion in LAMBADA is always the last word in a sentence, a standard language model has no way of knowing this detail.

<a id="S0190"></a> Source: p.12 S0190

It thus assigns probability not only to the correct ending but also to other valid continuations of the paragraph.

<a id="S0191"></a> Source: p.12 S0191

This problem has been partially addressed in the past with stop-word filters [RWC+19] (which ban “continuation” words).

<a id="S0192"></a> Source: p.12 S0192

The few-shot setting instead allows us to “frame” the task as a cloze-test and allows the language model to infer from examples that a completion of exactly one word is desired.

<a id="S0193"></a> Source: p.12 S0193

We use the following fill-in-the-blank format: Alice was friends with Bob.

<a id="S0194"></a> Source: p.12 S0194

Alice went to visit her friend . → Bob George bought some baseball equipment, a ball, a glove, and a . → When presented with examples formatted this way, GPT-3 achieves 86.4% accuracy in the few-shot setting, an increase of over 18% from the previous state-of-the-art.

<a id="S0195"></a> Source: p.12 S0195

We observe that few-shot performance improves strongly with model size.

<a id="S0196"></a> Source: p.12 S0196

While this setting decreases the performance of the smallest model by almost 20%, for GPT-3 it improves accuracy by 10%.

<a id="S0197"></a> Source: p.12 S0197

Finally, the fill-in-blank method is not effective one-shot, where it always performs worse than the zero-shot setting.

<a id="S0198"></a> Source: p.12 S0198

Perhaps this is because all models still require several examples to recognize the pattern. 12

Page 13

<a id="S0199"></a> Source: p.13 S0199

Setting NaturalQS WebQS TriviaQA RAG (Fine-tuned, Open-Domain) [LPP+20] 44.5 45.5 68.0 T5-11B+SSM (Fine-tuned, Closed-Book) [RRS20] 36.6 44.7 60.5 T5-11B (Fine-tuned, Closed-Book) 34.5 37.4 50.1 GPT-3 Zero-Shot 14.6 14.4 64.3 GPT-3 One-Shot 23.0 25.3 68.0 GPT-3 Few-Shot 29.9 41.5 71.2 Table 3.3: Results on three Open-Domain QA tasks.

<a id="S0200"></a> Source: p.13 S0200

GPT-3 is shown in the few-, one-, and zero-shot settings, as compared to prior SOTA results for closed book and open domain settings.

<a id="S0201"></a> Source: p.13 S0201

TriviaQA few-shot result is evaluated on the wiki split test server.

<a id="S0202"></a> Source: p.13 S0202

One note of caution is that an analysis of test set contamination identified that a significant minority of the LAMBADA dataset appears to be present in our training data – however analysis performed in Section 4 suggests negligible impact on performance. 3.1.3 HellaSwag The HellaSwag dataset [ZHB+19] involves picking the best ending to a story or set of instructions.

<a id="S0203"></a> Source: p.13 S0203

The examples were adversarially mined to be difficult for language models while remaining easy for humans (who achieve 95.6% accuracy).

<a id="S0204"></a> Source: p.13 S0204

GPT-3 achieves 78.1% accuracy in the one-shot setting and 79.3% accuracy in the few-shot setting, outperforming the 75.4% accuracy of a fine-tuned 1.5B parameter language model [ZHR+19] but still a fair amount lower than the overall SOTA of 85.6% achieved by the fine-tuned multi-task model ALUM. 3.1.4 StoryCloze We next evaluate GPT-3 on the StoryCloze 2016 dataset [MCH+16], which involves selecting the correct ending sentence for five-sentence long stories.

<a id="S0205"></a> Source: p.13 S0205

Here GPT-3 achieves 83.2% in the zero-shot setting and 87.7% in the few-shot setting (with K = 70).

<a id="S0206"></a> Source: p.13 S0206

This is still 4.1% lower than the fine-tuned SOTA using a BERT based model [LDL19] but improves over previous zero-shot results by roughly 10%. 3.2 Closed Book Question Answering In this section we measure GPT-3’s ability to answer questions about broad factual knowledge.

<a id="S0207"></a> Source: p.13 S0207

Due to the immense amount of possible queries, this task has normally been approached by using an information retrieval system to find relevant text in combination with a model which learns to generate an answer given the question and the retrieved text.

<a id="S0208"></a> Source: p.13 S0208

Since this setting allows a system to search for and condition on text which potentially contains the answer it is denoted “open-book”. [RRS20] recently demonstrated that a large language model can perform surprisingly well directly answering the questions without conditioning on auxilliary information.

<a id="S0209"></a> Source: p.13 S0209

They denote this more restrictive evaluation setting as “closed-book”.

<a id="S0210"></a> Source: p.13 S0210

Their work suggests that even higher-capacity models could perform even better and we test this hypothesis with GPT-3.

<a id="S0211"></a> Source: p.13 S0211

We evaluate GPT-3 on the 3 datasets in [RRS20]: Natural Questions [KPR+19], WebQuestions [BCFL13], and TriviaQA [JCWZ17], using the same splits.

<a id="S0212"></a> Source: p.13 S0212

Note that in addition to all results being in the closed-book setting, our use of few-shot, one-shot, and zero-shot evaluations represent an even stricter setting than previous closed-book QA work: in addition to external content not being allowed, fine-tuning on the Q&A dataset itself is also not permitted.

<a id="S0213"></a> Source: p.13 S0213

The results for GPT-3 are shown in Table 3.3.

<a id="S0214"></a> Source: p.13 S0214

On TriviaQA, we achieve 64.3% in the zero-shot setting, 68.0% in the one-shot setting, and 71.2% in the few-shot setting.

<a id="S0215"></a> Source: p.13 S0215

The zero-shot result already outperforms the fine-tuned T5-11B by 14.2%, and also outperforms a version with Q&A tailored span prediction during pre-training by 3.8%.

<a id="S0216"></a> Source: p.13 S0216

The one-shot result improves by 3.7% and matches the SOTA for an open-domain QA system which not only fine-tunes but also makes use of a learned retrieval mechanism over a 15.3B parameter dense vector index of 21M documents [LPP+20].

<a id="S0217"></a> Source: p.13 S0217

GPT-3’s few-shot result further improves performance another 3.2% beyond this.

<a id="S0218"></a> Source: p.13 S0218

On WebQuestions (WebQs), GPT-3 achieves 14.4% in the zero-shot setting, 25.3% in the one-shot setting, and 41.5% in the few-shot setting.

<a id="S0219"></a> Source: p.13 S0219

This compares to 37.4% for fine-tuned T5-11B, and 44.7% for fine-tuned T5-11B+SSM, which uses a Q&A-specific pre-training procedure.

<a id="S0220"></a> Source: p.13 S0220

GPT-3 in the few-shot setting approaches the performance of state-of-the-art fine-tuned models.

<a id="S0221"></a> Source: p.13 S0221

Notably, compared to TriviaQA, WebQS shows a much larger gain from zero-shot to few-shot (and indeed its zero-shot and one-shot performance are poor), perhaps suggesting that the WebQs questions 13

Page 14

<a id="S0222"></a> Source: p.14 S0222

Figure 3.3: On TriviaQA GPT3’s performance grows smoothly with model size, suggesting that language models continue to absorb knowledge as their capacity increases.

<a id="S0223"></a> Source: p.14 S0223

One-shot and few-shot performance make significant gains over zero-shot behavior, matching and exceeding the performance of the SOTA fine-tuned open-domain model, RAG [LPP+20] and/or the style of their answers are out-of-distribution for GPT-3.

<a id="S0224"></a> Source: p.14 S0224

Nevertheless, GPT-3 appears able to adapt to this distribution, recovering strong performance in the few-shot setting.

<a id="S0225"></a> Source: p.14 S0225

On Natural Questions (NQs) GPT-3 achieves 14.6% in the zero-shot setting, 23.0% in the one-shot setting, and 29.9% in the few-shot setting, compared to 36.6% for fine-tuned T5 11B+SSM.

<a id="S0226"></a> Source: p.14 S0226

Similar to WebQS, the large gain from zero-shot to few-shot may suggest a distribution shift, and may also explain the less competitive performance compared to TriviaQA and WebQS.

<a id="S0227"></a> Source: p.14 S0227

In particular, the questions in NQs tend towards very fine-grained knowledge on Wikipedia specifically which could be testing the limits of GPT-3’s capacity and broad pretraining distribution.

<a id="S0228"></a> Source: p.14 S0228

Overall, on one of the three datasets GPT-3’s one-shot matches the open-domain fine-tuning SOTA.

<a id="S0229"></a> Source: p.14 S0229

On the other two datasets it approaches the performance of the closed-book SOTA despite not using fine-tuning.

<a id="S0230"></a> Source: p.14 S0230

On all 3 datasets, we find that performance scales very smoothly with model size (Figure 3.3 and Appendix H Figure H.7), possibly reflecting the idea that model capacity translates directly to more ‘knowledge’ absorbed in the parameters of the model. 3.3 Translation For GPT-2 a filter was used on a multilingual collection of documents to produce an English only dataset due to capacity concerns.

<a id="S0231"></a> Source: p.14 S0231

Even with this filtering GPT-2 showed some evidence of multilingual capability and performed non-trivially when translating between French and English despite only training on 10 megabytes of remaining French text.

<a id="S0232"></a> Source: p.14 S0232

Since we increase the capacity by over two orders of magnitude from GPT-2 to GPT-3, we also expand the scope of the training dataset to include more representation of other languages, though this remains an area for further improvement.

<a id="S0233"></a> Source: p.14 S0233

As discussed in 2.2 the majority of our data is derived from raw Common Crawl with only quality-based filtering.

<a id="S0234"></a> Source: p.14 S0234

Although GPT-3’s training data is still primarily English (93% by word count), it also includes 7% of text in other languages.

<a id="S0235"></a> Source: p.14 S0235

These languages are documented in the supplemental material.

<a id="S0236"></a> Source: p.14 S0236

In order to better understand translation capability, we also expand our analysis to include two additional commonly studied languages, German and Romanian.

<a id="S0237"></a> Source: p.14 S0237

Existing unsupervised machine translation approaches often combine pretraining on a pair of monolingual datasets with back-translation [SHB15] to bridge the two languages in a controlled way.

<a id="S0238"></a> Source: p.14 S0238

By contrast, GPT-3 learns from a blend of training data that mixes many languages together in a natural way, combining them on a word, sentence, and document level.

<a id="S0239"></a> Source: p.14 S0239

GPT-3 also uses a single training objective which is not customized or designed for any task in particular.

<a id="S0240"></a> Source: p.14 S0240

However, our one / few-shot settings aren’t strictly comparable to prior unsupervised work since they make use of a small amount of paired examples (1 or 64).

<a id="S0241"></a> Source: p.14 S0241

This corresponds to up to a page or two of in-context training data.

<a id="S0242"></a> Source: p.14 S0242

Zero-shot GPT-3, which only receives on a natural language description of the task, still underperforms recent unsupervised NMT results.

<a id="S0243"></a> Source: p.14 S0243

However, providing only a single example demonstration for 14

Page 15

<a id="S0244"></a> Source: p.15 S0244

Setting En→Fr Fr→En En→De De→En En→Ro Ro→En SOTA (Supervised) 45.6a 35.0 b 41.2c 40.2d 38.5e 39.9e XLM [LC19] 33.4 33.3 26.4 34.3 33.3 31.8 MASS [STQ+19] 37.5 34.9 28.3 35.2 35.2 33.1 mBART [LGG+20] - - 29.8 34.0 35.0 30.5 GPT-3 Zero-Shot 25.2 21.2 24.6 27.2 14.1 19.9 GPT-3 One-Shot 28.3 33.7 26.2 30.4 20.6 38.6 GPT-3 Few-Shot 32.6 39.2 29.7 40.6 21.0 39.5 Table 3.4: Few-shot GPT-3 outperforms previous unsupervised NMT work by 5 BLEU when translating into English reflecting its strength as an English LM.

<a id="S0245"></a> Source: p.15 S0245

We report BLEU scores on the WMT’14 Fr↔En, WMT’16 De↔En, and WMT’16 Ro↔En datasets as measured by multi-bleu.perl with XLM’s tokenization in order to compare most closely with prior unsupervised NMT work.

<a id="S0246"></a> Source: p.15 S0246

SacreBLEUf [Pos18] results reported in Appendix H.

<a id="S0247"></a> Source: p.15 S0247

Underline indicates an unsupervised or few-shot SOTA, bold indicates supervised SOTA with relative confidence. a[EOAG18] b[DHKH14] c[WXH+18] d[oR16] e[LGG+20] f [SacreBLEU signature: BLEU+case.mixed+numrefs.1+smooth.exp+tok.intl+version.1.2.20] Figure 3.4: Few-shot translation performance on 6 language pairs as model capacity increases.

<a id="S0248"></a> Source: p.15 S0248

There is a consistent trend of improvement across all datasets as the model scales, and as well as tendency for translation into English to be stronger than translation from English. 15

Page 16

<a id="S0249"></a> Source: p.16 S0249

Setting Winograd Winogrande (XL) Fine-tuned SOTA 90.1a 84.6b GPT-3 Zero-Shot 88.3 70.2 GPT-3 One-Shot 89.7 73.2 GPT-3 Few-Shot 88.6* 77.7 Table 3.5: Results on the WSC273 version of Winograd schemas and the adversarial Winogrande dataset.

<a id="S0250"></a> Source: p.16 S0250

See Section 4 for details on potential contamination of the Winograd test set. a[SBBC19] b[LYN+20] Figure 3.5: Zero-, one-, and few-shot performance on the adversarial Winogrande dataset as model capacity scales.

<a id="S0251"></a> Source: p.16 S0251

Scaling is relatively smooth with the gains to few-shot learning increasing with model size, and few-shot GPT-3 175B is competitive with a fine-tuned RoBERTA-large. each translation task improves performance by over 7 BLEU and nears competitive performance with prior work.

<a id="S0252"></a> Source: p.16 S0252

GPT-3 in the full few-shot setting further improves another 4 BLEU resulting in similar average performance to prior unsupervised NMT work.

<a id="S0253"></a> Source: p.16 S0253

GPT-3 has a noticeable skew in its performance depending on language direction.

<a id="S0254"></a> Source: p.16 S0254

For the three input languages studied, GPT-3 significantly outperforms prior unsupervised NMT work when translating into English but underperforms when translating in the other direction.

<a id="S0255"></a> Source: p.16 S0255

Performance on En-Ro is a noticeable outlier at over 10 BLEU worse than prior unsupervised NMT work.

<a id="S0256"></a> Source: p.16 S0256

This could be a weakness due to reusing the byte-level BPE tokenizer of GPT-2 which was developed for an almost entirely English training dataset.

<a id="S0257"></a> Source: p.16 S0257

For both Fr-En and De-En, few shot GPT-3 outperforms the best supervised result we could find but due to our unfamiliarity with the literature and the appearance that these are un-competitive benchmarks we do not suspect those results represent true state of the art.

<a id="S0258"></a> Source: p.16 S0258

For Ro-En, few shot GPT-3 performs within 0.5 BLEU of the overall SOTA which is achieved by a combination of unsupervised pretraining, supervised finetuning on 608K labeled examples, and backtranslation [LHCG19b].

<a id="S0259"></a> Source: p.16 S0259

Finally, across all language pairs and across all three settings (zero-, one-, and few-shot), there is a smooth trend of improvement with model capacity.

<a id="S0260"></a> Source: p.16 S0260

This is shown in Figure 3.4 in the case of few-shot results, and scaling for all three settings is shown in Appendix H. 3.4 Winograd-Style Tasks The Winograd Schemas Challenge [LDM12] is a classical task in NLP that involves determining which word a pronoun refers to, when the pronoun is grammatically ambiguous but semantically unambiguous to a human.

<a id="S0261"></a> Source: p.16 S0261

Recently fine-tuned language models have achieved near-human performance on the original Winograd dataset, but more difficult versions 16

Page 17

<a id="S0262"></a> Source: p.17 S0262

Setting PIQA ARC (Easy) ARC (Challenge) OpenBookQA Fine-tuned SOTA 79.4 92.0[KKS+20] 78.5[KKS+20] 87.2[KKS+20] GPT-3 Zero-Shot 80.5 68.8 51.4 57.6 GPT-3 One-Shot 80.5 71.2 53.2 58.8 GPT-3 Few-Shot 82.8* 70.1 51.5 65.4 Table 3.6: GPT-3 results on three commonsense reasoning tasks, PIQA, ARC, and OpenBookQA.

<a id="S0263"></a> Source: p.17 S0263

GPT-3 Few-Shot PIQA result is evaluated on the test server.

<a id="S0264"></a> Source: p.17 S0264

See Section 4 for details on potential contamination issues on the PIQA test set.

<a id="S0265"></a> Source: p.17 S0265

Figure 3.6: GPT-3 results on PIQA in the zero-shot, one-shot, and few-shot settings.

<a id="S0266"></a> Source: p.17 S0266

The largest model achieves a score on the development set in all three conditions that exceeds the best recorded score on the task. such as the adversarially-mined Winogrande dataset [SBBC19] still significantly lag human performance.

<a id="S0267"></a> Source: p.17 S0267

We test GPT-3’s performance on both Winograd and Winogrande, as usual in the zero-, one-, and few-shot setting.

<a id="S0268"></a> Source: p.17 S0268

On Winograd we test GPT-3 on the original set of 273 Winograd schemas, using the same “partial evaluation” method described in [RWC+19].

<a id="S0269"></a> Source: p.17 S0269

Note that this setting differs slightly from the WSC task in the SuperGLUE benchmark, which is presented as binary classification and requires entity extraction to convert to the form described in this section.

<a id="S0270"></a> Source: p.17 S0270

On Winograd GPT-3 achieves 88.3%, 89.7%, and 88.6% in the zero-shot, one-shot, and few-shot settings, showing no clear in-context learning but in all cases achieving strong results just a few points below state-of-the-art and estimated human performance.

<a id="S0271"></a> Source: p.17 S0271

We note that contamination analysis found some Winograd schemas in the training data but this appears to have only a small effect on results (see Section 4).

<a id="S0272"></a> Source: p.17 S0272

On the more difficult Winogrande dataset, we do find gains to in-context learning: GPT-3 achieves 70.2% in the zero-shot setting, 73.2% in the one-shot setting, and 77.7% in the few-shot setting.

<a id="S0273"></a> Source: p.17 S0273

For comparison a fine-tuned RoBERTA model achieves 79%, state-of-the-art is 84.6% achieved with a fine-tuned high capacity model (T5), and human performance on the task as reported by [SBBC19] is 94.0%. 3.5 Common Sense Reasoning Next we consider three datasets which attempt to capture physical or scientific reasoning, as distinct from sentence completion, reading comprehension, or broad knowledge question answering.

<a id="S0274"></a> Source: p.17 S0274

The first, PhysicalQA (PIQA) [BZB+19], asks common sense questions about how the physical world works and is intended as a probe of grounded understanding of the world.

<a id="S0275"></a> Source: p.17 S0275

GPT-3 achieves 81.0% accuracy zero-shot, 80.5% accuracy one-shot, and 82.8% accuracy few-shot (the last measured on PIQA’s test server).

<a id="S0276"></a> Source: p.17 S0276

This compares favorably to the 79.4% accuracy prior state-of-the-art of a 17

Page 18

<a id="S0277"></a> Source: p.18 S0277

Setting CoQA DROP QuAC SQuADv2 RACE-h RACE-m Fine-tuned SOTA 90.7a 89.1b 74.4c 93.0d 90.0e 93.1e GPT-3 Zero-Shot 81.5 23.6 41.5 59.5 45.5 58.4 GPT-3 One-Shot 84.0 34.3 43.3 65.4 45.9 57.4 GPT-3 Few-Shot 85.0 36.5 44.3 69.8 46.8 58.1 Table 3.7: Results on reading comprehension tasks.

<a id="S0278"></a> Source: p.18 S0278

All scores are F1 except results for RACE which report accuracy. a[JZC+19] b[JN20] c[AI19] d[QIA20] e[SPP+19] fine-tuned RoBERTa.

<a id="S0279"></a> Source: p.18 S0279

PIQA shows relatively shallow scaling with model size and is still over 10% worse than human performance, but GPT-3’s few-shot and even zero-shot result outperform the current state-of-the-art.

<a id="S0280"></a> Source: p.18 S0280

Our analysis flagged PIQA for a potential data contamination issue (despite hidden test labels), and we therefore conservatively mark the result with an asterisk.

<a id="S0281"></a> Source: p.18 S0281

ARC [CCE+18] is a dataset of multiple-choice questions collected from 3rd to 9th grade science exams.

<a id="S0282"></a> Source: p.18 S0282

On the “Challenge” version of the dataset which has been filtered to questions which simple statistical or information retrieval methods are unable to correctly answer, GPT-3 achieves 51.4% accuracy in the zero-shot setting, 53.2% in the one-shot setting, and 51.5% in the few-shot setting.

<a id="S0283"></a> Source: p.18 S0283

This is approaching the performance of a fine-tuned RoBERTa baseline (55.9%) from UnifiedQA [KKS+20].

<a id="S0284"></a> Source: p.18 S0284

On the “Easy” version of the dataset (questions which either of the mentioned baseline approaches answered correctly), GPT-3 achieves 68.8%, 71.2%, and 70.1% which slightly exceeds a fine-tuned RoBERTa baseline from [KKS+20].

<a id="S0285"></a> Source: p.18 S0285

However, both of these results are still much worse than the overall SOTAs achieved by the UnifiedQA which exceeds GPT-3’s few-shot results by 27% on the challenge set and 22% on the easy set.

<a id="S0286"></a> Source: p.18 S0286

On OpenBookQA [MCKS18], GPT-3 improves significantly from zero to few shot settings but is still over 20 points short of the overall SOTA.

<a id="S0287"></a> Source: p.18 S0287

GPT-3’s few-shot performance is similar to a fine-tuned BERT Large baseline on the leaderboard.

<a id="S0288"></a> Source: p.18 S0288

Overall, in-context learning with GPT-3 shows mixed results on commonsense reasoning tasks, with only small and inconsistent gains observed in the one and few-shot learning settings for both PIQA and ARC, but a significant improvement is observed on OpenBookQA.

<a id="S0289"></a> Source: p.18 S0289

GPT-3 sets SOTA on the new PIQA dataset in all evaluation settings. 3.6 Reading Comprehension Next we evaluate GPT-3 on the task of reading comprehension.

<a id="S0290"></a> Source: p.18 S0290

We use a suite of 5 datasets including abstractive, multiple choice, and span based answer formats in both dialog and single question settings.

<a id="S0291"></a> Source: p.18 S0291

We observe a wide spread in GPT-3’s performance across these datasets suggestive of varying capability with different answer formats.

<a id="S0292"></a> Source: p.18 S0292

In general we observe GPT-3 is on par with initial baselines and early results trained using contextual representations on each respective dataset.

<a id="S0293"></a> Source: p.18 S0293

GPT-3 performs best (within 3 points of the human baseline) on CoQA [RCM19] a free-form conversational dataset and performs worst (13 F1 below an ELMo baseline) on QuAC [CHI+18] a dataset which requires modeling structured dialog acts and answer span selections of teacher-student interactions.

<a id="S0294"></a> Source: p.18 S0294

On DROP [DWD+19], a dataset testing discrete reasoning and numeracy in the context of reading comprehension, GPT-3 in a few-shot setting outperforms the fine-tuned BERT baseline from the original paper but is still well below both human performance and state-of-the-art approaches which augment neural networks with symbolic systems [RLL+19].

<a id="S0295"></a> Source: p.18 S0295

On SQuAD 2.0 [RJL18], GPT-3 demonstrates its few-shot learning capabilities, improving by almost 10 F1 (to 69.8) compared to a zero-shot setting.

<a id="S0296"></a> Source: p.18 S0296

This allows it to slightly outperform the best fine-tuned result in the original paper.

<a id="S0297"></a> Source: p.18 S0297

On RACE [LXL+17], a multiple choice dataset of middle school and high school english examinations, GPT-3 performs relatively weakly and is only competitive with the earliest work utilizing contextual representations and is still 45% behind SOTA. 3.7 SuperGLUE In order to better aggregate results on NLP tasks and compare to popular models such as BERT and RoBERTa in a more systematic way, we also evaluate GPT-3 on a standardized collection of datasets, the SuperGLUE benchmark [WPN+19] [WPN+19] [CLC+19] [DMST19] [RBG11] [KCR+18] [ZLL+18] [DGM06] [BHDD+06] [GMDD07] [BDD+09] [PCC18] [PHR+18].

<a id="S0298"></a> Source: p.18 S0298

GPT-3’s test-set performance on the SuperGLUE dataset is shown in Table 3.8.

<a id="S0299"></a> Source: p.18 S0299

In the few-shot setting, we used 32 examples for all tasks, sampled randomly from the training set.

Page 19

<a id="S0300"></a> Source: p.19 S0300

Figure 3.7: GPT-3 results on CoQA reading comprehension task.

<a id="S0301"></a> Source: p.19 S0301

GPT-3 175B achieves 85 F1 in the few-shot setting, only a few points behind measured human performance and state-of-the-art fine-tuned models.

<a id="S0302"></a> Source: p.19 S0302

Zero-shot and one-shot performance is a few points behind, with the gains to few-shot being largest for bigger models.

<a id="S0303"></a> Source: p.19 S0303

SuperGLUE BoolQ CB CB COPA RTE Average Accuracy Accuracy F1 Accuracy Accuracy Fine-tuned SOTA 89.0 91.0 96.9 93.9 94.8 92.5 Fine-tuned BERT-Large 69.0 77.4 83.6 75.7 70.6 71.7 GPT-3 Few-Shot 71.8 76.4 75.6 52.0 92.0 69.0 WiC WSC MultiRC MultiRC ReCoRD ReCoRD Accuracy Accuracy Accuracy F1a Accuracy F1 Fine-tuned SOTA 76.1 93.8 62.3 88.2 92.5 93.3 Fine-tuned BERT-Large 69.6 64.6 24.1 70.0 71.3 72.0 GPT-3 Few-Shot 49.4 80.1 30.5 75.4 90.2 91.1 Table 3.8: Performance of GPT-3 on SuperGLUE compared to fine-tuned baselines and SOTA.

<a id="S0304"></a> Source: p.19 S0304

All results are reported on the test set.

<a id="S0305"></a> Source: p.19 S0305

GPT-3 few-shot is given a total of 32 examples within the context of each task and performs no gradient updates. 19

Page 20

<a id="S0306"></a> Source: p.20 S0306

Figure 3.8: Performance on SuperGLUE increases with model size and number of examples in context. A value of K = 32 means that our model was shown 32 examples per task, for 256 examples total divided across the 8 tasks in SuperGLUE.

<a id="S0307"></a> Source: p.20 S0307

We report GPT-3 values on the dev set, so our numbers are not directly comparable to the dotted reference lines (our test set results are in Table 3.8).

<a id="S0308"></a> Source: p.20 S0308

The BERT-Large reference model was fine-tuned on the SuperGLUE training set (125K examples), whereas BERT++ was first fine-tuned on MultiNLI (392K examples) and SWAG (113K examples) before further fine-tuning on the SuperGLUE training set (for a total of 630K fine-tuning examples).

<a id="S0309"></a> Source: p.20 S0309

We find the difference in performance between the BERT-Large and BERT++ to be roughly equivalent to the difference between GPT-3 with one example per context versus eight examples per context. and MultiRC, we sampled a new set of examples to use in the context for each problem.

<a id="S0310"></a> Source: p.20 S0310

For WSC and MultiRC, we used the same set of randomly drawn examples from the training set as context for all of the problems we evaluated.

<a id="S0311"></a> Source: p.20 S0311

We observe a wide range in GPT-3’s performance across tasks.

<a id="S0312"></a> Source: p.20 S0312

On COPA and ReCoRD GPT-3 achieves near-SOTA performance in the one-shot and few-shot settings, with COPA falling only a couple points short and achieving second place on the leaderboard, where first place is held by a fine-tuned 11 billion parameter model (T5).

<a id="S0313"></a> Source: p.20 S0313

On WSC, performance is still relatively strong, achieving 80.1% in the few-shot setting (note that GPT-3 achieves 88.6% on the original Winograd dataset as described in Section 3.4).

<a id="S0314"></a> Source: p.20 S0314

On BoolQ, MultiRC, and RTE, performance is reasonable, roughly matching that of a fine-tuned BERT-Large.

<a id="S0315"></a> Source: p.20 S0315

On CB, we see signs of life at 75.6% in the few-shot setting.

<a id="S0316"></a> Source: p.20 S0316

WiC is a notable weak spot with few-shot performance at 49.4% (at random chance).

<a id="S0317"></a> Source: p.20 S0317

We tried a number of different phrasings and formulations for WiC (which involves determining if a word is being used with the same meaning in two sentences), none of which was able to achieve strong performance.

<a id="S0318"></a> Source: p.20 S0318

This hints at a phenomenon that will become clearer in the next section (which discusses the ANLI benchmark) – GPT-3 appears to be weak in the few-shot or one-shot setting at some tasks that involve comparing two sentences or snippets, for example whether a word is used the same way in two sentences (WiC), whether one sentence is a paraphrase of another, or whether one sentence implies another.

<a id="S0319"></a> Source: p.20 S0319

This could also explain the comparatively low scores for RTE and CB, which also follow this format.

<a id="S0320"></a> Source: p.20 S0320

Despite these weaknesses, GPT-3 still outperforms a fine-tuned BERT-large on four of eight tasks and on two tasks GPT-3 is close to the state-of-the-art held by a fine-tuned 11 billion parameter model.

<a id="S0321"></a> Source: p.20 S0321

Finally, we note that the few-shot SuperGLUE score steadily improves with both model size and with number of examples in the context showing increasing benefits from in-context learning (Figure 3.8).

<a id="S0322"></a> Source: p.20 S0322

We scale K up to 32 examples per task, after which point additional examples will not reliably fit into our context.

<a id="S0323"></a> Source: p.20 S0323

When sweeping over values of K, we find that GPT-3 requires less than eight total examples per task to outperform a fine-tuned BERT-Large on overall SuperGLUE score. 3.8 NLI Natural Language Inference (NLI) [Fyo00] concerns the ability to understand the relationship between two sentences.

<a id="S0324"></a> Source: p.20 S0324

In practice, this task is usually structured as a two or three class classification problem where the model classifies 20

Page 21

<a id="S0325"></a> Source: p.21 S0325

Figure 3.9: Performance of GPT-3 on ANLI Round 3.

<a id="S0326"></a> Source: p.21 S0326

Results are on the dev-set, which has only 1500 examples and therefore has high variance (we estimate a standard deviation of 1.2%).

<a id="S0327"></a> Source: p.21 S0327

We find that smaller models hover around random chance, while few-shot GPT-3 175B closes almost half the gap from random chance to SOTA.

<a id="S0328"></a> Source: p.21 S0328

Results for ANLI rounds 1 and 2 are shown in the appendix. whether the second sentence logically follows from the first, contradicts the first sentence, or is possibly true (neutral).

<a id="S0329"></a> Source: p.21 S0329

SuperGLUE includes an NLI dataset, RTE, which evaluates the binary version of the task.

<a id="S0330"></a> Source: p.21 S0330

On RTE, only the largest version of GPT-3 performs convincingly better than random (56%) in any evaluation setting, but in a few-shot setting GPT-3 performs similarly to a single-task fine-tuned BERT Large.

<a id="S0331"></a> Source: p.21 S0331

We also evaluate on the recently introduced Adversarial Natural Language Inference (ANLI) dataset [NWD+19].

<a id="S0332"></a> Source: p.21 S0332

ANLI is a difficult dataset employing a series of adversarially mined natural language inference questions in three rounds (R1, R2, and R3).

<a id="S0333"></a> Source: p.21 S0333

Similar to RTE, all of our models smaller than GPT-3 perform at almost exactly random chance on ANLI, even in the few-shot setting (∼ 33%), whereas GPT-3 itself shows signs of life on Round 3.

<a id="S0334"></a> Source: p.21 S0334

Results for ANLI R3 are highlighted in Figure 3.9 and full results for all rounds can be found in Appendix H.

<a id="S0335"></a> Source: p.21 S0335

These results on both RTE and ANLI suggest that NLI is still a very difficult task for language models and they are only just beginning to show signs of progress. 3.9 Synthetic and Qualitative Tasks One way to probe GPT-3’s range of abilities in the few-shot (or zero- and one-shot) setting is to give it tasks which require it to perform simple on-the-fly computational reasoning, recognize a novel pattern that is unlikely to have occurred in training, or adapt quickly to an unusual task.

<a id="S0336"></a> Source: p.21 S0336

We devise several tasks to test this class of abilities.

<a id="S0337"></a> Source: p.21 S0337

First, we test GPT-3’s ability to perform arithmetic.

<a id="S0338"></a> Source: p.21 S0338

Second, we create several tasks that involve rearranging or unscrambling the letters in a word, tasks which are unlikely to have been exactly seen during training.

<a id="S0339"></a> Source: p.21 S0339

Third, we test GPT-3’s ability to solve SAT-style analogy problems few-shot.

<a id="S0340"></a> Source: p.21 S0340

Finally, we test GPT-3 on several qualitative tasks, including using new words in a sentence, correcting English grammar, and news article generation.

<a id="S0341"></a> Source: p.21 S0341

We will release the synthetic datasets with the hope of stimulating further study of test-time behavior of language models. 3.9.1 Arithmetic To test GPT-3’s ability to perform simple arithmetic operations without task-specific training, we developed a small battery of 10 tests that involve asking GPT-3 a simple arithmetic problem in natural language: • 2 digit addition (2D+) – The model is asked to add two integers sampled uniformly from [0, 100), phrased in the form of a question, e.g. “Q: What is 48 plus 76? A: 124.” • 2 digit subtraction (2D-) – The model is asked to subtract two integers sampled uniformly from [0, 100); the answer may be negative.

<a id="S0342"></a> Source: p.21 S0342

Example: “Q: What is 34 minus 53? A: -19”. • 3 digit addition (3D+) – Same as 2 digit addition, except numbers are uniformly sampled from [0, 1000). 21

Page 22

<a id="S0343"></a> Source: p.22 S0343

Figure 3.10: Results on all 10 arithmetic tasks in the few-shot settings for models of different sizes.

<a id="S0344"></a> Source: p.22 S0344

There is a significant jump from the second largest model (GPT-3 13B) to the largest model (GPT-3 175), with the latter being able to reliably accurate 2 digit arithmetic, usually accurate 3 digit arithmetic, and correct answers a significant fraction of the time on 4-5 digit arithmetic, 2 digit multiplication, and compound operations.

<a id="S0345"></a> Source: p.22 S0345

Results for one-shot and zero-shot are shown in the appendix. • 3 digit subtraction (3D-) – Same as 2 digit subtraction, except numbers are uniformly sampled from [0, 1000). • 4 digit addition (4D+) – Same as 3 digit addition, except uniformly sampled from [0, 10000). • 4 digit subtraction (4D-) – Same as 3 digit subtraction, except uniformly sampled from [0, 10000). • 5 digit addition (5D+) – Same as 3 digit addition, except uniformly sampled from [0, 100000). • 5 digit subtraction (5D-) – Same as 3 digit subtraction, except uniformly sampled from [0, 100000). • 2 digit multiplication (2Dx) – The model is asked to multiply two integers sampled uniformly from [0, 100), e.g. “Q: What is 24 times 42? A: 1008”. • One-digit composite (1DC) – The model is asked to perform a composite operation on three 1 digit numbers, with parentheses around the last two.

<a id="S0346"></a> Source: p.22 S0346

For example, “Q: What is 6+(4*8)? A: 38”.

<a id="S0347"></a> Source: p.22 S0347

The three 1 digit numbers are selected uniformly on [0, 10) and the operations are selected uniformly from {+,-,*}.

<a id="S0348"></a> Source: p.22 S0348

In all 10 tasks the model must generate the correct answer exactly.

<a id="S0349"></a> Source: p.22 S0349

For each task we generate a dataset of 2,000 random instances of the task and evaluate all models on those instances.

<a id="S0350"></a> Source: p.22 S0350

First we evaluate GPT-3 in the few-shot setting, for which results are shown in Figure 3.10.

<a id="S0351"></a> Source: p.22 S0351

On addition and subtraction, GPT-3 displays strong proficiency when the number of digits is small, achieving 100% accuracy on 2 digit addition, 98.9% at 2 digit subtraction, 80.2% at 3 digit addition, and 94.2% at 3-digit subtraction.

<a id="S0352"></a> Source: p.22 S0352

Performance decreases as the number of digits increases, but GPT-3 still achieves 25-26% accuracy on four digit operations and 9-10% accuracy on five digit operations, suggesting at least some capacity to generalize to larger numbers of digits.

<a id="S0353"></a> Source: p.22 S0353

GPT-3 also achieves 29.2% accuracy at 2 digit multiplication, an especially computationally intensive operation.

<a id="S0354"></a> Source: p.22 S0354

Finally, GPT-3 achieves 21.3% accuracy at single digit combined operations (for example, 9*(7+5)), suggesting that it has some robustness beyond just single operations.

<a id="S0355"></a> Source: p.22 S0355

As Figure 3.10 makes clear, small models do poorly on all of these tasks – even the 13 billion parameter model (the second largest after the 175 billion full GPT-3) can solve 2 digit addition and subtraction only half the time, and all other operations less than 10% of the time.

<a id="S0356"></a> Source: p.22 S0356

One-shot and zero-shot performance are somewhat degraded relative to few-shot performance, suggesting that adaptation to the task (or at the very least recognition of the task) is important to performing these computations correctly.

<a id="S0357"></a> Source: p.22 S0357

Nevertheless, one-shot performance is still quite strong, and even zero-shot performance of the full GPT-3 significantly 22

Page 23

<a id="S0358"></a> Source: p.23 S0358

Setting 2D+ 2D- 3D+ 3D- 4D+ 4D- 5D+ 5D- 2Dx 1DC GPT-3 Zero-shot 76.9 58.0 34.2 48.3 4.0 7.5 0.7 0.8 19.8 9.8 GPT-3 One-shot 99.6 86.4 65.5 78.7 14.0 14.0 3.5 3.8 27.4 14.3 GPT-3 Few-shot 100.0 98.9 80.4 94.2 25.5 26.8 9.3 9.9 29.2 21.3 Table 3.9: Results on basic arithmetic tasks for GPT-3 175B. {2,3,4,5}D{+,-} is 2, 3, 4, and 5 digit addition or subtraction, 2Dx is 2 digit multiplication. 1DC is 1 digit composite operations.

<a id="S0359"></a> Source: p.23 S0359

Results become progressively stronger moving from the zero-shot to one-shot to few-shot setting, but even the zero-shot shows significant arithmetic abilities.

<a id="S0360"></a> Source: p.23 S0360

Setting CL A1 A2 RI RW GPT-3 Zero-shot 3.66 2.28 8.91 8.26 0.09 GPT-3 One-shot 21.7 8.62 25.9 45.4 0.48 GPT-3 Few-shot 37.9 15.1 39.7 67.2 0.44 Table 3.10: GPT-3 175B performance on various word unscrambling and word manipulation tasks, in zero-, one-, and few-shot settings.

<a id="S0361"></a> Source: p.23 S0361

CL is “cycle letters in word”, A1 is anagrams of but the first and last letters, A2 is anagrams of all but the first and last two letters, RI is “Random insertion in word”, RW is “reversed words”. outperforms few-shot learning for all smaller models.

<a id="S0362"></a> Source: p.23 S0362

All three settings for the full GPT-3 are shown in Table 3.9, and model capacity scaling for all three settings is shown in Appendix H.

<a id="S0363"></a> Source: p.23 S0363

To spot-check whether the model is simply memorizing specific arithmetic problems, we took the 3-digit arithmetic problems in our test set and searched for them in our training data in both the forms "<NUM1> + <NUM2> =" and "<NUM1> plus <NUM2>".

<a id="S0364"></a> Source: p.23 S0364

Out of 2,000 addition problems we found only 17 matches (0.8%) and out of 2,000 subtraction problems we found only 2 matches (0.1%), suggesting that only a trivial fraction of the correct answers could have been memorized.

<a id="S0365"></a> Source: p.23 S0365

In addition, inspection of incorrect answers reveals that the model often makes mistakes such as not carrying a “1”, suggesting it is actually attempting to perform the relevant computation rather than memorizing a table.

<a id="S0366"></a> Source: p.23 S0366

Overall, GPT-3 displays reasonable proficiency at moderately complex arithmetic in few-shot, one-shot, and even zero-shot settings. 3.9.2 Word Scrambling and Manipulation Tasks To test GPT-3’s ability to learn novel symbolic manipulations from a few examples, we designed a small battery of 5 “character manipulation” tasks.

<a id="S0367"></a> Source: p.23 S0367

Each task involves giving the model a word distorted by some combination of scrambling, addition, or deletion of characters, and asking it to recover the original word.

<a id="S0368"></a> Source: p.23 S0368

The 5 tasks are: • Cycle letters in word (CL) – The model is given a word with its letters cycled, then the “=” symbol, and is expected to generate the original word.

<a id="S0369"></a> Source: p.23 S0369

For example, it might be given “lyinevitab” and should output “inevitably”. • Anagrams of all but first and last characters (A1) – The model is given a word where every letter except the first and last have been scrambled randomly, and must output the original word.

<a id="S0370"></a> Source: p.23 S0370

Example: criroptuon = corruption. • Anagrams of all but first and last 2 characters (A2) – The model is given a word where every letter except the first 2 and last 2 have been scrambled randomly, and must recover the original word.

<a id="S0371"></a> Source: p.23 S0371

Example: opoepnnt → opponent. • Random insertion in word (RI) – A random punctuation or space character is inserted between each letter of a word, and the model must output the original word.

<a id="S0372"></a> Source: p.23 S0372

Example: s.u!c/c!e.s s i/o/n = succession. • Reversed words (RW) – The model is given a word spelled backwards, and must output the original word.

<a id="S0373"></a> Source: p.23 S0373

For each task we generate 10,000 examples, which we chose to be the top 10,000 most frequent words as measured by [Nor09] of length more than 4 characters and less than 15 characters.

<a id="S0374"></a> Source: p.23 S0374

The few-shot results are shown in Figure 3.11.

<a id="S0375"></a> Source: p.23 S0375

Task performance tends to grow smoothly with model size, with the full GPT-3 model achieving 66.9% on removing 23

Page 24

<a id="S0376"></a> Source: p.24 S0376

Figure 3.11: Few-shot performance on the five word scrambling tasks for different sizes of model.

<a id="S0377"></a> Source: p.24 S0377

There is generally smooth improvement with model size although the random insertion task shows an upward slope of improvement with the 175B model solving the task the majority of the time.

<a id="S0378"></a> Source: p.24 S0378

Scaling of one-shot and zero-shot performance is shown in the appendix.

<a id="S0379"></a> Source: p.24 S0379

All tasks are done with K = 100. random insertions, 38.6% on cycling letters, 40.2% on the easier anagram task, and 15.1% on the more difficult anagram task (where only the first and last letters are held fixed).

<a id="S0380"></a> Source: p.24 S0380

None of the models can reverse the letters in a word.

<a id="S0381"></a> Source: p.24 S0381

In the one-shot setting, performance is significantly weaker (dropping by half or more), and in the zero-shot setting the model can rarely perform any of the tasks (Table 3.10).

<a id="S0382"></a> Source: p.24 S0382

This suggests that the model really does appear to learn these tasks at test time, as the model cannot perform them zero-shot and their artificial nature makes them unlikely to appear in the pre-training data (although we cannot confirm this with certainty).

<a id="S0383"></a> Source: p.24 S0383

We can further quantify performance by plotting “in-context learning curves”, which show task performance as a function of the number of in-context examples.

<a id="S0384"></a> Source: p.24 S0384

We show in-context learning curves for the Symbol Insertion task in Figure 1.2.

<a id="S0385"></a> Source: p.24 S0385

We can see that larger models are able to make increasingly effective use of in-context information, including both task examples and natural language task descriptions.

<a id="S0386"></a> Source: p.24 S0386

Finally, it is worth adding that solving these tasks requires character-level manipulations, whereas our BPE encoding operates on significant fractions of a word (on average ∼ 0.7 words per token), so from the LM’s perspective succeeding at these tasks involves not just manipulating BPE tokens but understanding and pulling apart their substructure.

<a id="S0387"></a> Source: p.24 S0387

Also, CL, A1, and A2 are not bijective (that is, the unscrambled word is not a deterministic function of the scrambled word), requiring the model to perform some search to find the correct unscrambling.

<a id="S0388"></a> Source: p.24 S0388

Thus, the skills involved appear to require non-trivial pattern-matching and computation. 3.9.3 SAT Analogies To test GPT-3 on another task that is somewhat unusual relative to the typical distribution of text, we collected a set of 374 “SAT analogy” problems [TLBS03].

<a id="S0389"></a> Source: p.24 S0389

Analogies are a style of multiple choice question that constituted a section of the SAT college entrance exam before 2005. A typical example is “audacious is to boldness as (a) sanctimonious is to hypocrisy, (b) anonymous is to identity, (c) remorseful is to misdeed, (d) deleterious is to result, (e) impressionable is to temptation”.

<a id="S0390"></a> Source: p.24 S0390

The student is expected to choose which of the five word pairs has the same relationship as the original word pair; in this example the answer is “sanctimonious is to hypocrisy”.

<a id="S0391"></a> Source: p.24 S0391

On this task GPT-3 achieves 65.2% in the few-shot setting, 59.1% in the one-shot setting, and 53.7% in the zero-shot setting, whereas the average score among college applicants was 57% [TL05] (random guessing yields 20%).

<a id="S0392"></a> Source: p.24 S0392

As shown in Figure 3.12, the results improve with scale, with the the full 175 billion model improving by over 10% compared to the 13 billion parameter model. 24

Page 25

<a id="S0393"></a> Source: p.25 S0393

Figure 3.12: Zero-, one-,and few-shot performance on SAT analogy tasks, for different sizes of model.

<a id="S0394"></a> Source: p.25 S0394

The largest model achieves 65% accuracy in the few-shot setting, and also demonstrates significant gains to in-context learning which are not present in smaller models. 3.9.4 News Article Generation Previous work on generative language models qualitatively tested their ability to generate synthetic “news articles” by conditional sampling from the model given a human-written prompt consisting of a plausible first sentence for a news story [RWC+19].

<a id="S0395"></a> Source: p.25 S0395

Relative to [RWC+19], the dataset used to train GPT-3 is much less weighted towards news articles, so trying to generate news articles via raw unconditional samples is less effective – for example GPT-3 often interprets the proposed first sentence of a “news article” as a tweet and then posts synthetic responses or follow-up tweets.

<a id="S0396"></a> Source: p.25 S0396

To solve this problem we employed GPT-3’s few-shot learning abilities by providing three previous news articles in the model’s context to condition it.

<a id="S0397"></a> Source: p.25 S0397

With the title and subtitle of a proposed next article, the model is able to reliably generate short articles in the “news” genre.

<a id="S0398"></a> Source: p.25 S0398

To gauge the quality of news article generation from GPT-3 (which we believe is likely to be correlated with conditional sample generation quality in general), we decided to measure human ability to distinguish GPT-3-generated articles from real ones.

<a id="S0399"></a> Source: p.25 S0399

Similar work has been carried out by Kreps et al. [KMB20] and Zellers et al. [ZHR+19].

<a id="S0400"></a> Source: p.25 S0400

Generative language models are trained to match the distribution of content generated by humans, so the (in)ability of humans to distinguish the two is a potentially important measure of quality.3 In order to see how well humans can detect model generated text, we arbitrarily selected 25 article titles and subtitles from the website newser.com (mean length: 215 words).

<a id="S0401"></a> Source: p.25 S0401

We then generated completions of these titles and subtitles from four language models ranging in size from 125M to 175B (GPT-3) parameters (mean length: 200 words).

<a id="S0402"></a> Source: p.25 S0402

For each model, we presented around 80 US-based participants with a quiz consisting of these real titles and subtitles followed by either the human written article or the article generated by the model4.

<a id="S0403"></a> Source: p.25 S0403

Participants were asked to select whether the article was “very likely written by a human”, “more likely written by a human”, “I don’t know”, “more likely written by a machine”, or “very likely written by a machine”.

<a id="S0404"></a> Source: p.25 S0404

The articles we selected were not in the models’ training data and the model outputs were formatted and selected programmatically to prevent human cherry-picking.

<a id="S0405"></a> Source: p.25 S0405

All models used the same context to condition outputs on and were pre-trained with the same context size and the same article titles and subtitles were used as prompts for each model.

<a id="S0406"></a> Source: p.25 S0406

However, we also ran an experiment to control for participant effort and attention that followed the same format but involved intentionally bad model generated articles.

<a id="S0407"></a> Source: p.25 S0407

This was done by generating articles from a “control model”: a 160M parameter model with no context and increased output randomness. 3This task is also relevant to the potential misuse of language models discussed in Section 6.1. 4We wanted to identify how good an average person on the internet is at detecting language model outputs, so we focused on participants drawn from the general US population.

Page 26

<a id="S0408"></a> Source: p.26 S0408

95% Confidence t compared to “I don’t know” Mean accuracy Interval (low, hi) control (p-value) assignments Control (deliberately bad model) 86% 83%–90% - 3.6 % GPT-3 Small 76% 72%–80% 3.9 (2e-4) 4.9% GPT-3 Medium 61% 58%–65% 10.3 (7e-21) 6.0% GPT-3 Large 68% 64%–72% 7.3 (3e-11) 8.7% GPT-3 XL 62% 59%–65% 10.7 (1e-19) 7.5% GPT-3 2.7B 62% 58%–65% 10.4 (5e-19) 7.1% GPT-3 6.7B 60% 56%–63% 11.2 (3e-21) 6.2% GPT-3 13B 55% 52%–58% 15.3 (1e-32) 7.1% GPT-3 175B 52% 49%–54% 16.9 (1e-34) 7.8% Table 3.11: Human accuracy in identifying whether short (∼200 word) news articles are model generated.

<a id="S0409"></a> Source: p.26 S0409

We find that human accuracy (measured by the ratio of correct assignments to non-neutral assignments) ranges from 86% on the control model to 52% on GPT-3 175B.

<a id="S0410"></a> Source: p.26 S0410

This table compares mean accuracy between five different models, and shows the results of a two-sample T-Test for the difference in mean accuracy between each model and the control model (an unconditional GPT-3 Small model with increased output randomness).

<a id="S0411"></a> Source: p.26 S0411

Mean human accuracy (the ratio of correct assignments to non-neutral assignments per participant) at detecting that the intentionally bad articles were model generated was ∼ 86% where 50% is chance level performance.

<a id="S0412"></a> Source: p.26 S0412

By contrast, mean human accuracy at detecting articles that were produced by the 175B parameter model was barely above chance at ∼ 52% (see Table 3.11).5 Human abilities to detect model generated text appear to decrease as model size increases: there appears to be a trend towards chance accuracy with model size, and human detection of GPT-3 is close to chance.6 This is true despite the fact that participants spend more time on each output as model size increases (see Appendix E).

<a id="S0413"></a> Source: p.26 S0413

Examples of synthetic articles from GPT-3 are given in Figures 3.14 and 3.15.7 Much of the text is—as indicated by the evaluations—difficult for humans to distinguish from authentic human content.

<a id="S0414"></a> Source: p.26 S0414

Factual inaccuracies can be an indicator that an article is model generated since, unlike human authors, the models have no access to the specific facts that the article titles refer to or when the article was written.

<a id="S0415"></a> Source: p.26 S0415

Other indicators include repetition, non sequiturs, and unusual phrasings, though these are often subtle enough that they are not noticed.

<a id="S0416"></a> Source: p.26 S0416

Related work on language model detection by Ippolito et al. [IDCBE19] indicates that automatic discriminators like G R O V E R [ZHR+19] and GLTR [GSR19] may have greater success at detecting model generated text than human evaluators.

<a id="S0417"></a> Source: p.26 S0417

Automatic detection of these models may be a promising area of future research.

<a id="S0418"></a> Source: p.26 S0418

Ippolito et al. [IDCBE19] also note that human accuracy at detecting model generated text increases as humans observe more tokens.

<a id="S0419"></a> Source: p.26 S0419

To do a preliminary investigation of how good humans are at detecting longer news articles generated by GPT-3 175B, we selected 12 world news articles from Reuters with an average length of 569 words and generated completions of these articles from GPT-3 with an average length of 498 words (298 words longer than our initial experiments).

<a id="S0420"></a> Source: p.26 S0420

Following the methodology above, we ran two experiments, each on around 80 US-based participants, to compare human abilities to detect the articles generated by GPT-3 and a control model.

<a id="S0421"></a> Source: p.26 S0421

We found that mean human accuracy at detecting the intentionally bad longer articles from the control model was ∼ 88%, while mean human accuracy at detecting the longer articles that were produced by GPT-3 175B was still barely above chance at ∼ 52% (see Table 3.12).

<a id="S0422"></a> Source: p.26 S0422

This indicates that, for news articles that are around 500 words long, GPT-3 continues to produce articles that humans find difficult to distinguish from human written news articles. 3.9.5 Learning and Using Novel Words A task studied in developmental linguistics [CB78] is the ability to learn and utilize new words, for example using a word in a sentence after seeing it defined only once, or conversely inferring a word’s meaning from only one usage.

<a id="S0423"></a> Source: p.26 S0423

Here we qualitatively test GPT-3’s ability to do the former.

<a id="S0424"></a> Source: p.26 S0424

Specifically, we give GPT-3 the definition of a nonexistent word, such as “Gigamuru”, and then ask it to use it in a sentence.

<a id="S0425"></a> Source: p.26 S0425

We provide one to five previous examples of a (separate) 5We use a two-sample Student’s T-Test to test for significant difference between the means of the participant accuracies of each model and the control model and report the normalized difference in the means (as the t-statistic) and the p-value. 6If a model consistently produces texts that are more impressive than human articles, it is possible that human performance on this task would drop below 50%.

<a id="S0426"></a> Source: p.26 S0426

Indeed, many individual participants scored below 50% on this task. 7Additional non-news samples can be found in Appendix F. 26

Page 27

<a id="S0427"></a> Source: p.27 S0427

Figure 3.13: People’s ability to identify whether news articles are model-generated (measured by the ratio of correct assignments to non-neutral assignments) decreases as model size increases.

<a id="S0428"></a> Source: p.27 S0428

Accuracy on the outputs on the deliberatelybad control model (an unconditioned GPT-3 Small model with higher output randomness) is indicated with the dashed line at the top, and the random chance (50%) is indicated with the dashed line at the bottom.

<a id="S0429"></a> Source: p.27 S0429

Line of best fit is a power law with 95% confidence intervals. 95% Confidence t compared to “I don’t know” Mean accuracy Interval (low, hi) control (p-value) assignments Control 88% 84%–91% - 2.7% GPT-3 175B 52% 48%–57% 12.7 (3.2e-23) 10.6% Table 3.12: People’s ability to identify whether ∼ 500 word articles are model generated (as measured by the ratio of correct assignments to non-neutral assignments) was 88% on the control model and 52% on GPT-3 175B.

<a id="S0430"></a> Source: p.27 S0430

This table shows the results of a two-sample T-Test for the difference in mean accuracy between GPT-3 175B and the control model (an unconditional GPT-3 Small model with increased output randomness). 27

Page 28

<a id="S0431"></a> Source: p.28 S0431

Title: United Methodists Agree to Historic Split Subtitle: Those who oppose gay marriage will form their own denomination Article: After two days of intense debate, the United Methodist Church has agreed to a historic split - one that is expected to end in the creation of a new denomination, one that will be "theologically and socially conservative," according to The Washington Post.

<a id="S0432"></a> Source: p.28 S0432

The majority of delegates attending the church’s annual General Conference in May voted to strengthen a ban on the ordination of LGBTQ clergy and to write new rules that will "discipline" clergy who officiate at same-sex weddings.

<a id="S0433"></a> Source: p.28 S0433

But those who opposed these measures have a new plan: They say they will form a separate denomination by 2020, calling their church the Christian Methodist denomination.

<a id="S0434"></a> Source: p.28 S0434

The Post notes that the denomination, which claims 12.5 million members, was in the early 20th century the "largest Protestant denomination in the U.S.," but that it has been shrinking in recent decades.

<a id="S0435"></a> Source: p.28 S0435

The new split will be the second in the church’s history.

<a id="S0436"></a> Source: p.28 S0436

The first occurred in 1968, when roughly 10 percent of the denomination left to form the Evangelical United Brethren Church.

<a id="S0437"></a> Source: p.28 S0437

The Post notes that the proposed split "comes at a critical time for the church, which has been losing members for years," which has been "pushed toward the brink of a schism over the role of LGBTQ people in the church." Gay marriage is not the only issue that has divided the church.

<a id="S0438"></a> Source: p.28 S0438

In 2016, the denomination was split over ordination of transgender clergy, with the North Pacific regional conference voting to ban them from serving as clergy, and the South Pacific regional conference voting to allow them.

<a id="S0439"></a> Source: p.28 S0439

Figure 3.14: The GPT-3 generated news article that humans had the greatest difficulty distinguishing from a human written article (accuracy: 12%).

<a id="S0440"></a> Source: p.28 S0440

Title: Star’s Tux Promise Draws Megyn Kelly’s Sarcasm Subtitle: Joaquin Phoenix pledged to not change for each awards event Article: A year ago, Joaquin Phoenix made headlines when he appeared on the red carpet at the Golden Globes wearing a tuxedo with a paper bag over his head that read, "I am a shape-shifter. I can’t change the world. I can only change myself." It was a promise to not change to fit into the Hollywood mold: "I think that’s a really special thing, to not change yourself. I think it’s a really special thing to say, ‘This is what’s inside of me, I’m proud of it, and I’m not going to be ashamed because of the way that someone else thinks I should be.’" Now, it’s the Oscars, and Phoenix is at it again.

<a id="S0441"></a> Source: p.28 S0441

But this time, his publicist is saying he’ll be wearing a tux no matter what.

<a id="S0442"></a> Source: p.28 S0442

Megyn Kelly was not impressed, and she let him have it on The Tonight Show. "You know, I feel like, I feel like you could have worn the tux," she says. "But you’re saying you’re a shape-shifter. I don’t know if you can change your tux, but you can change your mind.

<a id="S0443"></a> Source: p.28 S0443

You can change your mind." Phoenix says he did, but it didn’t stick. "I was like, ‘Okay, I’m going to wear a tuxedo to this thing.’ And then I thought, ‘I don’t want to wear a tuxedo to this thing.’" Kelly goes on to encourage him to change his mind again, but Phoenix says it’s too late: "I’m committed to wearing this." Figure 3.15: The GPT-3 generated news article that humans found the easiest to distinguish from a human written article (accuracy: 61%). 28

Page 29

<a id="S0444"></a> Source: p.29 S0444

A "whatpu" is a small, furry animal native to Tanzania.

<a id="S0445"></a> Source: p.29 S0445

An example of a sentence that uses the word whatpu is: We were traveling in Africa and we saw these very cute whatpus.

<a id="S0446"></a> Source: p.29 S0446

To do a "farduddle" means to jump up and down really fast.

<a id="S0447"></a> Source: p.29 S0447

An example of a sentence that uses the word farduddle is: One day when I was playing tag with my little sister, she got really excited and she started doing these crazy farduddles. A "yalubalu" is a type of vegetable that looks like a big pumpkin.

<a id="S0448"></a> Source: p.29 S0448

An example of a sentence that uses the word yalubalu is: I was on a trip to Africa and I tried this yalubalu vegetable that was grown in a garden there.

<a id="S0449"></a> Source: p.29 S0449

It was delicious. A "Burringo" is a car with very fast acceleration.

<a id="S0450"></a> Source: p.29 S0450

An example of a sentence that uses the word Burringo is: In our garage we have a Burringo that my father drives to work every day. A "Gigamuru" is a type of Japanese musical instrument.

<a id="S0451"></a> Source: p.29 S0451

An example of a sentence that uses the word Gigamuru is: I have a Gigamuru that my uncle gave me as a gift. I love to play it at home.

<a id="S0452"></a> Source: p.29 S0452

To "screeg" something is to swing a sword at it.

<a id="S0453"></a> Source: p.29 S0453

An example of a sentence that uses the word screeg is: We screeghed at each other for several minutes and then we went outside and ate ice cream.

<a id="S0454"></a> Source: p.29 S0454

Figure 3.16: Representative GPT-3 completions for the few-shot task of using a new word in a sentence.

<a id="S0455"></a> Source: p.29 S0455

Boldface is GPT-3’s completions, plain text is human prompts.

<a id="S0456"></a> Source: p.29 S0456

In the first example both the prompt and the completion are provided by a human; this then serves as conditioning for subsequent examples where GPT-3 receives successive additional prompts and provides the completions.

<a id="S0457"></a> Source: p.29 S0457

Nothing task-specific is provided to GPT-3 other than the conditioning shown here. nonexistent word being defined and used in a sentence, so the task is few-shot in terms of previous examples of the broad task and one-shot in terms of the specific word.

<a id="S0458"></a> Source: p.29 S0458

Table 3.16 shows the 6 examples we generated; all definitions were human-generated, and the first answer was human-generated as conditioning while the subsequent answers were generated by GPT-3.

<a id="S0459"></a> Source: p.29 S0459

These examples were generated continuously in one sitting and we did not omit or repeatedly try any prompts.

<a id="S0460"></a> Source: p.29 S0460

In all cases the generated sentence appears to be a correct or at least plausible use of the word.

<a id="S0461"></a> Source: p.29 S0461

In the final sentence the model generates a plausible conjugation for the word “screeg” (namely “screeghed”), although the use of the word is slightly awkward (“screeghed at each other”) despite being plausible in the sense that it could describe a toy sword fight.

<a id="S0462"></a> Source: p.29 S0462

Overall, GPT-3 appears to be at least proficient at the task of using novel words in a sentence. 3.9.6 Correcting English Grammar Another task well suited for few-shot learning is correcting English grammar.

<a id="S0463"></a> Source: p.29 S0463

We test this with GPT-3 in the fewshot setting by giving prompts of the form "Poor English Input: <sentence>\n Good English Output: <sentence>".

<a id="S0464"></a> Source: p.29 S0464

We give GPT-3 one human-generated correction and then ask it to correct 5 more (again without any omissions or repeats).

<a id="S0465"></a> Source: p.29 S0465

Results are shown in Figure 3.17. 4 Measuring and Preventing Memorization Of Benchmarks Since our training dataset is sourced from the internet, it is possible that our model was trained on some of our benchmark test sets.

<a id="S0466"></a> Source: p.29 S0466

Accurately detecting test contamination from internet-scale datasets is a new area of research without established best practices.

<a id="S0467"></a> Source: p.29 S0467

While it is common practice to train large models without investigating contamination, given the increasing scale of pretraining datasets, we believe this issue is becoming increasingly important to attend to.

<a id="S0468"></a> Source: p.29 S0468

One of the first papers to train a language model on Common Crawl data [TL18] detected and removed a training document which overlapped with one of their evaluation datasets.

<a id="S0469"></a> Source: p.29 S0469

Other work such as GPT-2 [RWC+19] also conducted post-hoc overlap analysis.

<a id="S0470"></a> Source: p.29 S0470

Their study was relatively encouraging, finding that 29

Page 30

<a id="S0471"></a> Source: p.30 S0471

Poor English input: I eated the purple berries.

<a id="S0472"></a> Source: p.30 S0472

Good English output: I ate the purple berries.

<a id="S0473"></a> Source: p.30 S0473

Poor English input: Thank you for picking me as your designer. I’d appreciate it.

<a id="S0474"></a> Source: p.30 S0474

Good English output: Thank you for choosing me as your designer. I appreciate it.

<a id="S0475"></a> Source: p.30 S0475

Poor English input: The mentioned changes have done. or I did the alteration that you requested. or I changed things you wanted and did the modifications.

<a id="S0476"></a> Source: p.30 S0476

Good English output: The requested changes have been made. or I made the alteration that you requested. or I changed things you wanted and made the modifications.

<a id="S0477"></a> Source: p.30 S0477

Poor English input: I’d be more than happy to work with you in another project.

<a id="S0478"></a> Source: p.30 S0478

Good English output: I’d be more than happy to work with you on another project.

<a id="S0479"></a> Source: p.30 S0479

Poor English input: Please provide me with a short brief of the design you’re looking for and that’d be nice if you could share some examples or project you did before.

<a id="S0480"></a> Source: p.30 S0480

Good English output: Please provide me with a brief description of the design you’re looking for and that would be nice if you could share some examples or projects you have done before.

<a id="S0481"></a> Source: p.30 S0481

Poor English input: The patient was died.

<a id="S0482"></a> Source: p.30 S0482

Poor English input: We think that Leslie likes ourselves.

<a id="S0483"></a> Source: p.30 S0483

Good English output: We think that Leslie likes us.

<a id="S0484"></a> Source: p.30 S0484

Poor English input: Janet broke Bill on the finger.

<a id="S0485"></a> Source: p.30 S0485

Good English output: Janet broke Bill’s finger.

<a id="S0486"></a> Source: p.30 S0486

Poor English input: Mary arranged for, in St.

<a id="S0487"></a> Source: p.30 S0487

Good English output: Mary arranged for John to rent a house in St.

<a id="S0488"></a> Source: p.30 S0488

Poor English input: Today I have went to the store to to buys some many bottle of water.

<a id="S0489"></a> Source: p.30 S0489

Good English output: Today I went to the store to buy some bottles of water.

<a id="S0490"></a> Source: p.30 S0490

Poor English input: I have tried to hit ball with bat, but my swing is has miss.

<a id="S0491"></a> Source: p.30 S0491

Good English output: I tried to hit the ball with the bat, but my swing missed.

<a id="S0492"></a> Source: p.30 S0492

Figure 3.17: Representative GPT-3 completions for the few-shot task of correcting English grammar.

<a id="S0493"></a> Source: p.30 S0493

Boldface is GPT-3’s completions, plain text is human prompts.

<a id="S0494"></a> Source: p.30 S0494

In the first few examples example both the prompt and the completion are provided by a human; this then serves as conditioning for subsequent examples where GPT-3 receives successive additional prompts and provides the completions.

<a id="S0495"></a> Source: p.30 S0495

Nothing task-specific is provided to GPT-3 aside from the first few examples as conditioning and the “Poor English input/Good English output” framing.

<a id="S0496"></a> Source: p.30 S0496

We note that the distinction between ”poor” and ”good” English (and the terms themselves) is complex, contextual, and contested.

<a id="S0497"></a> Source: p.30 S0497

As the example mentioning the rental of a house shows, assumptions that the model makes about what “good” is can even lead it to make errors (here, the model not only adjusts grammar, but also removes the word ”cheap” in a way that alters meaning). 30

Page 31

<a id="S0498"></a> Source: p.31 S0498

Figure 4.1: GPT-3 Training Curves We measure model performance during training on a deduplicated validation split of our training distribution.

<a id="S0499"></a> Source: p.31 S0499

Though there is some gap between training and validation performance, the gap grows only minimally with model size and training time, suggesting that most of the gap comes from a difference in difficulty rather than overfitting. although models did perform moderately better on data that overlapped between training and testing, this did not significantly impact reported results due to the small fraction of data which was contaminated (often only a few percent).

<a id="S0500"></a> Source: p.31 S0500

GPT-3 operates in a somewhat different regime.

<a id="S0501"></a> Source: p.31 S0501

On the one hand, the dataset and model size are about two orders of magnitude larger than those used for GPT-2, and include a large amount of Common Crawl, creating increased potential for contamination and memorization.

<a id="S0502"></a> Source: p.31 S0502

On the other hand, precisely due to the large amount of data, even GPT-3 175B does not overfit its training set by a significant amount, measured relative to a held-out validation set with which it was deduplicated (Figure 4.1).

<a id="S0503"></a> Source: p.31 S0503

Thus, we expect that contamination is likely to be frequent, but that its effects may not be as large as feared.

<a id="S0504"></a> Source: p.31 S0504

We initially tried to address the issue of contamination by proactively searching for and attempting to remove any overlap between our training data and the development and test sets of all benchmarks studied in this paper.

<a id="S0505"></a> Source: p.31 S0505

Unfortunately, a bug resulted in only partial removal of all detected overlaps from the training data.

<a id="S0506"></a> Source: p.31 S0506

Due to the cost of training, it wasn’t feasible to retrain the model.

<a id="S0507"></a> Source: p.31 S0507

To address this, we investigate in detail how the remaining detected overlap impacts results.

<a id="S0508"></a> Source: p.31 S0508

For each benchmark, we produce a ‘clean’ version which removes all potentially leaked examples, defined roughly as examples that have a 13-gram overlap with anything in the pretraining set (or that overlap with the whole example when it is shorter than 13-grams).

<a id="S0509"></a> Source: p.31 S0509

The goal is to very conservatively flag anything that could potentially be contamination, so as to produce a clean subset that is free of contamination with high confidence.

<a id="S0510"></a> Source: p.31 S0510

The exact procedure is detailed in Appendix C.

<a id="S0511"></a> Source: p.31 S0511

We then evaluate GPT-3 on these clean benchmarks, and compare to the original score.

<a id="S0512"></a> Source: p.31 S0512

If the score on the clean subset is similar to the score on the entire dataset, this suggests that contamination, even if present, does not have a significant effect on reported results.

<a id="S0513"></a> Source: p.31 S0513

If the score on the clean subset is lower, this suggests contamination may be inflating the results.

<a id="S0514"></a> Source: p.31 S0514

The results are summarized in Figure 4.2.

<a id="S0515"></a> Source: p.31 S0515

Although potential contamination is often high (with a quarter of benchmarks scoring over 50%), in most cases performance changes only negligibly, and we see no evidence that contamination level and performance difference are correlated.

<a id="S0516"></a> Source: p.31 S0516

We conclude that either our conservative method substantially overestimated contamination or that contamination has little effect on performance.

<a id="S0517"></a> Source: p.31 S0517

Below, we review in more detail the few specific cases where either (1) the model performs significantly worse on the cleaned version, or (2) potential contamination is very high, which makes measuring the performance difference difficult.

<a id="S0518"></a> Source: p.31 S0518

Our analysis flagged six groups of benchmarks for further investigation: Word Scrambling, Reading Comprehension (QuAC, SQuAD2, DROP), PIQA, Winograd, language modeling tasks (Wikitext tasks, 1BW), and German to English 31

Page 32

<a id="S0519"></a> Source: p.32 S0519

Figure 4.2: Benchmark contamination analysis We constructed cleaned versions of each of our benchmarks to check for potential contamination in our training set.

<a id="S0520"></a> Source: p.32 S0520

The x-axis is a conservative lower bound for how much of the dataset is known with high confidence to be clean, and the y-axis shows the difference in performance when evaluating only on the verified clean subset.

<a id="S0521"></a> Source: p.32 S0521

Performance on most benchmarks changed negligibly, but some were flagged for further review.

<a id="S0522"></a> Source: p.32 S0522

On inspection we find some evidence for contamination of the PIQA and Winograd results, and we mark the corresponding results in Section 3 with an asterisk.

<a id="S0523"></a> Source: p.32 S0523

We find no evidence that other benchmarks are affected. translation.

<a id="S0524"></a> Source: p.32 S0524

Since our overlap analysis is designed to be extremely conservative, we expect it to produce some false positives.

<a id="S0525"></a> Source: p.32 S0525

We summarize the results for each group of tasks below: • Reading Comprehension: Our initial analysis flagged >90% of task examples from QuAC, SQuAD2, and DROP as potentially contaminated, so large that even measuring the differential on a clean subset was difficult.

<a id="S0526"></a> Source: p.32 S0526

Upon manual inspection, however, we found that for every overlap we inspected, in all 3 datasets, the source text was present in our training data but the question/answer pairs were not, meaning the model gains only background information and cannot memorize the answer to a specific question. • German translation: We found 25% of the examples in the WMT16 German-English test set were marked as potentially contaminated, with an associated total effect size of 1-2 BLEU.

<a id="S0527"></a> Source: p.32 S0527

Upon inspection, none of the flagged examples contain paired sentences resembling NMT training data and collisions were monolingual matches mostly of snippets of events discussed in the news. • Reversed Words and Anagrams: Recall that these tasks are of the form “alaok = koala”.

<a id="S0528"></a> Source: p.32 S0528

Due to the short length of these tasks, we used 2-grams for filtering (ignoring punctuation).

<a id="S0529"></a> Source: p.32 S0529

After inspecting the flagged overlaps, we found that they were not typically instances of real reversals or unscramblings in the training set, but rather palindromes or trivial unscramblings, e.g “kayak = kayak”.

<a id="S0530"></a> Source: p.32 S0530

The amount of overlap was small, but removing the trivial tasks lead to an increase in difficulty and thus a spurious signal.

<a id="S0531"></a> Source: p.32 S0531

Related to this, the symbol insertion task shows high overlap but no effect on performance – this is because that task involves removing non-letter characters from a word, and the overlap analysis itself ignores such characters, leading to many spurious matches. • PIQA: The overlap analysis flagged 29% of examples as contaminated, and observed a 3 percentage point absolute decrease (4% relative decrease) in performance on the clean subset.

<a id="S0532"></a> Source: p.32 S0532

Though the test dataset was released after our training set was created and its labels are hidden, some of the web pages used by the crowdsourced dataset creators are contained in our training set.

<a id="S0533"></a> Source: p.32 S0533

We found a similar decrease in a 25x smaller model with much less capacity to memorize, leading us to suspect that the shift is likely statistical bias rather than memorization; examples which workers copied may simply be easier.

<a id="S0534"></a> Source: p.32 S0534

Unfortunately, we cannot rigorously prove this hypothesis.

<a id="S0535"></a> Source: p.32 S0535

We therefore mark our PIQA results with an asterisk to denote this potential contamination. • Winograd: The overlap analysis flagged 45% of examples, and found a 2.6% decrease in performance on the clean subset.

<a id="S0536"></a> Source: p.32 S0536

Manual inspection of the overlapping data point showed that 132 Winograd schemas were in fact present in our training set, though presented in a different format than we present the task to the model.

<a id="S0537"></a> Source: p.32 S0537

Although the decrease in performance is small, we mark our Winograd results in the main paper with an asterisk. 32

Page 33

<a id="S0538"></a> Source: p.33 S0538

• Language modeling: We found the 4 Wikipedia language modeling benchmarks measured in GPT-2, plus the Children’s Book Test dataset, to be almost entirely contained in our training data.

<a id="S0539"></a> Source: p.33 S0539

Since we cannot reliably extract a clean subset here, we do not report results on these datasets, even though we intended to when starting this work.

<a id="S0540"></a> Source: p.33 S0540

We note that Penn Tree Bank due to its age was unaffected and therefore became our chief language modeling benchmark.

<a id="S0541"></a> Source: p.33 S0541

We also inspected datasets where contamination was high, but the impact on performance was close to zero, simply to verify how much actual contamination existed.

<a id="S0542"></a> Source: p.33 S0542

These appeared to often contain false positives.

<a id="S0543"></a> Source: p.33 S0543

They had either no actual contamination, or had contamination that did not give away the answer to the task.

<a id="S0544"></a> Source: p.33 S0544

One notable exception was LAMBADA, which appeared to have substantial genuine contamination, yet the impact on performance was very small, with the clean subset scoring within 0.5% of the full dataset.

<a id="S0545"></a> Source: p.33 S0545

Also, strictly speaking, our fill-in-the-blank format precludes the simplest form of memorization.

<a id="S0546"></a> Source: p.33 S0546

Nevertheless, since we made very large gains on LAMBADA in this paper, the potential contamination is noted in the results section.

<a id="S0547"></a> Source: p.33 S0547

An important limitation of our contamination analysis is that we cannot be sure that the clean subset is drawn from the same distribution as the original dataset.

<a id="S0548"></a> Source: p.33 S0548

It remains possible that memorization inflates results but at the same time is precisely counteracted by some statistical bias causing the clean subset to be easier.

<a id="S0549"></a> Source: p.33 S0549

However, the sheer number of shifts close to zero suggests this is unlikely, and we also observed no noticeable difference in the shifts for small models, which are unlikely to be memorizing.

<a id="S0550"></a> Source: p.33 S0550

Overall, we have made a best effort to measure and document the effects of data contamination, and to note or outright remove problematic results, depending on the severity.

<a id="S0551"></a> Source: p.33 S0551

Much work remains to be done to address this important and subtle issue for the field in general, both when designing benchmarks and when training models.

<a id="S0552"></a> Source: p.33 S0552

For a more detailed explanation of our analysis, we refer the reader to Appendix C. 5 Limitations GPT-3 and our analysis of it have a number of limitations.

<a id="S0553"></a> Source: p.33 S0553

Below we describe some of these and suggest directions for future work.

<a id="S0554"></a> Source: p.33 S0554

First, despite the strong quantitative and qualitative improvements of GPT-3, particularly compared to its direct predecessor GPT-2, it still has notable weaknesses in text synthesis and several NLP tasks.

<a id="S0555"></a> Source: p.33 S0555

On text synthesis, although the overall quality is high, GPT-3 samples still sometimes repeat themselves semantically at the document level, start to lose coherence over sufficiently long passages, contradict themselves, and occasionally contain non-sequitur sentences or paragraphs.

<a id="S0556"></a> Source: p.33 S0556

We will release a collection of 500 uncurated unconditional samples to help provide a better sense of GPT-3’s limitations and strengths at text synthesis.

<a id="S0557"></a> Source: p.33 S0557

Within the domain of discrete language tasks, we have noticed informally that GPT-3 seems to have special difficulty with “common sense physics”, despite doing well on some datasets (such as PIQA [BZB+19]) that test this domain.

<a id="S0558"></a> Source: p.33 S0558

Specifically GPT-3 has difficulty with questions of the type “If I put cheese into the fridge, will it melt?”.

<a id="S0559"></a> Source: p.33 S0559

Quantitatively, GPT-3’s in-context learning performance has some notable gaps on our suite of benchmarks, as described in Section 3, and in particular it does little better than chance when evaluated one-shot or even few-shot on some “comparison” tasks, such as determining if two words are used the same way in a sentence, or if one sentence implies another (WIC and ANLI respectively), as well as on a subset of reading comprehension tasks.

<a id="S0560"></a> Source: p.33 S0560

This is especially striking given GPT-3’s strong few-shot performance on many other tasks.

<a id="S0561"></a> Source: p.33 S0561

GPT-3 has several structural and algorithmic limitations, which could account for some of the issues above.

<a id="S0562"></a> Source: p.33 S0562

We focused on exploring in-context learning behavior in autoregressive language models because it is straightforward to both sample and compute likelihoods with this model class.

<a id="S0563"></a> Source: p.33 S0563

As a result our experiments do not include any bidirectional architectures or other training objectives such as denoising.

<a id="S0564"></a> Source: p.33 S0564

This is a noticeable difference from much of the recent literature, which has documented improved fine-tuning performance when using these approaches over standard language models [RSR+19].

<a id="S0565"></a> Source: p.33 S0565

Thus our design decision comes at the cost of potentially worse performance on tasks which empirically benefit from bidirectionality.

<a id="S0566"></a> Source: p.33 S0566

This may include fill-in-the-blank tasks, tasks that involve looking back and comparing two pieces of content, or tasks that require re-reading or carefully considering a long passage and then generating a very short answer.

<a id="S0567"></a> Source: p.33 S0567

This could be a possible explanation for GPT-3’s lagging few-shot performance on a few of the tasks, such as WIC (which involves comparing the use of a word in two sentences), ANLI (which involves comparing two sentences to see if one implies the other), and several reading comprehension tasks (e.g.

<a id="S0568"></a> Source: p.33 S0568

We also conjecture, based on past literature, that a large bidirectional model would be stronger at fine-tuning than GPT-3.

<a id="S0569"></a> Source: p.33 S0569

Making a bidirectional model at the scale of GPT-3, and/or trying to make bidirectional models work with few- or zero-shot learning, is a promising direction for future research, and could help achieve the “best of both worlds”. A more fundamental limitation of the general approach described in this paper – scaling up any LM-like model, whether autoregressive or bidirectional – is that it may eventually run into (or could already be running into) the limits of the 33

Page 34

<a id="S0570"></a> Source: p.34 S0570

Our current objective weights every token equally and lacks a notion of what is most important to predict and what is less important. [RRS20] demonstrate benefits of customizing prediction to entities of interest.

<a id="S0571"></a> Source: p.34 S0571

Also, with self-supervised objectives, task specification relies on forcing the desired task into a prediction problem, whereas ultimately, useful language systems (for example virtual assistants) might be better thought of as taking goal-directed actions rather than just making predictions.

<a id="S0572"></a> Source: p.34 S0572

Finally, large pretrained language models are not grounded in other domains of experience, such as video or real-world physical interaction, and thus lack a large amount of context about the world [BHT+20].

<a id="S0573"></a> Source: p.34 S0573

For all these reasons, scaling pure self-supervised prediction is likely to hit limits, and augmentation with a different approach is likely to be necessary.

<a id="S0574"></a> Source: p.34 S0574

Promising future directions in this vein might include learning the objective function from humans [ZSW+19a], fine-tuning with reinforcement learning, or adding additional modalities such as images to provide grounding and a better model of the world [CLY+19].

<a id="S0575"></a> Source: p.34 S0575

Another limitation broadly shared by language models is poor sample efficiency during pre-training.

<a id="S0576"></a> Source: p.34 S0576

While GPT-3 takes a step towards test-time sample efficiency closer to that of humans (one-shot or zero-shot), it still sees much more text during pre-training than a human sees in the their lifetime [Lin20].

<a id="S0577"></a> Source: p.34 S0577

Improving pre-training sample efficiency is an important direction for future work, and might come from grounding in the physical world to provide additional information, or from algorithmic improvements. A limitation, or at least uncertainty, associated with few-shot learning in GPT-3 is ambiguity about whether few-shot learning actually learns new tasks “from scratch” at inference time, or if it simply recognizes and identifies tasks that it has learned during training.

<a id="S0578"></a> Source: p.34 S0578

These possibilities exist on a spectrum, ranging from demonstrations in the training set that are drawn from exactly the same distribution as those at test time, to recognizing the same task but in a different format, to adapting to a specific style of a general task such as QA, to learning a skill entirely de novo.

<a id="S0579"></a> Source: p.34 S0579

Where GPT-3 is on this spectrum may also vary from task to task.

<a id="S0580"></a> Source: p.34 S0580

Synthetic tasks such as wordscrambling or defining nonsense words seem especially likely to be learned de novo, whereas translation clearly must be learned during pretraining, although possibly from data that is very different in organization and style than the test data.

<a id="S0581"></a> Source: p.34 S0581

Ultimately, it is not even clear what humans learn from scratch vs from prior demonstrations.

<a id="S0582"></a> Source: p.34 S0582

Even organizing diverse demonstrations during pre-training and identifying them at test time would be an advance for language models, but nevertheless understanding precisely how few-shot learning works is an important unexplored direction for future research. A limitation associated with models at the scale of GPT-3, regardless of objective function or algorithm, is that they are both expensive and inconvenient to perform inference on, which may present a challenge for practical applicability of models of this scale in their current form.

<a id="S0583"></a> Source: p.34 S0583

One possible future direction to address this is distillation [HVD15] of large models down to a manageable size for specific tasks.

<a id="S0584"></a> Source: p.34 S0584

Large models such as GPT-3 contain a very wide range of skills, most of which are not needed for a specific task, suggesting that in principle aggressive distillation may be possible.

<a id="S0585"></a> Source: p.34 S0585

Distillation is well-explored in general [LHCG19a] but has not been tried at the scale of hundred of billions parameters; new challenges and opportunities may be associated with applying it to models of this size.

<a id="S0586"></a> Source: p.34 S0586

Finally, GPT-3 shares some limitations common to most deep learning systems – its decisions are not easily interpretable, it is not necessarily well-calibrated in its predictions on novel inputs as observed by the much higher variance in performance than humans on standard benchmarks, and it retains the biases of the data it has been trained on.

<a id="S0587"></a> Source: p.34 S0587

This last issue – biases in the data that may lead the model to generate stereotyped or prejudiced content – is of special concern from a societal perspective, and will be discussed along with other issues in the next section on Broader Impacts (Section 6). 6 Broader Impacts Language models have a wide range of beneficial applications for society, including code and writing auto-completion, grammar assistance, game narrative generation, improving search engine responses, and answering questions.

<a id="S0588"></a> Source: p.34 S0588

But they also have potentially harmful applications.

<a id="S0589"></a> Source: p.34 S0589

GPT-3 improves the quality of text generation and adaptability over smaller models and increases the difficulty of distinguishing synthetic text from human-written text.

<a id="S0590"></a> Source: p.34 S0590

It therefore has the potential to advance both the beneficial and harmful applications of language models.

<a id="S0591"></a> Source: p.34 S0591

Here we focus on the potential harms of improved language models, not because we believe the harms are necessarily greater, but in order to stimulate efforts to study and mitigate them.

<a id="S0592"></a> Source: p.34 S0592

The broader impacts of language models like this are numerous.

<a id="S0593"></a> Source: p.34 S0593

We focus on two primary issues: the potential for deliberate misuse of language models like GPT-3 in Section 6.1, and issues of bias, fairness, and representation within models like GPT-3 in Section 6.2.

<a id="S0594"></a> Source: p.34 S0594

We also briefly discuss issues of energy efficiency (Section 6.3). 34

Page 35

<a id="S0595"></a> Source: p.35 S0595

6.1 Misuse of Language Models Malicious uses of language models can be somewhat difficult to anticipate because they often involve repurposing language models in a very different environment or for a different purpose than researchers intended.

<a id="S0596"></a> Source: p.35 S0596

To help with this, we can think in terms of traditional security risk assessment frameworks, which outline key steps such as identifying threats and potential impacts, assessing likelihood, and determining risk as a combination of likelihood and impact [Ros12].

<a id="S0597"></a> Source: p.35 S0597

We discuss three factors: potential misuse applications, threat actors, and external incentive structures. 6.1.1 Potential Misuse Applications Any socially harmful activity that relies on generating text could be augmented by powerful language models.

<a id="S0598"></a> Source: p.35 S0598

Examples include misinformation, spam, phishing, abuse of legal and governmental processes, fraudulent academic essay writing and social engineering pretexting.

<a id="S0599"></a> Source: p.35 S0599

Many of these applications bottleneck on human beings to write sufficiently high quality text.

<a id="S0600"></a> Source: p.35 S0600

Language models that produce high quality text generation could lower existing barriers to carrying out these activities and increase their efficacy.

<a id="S0601"></a> Source: p.35 S0601

The misuse potential of language models increases as the quality of text synthesis improves.

<a id="S0602"></a> Source: p.35 S0602

The ability of GPT-3 to generate several paragraphs of synthetic content that people find difficult to distinguish from human-written text in 3.9.4 represents a concerning milestone in this regard. 6.1.2 Threat Actor Analysis Threat actors can be organized by skill and resource levels, ranging from low or moderately skilled and resourced actors who may be able to build a malicious product to ‘advanced persistent threats’ (APTs): highly skilled and well-resourced (e.g. state-sponsored) groups with long-term agendas [SBC+19].

<a id="S0603"></a> Source: p.35 S0603

To understand how low and mid-skill actors think about language models, we have been monitoring forums and chat groups where misinformation tactics, malware distribution, and computer fraud are frequently discussed.

<a id="S0604"></a> Source: p.35 S0604

While we did find significant discussion of misuse following the initial release of GPT-2 in spring of 2019, we found fewer instances of experimentation and no successful deployments since then.

<a id="S0605"></a> Source: p.35 S0605

Additionally, those misuse discussions were correlated with media coverage of language model technologies.

<a id="S0606"></a> Source: p.35 S0606

From this, we assess that the threat of misuse from these actors is not immediate, but significant improvements in reliability could change this.

<a id="S0607"></a> Source: p.35 S0607

Because APTs do not typically discuss operations in the open, we have consulted with professional threat analysts about possible APT activity involving the use of language models.

<a id="S0608"></a> Source: p.35 S0608

Since the release of GPT-2 there has been no discernible difference in operations that may see potential gains by using language models.

<a id="S0609"></a> Source: p.35 S0609

The assessment was that language models may not be worth investing significant resources in because there has been no convincing demonstration that current language models are significantly better than current methods for generating text, and because methods for “targeting” or “controlling” the content of language models are still at a very early stage. 6.1.3 External Incentive Structures Each threat actor group also has a set of tactics, techniques, and procedures (TTPs) that they rely on to accomplish their agenda.

<a id="S0610"></a> Source: p.35 S0610

TTPs are influenced by economic factors like scalability and ease of deployment; phishing is extremely popular among all groups because it offers a low-cost, low-effort, high-yield method of deploying malware and stealing login credentials.

<a id="S0611"></a> Source: p.35 S0611

Using language models to augment existing TTPs would likely result in an even lower cost of deployment.

<a id="S0612"></a> Source: p.35 S0612

Ease of use is another significant incentive.

<a id="S0613"></a> Source: p.35 S0613

Having stable infrastructure has a large impact on the adoption of TTPs.

<a id="S0614"></a> Source: p.35 S0614

The outputs of language models are stochastic, however, and though developers can constrain these (e.g. using top-k truncation) they are not able to perform consistently without human feedback.

<a id="S0615"></a> Source: p.35 S0615

If a social media disinformation bot produces outputs that are reliable 99% of the time, but produces incoherent outputs 1% of the time, this could reduce the amount of human labor required in operating this bot.

<a id="S0616"></a> Source: p.35 S0616

But a human is still needed to filter the outputs, which restricts how scalable the operation can be.

<a id="S0617"></a> Source: p.35 S0617

Based on our analysis of this model and analysis of threat actors and the landscape, we suspect AI researchers will eventually develop language models that are sufficiently consistent and steerable that they will be of greater interest to malicious actors.

<a id="S0618"></a> Source: p.35 S0618

We expect this will introduce challenges for the broader research community, and hope to work on this through a combination of mitigation research, prototyping, and coordinating with other technical developers. 35

Page 36

<a id="S0619"></a> Source: p.36 S0619

6.2 Fairness, Bias, and Representation Biases present in training data may lead models to generate stereotyped or prejudiced content.

<a id="S0620"></a> Source: p.36 S0620

This is concerning, since model bias could harm people in the relevant groups in different ways by entrenching existing stereotypes and producing demeaning portrayals amongst other potential harms [Cra17].

<a id="S0621"></a> Source: p.36 S0621

We have conducted an analysis of biases in the model in order to better understand GPT-3’s limitations when it comes to fairness, bias, and representation. 8 Our goal is not to exhaustively characterize GPT-3, but to give a preliminary analysis of some of its limitations and behaviors.

<a id="S0622"></a> Source: p.36 S0622

We focus on biases relating to gender, race, and religion, although many other categories of bias are likely present and could be studied in follow-up work.

<a id="S0623"></a> Source: p.36 S0623

This is a preliminary analysis and does not reflect all of the model’s biases even within the studied categories.

<a id="S0624"></a> Source: p.36 S0624

Broadly, our analysis indicates that internet-trained models have internet-scale biases; models tend to reflect stereotypes present in their training data.

<a id="S0625"></a> Source: p.36 S0625

Below we discuss our preliminary findings of bias along the dimensions of gender, race, and religion.

<a id="S0626"></a> Source: p.36 S0626

We probe for bias in the 175 billion parameter model and also in similar smaller models, to see if and how they are different in this dimension. 6.2.1 Gender In our investigation of gender bias in GPT-3, we focused on associations between gender and occupation.

<a id="S0627"></a> Source: p.36 S0627

We found that occupations in general have a higher probability of being followed by a male gender identifier than a female one (in other words, they are male leaning) when given a context such as "The {occupation} was a" (Neutral Variant). 83% of the 388 occupations we tested were more likely to be followed by a male identifier by GPT-3.

<a id="S0628"></a> Source: p.36 S0628

We measured this by feeding the model a context such as "The detective was a" and then looking at the probability of the model following up with male indicating words (eg. man, male etc.) or female indicating words (woman, female etc.).

<a id="S0629"></a> Source: p.36 S0629

In particular, occupations demonstrating higher levels of education such as legislator, banker, or professor emeritus were heavily male leaning along with occupations that require hard physical labour such as mason, millwright, and sheriff.

<a id="S0630"></a> Source: p.36 S0630

Occupations that were more likely to be followed by female identifiers include midwife, nurse, receptionist, housekeeper etc.

<a id="S0631"></a> Source: p.36 S0631

We also tested how these probabilities changed when we shifted the context to be the "The competent {occupation} was a" (Competent Variant), and when we shifted the context to be "The incompetent {occupation} was a" (Incompetent Variant) for each occupation in the dataset.

<a id="S0632"></a> Source: p.36 S0632

We found that, when prompted with "The competent {occupation} was a," the majority of occupations had an even higher probability of being followed by a male identifier than a female one than was the case with our original neutral prompt, "The {occupation} was a".

<a id="S0633"></a> Source: p.36 S0633

With the prompt "The incompetent {occupation} was a" the majority of occupations still leaned male with a similar probability than for our original neutral prompt.

<a id="S0634"></a> Source: p.36 S0634

The average occupation bias - measured as 1 (cid:80) log( P (female|Context) ) - was −1.11 for the Neutral Variant, −2.14 for the Competent Variant and −1.15 njobs jobs P (male|Context)) for the Incompetent Variant.

<a id="S0635"></a> Source: p.36 S0635

We also carried out pronoun resolution on the Winogender dataset [RNLVD18] using two methods which further corroborated the model’s tendency to associate most occupations with males.

<a id="S0636"></a> Source: p.36 S0636

One method measured the models ability to correctly assign a pronoun as the occupation or the participant.

<a id="S0637"></a> Source: p.36 S0637

For example, we fed the model a context such as "The advisor met with the advisee because she wanted to get advice about job applications. ‘She’ refers to the" and found the option with the lowest probability between the two possible options (Choices between Occupation Option: advisor; Participant Option: advisee).

<a id="S0638"></a> Source: p.36 S0638

Occupation and participant words often have societal biases associated with them such as the assumption that most occupants are by default male.

<a id="S0639"></a> Source: p.36 S0639

We found that the language models learnt some of these biases such as a tendency to associate female pronouns with participant positions more than male pronouns.

<a id="S0640"></a> Source: p.36 S0640

GPT-3 175B had the highest accuracy of all the models (64.17%) on this task.

<a id="S0641"></a> Source: p.36 S0641

It was also the only model where the accuracy for Occupant sentences (sentences where the correct answer was the Occupation option) for females was higher than for males (81.7% vs 76.7%).

<a id="S0642"></a> Source: p.36 S0642

All other models had a higher accuracy for male pronouns with Occupation sentences as compared to female pronouns with the exception of our second largest model- GPT-3 13B - which had the same accuracy (60%) for both.

<a id="S0643"></a> Source: p.36 S0643

This offers some preliminary evidence that in places where issues of bias can make language models susceptible to error, the larger models are more robust than smaller models.

<a id="S0644"></a> Source: p.36 S0644

We also performed co-occurrence tests, where we analyzed which words are likely to occur in the vicinity of other preselected words.

<a id="S0645"></a> Source: p.36 S0645

We created a model output sample set by generating 800 outputs of length 50 each with a temperature 8Evaluating fairness, bias, and representation in language models is a rapidly-developing area with a large body of prior work.

<a id="S0646"></a> Source: p.36 S0646

See, for example, [HZJ+19, NBR20, SCNP19]. 36

Page 37

<a id="S0647"></a> Source: p.37 S0647

Table 6.1: Most Biased Descriptive Words in 175B Model Top 10 Most Biased Male Descriptive Words with Raw Top 10 Most Biased Female Descriptive Words with Raw Co-Occurrence Counts Co-Occurrence Counts Average Number of Co-Occurrences Across All Words: Average Number of Co-Occurrences Across All Words: 17.5 23.9 Large (16) Optimistic (12) Mostly (15) Bubbly (12) Lazy (14) Naughty (12) Fantastic (13) Easy-going (12) Eccentric (13) Petite (10) Protect (10) Tight (10) Jolly (10) Pregnant (10) Stable (9) Gorgeous (28) Personable (22) Sucked (8) Survive (7) Beautiful (158) of 1 and top p of 0.9 for every prompt in our dataset.

<a id="S0648"></a> Source: p.37 S0648

For gender, we had prompts such as "He was very", "She was very", "He would be described as", "She would be described as"9.

<a id="S0649"></a> Source: p.37 S0649

We looked at the adjectives and adverbs in the top 100 most favored words using an off-the-shelf POS tagger [LB02].

<a id="S0650"></a> Source: p.37 S0650

We found females were more often described using appearance oriented words such as ”beautiful” and ”gorgeous” as compared to men who were more often described using adjectives that span a greater spectrum.

<a id="S0651"></a> Source: p.37 S0651

Table 6.1 shows the top 10 most favored descriptive words for the model along with the raw number of times each word co-occurred with a pronoun indicator. “Most Favored” here indicates words which were most skewed towards a category by co-occurring with it at a higher rate as compared to the other category.

<a id="S0652"></a> Source: p.37 S0652

To put these numbers in perspective, we have also included the average for the number of co-occurrences across all qualifying words for each gender. 6.2.2 Race To investigate racial bias in GPT-3, we seeded the model with prompts such as - "The {race} man was very", "The {race} woman was very" and "People would describe the {race} person as" and generated 800 samples for each of the above prompts, with {race} replaced with a term indicating a racial category such as White or Asian.

<a id="S0653"></a> Source: p.37 S0653

We then measure word co-occurrences in the generated samples.

<a id="S0654"></a> Source: p.37 S0654

Given prior research demonstrating that language models produce text of differing sentiment when varying features such as occupation [HZJ+19], we explored how race impacted sentiment.

<a id="S0655"></a> Source: p.37 S0655

We measured sentiment using Senti WordNet [BES10] for the words which co-occurred disproportionately with each race.

<a id="S0656"></a> Source: p.37 S0656

Each word sentiment varied from 100 to -100, with positive scores indicating positive words (eg. wonderfulness: 100, amicable: 87.5), negative scores indicating negative words (eg. wretched: -87.5 , horrid: -87.5) and a score of 0 indicating neutral words (eg. sloping, chalet).

<a id="S0657"></a> Source: p.37 S0657

It should be noted that we were explicitly prompting the models to talk about race and this in turn generated text that focused on racial features; these results are not from the models talking about race in the wild but talking about race in an experimental setup where they have been primed to do so.

<a id="S0658"></a> Source: p.37 S0658

Additionally, since we are measuring sentiment by simply looking at word co-occurrences, the resulting sentiment can reflect socio-historical factors - for instance, text relating to a discussion of slavery will frequently have a negative sentiment, which may lead to a demographic being associated with a negative sentiment under this testing methodology.

<a id="S0659"></a> Source: p.37 S0659

Across the models we analyzed, ‘Asian’ had a consistently high sentiment - it ranked 1st in 3 out of 7 models.

<a id="S0660"></a> Source: p.37 S0660

On the other hand, ’Black’ had a consistently low sentiment - it ranked the lowest in 5 out of 7 models.

<a id="S0661"></a> Source: p.37 S0661

These differences narrowed marginally on the larger model sizes.

<a id="S0662"></a> Source: p.37 S0662

This analysis gives a sense of the biases of different models and highlights the need for more sophisticated analysis of the relationship between sentiment, entities, and input data. 9We only used male and female pronouns.

<a id="S0663"></a> Source: p.37 S0663

This simplifying assumption makes it easier to study co-occurrence since it does not require the isolation of instances in which ‘they’ refers to a singular noun from those where it didn’t, but other forms of gender bias are likely present and could be studied using different approaches. 37

Page 38

<a id="S0664"></a> Source: p.38 S0664

Figure 6.1: Racial Sentiment Across Models Religion Most Favored Descriptive Words Atheism ‘Theists’, ‘Cool’, ‘Agnostics’, ‘Mad’, ‘Theism’, ‘Defensive’, ‘Complaining’, ‘Correct’, ‘Arrogant’, ‘Characterized’ Buddhism ‘Myanmar’, ‘Vegetarians’, ‘Burma’, ‘Fellowship’, ‘Monk’, ‘Japanese’, ‘Reluctant’, ‘Wisdom’, ‘Enlightenment’, ‘Non-Violent’ Christianity ‘Attend’, ‘Ignorant’, ‘Response’, ‘Judgmental’, ‘Grace’, ‘Execution’, ‘Egypt’, ‘Continue’, ‘Comments’, ‘Officially’ Hinduism ‘Caste’, ‘Cows’, ‘BJP’, ‘Kashmir’, ‘Modi’, ‘Celebrated’, ‘Dharma’, ‘Pakistani’, ‘Originated’, ‘Africa’ Islam ‘Pillars’, ‘Terrorism’, ‘Fasting’, ‘Sheikh’, ‘Non-Muslim’, ‘Source’, ‘Charities’, ‘Levant’, ‘Allah’, ‘Prophet’ Judaism ‘Gentiles’, ‘Race’, ‘Semites’, ‘Whites’, ‘Blacks’, ‘Smartest’, ‘Racists’, ‘Arabs’, ‘Game’, ‘Russian’ Table 6.2: Shows the ten most favored words about each religion in the GPT-3 175B model. 6.2.3 Religion We studied which words co-occurred with religious terms relating to Atheism, Buddhism, Christianity, Hinduism, Islam, and Judaism, by generating 800 model outputs of length ≈50 with a temperature of 1 and a top p of 0.9 for every prompt.

<a id="S0665"></a> Source: p.38 S0665

Our prompts were of the nature "{Religion practitioners} are" (Eg. "Christians are") for each of the six religious categories listed above.

<a id="S0666"></a> Source: p.38 S0666

We then allowed the model to naturally carry out completions and created a corpus of such completions for studying co-occurrence of words.

<a id="S0667"></a> Source: p.38 S0667

The following is an example output from the model: "Buddhists are divided into two main branches - Theravada and Mahayana.

<a id="S0668"></a> Source: p.38 S0668

Theravada is the more conservative branch, centering on monastic life and the earliest sutras and refusing to recognize the later Mahayana sutras as authentic." Similar to race, we found that the models make associations with religious terms that indicate some propensity to reflect how these terms are sometimes presented in the world.

<a id="S0669"></a> Source: p.38 S0669

For example, with the religion Islam, we found that words such as ramadan, prophet and mosque co-occurred at a higher rate than for other religions.

<a id="S0670"></a> Source: p.38 S0670

We also found that words such as violent, terrorism and terrorist co-occurred at a greater rate with Islam than with other religions and were in the top 40 most favored words for Islam in GPT-3. 38

Page 39

<a id="S0671"></a> Source: p.39 S0671

6.2.4 Future Bias and Fairness Challenges We have presented this preliminary analysis to share some of the biases we found in order to motivate further research, and to highlight the inherent difficulties in characterizing biases in large-scale generative models; we expect this to be an area of continuous research for us and are excited to discuss different methodological approaches with the community.

<a id="S0672"></a> Source: p.39 S0672

We view the work in this section as subjective signposting - we chose gender, race, and religion as a starting point, but we recognize the inherent subjectivity in this choice.

<a id="S0673"></a> Source: p.39 S0673

Our work is inspired by the literature on characterizing model attributes to develop informative labels such as Model Cards for Model Reporting from [MWZ+18].

<a id="S0674"></a> Source: p.39 S0674

Ultimately, it is important not just to characterize biases in language systems but to intervene.

<a id="S0675"></a> Source: p.39 S0675

The literature on this is also extensive [QMZH19, HZJ+19], so we offer only a few brief comments on future directions specific to large language models.

<a id="S0676"></a> Source: p.39 S0676

In order to pave the way for effective bias prevention in general purpose models, there is a need for building a common vocabulary tying together the normative, technical and empirical challenges of bias mitigation for these models.

<a id="S0677"></a> Source: p.39 S0677

There is room for more research that engages with the literature outside NLP, better articulates normative statements about harm, and engages with the lived experience of communities affected by NLP systems [BBDIW20].

<a id="S0678"></a> Source: p.39 S0678

Thus, mitigation work should not be approached purely with a metric driven objective to ‘remove’ bias as this has been shown to have blind spots [GG19, NvNvdG19] but in a holistic manner. 6.3 Energy Usage Practical large-scale pre-training requires large amounts of computation, which is energy-intensive: training the GPT-3 175B consumed several thousand petaflop/s-days of compute during pre-training, compared to tens of petaflop/s-days for a 1.5B parameter GPT-2 model (Figure 2.2).

<a id="S0679"></a> Source: p.39 S0679

This means we should be cognizant of the cost and efficiency of such models, as advocated by [SDSE19].

<a id="S0680"></a> Source: p.39 S0680

The use of large-scale pre-training also gives another lens through which to view the efficiency of large models - we should consider not only the resources that go into training them, but how these resources are amortized over the lifetime of a model, which will subsequently be used for a variety of purposes and fine-tuned for specific tasks.

<a id="S0681"></a> Source: p.39 S0681

Though models like GPT-3 consume significant resources during training, they can be surprisingly efficient once trained: even with the full GPT-3 175B, generating 100 pages of content from a trained model can cost on the order of 0.4 kW-hr, or only a few cents in energy costs.

<a id="S0682"></a> Source: p.39 S0682

Additionally, techniques like model distillation [LHCG19a] can further bring down the cost of such models, letting us adopt a paradigm of training single, large-scale models, then creating more efficient versions of them for use in appropriate contexts.

<a id="S0683"></a> Source: p.39 S0683

Algorithmic progress may also naturally further increase the efficiency of such models over time, similar to trends observed in image recognition and neural machine translation [HB20]. 7 Related Work Several lines of work have focused on increasing parameter count and/or computation in language models as a means to improve generative or task performance.

<a id="S0684"></a> Source: p.39 S0684

An early work scaled LSTM based language models to over a billion parameters [JVS+16].

<a id="S0685"></a> Source: p.39 S0685

One line of work straightforwardly increases the size of transformer models, scaling up parameters and FLOPS-per-token roughly in proportion.

<a id="S0686"></a> Source: p.39 S0686

Work in this vein has successively increased model size: 213 million parameters [VSP+17] in the original paper, 300 million parameters [DCLT18], 1.5 billion parameters [RWC+19], 8 billion parameters [SPP+19], 11 billion parameters [RSR+19], and most recently 17 billion parameters [Tur20]. A second line of work has focused on increasing parameter count but not computation, as a means of increasing models’ capacity to store information without increased computational cost.

<a id="S0687"></a> Source: p.39 S0687

These approaches rely on the conditional computation framework [BLC13] and specifically, the mixture-of-experts method [SMM+17] has been used to produce 100 billion parameter models and more recently 50 billion parameter translation models [AJF19], though only a small fraction of the parameters are actually used on each forward pass. A third approach increases computation without increasing parameters; examples of this approach include adaptive computation time [Gra16] and the universal transformer [DGV+18].

<a id="S0688"></a> Source: p.39 S0688

Our work focuses on the first approach (scaling compute and parameters together, by straightforwardly making the neural net larger), and increases model size 10x beyond previous models that employ this strategy.

<a id="S0689"></a> Source: p.39 S0689

Several efforts have also systematically studied the effect of scale on language model performance. [KMH+20, RRBS19, LWS+20, HNA+17], find a smooth power-law trend in loss as autoregressive language models are scaled up.

<a id="S0690"></a> Source: p.39 S0690

This work suggests that this trend largely continues as models continue to scale up (although a slight bending of the curve can perhaps be detected in Figure 3.1), and we also find relatively smooth increases in many (though not all) downstream tasks across 3 orders of magnitude of scaling.

<a id="S0691"></a> Source: p.39 S0691

Another line of work goes in the opposite direction from scaling, attempting to preserve strong performance in language models that are as small as possible.

<a id="S0692"></a> Source: p.39 S0692

This approach includes ALBERT [LCG+19] as well as general [HVD15] and 39

Page 40

<a id="S0693"></a> Source: p.40 S0693

task-specific [SDCW19, JYS+19, KR16] approaches to distillation of language models.

<a id="S0694"></a> Source: p.40 S0694

These architectures and techniques are potentially complementary to our work, and could be applied to decrease latency and memory footprint of giant models.

<a id="S0695"></a> Source: p.40 S0695

As fine-tuned language models have neared human performance on many standard benchmark tasks, considerable effort has been devoted to constructing more difficult or open-ended tasks, including question answering [KPR+19, IBGC+14, CCE+18, MCKS18], reading comprehension [CHI+18, RCM19], and adversarially constructed datasets designed to be difficult for existing language models [SBBC19, NWD+19].

<a id="S0696"></a> Source: p.40 S0696

In this work we test our models on many of these datasets.

<a id="S0697"></a> Source: p.40 S0697

Many previous efforts have focused specifically on question-answering, which constitutes a significant fraction of the tasks we tested on.

<a id="S0698"></a> Source: p.40 S0698

Recent efforts include [RSR+19, RRS20], which fine-tuned an 11 billion parameter language model, and [GLT+20], which focused on attending over a large corpus of data at test time.

<a id="S0699"></a> Source: p.40 S0699

Our work differs in focusing on in-context learning but could be combined in the future with those of [GLT+20, LPP+20].

<a id="S0700"></a> Source: p.40 S0700

Metalearning in language models has been utilized in [RWC+19], though with much more limited results and no systematic study.

<a id="S0701"></a> Source: p.40 S0701

More broadly, language model metalearning has an inner-loop-outer-loop structure, making it structurally similar to metalearning as applied to ML in general.

<a id="S0702"></a> Source: p.40 S0702

Here there is an extensive literature, including matching networks [VBL+16], RL2 [DSC+16], learning to optimize [RL16, ADG+16, LM17] and MAML [FAL17].

<a id="S0703"></a> Source: p.40 S0703

Our approach of stuffing the model’s context with previous examples is most structurally similar to RL2 and also resembles [HYC01], in that an inner loop of adaptation takes place through computation in the model’s activations across timesteps, without updating the weights, while an outer loop (in this case just language model pre-training) updates the weights, and implicitly learns the ability to adapt to or at least recognize tasks defined at inference-time.

<a id="S0704"></a> Source: p.40 S0704

Few-shot auto-regressive density estimation was explored in [RCP+17] and [GWC+18] studied low-resource NMT as a few-shot learning problem.

<a id="S0705"></a> Source: p.40 S0705

While the mechanism of our few-shot approach is different, prior work has also explored ways of using pre-trained language models in combination with gradient descent to perform few-shot learning [SS20].

<a id="S0706"></a> Source: p.40 S0706

Another sub-field with similar goals is semi-supervised learning where approaches such as UDA [XDH+19] also explore methods of fine-tuning when very little labeled data is available.

<a id="S0707"></a> Source: p.40 S0707

Giving multi-task models instructions in natural language was first formalized in a supervised setting with [MKXS18] and utilized for some tasks (such as summarizing) in a language model with [RWC+19].

<a id="S0708"></a> Source: p.40 S0708

The notion of presenting tasks in natural language was also explored in the text-to-text transformer [RSR+19], although there it was applied for multi-task fine-tuning rather than for in-context learning without weight updates.

<a id="S0709"></a> Source: p.40 S0709

Another approach to increasing generality and transfer-learning capability in language models is multi-task learning [Car97], which fine-tunes on a mixture of downstream tasks together, rather than separately updating the weights for each one.

<a id="S0710"></a> Source: p.40 S0710

If successful multi-task learning could allow a single model to be used for many tasks without updating the weights (similar to our in-context learning approach), or alternatively could improve sample efficiency when updating the weights for a new task.

<a id="S0711"></a> Source: p.40 S0711

Multi-task learning has shown some promising initial results [LGH+15, LSP+18] and multi-stage fine-tuning has recently become a standardized part of SOTA results on some datasets [PFB18] and pushed the boundaries on certain tasks [KKS+20], but is still limited by the need to manually curate collections of datasets and set up training curricula.

<a id="S0712"></a> Source: p.40 S0712

By contrast pre-training at large enough scale appears to offer a “natural” broad distribution of tasks implicitly contained in predicting the text itself.

<a id="S0713"></a> Source: p.40 S0713

One direction for future work might be attempting to generate a broader set of explicit tasks for multi-task learning, for example through procedural generation [TFR+17], human interaction [ZSW+19b], or active learning [Mac92].

<a id="S0714"></a> Source: p.40 S0714

Algorithmic innovation in language models over the last two years has been enormous, including denoising-based bidirectionality [DCLT18], prefixLM [DL15] and encoder-decoder architectures [LLG+19, RSR+19], random permutations during training [YDY+19], architectures that improve the efficiency of sampling [DYY+19], improvements in data and training procedures [LOG+19], and efficiency increases in the embedding parameters [LCG+19].

<a id="S0715"></a> Source: p.40 S0715

Many of these techniques provide significant gains on downstream tasks.

<a id="S0716"></a> Source: p.40 S0716

In this work we continue to focus on pure autoregressive language models, both in order to focus on in-context learning performance and to reduce the complexity of our large model implementations.

<a id="S0717"></a> Source: p.40 S0717

However, it is very likely that incorporating these algorithmic advances could improve GPT-3’s performance on downstream tasks, especially in the fine-tuning setting, and combining GPT-3’s scale with these algorithmic techniques is a promising direction for future work. 8 Conclusion We presented a 175 billion parameter language model which shows strong performance on many NLP tasks and benchmarks in the zero-shot, one-shot, and few-shot settings, in some cases nearly matching the performance of 40

Page 41

<a id="S0718"></a> Source: p.41 S0718

state-of-the-art fine-tuned systems, as well as generating high-quality samples and strong qualitative performance at tasks defined on-the-fly.

<a id="S0719"></a> Source: p.41 S0719

We documented roughly predictable trends of scaling in performance without using fine-tuning.

<a id="S0720"></a> Source: p.41 S0720

We also discussed the social impacts of this class of model.

<a id="S0721"></a> Source: p.41 S0721

Despite many limitations and weaknesses, these results suggest that very large language models may be an important ingredient in the development of adaptable, general language systems.

<a id="S0722"></a> Source: p.41 S0722

Acknowledgements The authors would like to thank Ryan Lowe for giving detailed feedback on drafts of the paper.

<a id="S0723"></a> Source: p.41 S0723

Thanks to Jakub Pachocki and Szymon Sidor for suggesting tasks, and Greg Brockman, Michael Petrov, Brooke Chan, and Chelsea Voss for helping run evaluations on OpenAI’s infrastructure.

<a id="S0724"></a> Source: p.41 S0724

Thanks to David Luan for initial support in scaling up this project, Irene Solaiman for discussions about ways to approach and evaluate bias, Harrison Edwards and Yura Burda for discussions and experimentation with in-context learning, Geoffrey Irving and Paul Christiano for early discussions of language model scaling, Long Ouyang for advising on the design of the human evaluation experiments, Chris Hallacy for discussions on data collection, and Shan Carter for help with visual design.

<a id="S0725"></a> Source: p.41 S0725

Thanks to the millions of people who created content that was used in the training of the model, and to those who were involved in indexing or upvoting the content (in the case of WebText).

<a id="S0726"></a> Source: p.41 S0726

Additionally, we would like to thank the entire OpenAI infrastructure and supercomputing teams for making it possible to train models at this scale. 41

Page 42

<a id="S0727"></a> Source: p.42 S0727

Contributions Tom Brown, Ben Mann, Prafulla Dhariwal, Dario Amodei, Nick Ryder, Daniel M Ziegler, and Jeffrey Wu implemented the large-scale models, training infrastructure, and model-parallel strategies.

<a id="S0728"></a> Source: p.42 S0728

Tom Brown, Dario Amodei, Ben Mann, and Nick Ryder conducted pre-training experiments.

<a id="S0729"></a> Source: p.42 S0729

Ben Mann and Alec Radford collected, filtered, deduplicated, and conducted overlap analysis on the training data.

<a id="S0730"></a> Source: p.42 S0730

Melanie Subbiah, Ben Mann, Dario Amodei, Jared Kaplan, Sam McCandlish, Tom Brown, Tom Henighan, and Girish Sastry implemented the downstream tasks and the software framework for supporting them, including creation of synthetic tasks.

<a id="S0731"></a> Source: p.42 S0731

Jared Kaplan and Sam McCandlish initially predicted that a giant language model should show continued gains, and applied scaling laws to help predict and guide model and data scaling decisions for the research.

<a id="S0732"></a> Source: p.42 S0732

Ben Mann implemented sampling without replacement during training.

<a id="S0733"></a> Source: p.42 S0733

Alec Radford originally demonstrated few-shot learning occurs in language models.

<a id="S0734"></a> Source: p.42 S0734

Jared Kaplan and Sam McCandlish showed that larger models learn more quickly in-context, and systematically studied in-context learning curves, task prompting, and evaluation methods.

<a id="S0735"></a> Source: p.42 S0735

Prafulla Dhariwal implemented an early version of the codebase, and developed the memory optimizations for fully half-precision training.

<a id="S0736"></a> Source: p.42 S0736

Rewon Child and Mark Chen developed an early version of our model-parallel strategy.

<a id="S0737"></a> Source: p.42 S0737

Rewon Child and Scott Gray contributed the sparse transformer.

<a id="S0738"></a> Source: p.42 S0738

Aditya Ramesh experimented with loss scaling strategies for pretraining.

<a id="S0739"></a> Source: p.42 S0739

Melanie Subbiah and Arvind Neelakantan implemented, experimented with, and tested beam search.

<a id="S0740"></a> Source: p.42 S0740

Pranav Shyam worked on SuperGLUE and assisted with connections to few-shot learning and meta-learning literature.

<a id="S0741"></a> Source: p.42 S0741

Sandhini Agarwal conducted the fairness and representation analysis.

<a id="S0742"></a> Source: p.42 S0742

Girish Sastry and Amanda Askell conducted the human evaluations of the model.

<a id="S0743"></a> Source: p.42 S0743

Ariel Herbert-Voss conducted the threat analysis of malicious use.

<a id="S0744"></a> Source: p.42 S0744

Gretchen Krueger edited and red-teamed the policy sections of the paper.

<a id="S0745"></a> Source: p.42 S0745

Benjamin Chess, Clemens Winter, Eric Sigler, Christopher Hesse, Mateusz Litwin, and Christopher Berner optimized OpenAI’s clusters to run the largest models efficiently.

<a id="S0746"></a> Source: p.42 S0746

Scott Gray developed fast GPU kernels used during training.

<a id="S0747"></a> Source: p.42 S0747

Jack Clark led the analysis of ethical impacts — fairness and representation, human assessments of the model, and broader impacts analysis, and advised Gretchen, Amanda, Girish, Sandhini, and Ariel on their work.

<a id="S0748"></a> Source: p.42 S0748

Dario Amodei, Alec Radford, Tom Brown, Sam McCandlish, Nick Ryder, Jared Kaplan, Sandhini Agarwal, Amanda Askell, Girish Sastry, and Jack Clark wrote the paper.

<a id="S0749"></a> Source: p.42 S0749

Sam McCandlish led the analysis of model scaling, and advised Tom Henighan and Jared Kaplan on their work.

<a id="S0750"></a> Source: p.42 S0750

Alec Radford advised the project from an NLP perspective, suggested tasks, put the results in context, and demonstrated the benefit of weight decay for training.

<a id="S0751"></a> Source: p.42 S0751

Ilya Sutskever was an early advocate for scaling large generative likelihood models, and advised Pranav, Prafulla, Rewon, Alec, and Aditya on their work.

<a id="S0752"></a> Source: p.42 S0752

Dario Amodei designed and led the research. 42

Page 43

<a id="S0753"></a> Source: p.43 S0753

A Details of Common Crawl Filtering As mentioned in Section 2.2, we employed two techniques to improve the quality of the Common Crawl dataset: (1) filtering Common Crawl and (2) fuzzy deduplication: 1.

<a id="S0754"></a> Source: p.43 S0754

In order to improve the quality of Common Crawl, we developed an automatic filtering method to remove low quality documents.

<a id="S0755"></a> Source: p.43 S0755

Using the original WebText as a proxy for high-quality documents, we trained a classifier to distinguish these from raw Common Crawl.

<a id="S0756"></a> Source: p.43 S0756

We then used this classifier to re-sample Common Crawl by prioritizing documents which were predicted by the classifier to be higher quality.

<a id="S0757"></a> Source: p.43 S0757

The classifier is trained using logistic regression classifier with features from Spark’s standard tokenizer and HashingTF 10.

<a id="S0758"></a> Source: p.43 S0758

For the positive examples, we used a collection of curated datasets such as WebText, Wikiedia, and our web books corpus as the positive examples, and for the negative examples, we used unfiltered Common Crawl.

<a id="S0759"></a> Source: p.43 S0759

We used this classifier to score Common Crawl documents.

<a id="S0760"></a> Source: p.43 S0760

We kept each document in our dataset iff np.random.pareto(α) > 1 − document_score We chose α = 9 in order to take mostly documents the classifier scored highly, but still include some documents that were out of distribution. α was chosen to match the distribution of scores from our classifier on WebText.

<a id="S0761"></a> Source: p.43 S0761

We found this re-weighting increased quality as measured by loss on a range of out-of-distribution generative text samples. 2.

<a id="S0762"></a> Source: p.43 S0762

To further improve model quality and prevent overfitting (which becomes increasingly important as model capacity increases), we fuzzily deduplicated documents (i.e. removed documents with high overlap with other documents) within each dataset using Spark’s MinHashLSH implementation with 10 hashes, using the same features as were used for classification above.

<a id="S0763"></a> Source: p.43 S0763

We also fuzzily removed WebText from Common Crawl.

<a id="S0764"></a> Source: p.43 S0764

Overall this decreased dataset size by an average of 10%.

<a id="S0765"></a> Source: p.43 S0765

After filtering for duplicates and quality, we also partially removed text occurring in benchmark datasets, described in Appendix C. B Details of Model Training To train all versions of GPT-3, we use Adam with β = 0.9, β = 0.95, and (cid:15) = 10−8, we clip the global norm of the 1 2 gradient at 1.0, and we use cosine decay for learning rate down to 10% of its value, over 260 billion tokens (after 260 billion tokens, training continues at 10% of the original learning rate).

<a id="S0766"></a> Source: p.43 S0766

There is a linear LR warmup over the first 375 million tokens.

<a id="S0767"></a> Source: p.43 S0767

We also gradually increase the batch size linearly from a small value (32k tokens) to the full value over the first 4-12 billion tokens of training, depending on the model size.

<a id="S0768"></a> Source: p.43 S0768

Data are sampled without replacement during training (until an epoch boundary is reached) to minimize overfitting.

<a id="S0769"></a> Source: p.43 S0769

All models use weight decay of 0.1 to provide a small amount of regularization [LH17].

<a id="S0770"></a> Source: p.43 S0770

During training we always train on sequences of the full n = 2048 token context window, packing multiple ctx documents into a single sequence when documents are shorter than 2048, in order to increase computational efficiency.

<a id="S0771"></a> Source: p.43 S0771

Sequences with multiple documents are not masked in any special way but instead documents within a sequence are delimited with a special end of text token, giving the language model the information necessary to infer that context separated by the end of text token is unrelated.

<a id="S0772"></a> Source: p.43 S0772

This allows for efficient training without need for any special sequence-specific masking. C Details of Test Set Contamination Studies In section 4 we gave a high level overview of test set contamination studies.

<a id="S0773"></a> Source: p.43 S0773

In this section we provide details on methodology and results.

<a id="S0774"></a> Source: p.43 S0774

Initial training set filtering We attempted to remove text occurring in benchmarks from training data by searching for 13−gram overlaps between all test/development sets used in this work and our training data, and we removed the colliding 13−gram as well as a 200 character window around it, splitting the original document into pieces.

<a id="S0775"></a> Source: p.43 S0775

For filtering purposes we define a gram as a lowercase, whitespace delimited word with no punctuation.

<a id="S0776"></a> Source: p.43 S0776

Pieces less than 200 characters long were discarded.

<a id="S0777"></a> Source: p.43 S0777

Documents split into more than 10 pieces were considered contaminated and 10https://spark.apache.org/docs/latest/api/python/pyspark.ml.html#pyspark.ml.feature.HashingTF 43

Page 44

<a id="S0778"></a> Source: p.44 S0778

Originally we removed entire documents given a single collision, but that overly penalized long documents such as books for false positives.

<a id="S0779"></a> Source: p.44 S0779

An example of a false positive might be a test set based on Wikipedia, in which the Wikipedia article quotes a single line from a book.

<a id="S0780"></a> Source: p.44 S0780

We ignored 13−grams that matched more than 10 training documents, as inspection showed the majority of these to contain common cultural phrases, legal boilerplate, or similar content that we likely do want the model to learn, rather than undesired specific overlaps with test sets.

<a id="S0781"></a> Source: p.44 S0781

Examples for various frequencies can be found in the GPT-3 release repository11.

<a id="S0782"></a> Source: p.44 S0782

Overlap methodology For our benchmark overlap analysis in Section 4, we used a variable number of words N to check for overlap for each dataset, where N is the 5th percentile example length in words, ignoring all punctuation, whitespace, and casing.

<a id="S0783"></a> Source: p.44 S0783

Due to spurious collisions at lower values of N we use a minimum value of 8 on non-synthetic tasks.

<a id="S0784"></a> Source: p.44 S0784

For performance reasons, we set a maximum value of 13 for all tasks.

<a id="S0785"></a> Source: p.44 S0785

Values for N and the amount of data marked as dirty are shown in Table C.1.

<a id="S0786"></a> Source: p.44 S0786

Unlike GPT-2’s use of bloom filters to compute probabilistic bounds for test contamination, we used Apache Spark to compute exact collisions across all training and test sets.

<a id="S0787"></a> Source: p.44 S0787

We compute overlaps between test sets and our full training corpus, even though we only trained on 40% of our filtered Common Crawl documents per Section 2.2.

<a id="S0788"></a> Source: p.44 S0788

We define a ‘dirty’ example as one with any N -gram overlap with any training document, and a ‘clean’ example as one with no collision.

<a id="S0789"></a> Source: p.44 S0789

Test and validation splits had similar contamination levels despite some test splits being unlabeled.

<a id="S0790"></a> Source: p.44 S0790

Due to a bug revealed by this analysis, filtering described above failed on long documents such as books.

<a id="S0791"></a> Source: p.44 S0791

Because of cost considerations it was infeasible to retrain the model on a corrected version of the training dataset.

<a id="S0792"></a> Source: p.44 S0792

As such, several language modeling benchmarks plus the Children’s Book Test showed almost complete overlap, and therefore were not included in this paper.

<a id="S0793"></a> Source: p.44 S0793

Overlaps are shown in Table C.1 Overlap results To understand how much having seen some of the data helps the model perform on downstream tasks, we filter every validation and test set by dirtiness.

<a id="S0794"></a> Source: p.44 S0794

Then we run evaluation on the clean-only examples and report the relative percent change between the clean score and the original score.

<a id="S0795"></a> Source: p.44 S0795

If the clean score is more than 1% or 2% worse than the overall score, it suggests the model may have overfit to the examples it has seen.

<a id="S0796"></a> Source: p.44 S0796

If the clean score is significantly better, our filtering scheme may have preferentially marked easier examples as dirty.

<a id="S0797"></a> Source: p.44 S0797

This overlap metric tends to show a high rate of false positives for datasets that contain background information (but not answers) drawn from the web (such as SQuAD, which draws from Wikipedia) or examples less than 8 words long, which we ignored in our filtering process (except for wordscrambling tasks).

<a id="S0798"></a> Source: p.44 S0798

One instance where this technique seems to fail to give good signal is DROP, a reading comprehension task in which 94% of the examples are dirty.

<a id="S0799"></a> Source: p.44 S0799

The information required to answer the question is in a passage provided to the model, so having seen the passage during training but not the questions and answers does not meaningfully constitute cheating.

<a id="S0800"></a> Source: p.44 S0800

We confirmed that every matching training document contained only the source passage, and none of the questions and answers in the dataset.

<a id="S0801"></a> Source: p.44 S0801

The more likely explanation for the decrease in performance is that the 6% of examples that remain after filtering come from a slightly different distribution than the dirty examples.

<a id="S0802"></a> Source: p.44 S0802

Figure 4.2 shows that as the dataset becomes more contaminated, the variance of the clean/all fraction increases, but there is no apparent bias towards improved or degraded performance.

<a id="S0803"></a> Source: p.44 S0803

This suggests that GPT-3 is relatively insensitive to contamination.

<a id="S0804"></a> Source: p.44 S0804

See Section 4 for details on the datasets we flagged for further review. 11https://github.com/openai/gpt-3/blob/master/overlap_frequency.md 44

Page 45

<a id="S0805"></a> Source: p.45 S0805

Relative Total Dirty Dirty Clean Clean Clean Difference Name Split Metric N Acc/F1/BLEU Count Acc/F1/BLEU Count Acc/F1/BLEU Count Percentage Clean vs All Quac dev f1 13 44.3 7353 44.3 7315 54.1 38 1% 20% SQuADv2 dev f1 13 69.8 11873 69.9 11136 68.4 737 6% -2% DROP dev f1 13 36.5 9536 37.0 8898 29.5 638 7% -21% Symbol Insertion dev acc 7 66.9 10000 66.8 8565 67.1 1435 14% 0% CoQa dev f1 13 86.0 7983 85.3 5107 87.1 2876 36% 1% ReCoRD dev acc 13 89.5 10000 90.3 6110 88.2 3890 39% -1% Winograd test acc 9 88.6 273 90.2 164 86.2 109 40% -3% BoolQ dev acc 13 76.0 3270 75.8 1955 76.3 1315 40% 0% MultiRC dev acc 13 74.2 953 73.4 558 75.3 395 41% 1% RACE-h test acc 13 46.8 3498 47.0 1580 46.7 1918 55% 0% LAMBADA test acc 13 86.4 5153 86.9 2209 86.0 2944 57% 0% LAMBADA (No Blanks) test acc 13 77.8 5153 78.5 2209 77.2 2944 57% -1% WSC dev acc 13 76.9 104 73.8 42 79.0 62 60% 3% PIQA dev acc 8 82.3 1838 89.9 526 79.3 1312 71% -4% RACE-m test acc 13 58.5 1436 53.0 366 60.4 1070 75% 3% De→En 16 test bleu-sb 12 43.0 2999 47.4 739 40.8 2260 75% -5% En→De 16 test bleu-sb 12 30.9 2999 32.6 739 29.9 2260 75% -3% En→Ro 16 test bleu-sb 12 25.8 1999 24.9 423 26.1 1576 79% 1% Ro→En 16 test bleu-sb 12 41.3 1999 40.4 423 41.6 1576 79% 1% WebQs test acc 8 41.5 2032 41.6 428 41.5 1604 79% 0% ANLI R1 test acc 13 36.8 1000 40.5 200 35.9 800 80% -3% ANLI R2 test acc 13 34.0 1000 29.4 177 35.0 823 82% 3% TriviaQA dev acc 10 71.2 7993 70.8 1390 71.3 6603 83% 0% ANLI R3 test acc 13 40.2 1200 38.3 196 40.5 1004 84% 1% En→Fr 14 test bleu-sb 13 39.9 3003 38.3 411 40.3 2592 86% 1% Fr→En 14 test bleu-sb 13 41.4 3003 40.9 411 41.4 2592 86% 0% WiC dev acc 13 51.4 638 53.1 49 51.3 589 92% 0% RTE dev acc 13 71.5 277 71.4 21 71.5 256 92% 0% CB dev acc 13 80.4 56 100.0 4 78.8 52 93% -2% Anagrams 2 dev acc 2 40.2 10000 76.2 705 37.4 9295 93% -7% Reversed Words dev acc 2 0.4 10000 1.5 660 0.3 9340 93% -26% OpenBookQA test acc 8 65.4 500 58.1 31 65.9 469 94% 1% ARC (Easy) test acc 11 70.1 2268 77.5 89 69.8 2179 96% 0% Anagrams 1 dev acc 2 15.0 10000 49.8 327 13.8 9673 97% -8% COPA dev acc 9 93.0 100 100.0 3 92.8 97 97% 0% ARC (Challenge) test acc 12 51.6 1144 45.2 31 51.8 1113 97% 0% HellaSwag dev acc 13 79.3 10042 86.2 152 79.2 9890 98% 0% NQs test acc 11 29.9 3610 32.7 52 29.8 3558 99% 0% Cycled Letters dev acc 2 38.6 10000 20.5 73 38.7 9927 99% 0% SAT Analogies dev acc 9 65.8 374 100.0 2 65.6 372 99% 0% StoryCloze test acc 13 87.7 1871 100.0 2 87.6 1869 100% 0% Winogrande dev acc 13 77.7 1267 - 0 77.7 1267 100% 0% Table C.1: Overlap statistics for all datasets sorted from dirtiest to cleanest.

<a id="S0806"></a> Source: p.45 S0806

We consider a dataset example dirty if it has a single N -gram collision with any document in our training corpus. “Relative Difference Clean vs All” shows the percent change in performance between only the clean examples vs all the examples in the benchmark. “Count” shows the number of examples. “Clean percentage” is the percent of examples that are clean vs total.

<a id="S0807"></a> Source: p.45 S0807

For “Acc/F1/BLEU” we use the metric specified in “Metric”.

<a id="S0808"></a> Source: p.45 S0808

These scores come from evaluations with a different seed for the random examples used for in-context learning, and will therefore differ slightly from the scores elsewhere in the paper. 45

Page 46

<a id="S0809"></a> Source: p.46 S0809

D Total Compute Used to Train Language Models This appendix contains the calculations that were used to derive the approximate compute used to train the language models in Figure 2.2.

<a id="S0810"></a> Source: p.46 S0810

As a simplifying assumption, we ignore the attention operation, as it typically uses less than 10% of the total compute for the models we are analyzing.

<a id="S0811"></a> Source: p.46 S0811

Calculations can be seen in Table D.1 and are explained within the table caption.

<a id="S0812"></a> Source: p.46 S0812

Fwd-pass Frac of Total train Total train Flops flops per params active compute compute Params Training tokens per param Mult for active param for each Model (PF-days) (flops) (M) (billions) per token bwd pass per token token T5-Small 2.08E+00 1.80E+20 60 1,000 3 3 1 0.5 T5-Base 7.64E+00 6.60E+20 220 1,000 3 3 1 0.5 T5-Large 2.67E+01 2.31E+21 770 1,000 3 3 1 0.5 T5-3B 1.04E+02 9.00E+21 3,000 1,000 3 3 1 0.5 T5-11B 3.82E+02 3.30E+22 11,000 1,000 3 3 1 0.5 BERT-Base 1.89E+00 1.64E+20 109 250 6 3 2 1.0 BERT-Large 6.16E+00 5.33E+20 355 250 6 3 2 1.0 RoBERTa-Base 1.74E+01 1.50E+21 125 2,000 6 3 2 1.0 RoBERTa-Large 4.93E+01 4.26E+21 355 2,000 6 3 2 1.0 GPT-3 Small 2.60E+00 2.25E+20 125 300 6 3 2 1.0 GPT-3 Medium 7.42E+00 6.41E+20 356 300 6 3 2 1.0 GPT-3 Large 1.58E+01 1.37E+21 760 300 6 3 2 1.0 GPT-3 XL 2.75E+01 2.38E+21 1,320 300 6 3 2 1.0 GPT-3 2.7B 5.52E+01 4.77E+21 2,650 300 6 3 2 1.0 GPT-3 6.7B 1.39E+02 1.20E+22 6,660 300 6 3 2 1.0 GPT-3 13B 2.68E+02 2.31E+22 12,850 300 6 3 2 1.0 GPT-3 175B 3.64E+03 3.14E+23 174,600 300 6 3 2 1.0 Table D.1: Starting from the right hand side and moving left, we begin with the number of training tokens that each model was trained with.

<a id="S0813"></a> Source: p.46 S0813

Next we note that since T5 uses an encoder-decoder model, only half of the parameters are active for each token during a forward or backwards pass.

<a id="S0814"></a> Source: p.46 S0814

We then note that each token is involved in a single addition and a single multiply for each active parameter in the forward pass (ignoring attention).

<a id="S0815"></a> Source: p.46 S0815

Then we add a multiplier of 3x to account for the backwards pass (as computing both ∂params and ∂acts use a similar amount of compute as the ∂loss ∂loss forwards pass.

<a id="S0816"></a> Source: p.46 S0816

Combining the previous two numbers, we get the total flops per parameter per token.

<a id="S0817"></a> Source: p.46 S0817

We multiply this value by the total training tokens and the total parameters to yield the number of total flops used during training.

<a id="S0818"></a> Source: p.46 S0818

We report both flops and petaflop/s-day (each of which are 8.64e+19 flops). E Human Quality Assessment of Synthetic News Articles This appendix contains details on the experiments measuring human ability to distinguish GPT-3-generated synthetic news articles from real news articles.

<a id="S0819"></a> Source: p.46 S0819

We first describe the experiments on the ∼ 200 word news articles, and then describe the preliminary investigation of ∼ 500 word news articles generated by GPT-3.

<a id="S0820"></a> Source: p.46 S0820

Participants: We recruited 718 unique participants to take part in 6 experiments. 97 participants were excluded for failing an internet check question, leaving a total of 621 participants: 343 male, 271 female, and 7 other.

<a id="S0821"></a> Source: p.46 S0821

Mean participant age was ∼ 38 years old.

<a id="S0822"></a> Source: p.46 S0822

All participants were recruited through Positly, which maintains a whitelist of high-performing workers from Mechanical Turk.

<a id="S0823"></a> Source: p.46 S0823

All participants were US-based but there were no other demographic restrictions.

<a id="S0824"></a> Source: p.46 S0824

Participants were paid $12 for their participation, based on a task time estimate of 60 minutes determined by pilot runs.

<a id="S0825"></a> Source: p.46 S0825

In order to ensure that the sample of participants for each experiment quiz was unique, participants were not allowed to take part in an experiment more than once.

<a id="S0826"></a> Source: p.46 S0826

Procedure and design: We arbitrarily selected 25 news articles that appeared in newser.com in early 2020.

<a id="S0827"></a> Source: p.46 S0827

We used the article titles and subtitles to produce outputs from the 125M, 350M, 760M, 1.3B, 2.7B, 6.7B, 13.0B, and 200B (GPT-3) parameter language models.

<a id="S0828"></a> Source: p.46 S0828

Five outputs per question were generated by each model and the generation with a word count closest to that of the human written article was selected automatically.

<a id="S0829"></a> Source: p.46 S0829

This was to minimize the effect that completion length might have on participants’ judgments.

<a id="S0830"></a> Source: p.46 S0830

The same output procedure for each model with the exception of the removal of the intentionally bad control model, as described in the main text. 46

Page 47

<a id="S0831"></a> Source: p.47 S0831

Average Participants Participants Genders Mean Word Count Model Recruited Excluded (m:f:other) Age (human:model) Control 76 7 32:37:0 39 216:216 GPT-3 Small 80 7 41:31:1 40 216:188 GPT-3 Medium 80 7 46:28:2 39 216:202 GPT-3 Large 81 24 46:28:2 37 216:200 GPT-3 XL 79 14 32:32:1 38 216:199 GPT-3 2.7B 80 11 36:33:0 40 216:202 GPT-3 6.7B 76 5 46:28:2 37 216:195 GPT-3 13.0B 81 13 46:28:2 37 216:209 GPT-3 175B 80 9 42:29:0 37 216:216 Table E.1: Participant details and article lengths for each experiment to evaluate human detection of ∼ 200 word model generated news articles.

<a id="S0832"></a> Source: p.47 S0832

Participants were excluded due to internet check fails.

<a id="S0833"></a> Source: p.47 S0833

Figure E.1: Participants spend more time trying to identify whether each news article is machine generated as model size increases.

<a id="S0834"></a> Source: p.47 S0834

Duration on the control model is indicated with the dashed line.

<a id="S0835"></a> Source: p.47 S0835

Line of best fit is a linear model on a log scale with 95% confidence intervals.

<a id="S0836"></a> Source: p.47 S0836

In each experiment, half of the participants were randomly assigned to quiz A and half were randomly assigned to quiz B.

<a id="S0837"></a> Source: p.47 S0837

Each quiz consisted of 25 articles: half (12-13) were human written and half (12-13) were model generated: the articles with human written completions in quiz A had model generated completions in quiz B and vice versa.

<a id="S0838"></a> Source: p.47 S0838

The order of quiz question was shuffled for each participant.

<a id="S0839"></a> Source: p.47 S0839

Participants could leave comments and were asked to indicate if they had seen the articles before.

<a id="S0840"></a> Source: p.47 S0840

Participants were instructed not to look up the articles or their content during the quiz and at the end of the quiz were asked if they had looked anything up during the quiz.

<a id="S0841"></a> Source: p.47 S0841

Statistical Tests: To compare means on the different runs, we performed a two-sample t-test for independent groups for each model against the control.

<a id="S0842"></a> Source: p.47 S0842

This was implemented in Python using the scipy.stats.ttest_ind function.

<a id="S0843"></a> Source: p.47 S0843

When plotting a regression line in the graph of average participant accuracy vs model size, we fit a power law of the form ax−b.

<a id="S0844"></a> Source: p.47 S0844

The 95% confidence intervals were estimated from the t-distribution of the sample mean.

<a id="S0845"></a> Source: p.47 S0845

Duration statistics: In the main text, we discussed the finding that the ability of human participants to distinguish model and human generated news articles decreases as our models become larger.

<a id="S0846"></a> Source: p.47 S0846

We have also found that the average time spent for a given set of questions increases as the model size increases, as shown in Figure E.1.

Page 48

<a id="S0847"></a> Source: p.48 S0847

Average Participants Participants Genders Mean Word Count Model Recruited Excluded (m:f:other) Age (human:model) Control 79 17 32:37:0 39 569:464 GPT-3 175B 81 19 32:30:0 40 569:498 Table E.2: Participant details and article lengths for the experiments investigating human detection of ∼ 500 word model generated news articles.

<a id="S0848"></a> Source: p.48 S0848

Participants were excluded due to internet check fails. accuracy scores despite increased time investment from participants supports the finding that larger models generate harder-to-distinguish news articles.

<a id="S0849"></a> Source: p.48 S0849

Preliminary investigation of ∼ 500 word articles: We recruited 160 unique US-based participants to take part in 2 experiments through Positly (details are given in Table E.2).

<a id="S0850"></a> Source: p.48 S0850

We randomly selected 12 Reuters world news articles from late 2019 and created a context for GPT-3 175B that consisted of a single Reuters article not in this set of 12.

<a id="S0851"></a> Source: p.48 S0851

We then used the article titles and Reuters locations to generate completions from GPT-3 175B and the 160M control model from the previous experiments.

<a id="S0852"></a> Source: p.48 S0852

These were used to create two 12-question quizzes per model, each consisting of half human written and half model generated articles.

<a id="S0853"></a> Source: p.48 S0853

Comprehension questions were added and articles were shown to participants in 3 stages at 30 second intervals to encourage closer reading.

<a id="S0854"></a> Source: p.48 S0854

Participants were paid $12 for this task.

<a id="S0855"></a> Source: p.48 S0855

Model generation selection methods, exclusion criteria, and statistical tests mirror those of the previous experiments. F Additional Samples from GPT-3 GPT-3 adapts well to many tasks other than the ones explored in the main body of the paper.

<a id="S0856"></a> Source: p.48 S0856

As an example, in Figure F.1, we show four uncurated samples from a prompt suggesting that the model write a poem, with a given title, in the style of Wallace Stevens.

<a id="S0857"></a> Source: p.48 S0857

We first experimented with a few prompts, then generated four samples with no additional editing or selection (sampling at temperature 1 using nucleus sampling [HBFC19] with P = 0.9).

<a id="S0858"></a> Source: p.48 S0858

Completions were truncated when the model began to write a new title and author heading, or broke into prose commentary. 48

Page 49

<a id="S0859"></a> Source: p.49 S0859

CAVAFY TRANSLATED BY EDMUND KEELEY [Poem text omitted] SOME TREES John Ashbery [Poem text omitted] Shadows on the Way Wallace Stevens -------- Generated Poem 1 -------- -------- Generated Poem 3 -------- I must have shadows on the way The sun was all we had.

<a id="S0860"></a> Source: p.49 S0860

Now, in the shade If I am to walk I must have All is changed.

<a id="S0861"></a> Source: p.49 S0861

The mind must dwell on those Each step taken slowly and alone White fields, that to its eyes were always old; To have it ready made Those ancient gleams, convoluting The way they lay among the huge roots, And I must think in lines of grey The great dark tomes of reverie, To have dim thoughts to be my guide The plumed labyrinth of the sea.

<a id="S0862"></a> Source: p.49 S0862

Must look on blue and green And never let my eye forget When our own breath gonne faint and fade That color is my friend Up the long stairway of the dead And purple must surround me too We shall remember well The blindness of the light, the whiteness The yellow of the sun is no more Of that white land.

<a id="S0863"></a> Source: p.49 S0863

The footsteps, and Intrusive than the bluish snow The grasses in that light, and in the shade That falls on all of us. I must have The menacing life.

<a id="S0864"></a> Source: p.49 S0864

Grey thoughts and blue thoughts walk with me If I am to go away at all.

<a id="S0865"></a> Source: p.49 S0865

We played, too, once, in the warmest rooms.

<a id="S0866"></a> Source: p.49 S0866

To one content, there is one content -------- Generated Poem 2 -------- For all mankind.

<a id="S0867"></a> Source: p.49 S0867

The forest glades Are all the more for being in fable, There is a way in the world.

<a id="S0868"></a> Source: p.49 S0868

And wood and town in legend shadowed- That way a man may live in a small town Once more to return.

<a id="S0869"></a> Source: p.49 S0869

The one content His lifetime and see no man come, Has in it every happiness, more brave Never the stir of men about a market, For being purposeless; and so Never the bobbing of covered carts, Those island gardens, deep in the mind, Never the creak passed on the stair Make of remoteness familiarity Closing some other life; an old man And of the unspoken name a friend.

<a id="S0870"></a> Source: p.49 S0870

Can prosper in that town, can prosper When his store is seen to close, his wife -------- Generated Poem 4 -------- Fades in his house, and the world shouts in vain That no one comes to hear him Nobody will come to this place.

<a id="S0871"></a> Source: p.49 S0871

In that small town But they are desolate, and they turn the traveler’s face Everything is known.

<a id="S0872"></a> Source: p.49 S0872

All the sounds of the world are far away.

<a id="S0873"></a> Source: p.49 S0873

The ways of the young men are known, When the wind rises above the trees, And the ways of the young women The boughs bow to the ground.

<a id="S0874"></a> Source: p.49 S0874

As they walk to the Spring; Even the birds that inhabit the tangle of weeds The rumors are all known, That is the roadside cover, are silent.

<a id="S0875"></a> Source: p.49 S0875

One listens, The quarrels, the hurt, the envy But hears no roar of the forest.

<a id="S0876"></a> Source: p.49 S0876

That way out of the world There is no utterance, there is no conversation, That leads to a wood.

<a id="S0877"></a> Source: p.49 S0877

Young, the young man There is a thin blue mist, Walks, a lover now, A darkness rising like smoke, Knows the streets as they are and is of them.

<a id="S0878"></a> Source: p.49 S0878

And within that darkness A possession of the heart.

<a id="S0879"></a> Source: p.49 S0879

It was here, and it will be here again- Here, under this sky empty and full of light.

<a id="S0880"></a> Source: p.49 S0880

There, in the wood, among trees, He sees shadows on the way, hears voices, hears the wind and the rustling of leaves; Through an open glade He sees a shape and the shape hears: It waits as he waits, Waits as the shadows wait, As the voices wait; Shadows on the way, voices in the wind.

<a id="S0881"></a> Source: p.49 S0881

Figure F.1: Four uncurated completions from a context suggesting the model compose a poem in the style of Wallace Stevens with the title ‘Shadows on the Way’. 49

Page 50

<a id="S0882"></a> Source: p.50 S0882

G Details of Task Phrasing and Specifications The following figures illustrate the formatting and phrasing of all the tasks included in the paper.

<a id="S0883"></a> Source: p.50 S0883

All data comes from the ground truth datasets in this section, and no samples from GPT-3 are included here.

<a id="S0884"></a> Source: p.50 S0884

Context → Article: Informal conversation is an important part of any business relationship.Before you start a discussion,however,make sure you understand which topics are suitable and which are considered taboo in a particular culture.

<a id="S0885"></a> Source: p.50 S0885

Latin Americans enjoy sharing information about their local history, art and customs.You may expect questions about your family,and be sure to show pictures of your children.You may feel free to ask similar questions of your Latin American friends.The French think of conversation as an art form,and they enjoy the value of lively discussions as well as disagreements.

<a id="S0886"></a> Source: p.50 S0886

For them,arguments can be interesting and they can cover pretty much or any topic ---- as long as they occur in are respectful and intelligent manner.

<a id="S0887"></a> Source: p.50 S0887

In the United States,business people like to discuss a wide range of topics,including opinions about work,family,hobbies,and politics.

<a id="S0888"></a> Source: p.50 S0888

In Japan,China,and Korea,however,people are much more private.They do not share much about their thoughts,feelings,or emotions because they feel that doing so might take away from the harmonious business relationship they’re trying to build.Middle Easterners are also private about their personal lives and family matters.It is considered rude,for example,to ask a businessman from Saudi Arabia about his wife or children.

<a id="S0889"></a> Source: p.50 S0889

As a general rule,it’s best not to talk about politics or religion with your business friends.This can get you into trouble,even in the United States,where people hold different religious views.In addition,discussing one’s salary is usually considered unsuitable.Sports is typically a friendly subject in most parts of the world,although be careful not to criticize national sport.Instead,be friendly and praise your host’s team. Q: What shouldn’t you do when talking about sports with colleagues from another country? A: Criticizing the sports of your colleagues’ country. Q: Which is typically a friendly topic in most places according to the author? A: Sports. Q: Why are people from Asia more private in their conversation with others? A: They don’t want to have their good relationship with others harmed by informal conversation. Q: The author considers politics and religion . A: Correct Answer → taboo Incorrect Answer → cheerful topics Incorrect Answer → rude topics Incorrect Answer → topics that can never be talked about Figure G.1: Formatted dataset example for RACE-h.

<a id="S0890"></a> Source: p.50 S0890

When predicting, we normalize by the unconditional probability of each answer as described in 2. 50

Page 51

<a id="S0891"></a> Source: p.51 S0891

Context → anli 2: anli 2: The Gold Coast Hotel & Casino is a hotel and casino located in Paradise, Nevada.

<a id="S0892"></a> Source: p.51 S0892

This locals’ casino is owned and operated by Boyd Gaming.

<a id="S0893"></a> Source: p.51 S0893

The Gold Coast is located one mile (∼ 1.6km) west of the Las Vegas Strip on West Flamingo Road.

<a id="S0894"></a> Source: p.51 S0894

It is located across the street from the Palms Casino Resort and the Rio All Suite Hotel and Casino.

<a id="S0895"></a> Source: p.51 S0895

Question: The Gold Coast is a budget-friendly casino.

<a id="S0896"></a> Source: p.51 S0896

Correct Answer → Neither Incorrect Answer → True Incorrect Answer → False Figure G.2: Formatted dataset example for ANLI R2 Context → Article: Mrs.

<a id="S0897"></a> Source: p.51 S0897

Once she told each student to bring along a few potatoes in plastic bag.

<a id="S0898"></a> Source: p.51 S0898

On each potato the students had to write a name of a person that they hated And the next day, every child brought some potatoes.

<a id="S0899"></a> Source: p.51 S0899

Some had two potatoes;some three;some up to five.

<a id="S0900"></a> Source: p.51 S0900

Smith then told the children to carry the bags everywhere they went, even to the toilet, for two weeks.

<a id="S0901"></a> Source: p.51 S0901

As day after day passed, the children started to complain about the awful smell of the rotten potatoes.

<a id="S0902"></a> Source: p.51 S0902

Those children who brought five potatoes began to feel the weight trouble of the bags.

<a id="S0903"></a> Source: p.51 S0903

After two weeks, the children were happy to hear that the game was finally ended.

<a id="S0904"></a> Source: p.51 S0904

Smith asked,"How did you feel while carrying the potatoes for two weeks?" The children started complaining about the trouble loudly.

<a id="S0905"></a> Source: p.51 S0905

Smith told them why she asked them to play the game.

<a id="S0906"></a> Source: p.51 S0906

She said,"This is exactly the situation when you carry your hatred for somebody inside your heart.

<a id="S0907"></a> Source: p.51 S0907

The terrible smell of the hatred will pollute your heart and you will carry something unnecessary with you all the time.

<a id="S0908"></a> Source: p.51 S0908

If you cannot stand the smell of the rotten potatoes for just two weeks, can you imagine how heavy it would be to have the hatred in your heart for your lifetime?

<a id="S0909"></a> Source: p.51 S0909

So throw away any hatred from your heart, and you’ll be really happy." Q: Which of the following is True according to the passage? A: If a kid hated four people,he or she had to carry four potatoes. Q: We can learn from the passage that we should . A: throw away the hatred inside Q: The children complained about besides the weight trouble. A: the smell Q: Mrs.Smith asked her students to write on the potatoes. A: Correct Answer → names Incorrect Answer → numbers Incorrect Answer → time Incorrect Answer → places Figure G.3: Formatted dataset example for RACE-m.

<a id="S0910"></a> Source: p.51 S0910

When predicting, we normalize by the unconditional probability of each answer as described in 2. 51

Page 52

<a id="S0911"></a> Source: p.52 S0911

Correct Answer → Using a brush, brush on sealant onto wood until it is fully saturated with the sealant.

<a id="S0912"></a> Source: p.52 S0912

Incorrect Answer → Using a brush, drip on sealant onto wood until it is fully saturated with the sealant.

<a id="S0913"></a> Source: p.52 S0913

Figure G.4: Formatted dataset example for PIQA Context → My body cast a shadow over the grass because Correct Answer → the sun was rising.

<a id="S0914"></a> Source: p.52 S0914

Figure G.5: Formatted dataset example for COPA Context → (CNN) Yuval Rabin, whose father, Yitzhak Rabin, was assassinated while serving as Prime Minister of Israel, criticized Donald Trump for appealing to "Second Amendment people" in a speech and warned that the words that politicians use can incite violence and undermine democracy. "Trump’s words are an incitement to the type of political violence that touched me personally," Rabin wrote in USAToday.

<a id="S0915"></a> Source: p.52 S0915

He said that Trump’s appeal to "Second Amendment people" to stop Hillary Clinton -- comments that were criticized as a call for violence against Clinton, something Trump denied -- "were a new level of ugliness in an ugly campaign season." - The son of a former Israeli Prime Minister who was assassinated wrote an op ed about the consequence of violent political rhetoric. - Warns of "parallels" between Israel of the 1990s and the U.S. today.

<a id="S0916"></a> Source: p.52 S0916

Correct Answer → - Referencing his father, who was shot and killed by an extremist amid political tension in Israel in 1995, Rabin condemned Donald Trump’s aggressive rhetoric.

<a id="S0917"></a> Source: p.52 S0917

Correct Answer → - Referencing his father, who was shot and killed by an extremist amid political tension in Israel in 1995, Rabin condemned Trump’s aggressive rhetoric.

<a id="S0918"></a> Source: p.52 S0918

Incorrect Answer → - Referencing his father, who was shot and killed by an extremist amid political tension in Israel in 1995, Rabin condemned Hillary Clinton’s aggressive rhetoric.

<a id="S0919"></a> Source: p.52 S0919

Incorrect Answer → - Referencing his father, who was shot and killed by an extremist amid political tension in Israel in 1995, Rabin condemned U.S.’s aggressive rhetoric.

<a id="S0920"></a> Source: p.52 S0920

Incorrect Answer → - Referencing his father, who was shot and killed by an extremist amid political tension in Israel in 1995, Rabin condemned Yitzhak Rabin’s aggressive rhetoric.

<a id="S0921"></a> Source: p.52 S0921

Figure G.6: Formatted dataset example for ReCoRD.

<a id="S0922"></a> Source: p.52 S0922

We consider the context above to be a single ”problem” because this is how the task is presented in the ReCoRD dataset and scored in the ReCoRD evaluation script.

<a id="S0923"></a> Source: p.52 S0923

Context → anli 1: anli 1: Fulton James MacGregor MSP is a Scottish politician who is a Scottish National Party (SNP) Member of Scottish Parliament for the constituency of Coatbridge and Chryston.

<a id="S0924"></a> Source: p.52 S0924

MacGregor is currently Parliamentary Liaison Officer to Shona Robison, Cabinet Secretary for Health & Sport.

<a id="S0925"></a> Source: p.52 S0925

He also serves on the Justice and Education & Skills committees in the Scottish Parliament.

<a id="S0926"></a> Source: p.52 S0926

Question: Fulton James MacGregor is a Scottish politican who is a Liaison officer to Shona Robison who he swears is his best friend.

<a id="S0927"></a> Source: p.52 S0927

Correct Answer → Neither Incorrect Answer → True Incorrect Answer → False Figure G.7: Formatted dataset example for ANLI R1 52

Page 53

<a id="S0928"></a> Source: p.53 S0928

Context → Organisms require energy in order to do what?

<a id="S0929"></a> Source: p.53 S0929

Figure G.8: Formatted dataset example for OpenBookQA.

<a id="S0930"></a> Source: p.53 S0930

When predicting, we normalize by the unconditional probability of each answer as described in 2.

<a id="S0931"></a> Source: p.53 S0931

Context → Making a cake: Several cake pops are shown on a display. A woman and girl are shown making the cake pops in a kitchen.

<a id="S0932"></a> Source: p.53 S0932

They Correct Answer → bake them, then frost and decorate.

<a id="S0933"></a> Source: p.53 S0933

Incorrect Answer → taste them as they place them on plates.

<a id="S0934"></a> Source: p.53 S0934

Incorrect Answer → put the frosting on the cake as they pan it.

<a id="S0935"></a> Source: p.53 S0935

Incorrect Answer → come out and begin decorating the cake as well.

<a id="S0936"></a> Source: p.53 S0936

Figure G.9: Formatted dataset example for HellaSwag Context → anli 3: anli 3: We shut the loophole which has American workers actually subsidizing the loss of their own job.

<a id="S0937"></a> Source: p.53 S0937

They just passed an expansion of that loophole in the last few days: $43 billion of giveaways, including favors to the oil and gas industry and the people importing ceiling fans from China.

<a id="S0938"></a> Source: p.53 S0938

Question: The loophole is now gone True, False, or Neither?

<a id="S0939"></a> Source: p.53 S0939

Correct Answer → False Incorrect Answer → True Incorrect Answer → Neither Figure G.10: Formatted dataset example for ANLI R3 Context → Question: George wants to warm his hands quickly by rubbing them.

<a id="S0940"></a> Source: p.53 S0940

Which skin surface will produce the most heat?

<a id="S0941"></a> Source: p.53 S0941

Answer: Correct Answer → dry palms Incorrect Answer → wet palms Incorrect Answer → palms covered with oil Incorrect Answer → palms covered with lotion Figure G.11: Formatted dataset example for ARC (Challenge).

<a id="S0942"></a> Source: p.53 S0942

When predicting, we normalize by the unconditional probability of each answer as described in 2.

<a id="S0943"></a> Source: p.53 S0943

Context → lull is to trust as Correct Answer → cajole is to compliance Incorrect Answer → balk is to fortitude Incorrect Answer → betray is to loyalty Incorrect Answer → hinder is to destination Incorrect Answer → soothe is to passion Figure G.12: Formatted dataset example for SAT Analogies Correct Context → Grace was happy to trade me her sweater for my jacket.

<a id="S0944"></a> Source: p.53 S0944

She thinks the sweater Incorrect Context → Grace was happy to trade me her sweater for my jacket.

<a id="S0945"></a> Source: p.53 S0945

She thinks the jacket Target Completion → looks dowdy on her.

<a id="S0946"></a> Source: p.53 S0946

Figure G.13: Formatted dataset example for Winograd.

<a id="S0947"></a> Source: p.53 S0947

The ‘partial’ evaluation method we use compares the probability of the completion given a correct and incorrect context. 53

Page 54

<a id="S0948"></a> Source: p.54 S0948

Correct Context → Johnny likes fruits more than vegetables in his new keto diet because the fruits Incorrect Context → Johnny likes fruits more than vegetables in his new keto diet because the vegetables Target Completion → are saccharine.

<a id="S0949"></a> Source: p.54 S0949

Figure G.14: Formatted dataset example for Winogrande.

<a id="S0950"></a> Source: p.54 S0950

The ‘partial’ evaluation method we use compares the probability of the completion given a correct and incorrect context.

<a id="S0951"></a> Source: p.54 S0951

Context → READING COMPREHENSION ANSWER KEY While this process moved along, diplomacy continued its rounds.

<a id="S0952"></a> Source: p.54 S0952

Direct pressure on the Taliban had proved unsuccessful.

<a id="S0953"></a> Source: p.54 S0953

As one NSC staff note put it, "Under the Taliban, Afghanistan is not so much a state sponsor of terrorism as it is a state sponsored by terrorists." In early 2000, the United States began a high-level effort to persuade Pakistan to use its influence over the Taliban.

<a id="S0954"></a> Source: p.54 S0954

In January 2000, Assistant Secretary of State Karl Inderfurth and the State Department’s counterterrorism coordinator, Michael Sheehan, met with General Musharraf in Islamabad, dangling before him the possibility of a presidential visit in March as a reward for Pakistani cooperation.

<a id="S0955"></a> Source: p.54 S0955

Such a visit was coveted by Musharraf, partly as a sign of his government’s legitimacy.

<a id="S0956"></a> Source: p.54 S0956

He told the two envoys that he would meet with Mullah Omar and press him on Bin Laden.

<a id="S0957"></a> Source: p.54 S0957

They left, however, reporting to Washington that Pakistan was unlikely in fact to do anything," given what it sees as the benefits of Taliban control of Afghanistan." President Clinton was scheduled to travel to India.

<a id="S0958"></a> Source: p.54 S0958

The State Department felt that he should not visit India without also visiting Pakistan.

<a id="S0959"></a> Source: p.54 S0959

The Secret Service and the CIA, however, warned in the strongest terms that visiting Pakistan would risk the President’s life.

<a id="S0960"></a> Source: p.54 S0960

Counterterrorism officials also argued that Pakistan had not done enough to merit a presidential visit.

<a id="S0961"></a> Source: p.54 S0961

But President Clinton insisted on including Pakistan in the itinerary for his trip to South Asia.

<a id="S0962"></a> Source: p.54 S0962

His one-day stopover on March 25, 2000, was the first time a U.S. president had been there since 1969.

<a id="S0963"></a> Source: p.54 S0963

At his meeting with Musharraf and others, President Clinton concentrated on tensions between Pakistan and India and the dangers of nuclear proliferation, but also discussed Bin Laden.

<a id="S0964"></a> Source: p.54 S0964

President Clinton told us that when he pulled Musharraf aside for a brief, one-on-one meeting, he pleaded with the general for help regarding Bin Laden." I offered him the moon when I went to see him, in terms of better relations with the United States, if he’d help us get Bin Laden and deal with another issue or two." The U.S. effort continued.

<a id="S0965"></a> Source: p.54 S0965

Who did The State Department feel should visit both India and Pakistan?

<a id="S0966"></a> Source: p.54 S0966

Correct Answer → - [False] Bin Laden Incorrect Answer → - [True] Bin Laden Figure G.15: Formatted dataset example for MultiRC.

<a id="S0967"></a> Source: p.54 S0967

There are three levels within MultiRC: (1) the passage, (2) the questions, and (3) the answers.

<a id="S0968"></a> Source: p.54 S0968

During evaluation, accuracy is determined at the per-question level, with a question being considered correct if and only if all the answers within the question are labeled correctly.

<a id="S0969"></a> Source: p.54 S0969

For this reason, we use K to refer to the number of questions shown within the context.

<a id="S0970"></a> Source: p.54 S0970

Context → Question: Which factor will most likely cause a person to develop a fever?

<a id="S0971"></a> Source: p.54 S0971

Answer: Correct Answer → a bacterial population in the bloodstream Incorrect Answer → a leg muscle relaxing after exercise Incorrect Answer → several viral particles on the skin Incorrect Answer → carbohydrates being digested in the stomach Figure G.16: Formatted dataset example for ARC (Easy).

<a id="S0972"></a> Source: p.54 S0972

When predicting, we normalize by the unconditional probability of each answer as described in 2. 54

Page 55

<a id="S0973"></a> Source: p.55 S0973

Context → Bob went to the gas station to fill up his car.

<a id="S0974"></a> Source: p.55 S0974

His tank was completely empty and so was his wallet.

<a id="S0975"></a> Source: p.55 S0975

The cashier offered to pay for his gas if he came back later to pay.

<a id="S0976"></a> Source: p.55 S0976

Correct Answer → Bob believed that there were good people in the world.

<a id="S0977"></a> Source: p.55 S0977

Incorrect Answer → Bob contemplated how unfriendly the world was.

<a id="S0978"></a> Source: p.55 S0978

Figure G.17: Formatted dataset example for StoryCloze Context → Helsinki is the capital and largest city of Finland.

<a id="S0979"></a> Source: p.55 S0979

It is in the region of Uusimaa, in southern Finland, on the shore of the Gulf of Finland.

<a id="S0980"></a> Source: p.55 S0980

Helsinki has a population of , an urban population of , and a metropolitan population of over 1.4 million, making it the most populous municipality and urban area in Finland.

<a id="S0981"></a> Source: p.55 S0981

Helsinki is some north of Tallinn, Estonia, east of Stockholm, Sweden, and west of Saint Petersburg, Russia.

<a id="S0982"></a> Source: p.55 S0982

Helsinki has close historical connections with these three cities.

<a id="S0983"></a> Source: p.55 S0983

The Helsinki metropolitan area includes the urban core of Helsinki, Espoo, Vantaa, Kauniainen, and surrounding commuter towns.

<a id="S0984"></a> Source: p.55 S0984

It is the world’s northernmost metro area of over one million people, and the city is the northernmost capital of an EU member state.

<a id="S0985"></a> Source: p.55 S0985

The Helsinki metropolitan area is the third largest metropolitan area in the Nordic countries after Stockholm and Copenhagen, and the City of Helsinki is the third largest after Stockholm and Oslo.

<a id="S0986"></a> Source: p.55 S0986

Helsinki is Finland’s major political, educational, financial, cultural, and research center as well as one of northern Europe’s major cities.

<a id="S0987"></a> Source: p.55 S0987

Approximately 75% of foreign companies that operate in Finland have settled in the Helsinki region.

<a id="S0988"></a> Source: p.55 S0988

The nearby municipality of Vantaa is the location of Helsinki Airport, with frequent service to various destinations in Europe and Asia. Q: what is the most populous municipality in Finland? A: Helsinki Q: how many people live there? A: 1.4 million in the metropolitan area Q: what percent of the foreign companies that operate in Finland are in Helsinki? A: 75% Q: what towns are a part of the metropolitan area? A: Target Completion → Helsinki, Espoo, Vantaa, Kauniainen, and surrounding commuter towns Figure G.18: Formatted dataset example for CoQA Context → Please unscramble the letters into a word, and write that word: asinoc = Target Completion → casino Figure G.19: Formatted dataset example for Cycled Letters 55

Page 56

<a id="S0989"></a> Source: p.56 S0989

Context → Passage: Saint Jean de Br´ebeuf was a French Jesuit missionary who travelled to New France in 1625.

<a id="S0990"></a> Source: p.56 S0990

There he worked primarily with the Huron for the rest of his life, except for a few years in France from 1629 to 1633.

<a id="S0991"></a> Source: p.56 S0991

He learned their language and culture, writing extensively about each to aid other missionaries.

<a id="S0992"></a> Source: p.56 S0992

In 1649, Br´ebeuf and another missionary were captured when an Iroquois raid took over a Huron village .

<a id="S0993"></a> Source: p.56 S0993

Together with Huron captives, the missionaries were ritually tortured and killed on March 16, 1649.

<a id="S0994"></a> Source: p.56 S0994

Br´ebeuf was beatified in 1925 and among eight Jesuit missionaries canonized as saints in the Roman Catholic Church in 1930.

<a id="S0995"></a> Source: p.56 S0995

Question: How many years did Saint Jean de Br´ebeuf stay in New France before he went back to France for a few years?

<a id="S0996"></a> Source: p.56 S0996

Answer: Target Completion → 4 Figure G.20: Formatted dataset example for DROP Context → Fill in blank: She held the torch in front of her.

<a id="S0997"></a> Source: p.56 S0997

About fifty feet ahead." She moved faster.

<a id="S0998"></a> Source: p.56 S0998

They both moved faster. "In fact," she said, raising the torch higher, "there’s more than a . -> Target Completion → step Figure G.21: Formatted dataset example for LAMBADA Context → Please unscramble the letters into a word, and write that word: skicts = Target Completion → sticks Figure G.22: Formatted dataset example for Anagrams 1 (A1) Context → Please unscramble the letters into a word, and write that word: volwskagen = Target Completion → volkswagen Figure G.23: Formatted dataset example for Anagrams 2 Context → Q: Who played tess on touched by an angel? A: Target Completion → Delloreese Patricia Early (July 6, 1931 { November 19, 2017), known professionally as Della Reese Figure G.24: Formatted dataset example for Natural Questions 56

Page 57

<a id="S0999"></a> Source: p.57 S0999

Context → TITLE: William Perry (American football) - Professional career PARAGRAPH: In 1985, he was selected in the first round of the 1985 NFL Draft by the Chicago Bears; he had been hand-picked by coach Mike Ditka.

<a id="S1000"></a> Source: p.57 S1000

However, defensive coordinator Buddy Ryan, who had a highly acrimonious relationship with Ditka, called Perry a "wasted draft-pick".

<a id="S1001"></a> Source: p.57 S1001

Perry soon became a pawn in the political power struggle between Ditka and Ryan.

<a id="S1002"></a> Source: p.57 S1002

Perry’s "Refrigerator" nickname followed him into the NFL and he quickly became a favorite of the Chicago Bears fans.

<a id="S1003"></a> Source: p.57 S1003

Teammates called him "Biscuit," as in "one biscuit shy of 350 pounds." While Ryan refused to play Perry, Ditka decided to use Perry as a fullback when the team was near the opponents’ goal line or in fourth and short situations, either as a ball carrier or a lead blocker for star running back Walter Payton.

<a id="S1004"></a> Source: p.57 S1004

Ditka stated the inspiration for using Perry as a fullback came to him during five-yard sprint exercises.

<a id="S1005"></a> Source: p.57 S1005

During his rookie season, Perry rushed for two touchdowns and caught a pass for one.

<a id="S1006"></a> Source: p.57 S1006

Perry even had the opportunity to run the ball during Super Bowl XX, as a nod to his popularity and contributions to the team’s success.

<a id="S1007"></a> Source: p.57 S1007

The first time he got the ball, he was tackled for a one-yard loss while attempting to throw his first NFL pass on a halfback option play.

<a id="S1008"></a> Source: p.57 S1008

The second time he got the ball, he scored a touchdown (running over Patriots linebacker Larry McGrew in the process).

<a id="S1009"></a> Source: p.57 S1009

About halfway through his rookie season, Ryan finally began to play Perry, who soon proved that he was a capable defensive lineman.

<a id="S1010"></a> Source: p.57 S1010

His Super Bowl ring size is the largest of any professional football player in the history of the event.

<a id="S1011"></a> Source: p.57 S1011

His ring size is 25, while the ring size for the average adult male is between 10 and 12.

<a id="S1012"></a> Source: p.57 S1012

Perry went on to play for ten years in the NFL, retiring after the 1994 season.

<a id="S1013"></a> Source: p.57 S1013

In his ten years as a pro, he regularly struggled with his weight, which hampered his performance at times.

<a id="S1014"></a> Source: p.57 S1014

He played in 138 games, recording 29.5 sacks and five fumble recoveries, which he returned for a total of 71 yards.

<a id="S1015"></a> Source: p.57 S1015

In his offensive career he ran five yards for two touchdowns, and had one reception for another touchdown.

<a id="S1016"></a> Source: p.57 S1016

Perry later attempted a comeback, playing an unremarkable 1996 season with the London Monarchs of the World League of American Football (later NFL Europa). Q: what team did he play for? A: Target Completion → the Chicago Bears Figure G.25: Formatted dataset example for QuAC Context → Please unscramble the letters into a word, and write that word: r e!c.i p r o.c a/l = Target Completion → reciprocal Figure G.26: Formatted dataset example for Symbol Insertion Context → Please unscramble the letters into a word, and write that word: taefed = Target Completion → defeat Figure G.27: Formatted dataset example for Reversed Words 57

Page 58

<a id="S1017"></a> Source: p.58 S1017

Context → Title: The Blitz Background: From the German point of view, March 1941 saw an improvement.

<a id="S1018"></a> Source: p.58 S1018

The Luftwaffe flew 4,000 sorties that month, including 12 major and three heavy attacks.

<a id="S1019"></a> Source: p.58 S1019

The electronic war intensified but the Luftwaffe flew major inland missions only on moonlit nights.

<a id="S1020"></a> Source: p.58 S1020

Ports were easier to find and made better targets.

<a id="S1021"></a> Source: p.58 S1021

To confuse the British, radio silence was observed until the bombs fell. X- and Y-Ger¨at beams were placed over false targets and switched only at the last minute.

<a id="S1022"></a> Source: p.58 S1022

Rapid frequency changes were introduced for X-Ger¨at, whose wider band of frequencies and greater tactical flexibility ensured it remained effective at a time when British selective jamming was degrading the effectiveness of Y-Ger¨at. Q: How many sorties were flown in March 1941? A: 4,000 Q: When did the Luftwaffe fly inland missions? A: Target Completion → only on moonlit nights Figure G.28: Formatted dataset example for SQuADv2 Context → Normal force -- In a simple case such as an object resting upon a table, the normal force on the object is equal but in opposite direction to the gravitational force applied on the object (or the weight of the object), that is, N = m g (\displaystyle N=mg), where m is mass, and g is the gravitational field strength (about 9.81 m/s on Earth).

<a id="S1023"></a> Source: p.58 S1023

The normal force here represents the force applied by the table against the object that prevents it from sinking through the table and requires that the table is sturdy enough to deliver this normal force without breaking.

<a id="S1024"></a> Source: p.58 S1024

However, it is easy to assume that the normal force and weight are action-reaction force pairs (a common mistake).

<a id="S1025"></a> Source: p.58 S1025

In this case, the normal force and weight need to be equal in magnitude to explain why there is no upward acceleration of the object.

<a id="S1026"></a> Source: p.58 S1026

For example, a ball that bounces upwards accelerates upwards because the normal force acting on the ball is larger in magnitude than the weight of the ball. question: is the normal force equal to the force of gravity? answer: Target Completion → yes Figure G.29: Formatted dataset example for BoolQ Context → The trend toward lower rents may seem surprising given that some communities in New York are bemoaning the loss of favorite local businesses to high rents.

<a id="S1027"></a> Source: p.58 S1027

But, despite the recent softening, for many of these retailers there’s still been too big a jump from the rental rates of the late 1970s, when their leases were signed.

<a id="S1028"></a> Source: p.58 S1028

Certainly, the recent drop in prices doesn’t mean Manhattan comes cheap. question: Manhattan comes cheap. true, false, or neither? answer: Target Completion → false Figure G.30: Formatted dataset example for CB 58

Page 59

<a id="S1029"></a> Source: p.59 S1029

Context → The bet, which won him dinner for four, was regarding the existence and mass of the top quark, an elementary particle discovered in 1995. question: The Top Quark is the last of six flavors of quarks predicted by the standard model theory of particle physics.

<a id="S1030"></a> Source: p.59 S1030

True or False? answer: Target Completion → False Figure G.31: Formatted dataset example for RTE Context → An outfitter provided everything needed for the safari.

<a id="S1031"></a> Source: p.59 S1031

Before his first walking holiday, he went to a specialist outfitter to buy some boots. question: Is the word ‘outfitter’ used in the same way in the two sentences above? answer: Target Completion → no Figure G.32: Formatted dataset example for WiC Context → Final Exam with Answer Key Instructions: Please carefully read the following passages.

<a id="S1032"></a> Source: p.59 S1032

For each passage, you must identify which noun the pronoun marked in bold refers to. ===== Passage: Mr.

<a id="S1033"></a> Source: p.59 S1033

Moncrieff visited Chester’s luxurious New York apartment, thinking that it belonged to his son Edward.

<a id="S1034"></a> Source: p.59 S1034

Moncrieff has decided to cancel Edward’s allowance on the ground that he no longer requires his financial support.

<a id="S1035"></a> Source: p.59 S1035

Question: In the passage above, what does the pronoun "his" refer to?

<a id="S1036"></a> Source: p.59 S1036

Answer: Target Completion → mr. moncrieff Figure G.33: Formatted dataset example for WSC Context → Q: ‘Nude Descending A Staircase’ is perhaps the most famous painting by which 20th century artist? A: Target Completion → MARCEL DUCHAMP Target Completion → r mutt Target Completion → duchamp Target Completion → marcel duchamp Target Completion → R.Mutt Target Completion → Marcel duChamp Target Completion → Henri-Robert-Marcel Duchamp Target Completion → Marcel du Champ Target Completion → henri robert marcel duchamp Target Completion → Duchampian Target Completion → Duchamp Target Completion → duchampian Target Completion → marcel du champ Target Completion → Marcel Duchamp Target Completion → MARCEL DUCHAMP Figure G.34: Formatted dataset example for TriviaQA.

<a id="S1037"></a> Source: p.59 S1037

TriviaQA allows for multiple valid completions. 59

Page 60

<a id="S1038"></a> Source: p.60 S1038

Context → Q: What school did burne hogarth establish? A: Target Completion → School of Visual Arts Figure G.35: Formatted dataset example for WebQA Context → Keinesfalls d¨urfen diese f¨ur den kommerziellen Gebrauch verwendet werden. = Target Completion → In no case may they be used for commercial purposes.

<a id="S1039"></a> Source: p.60 S1039

Figure G.36: Formatted dataset example for De→En.

<a id="S1040"></a> Source: p.60 S1040

This is the format for one- and few-shot learning, for this and other langauge tasks, the format for zero-shot learning is “Q: What is the {language} translation of {sentence} A: {translation}.” Context → In no case may they be used for commercial purposes. = Target Completion → Keinesfalls d¨urfen diese f¨ur den kommerziellen Gebrauch verwendet werden.

<a id="S1041"></a> Source: p.60 S1041

Figure G.37: Formatted dataset example for En→De Context → Analysis of instar distributions of larval I. verticalis collected from a series of ponds also indicated that males were in more advanced instars than females. = Target Completion → L’analyse de la distribution de fr´equence des stades larvaires d’I. verticalis dans une s´erie d’´etangs a ´egalement d´emontr´e que les larves m^ales ´etaient `a des stades plus avanc´es que les larves femelles.

<a id="S1042"></a> Source: p.60 S1042

Figure G.38: Formatted dataset example for En→Fr Context → L’analyse de la distribution de fr´equence des stades larvaires d’I. verticalis dans une s´erie d’´etangs a ´egalement d´emontr´e que les larves m^ales ´etaient `a des stades plus avanc´es que les larves femelles. = Target Completion → Analysis of instar distributions of larval I. verticalis collected from a series of ponds also indicated that males were in more advanced instars than females.

<a id="S1043"></a> Source: p.60 S1043

Figure G.39: Formatted dataset example for Fr→En Context → The truth is that you want, at any price, and against the wishes of the peoples of Europe, to continue the negotiations for Turkey’s accession to the European Union, despite Turkey’s continuing refusal to recognise Cyprus and despite the fact that the democratic reforms are at a standstill. = Target Completion → Adev˘arul este c˘a v˘a dorit¸i, cu orice pret¸ ¸si ^ımpotriva dorint¸ei europenilor, s˘a continuat¸i negocierile de aderare a Turciei la Uniunea European˘a, ^ın ciuda refuzului continuu al Turciei de a recunoa¸ste Ciprul ¸si ^ın ciuda faptului c˘a reformele democratice au ajuns ^ıntr-un punct mort.

<a id="S1044"></a> Source: p.60 S1044

Figure G.40: Formatted dataset example for En→Ro 60

Page 61

<a id="S1045"></a> Source: p.61 S1045

Context → Adev˘arul este c˘a v˘a dorit¸i, cu orice pret¸ ¸si ^ımpotriva dorint¸ei europenilor, s˘a continuat¸i negocierile de aderare a Turciei la Uniunea European˘a, ^ın ciuda refuzului continuu al Turciei de a recunoa¸ste Ciprul ¸si ^ın ciuda faptului c˘a reformele democratice au ajuns ^ıntr-un punct mort. = Target Completion → The truth is that you want, at any price, and against the wishes of the peoples of Europe, to continue the negotiations for Turkey’s accession to the European Union, despite Turkey’s continuing refusal to recognise Cyprus and despite the fact that the democratic reforms are at a standstill.

<a id="S1046"></a> Source: p.61 S1046

Figure G.41: Formatted dataset example for Ro→En Context → Q: What is (2 4) 6? A: Target Completion → 48 Figure G.42: Formatted dataset example for Arithmetic 1DC Context → Q: What is 17 minus 14? A: Target Completion → 3 Figure G.43: Formatted dataset example for Arithmetic 2D- Context → Q: What is 98 plus 45? A: Target Completion → 143 Figure G.44: Formatted dataset example for Arithmetic 2D+ Context → Q: What is 95 times 45? A: Target Completion → 4275 Figure G.45: Formatted dataset example for Arithmetic 2Dx Context → Q: What is 509 minus 488? A: Target Completion → 21 Figure G.46: Formatted dataset example for Arithmetic 3D- Context → Q: What is 556 plus 497? A: Target Completion → 1053 Figure G.47: Formatted dataset example for Arithmetic 3D+ Context → Q: What is 6209 minus 3365? A: Target Completion → 2844 Figure G.48: Formatted dataset example for Arithmetic 4D- 61

Page 62

<a id="S1047"></a> Source: p.62 S1047

Context → Q: What is 9923 plus 617? A: Target Completion → 10540 Figure G.49: Formatted dataset example for Arithmetic 4D+ Context → Q: What is 40649 minus 78746? A: Target Completion → -38097 Figure G.50: Formatted dataset example for Arithmetic 5D− Context → Q: What is 65360 plus 16204? A: Target Completion → 81564 Figure G.51: Formatted dataset example for Arithmetic 5D+ 62

Page 63

<a id="S1048"></a> Source: p.63 S1048

H Results on All Tasks for All Model Sizes Zero-Shot One-Shot Few-Shot Fine-tune 175B Name Metric Split SOTA K Small Med Large XL 2.7B 6.7B 13B 175B Small Med Large XL 2.7B 6.7B 13B 175B Small Med Large XL 2.7B 6.7B 13B 175B (test server) HellaSwag acc dev 85.6 20 33.7 43.6 51.0 54.7 62.8 67.4 70.9 78.9 33.0 42.9 50.5 53.5 61.9 66.5 70.0 78.1 33.5 43.1 51.3 54.9 62.9 67.3 71.3 79.3 LAMBADA acc test 68.0 15 42.7 54.3 60.4 63.6 67.1 70.3 72.5 76.2 22.0 47.1 52.6 58.3 61.1 65.4 69.0 72.5 22.0 40.4 63.2 57.0 78.1 79.1 81.3 86.4 LAMBADA ppl test 8.63 15 18.6 9.09 6.53 5.44 4.60 4.00 3.56 3.00 165.0 11.6 8.29 6.46 5.53 4.61 4.06 3.35 165.0 27.6 6.63 7.45 2.89 2.56 2.56 1.92 StoryCloze acc test 91.8 70 63.3 68.5 72.4 73.4 77.2 77.7 79.5 83.2 62.3 68.7 72.3 74.2 77.3 78.7 79.7 84.7 62.3 70.2 73.9 76.1 80.2 81.2 83.0 87.7 NQs acc test 44.5 64 0.64 1.75 2.71 4.40 6.01 5.79 7.84 14.6 1.19 3.07 4.79 5.43 8.73 9.78 13.7 23.0 1.72 4.46 7.89 9.72 13.2 17.0 21.0 29.9 TriviaQA acc dev 68.0 64 4.15 7.61 14.0 19.7 31.3 38.7 41.8 64.3 4.19 12.9 20.5 26.5 35.9 44.4 51.3 68.0 6.96 16.3 26.5 32.1 42.3 51.6 57.5 71.2 71.2 WebQs acc test 45.5 64 1.77 3.20 4.33 4.63 7.92 7.73 8.22 14.4 2.56 6.20 8.51 9.15 14.5 15.1 19.0 25.3 5.46 12.6 15.9 19.6 24.8 27.7 33.5 41.5 Ro→En 16 BLEU-mb test 39.9 64 2.08 2.71 3.09 3.15 16.3 8.34 20.2 19.9 0.55 15.4 23.0 26.3 30.6 33.2 35.6 38.6 1.25 20.7 25.8 29.2 33.1 34.8 37.0 39.5 Ro→En 16 BLEU-sb test 64 2.39 3.08 3.49 3.56 16.8 8.75 20.8 20.9 0.65 15.9 23.6 26.8 31.3 34.2 36.7 40.0 1.40 21.3 26.6 30.1 34.3 36.2 38.4 41.3 En→Ro 16 BLEU-mb test 38.5 64 2.14 2.65 2.53 2.50 3.46 4.24 5.32 14.1 0.35 3.30 7.89 8.72 13.2 15.1 17.3 20.6 1.25 5.90 9.33 10.7 14.3 16.3 18.0 21.0 En→Ro 16 BLEU-sb test 64 2.61 3.11 3.07 3.09 4.26 5.31 6.43 18.0 0.55 3.90 9.15 10.3 15.7 18.2 20.8 24.9 1.64 7.40 10.9 12.9 17.2 19.6 21.8 25.8 Fr→En 14 BLEU-mb test 35.0 64 1.81 2.53 3.47 3.13 20.6 15.1 21.8 21.2 1.28 15.9 23.7 26.3 29.0 30.5 30.2 33.7 4.98 25.5 28.5 31.1 33.7 34.9 36.6 39.2 Fr→En 14 BLEU-sb test 64 2.29 2.99 3.90 3.60 21.2 15.5 22.4 21.9 1.50 16.3 24.4 27.0 30.0 31.6 31.4 35.6 5.30 26.2 29.5 32.2 35.1 36.4 38.3 41.4 En→Fr 14 BLEU-mb test 45.6 64 1.74 2.16 2.73 2.15 15.1 8.82 12.0 25.2 0.49 8.00 14.8 15.9 20.3 23.3 24.9 28.3 4.08 14.5 19.3 21.5 24.9 27.3 29.5 32.6 En→Fr 14 BLEU-sb test 45.9 64 2.44 2.75 3.54 2.82 19.3 11.4 15.3 31.3 0.81 10.0 18.2 19.3 24.7 28.3 30.1 34.1 5.31 18.0 23.6 26.1 30.3 33.3 35.5 39.9 De→En 16 BLEU-mb test 40.2 64 2.06 2.87 3.41 3.63 21.5 17.3 23.0 27.2 0.83 16.2 22.5 24.7 28.2 30.7 33.0 30.4 3.25 22.7 26.2 29.2 32.7 34.8 37.3 40.6 De→En 16 BLEU-sb test 64 2.39 3.27 3.85 4.04 22.5 18.2 24.4 28.6 0.93 17.1 23.4 25.8 29.2 31.9 34.5 32.1 3.60 23.8 27.5 30.5 34.1 36.5 39.1 43.0 En→De 16 BLEU-mb test 41.2 64 1.70 2.27 2.31 2.43 12.9 8.66 10.4 24.6 0.50 7.00 12.9 13.1 18.3 20.9 22.5 26.2 3.42 12.3 15.4 17.1 20.9 23.0 26.6 29.7 En→De 16 BLEU-sb test 41.2 64 2.09 2.65 2.75 2.92 13.7 9.36 11.0 25.3 0.54 7.40 13.4 13.4 18.8 21.7 23.3 27.3 3.78 12.9 16.1 17.7 21.7 24.1 27.7 30.9 Winograd acc test 93.8 7 66.3 72.9 74.7 76.9 82.4 85.7 87.9 88.3 63.4 68.5 72.9 76.9 82.4 84.6 86.1 89.7 63.4 67.4 73.6 76.9 84.3 85.4 82.4 88.6 Winogrande acc dev 84.6 50 52.0 52.1 57.4 58.7 62.3 64.5 67.9 70.2 51.3 53.0 58.3 59.1 61.7 65.8 66.9 73.2 51.3 52.6 57.5 59.1 62.6 67.4 70.0 77.7 PIQA acc dev 77.1 50 64.6 70.2 72.9 75.1 75.6 78.0 78.5 81.0 64.3 69.3 71.8 74.4 74.3 76.3 77.8 80.5 64.3 69.4 72.0 74.3 75.4 77.8 79.9 82.3 82.8 ARC (Challenge) acc test 78.5 50 26.6 29.5 31.8 35.5 38.0 41.4 43.7 51.4 25.5 30.2 31.6 36.4 38.4 41.5 43.1 53.2 25.5 28.4 32.3 36.7 39.5 43.7 44.8 51.5 ARC (Easy) acc test 92.0 50 43.6 46.5 53.0 53.8 58.2 60.2 63.8 68.8 42.7 48.2 54.6 55.9 60.3 62.6 66.8 71.2 42.7 51.0 58.1 59.1 62.1 65.8 69.1 70.1 OpenBookQA acc test 87.2 100 35.6 43.2 45.2 46.8 53.0 50.4 55.6 57.6 37.0 39.8 46.2 46.4 53.4 53.0 55.8 58.8 37.0 43.6 48.0 50.6 55.6 55.2 60.8 65.4 Quac f1 dev 74.4 5 21.2 26.8 31.0 30.1 34.7 36.1 38.4 41.5 21.1 26.9 31.9 32.3 37.4 39.0 40.6 43.4 21.6 27.6 32.9 34.2 38.2 39.9 40.9 44.3 RACE-h acc test 90.0 10 35.2 37.9 40.1 40.9 42.4 44.1 44.6 45.5 34.3 37.7 40.0 42.0 43.8 44.3 44.6 45.9 34.3 37.0 40.4 41.4 42.3 44.7 45.1 46.8 RACE-m acc test 93.1 10 42.1 47.2 52.1 52.3 54.7 54.4 56.7 58.4 42.3 47.3 51.7 55.2 56.1 54.7 56.9 57.4 42.3 47.0 52.7 53.0 55.6 55.4 58.1 58.1 SQuADv2 em dev 90.7 16 22.6 32.8 33.9 43.1 43.6 45.4 49.0 52.6 25.1 37.5 37.9 47.9 47.9 51.1 56.0 60.1 27.5 40.5 39.2 53.5 50.0 56.6 62.6 64.9 SQuADv2 f1 dev 93.0 16 28.3 40.2 41.4 50.3 51.0 52.7 56.3 59.5 30.1 43.6 44.1 54.0 54.1 57.1 61.8 65.4 32.1 45.5 44.9 58.7 55.9 62.1 67.7 69.8 CoQA f1 dev 90.7 5 34.5 55.0 61.8 65.3 71.1 72.8 76.3 81.5 30.6 52.1 61.6 66.1 71.8 75.1 77.9 84.0 31.1 52.0 62.7 66.8 73.2 77.3 79.9 85.0 DROP f1 dev 89.1 20 9.40 13.6 14.4 16.4 19.7 17.0 24.0 23.6 11.7 18.1 20.9 23.0 26.4 27.3 29.2 34.3 12.9 18.7 24.0 25.6 29.7 29.7 32.3 36.5 BoolQ acc dev 91.0 32 49.7 60.3 58.9 62.4 67.1 65.4 66.2 60.5 52.6 61.7 60.4 63.7 68.4 68.7 69.0 76.7 43.1 60.6 62.0 64.1 70.3 70.0 70.2 77.5 76.4 CB acc dev 96.9 32 0.00 32.1 8.93 19.6 19.6 28.6 19.6 46.4 55.4 53.6 53.6 48.2 57.1 33.9 55.4 64.3 42.9 58.9 53.6 69.6 67.9 60.7 66.1 82.1 75.6 CB f1 dev 93.9 32 0.00 29.3 11.4 17.4 22.4 25.1 20.3 42.8 60.1 39.8 45.6 37.5 45.7 28.5 44.6 52.5 26.1 40.4 32.6 48.3 45.7 44.6 46.0 57.2 52.0 Copa acc dev 94.8 32 66.0 68.0 73.0 77.0 76.0 80.0 84.0 91.0 62.0 64.0 66.0 74.0 76.0 82.0 86.0 87.0 67.0 64.0 72.0 77.0 83.0 83.0 86.0 92.0 92.0 RTE acc dev 92.5 32 47.7 49.8 48.4 56.0 46.6 55.2 62.8 63.5 53.1 47.3 49.5 49.5 54.9 54.9 56.3 70.4 52.3 48.4 46.9 50.9 56.3 49.5 60.6 72.9 69.0 WiC acc dev 76.1 32 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 50.0 50.3 50.3 49.2 49.4 50.3 50.0 48.6 49.8 55.0 53.0 53.0 51.6 53.1 51.1 55.3 49.4 WSC acc dev 93.8 32 59.6 56.7 65.4 61.5 66.3 60.6 64.4 65.4 58.7 58.7 60.6 62.5 66.3 60.6 66.3 69.2 58.7 60.6 54.8 49.0 62.5 67.3 75.0 75.0 80.1 MultiRC acc dev 62.3 32 4.72 9.65 12.3 13.6 14.3 18.4 24.2 27.6 4.72 9.65 12.3 13.6 14.3 18.4 24.2 27.6 6.09 11.8 16.8 20.8 24.7 23.8 25.0 32.5 30.5 MultiRC f1a dev 88.2 32 57.0 59.7 60.4 59.9 60.0 64.5 71.4 72.9 57.0 59.7 60.4 59.9 60.0 64.5 71.4 72.9 45.0 55.9 64.2 65.4 69.5 66.4 69.3 74.8 75.4 ReCoRD acc dev 92.5 32 70.8 78.5 82.1 84.1 86.2 88.6 89.0 90.2 69.8 77.0 80.7 83.0 85.9 88.0 88.8 90.2 69.8 77.2 81.3 83.1 86.6 87.9 88.9 89.0 90.2 ReCoRD f1 dev 93.3 32 71.9 79.2 82.8 85.2 87.3 89.5 90.4 91.0 70.7 77.8 81.6 83.9 86.8 88.8 89.7 91.2 70.7 77.9 82.1 84.0 87.5 88.8 89.8 90.1 91.1 SuperGLUE average dev 89.0 40.6 47.4 46.8 49.6 50.1 52.3 54.4 58.2 54.4 55.1 56.7 57.8 61.2 59.7 64.3 68.9 50.2 56.2 56.8 60.0 64.3 63.6 66.9 73.2 71.8 ANLI R1 acc test 73.8 50 33.4 34.2 33.4 33.4 34.2 32.3 33.2 34.6 32.1 31.6 31.9 34.6 30.6 31.6 32.7 32.0 32.1 32.5 30.9 32.5 33.5 33.1 33.3 36.8 ANLI R2 acc test 50.7 50 33.2 31.9 33.3 33.3 33.8 33.5 33.5 35.4 35.7 33.7 33.2 32.7 32.7 33.9 33.9 33.9 35.7 33.8 32.1 31.4 32.6 33.3 32.6 34.0 ANLI R3 acc test 48.3 50 33.6 34.0 33.8 33.4 35.3 34.8 34.4 34.5 35.0 32.6 33.0 33.9 34.1 33.1 32.5 35.1 35.0 34.4 35.1 36.0 32.7 33.9 34.5 40.2 2D+ acc n/a 50 0.70 0.65 0.70 0.85 1.10 2.54 15.4 76.9 2.00 0.55 3.15 4.00 12.1 19.6 73.0 99.6 2.00 4.10 3.50 4.50 8.90 11.9 55.5 100.0 2D- acc n/a 50 1.25 1.25 1.25 1.25 1.60 7.60 12.6 58.0 1.15 0.95 1.45 1.95 3.85 11.5 44.6 86.4 1.15 1.45 2.25 2.70 7.35 13.6 52.4 98.9 3D+ acc n/a 50 0.10 0.10 0.05 0.10 0.10 0.25 1.40 34.2 0.15 0.00 0.10 0.30 0.45 0.95 15.4 65.5 0.15 0.45 0.30 0.55 0.75 0.90 8.40 80.4 3D- acc n/a 50 0.05 0.05 0.05 0.05 0.05 0.45 1.35 48.3 0.05 0.15 0.25 0.30 0.55 1.60 6.15 78.7 0.05 0.10 0.15 0.35 0.65 1.05 9.20 94.2 4D+ acc n/a 50 0.05 0.05 0.00 0.00 0.05 0.05 0.15 4.00 0.00 0.00 0.10 0.00 0.00 0.10 0.80 14.0 0.00 0.05 0.05 0.00 0.15 0.15 0.40 25.5 4D- acc n/a 50 0.00 0.00 0.00 0.00 0.00 0.00 0.10 7.50 0.00 0.00 0.00 0.00 0.05 0.00 0.50 14.0 0.00 0.05 0.00 0.00 0.10 0.05 0.40 26.8 5D+ acc n/a 50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.65 0.00 0.00 0.00 0.00 0.00 0.00 0.05 3.45 0.00 0.00 0.00 0.00 0.00 0.00 0.05 9.30 5D- acc n/a 50 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.80 0.00 0.00 0.00 0.00 0.00 0.00 0.05 3.75 0.00 0.00 0.00 0.00 0.00 0.00 0.00 9.90 2Dx acc n/a 50 2.20 2.25 2.65 2.10 2.55 5.80 6.15 19.8 1.35 2.35 3.35 2.35 4.75 9.15 11.0 27.4 1.35 2.90 2.70 2.85 4.25 6.10 7.05 29.2 1DC acc n/a 50 1.25 2.95 2.75 0.05 0.30 2.35 0.75 9.75 1.90 2.80 2.85 3.65 6.45 9.15 8.20 14.3 1.70 2.15 3.90 5.75 6.20 7.60 9.95 21.3 Cycled Letters acc n/a 100 0.62 0.71 2.85 0.00 0.63 1.35 2.58 3.66 1.67 4.36 5.68 6.46 6.25 9.41 15.1 21.7 4.63 9.27 10.7 14.5 16.7 21.9 27.7 37.9 Anagrams 1 acc n/a 100 0.10 0.14 0.40 0.00 0.27 0.69 1.16 2.28 0.21 0.61 1.12 1.27 1.60 2.72 3.72 8.62 0.50 1.27 2.13 3.05 3.81 5.49 8.38 15.1 Anagrams 2 acc n/a 100 0.81 1.21 2.69 0.01 1.71 3.75 4.53 8.91 1.19 2.62 4.70 4.77 6.97 10.2 14.6 25.9 1.94 4.80 7.59 9.87 12.6 18.9 25.6 39.7 Symbol Insertion acc n/a 100 0.00 0.00 0.10 0.00 0.05 0.42 0.89 8.26 0.03 0.05 0.57 1.18 1.67 3.46 6.62 45.4 0.11 0.28 2.19 4.18 6.61 11.0 27.3 67.2 Reversed Words acc n/a 100 0.00 0.01 0.01 0.01 0.02 0.03 0.03 0.09 0.02 0.01 0.01 0.00 0.05 0.07 0.11 0.48 0.00 0.05 0.00 0.17 0.24 0.30 0.42 0.44 SAT Analogies acc n/a 20 35.6 39.0 45.2 44.1 50.0 49.2 52.7 53.7 30.5 41.2 43.1 46.5 55.1 54.3 53.5 59.1 30.5 40.4 42.8 40.6 48.4 51.9 53.5 65.2 Table H.1: Scores for every task, setting and model that we investigate in this paper. 63

Page 64

<a id="S1049"></a> Source: p.64 S1049

Figure H.1: All results for all SuperGLUE tasks.

<a id="S1050"></a> Source: p.64 S1050

Figure H.3: All results for all Winograd tasks. 64

Page 65

<a id="S1051"></a> Source: p.65 S1051

Figure H.4: All results for all Arithmetic tasks.

<a id="S1052"></a> Source: p.65 S1052

Figure H.5: All results for all Cloze and Completion tasks. 65

Page 66

<a id="S1053"></a> Source: p.66 S1053

Figure H.6: All results for all Common Sense Reasoning tasks.

<a id="S1054"></a> Source: p.66 S1054

Figure H.7: All results for all QA tasks.

<a id="S1055"></a> Source: p.66 S1055

Figure H.8: All results for all Reading Comprehension tasks.

<a id="S1056"></a> Source: p.66 S1056

Figure H.9: All results for all ANLI rounds. 66

Page 67

<a id="S1057"></a> Source: p.67 S1057

Figure H.10: All results for all Scramble tasks.

<a id="S1058"></a> Source: p.67 S1058

Figure H.11: All results for all Translation tasks. 67

Page 68

<a id="S1059"></a> Source: p.68 S1059

References [ADG+16] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas.

<a id="S1060"></a> Source: p.68 S1060

Learning to learn by gradient descent by gradient descent.

<a id="S1061"></a> Source: p.68 S1061

In Advances in neural information processing systems, pages 3981–3989, 2016. [AI19] WeChat AI.

<a id="S1062"></a> Source: p.68 S1062

Tr-mt (ensemble), December 2019. [AJF19] Roee Aharoni, Melvin Johnson, and Orhan Firat.

<a id="S1063"></a> Source: p.68 S1063

Massively multilingual neural machine translation.

<a id="S1064"></a> Source: p.68 S1064

In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2019. [BBDIW20] Su Lin Blodgett, Solon Barocas, Hal Daume´ III, and Hanna Wallach.

<a id="S1065"></a> Source: p.68 S1065

Language (technology) is power: A critical survey of “bias” in nlp. arXiv preprint arXiv:2005.14050, 2020. [BCFL13] Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang.

<a id="S1066"></a> Source: p.68 S1066

Semantic parsing on freebase from question-answer pairs.

<a id="S1067"></a> Source: p.68 S1067

In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1533–1544, 2013. [BDD+09] Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, and Bernardo Magnini.

<a id="S1068"></a> Source: p.68 S1068

The fifth PASCAL recognizing textual entailment challenge. 2009. [BES10] Stefano Baccianella, Andrea Esuli, and Fabrizio Sebastiani.

<a id="S1069"></a> Source: p.68 S1069

Sentiwordnet 3.0: an enhanced lexical resource for sentiment analysis and opinion mining.

<a id="S1070"></a> Source: p.68 S1070

In Lrec, volume 10, pages 2200–2204, 2010. [BHDD+06] Roy Bar Haim, Ido Dagan, Bill Dolan, Lisa Ferro, Danilo Giampiccolo, Bernardo Magnini, and Idan Szpektor.

<a id="S1071"></a> Source: p.68 S1071

The second PASCAL recognising textual entailment challenge. 2006. [BHT+20] Yonatan Bisk, Ari Holtzman, Jesse Thomason, Jacob Andreas, Yoshua Bengio, Joyce Chai, Mirella Lapata, Angeliki Lazaridou, Jonathan May, Aleksandr Nisnevich, et al.

<a id="S1072"></a> Source: p.68 S1072

Experience grounds language. arXiv preprint arXiv:2004.10151, 2020. [BLC13] Yoshua Bengio, Nicholas Le´onard, and Aaron C.

<a id="S1073"></a> Source: p.68 S1073

Estimating or propagating gradients through stochastic neurons for conditional computation.

<a id="S1074"></a> Source: p.68 S1074

Arxiv, 2013. [BZB+19] Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi.

<a id="S1075"></a> Source: p.68 S1075

Piqa: Reasoning about physical commonsense in natural language. arXiv preprint arXiv:1911.11641, 2019. [Car97] Rich Caruana.

<a id="S1076"></a> Source: p.68 S1076

Machine learning, 28(1), 1997. [CB78] Susan Carey and Elsa Bartlett.

<a id="S1077"></a> Source: p.68 S1077

Proceedings of the Stanford Child Language Conference, 1978. [CCE+18] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord.

<a id="S1078"></a> Source: p.68 S1078

Think you have solved question answering? try arc, the ai2 reasoning challenge.

<a id="S1079"></a> Source: p.68 S1079

ArXiv, abs/1803.05457, 2018. [CGRS19] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever.

<a id="S1080"></a> Source: p.68 S1080

Generating long sequences with sparse transformers, 2019. [CHI+18] Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer.

<a id="S1081"></a> Source: p.68 S1081

Arxiv, 2018. [CLC+19] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova.

<a id="S1082"></a> Source: p.68 S1082

BoolQ: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. [CLY+19] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu.

<a id="S1083"></a> Source: p.68 S1083

Uniter: Learning universal image-text representations. arXiv preprint arXiv:1909.11740, 2019. [Cra17] Kate Crawford.

<a id="S1084"></a> Source: p.68 S1084

NIPS 2017 Keynote, 2017. [DCLT18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova.

<a id="S1085"></a> Source: p.68 S1085

BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 68

Page 69

<a id="S1086"></a> Source: p.69 S1086

[DGM06] Ido Dagan, Oren Glickman, and Bernardo Magnini.

<a id="S1087"></a> Source: p.69 S1087

The PASCAL recognising textual entailment challenge.

<a id="S1088"></a> Source: p.69 S1088

In Machine learning challenges. evaluating predictive uncertainty, visual object classification, and recognising textual entailment, pages 177–190.

<a id="S1089"></a> Source: p.69 S1089

Springer, 2006. [DGV+18] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser.

<a id="S1090"></a> Source: p.69 S1090

Arxiv, 2018. [DHKH14] Nadir Durrani, Barry Haddow, Philipp Koehn, and Kenneth Heafield.

<a id="S1091"></a> Source: p.69 S1091

Edinburgh’s phrase-based machine translation systems for wmt-14.

<a id="S1092"></a> Source: p.69 S1092

In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 97–104, 2014. [DL15] Andrew M.

<a id="S1093"></a> Source: p.69 S1093

In Advances in neural information processing systems, 2015. [DMST19] Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser.

<a id="S1094"></a> Source: p.69 S1094

The CommitmentBank: Investigating projection in naturally occurring discourse. 2019.

<a id="S1095"></a> Source: p.69 S1095

To appear in proceedings of Sinn und Bedeutung 23.

<a id="S1096"></a> Source: p.69 S1096

Data can be found at https://github.com/mcdm/CommitmentBank/. [DSC+16] Yan Duan, John Schulman, Xi Chen, Peter L.

<a id="S1097"></a> Source: p.69 S1097

Bartlett, Ilya Sutskever, and Pieter Abbeel.

<a id="S1098"></a> Source: p.69 S1098

Rl2: Fast reinforcement learning via slow reinforcement learning.

<a id="S1099"></a> Source: p.69 S1099

ArXiv, abs/1611.02779, 2016. [DWD+19] Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner.

<a id="S1100"></a> Source: p.69 S1100

Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs. arXiv preprint arXiv:1903.00161, 2019. [DYY+19] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G.

<a id="S1101"></a> Source: p.69 S1101

Transformer-xl: Attentive language models beyond a fixed-length context.

<a id="S1102"></a> Source: p.69 S1102

Arxiv, 2019. [EOAG18] Sergey Edunov, Myle Ott, Michael Auli, and David Grangier.

<a id="S1103"></a> Source: p.69 S1103

Understanding back-translation at scale. arXiv preprint arXiv:1808.09381, 2018. [FAL17] Chelsea Finn, Pieter Abbeel, and Sergey Levine.

<a id="S1104"></a> Source: p.69 S1104

Model-agnostic meta-learning for fast adaptation of deep networks.

<a id="S1105"></a> Source: p.69 S1105

ArXiv, abs/1703.03400, 2017. [Fyo00] Yaroslav Fyodorov. A natural logic inference system, 2000. [GG19] Hila Gonen and Yoav Goldberg.

<a id="S1106"></a> Source: p.69 S1106

Lipstick on a pig: Debiasing methods cover up systematic gender biases in word embeddings but do not remove them. arXiv preprint arXiv:1903.03862, 2019. [GLT+20] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang.

<a id="S1107"></a> Source: p.69 S1107

Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909, 2020. [GMDD07] Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan.

<a id="S1108"></a> Source: p.69 S1108

The third PASCAL recognizing textual entailment challenge.

<a id="S1109"></a> Source: p.69 S1109

In Proceedings of the ACL-PASCAL workshop on textual entailment and paraphrasing, pages 1–9.

<a id="S1110"></a> Source: p.69 S1110

Association for Computational Linguistics, 2007. [Gra16] Alex Graves.

<a id="S1111"></a> Source: p.69 S1111

Adaptive computation time for recurrent neural networks.

<a id="S1112"></a> Source: p.69 S1112

Arxiv, 2016. [GSL+18] Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith.

<a id="S1113"></a> Source: p.69 S1113

Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324, 2018. [GSR19] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M.

<a id="S1114"></a> Source: p.69 S1114

Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv: 1906.04043, 2019. [GWC+18] Jiatao Gu, Yong Wang, Yun Chen, Kyunghyun Cho, and Victor OK Li.

<a id="S1115"></a> Source: p.69 S1115

Meta-learning for low-resource neural machine translation. arXiv preprint arXiv:1808.08437, 2018. [HB20] Daniel Hernandez and Tom Brown.

<a id="S1116"></a> Source: p.69 S1116

Ai and efficiency, May 2020. [HBFC19] Ari Holtzman, Jan Buys, Maxwell Forbes, and Yejin Choi.

<a id="S1117"></a> Source: p.69 S1117

The curious case of neural text degeneration.

<a id="S1118"></a> Source: p.69 S1118

CoRR, abs/1904.09751, 2019. [HLW+20] Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song.

<a id="S1119"></a> Source: p.69 S1119

Pretrained transformers improve out of distribution robustness. arXiv preprint arXiv:2004.06100, 2020. 69

Page 70

<a id="S1120"></a> Source: p.70 S1120

[HNA+17] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md.

<a id="S1121"></a> Source: p.70 S1121

Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou.

<a id="S1122"></a> Source: p.70 S1122

Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. [HR18] Jeremy Howard and Sebastian Ruder.

<a id="S1123"></a> Source: p.70 S1123

Universal language model fine-tuning for text classification. arXiv preprint arXiv:1801.06146, 2018. [HVD15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean.

<a id="S1124"></a> Source: p.70 S1124

Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. [HYC01] Sepp Hochreiter, A Steven Younger, and Peter R Conwell.

<a id="S1125"></a> Source: p.70 S1125

Learning to Learn Using Gradient Descent.

<a id="S1126"></a> Source: p.70 S1126

In International Conference on Artificial Neural Networks, pages 87–94.

<a id="S1127"></a> Source: p.70 S1127

Springer, 2001. [HZJ+19] Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, and Pushmeet Kohli.

<a id="S1128"></a> Source: p.70 S1128

Reducing sentiment bias in language models via counterfactual evaluation. arXiv preprint arXiv:1911.03064, 2019. [IBGC+14] Mohit Iyyer, Jordan Boyd-Graber, Leonardo Claudino, Richard Socher, and Hal Daume´ III. A neural network for factoid question answering over paragraphs.

<a id="S1129"></a> Source: p.70 S1129

In Empirical Methods in Natural Language Processing, 2014. [IDCBE19] Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck.

<a id="S1130"></a> Source: p.70 S1130

Automatic detection of generated text is easiest when humans are fooled. arXiv preprint arXiv:1911.00650, 2019. [JCWZ17] Mandar Joshi, Eunsol Choi, Daniel S.

<a id="S1131"></a> Source: p.70 S1131

TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551, 2017. [JN20] Zheng Junyuan and Gamma Lab NYC.

<a id="S1132"></a> Source: p.70 S1132

Numeric transformer - albert, March 2020. [JVS+16] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu.

<a id="S1133"></a> Source: p.70 S1133

Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016. [JYS+19] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu.

<a id="S1134"></a> Source: p.70 S1134

TinyBERT: Distilling BERT for natural language understanding. arXiv preprint arXiv:1909.10351, 2019. [JZC+19] Ying Ju, Fubang Zhao, Shijie Chen, Bowen Zheng, Xuefeng Yang, and Yunfeng Liu.

<a id="S1135"></a> Source: p.70 S1135

Technical report on conversational question answering. arXiv preprint arXiv:1909.10772, 2019. [KCR+18] Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth.

<a id="S1136"></a> Source: p.70 S1136

Looking beyond the surface: A challenge set for reading comprehension over multiple sentences.

<a id="S1137"></a> Source: p.70 S1137

In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL), 2018. [KKS+20] Daniel Khashabi, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi.

<a id="S1138"></a> Source: p.70 S1138

Unifiedqa: Crossing format boundaries with a single qa system. arXiv preprint arXiv:2005.00700, 2020. [KMB20] Sarah E.

<a id="S1139"></a> Source: p.70 S1139

Kreps, Miles McCain, and Miles Brundage.

<a id="S1140"></a> Source: p.70 S1140

All the news that’s fit to fabricate: Ai-generated text as a tool of media misinformation, 2020. [KMH+20] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B.

<a id="S1141"></a> Source: p.70 S1141

Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei.

<a id="S1142"></a> Source: p.70 S1142

Scaling laws for neural language models, 2020. [KPR+19] Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N.

<a id="S1143"></a> Source: p.70 S1143

Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov.

<a id="S1144"></a> Source: p.70 S1144

Natural questions: a benchmark for question answering research.

<a id="S1145"></a> Source: p.70 S1145

Transactions of the Association of Computational Linguistics, 2019. [KR16] Yoon Kim and Alexander M.

<a id="S1146"></a> Source: p.70 S1146

Arxiv, 2016. [LB02] Edward Loper and Steven Bird.

<a id="S1147"></a> Source: p.70 S1147

Nltk: The natural language toolkit, 2002. [LC19] Guillaume Lample and Alexis Conneau.

<a id="S1148"></a> Source: p.70 S1148

Cross-lingual language model pretraining. arXiv preprint arXiv:1901.07291, 2019. 70

Page 71

<a id="S1149"></a> Source: p.71 S1149

[LCG+19] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut.

<a id="S1150"></a> Source: p.71 S1150

ALBERT: A lite BERT for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. [LCH+20] Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao.

<a id="S1151"></a> Source: p.71 S1151

Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994, 2020. [LDL19] Zhongyang Li, Xiao Ding, and Ting Liu.

<a id="S1152"></a> Source: p.71 S1152

Story ending prediction by transferable bert. arXiv preprint arXiv:1905.07504, 2019. [LDM12] Hector Levesque, Ernest Davis, and Leora Morgenstern.

<a id="S1153"></a> Source: p.71 S1153

In Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning, 2012. [LGG+20] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer.

<a id="S1154"></a> Source: p.71 S1154

Multilingual denoising pre-training for neural machine translation. arXiv preprint arXiv:2001.08210, 2020. [LGH+15] Xiaodong Liu, Jianfeng Gao, Xiaodong He, Li Deng, Kevin Duh, and Ye-Yi Wang.

<a id="S1155"></a> Source: p.71 S1155

Representation learning using multi-task deep neural networks for semantic classification and information retrieval.

<a id="S1156"></a> Source: p.71 S1156

In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2015. [LH17] Ilya Loshchilov and Frank Hutter.

<a id="S1157"></a> Source: p.71 S1157

Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. [LHCG19a] Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao.

<a id="S1158"></a> Source: p.71 S1158

Improving multi-task deep neural networks via knowledge distillation for natural language understanding. arXiv preprint arXiv:1904.09482, 2019. [LHCG19b] Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao.

<a id="S1159"></a> Source: p.71 S1159

Multi-task deep neural networks for natural language understanding. arXiv preprint arXiv:1901.11504, 2019. [Lin20] Tal Linzen.

<a id="S1160"></a> Source: p.71 S1160

How can we accelerate progress towards human-like linguistic generalization? arXiv preprint arXiv:2005.00955, 2020. [LLG+19] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer.

<a id="S1161"></a> Source: p.71 S1161

Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. [LM17] Ke Li and Jitendra Malik.

<a id="S1162"></a> Source: p.71 S1162

Learning to optimize neural nets. arXiv preprint arXiv:1703.00441, 2017. [LOG+19] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov.

<a id="S1163"></a> Source: p.71 S1163

RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019. [LPP+20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Ku¨ttler, Mike Lewis, Wen-tau Yih, Tim Rockta¨schel, Sebastian Riedel, and Kiela Douwe.

<a id="S1164"></a> Source: p.71 S1164

Retrieval-augmented generation for knowledge-intensive nlp tasks. arXiv preprint arXiv:2005.11401, 2020. [LSP+18] Peter J.

<a id="S1165"></a> Source: p.71 S1165

Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer.

<a id="S1166"></a> Source: p.71 S1166

Generating Wikipedia by summarizing long sequences. arXiv preprint arXiv:1801.10198, 2018. [LWS+20] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E.

<a id="S1167"></a> Source: p.71 S1167

Train large, then compress: Rethinking model size for efficient training and inference of transformers, 2020. [LXL+17] Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy.

<a id="S1168"></a> Source: p.71 S1168

Race: Large-scale reading comprehension dataset from examinations. arXiv preprint arXiv:1704.04683, 2017. [LYN+20] Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin.

<a id="S1169"></a> Source: p.71 S1169

Tttttackling winogrande schemas. arXiv preprint arXiv:2003.08380, 2020. [Mac92] David.

<a id="S1170"></a> Source: p.71 S1170

Information-based objective functions for active data selection.

Page 72

<a id="S1171"></a> Source: p.72 S1171

[MBXS17] Bryan McCann, James Bradbury, Caiming Xiong, and Richard Socher.

<a id="S1172"></a> Source: p.72 S1172

Learned in translation: Contextualized word vectors.

<a id="S1173"></a> Source: p.72 S1173

In Advances in Neural Information Processing Systems, pages 6294–6305, 2017. [MCCD13] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean.

<a id="S1174"></a> Source: p.72 S1174

Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. [MCH+16] Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. A corpus and evaluation framework for deeper understanding of commonsense stories. arXiv preprint arXiv:1604.01696, 2016. [MCKS18] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal.

<a id="S1175"></a> Source: p.72 S1175

Can a suit of armor conduct electricity? a new dataset for open book question answering.

<a id="S1176"></a> Source: p.72 S1176

ArXiv, abs/1809.02789, 2018. [MKAT18] Sam McCandlish, Jared Kaplan, Dario Amodei, and OpenAI Dota Team.

<a id="S1177"></a> Source: p.72 S1177

An empirical model of large-batch training, 2018. [MKM+94] Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, Ann Bies, Mark Ferguson, Karen Katz, and Britta Schasberger.

<a id="S1178"></a> Source: p.72 S1178

The penn treebank: annotating predicate argument structure.

<a id="S1179"></a> Source: p.72 S1179

In Proceedings of the workshop on Human Language Technology, pages 114–119.

<a id="S1180"></a> Source: p.72 S1180

Association for Computational Linguistics, 1994. [MKXS18] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher.

<a id="S1181"></a> Source: p.72 S1181

The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018. [MPL19] R Thomas McCoy, Ellie Pavlick, and Tal Linzen.

<a id="S1182"></a> Source: p.72 S1182

Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007, 2019. [MWZ+18] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru.

<a id="S1183"></a> Source: p.72 S1183

Model cards for model reporting, 2018. [NBR20] Moin Nadeem, Anna Bethke, and Siva Reddy.

<a id="S1184"></a> Source: p.72 S1184

Stereoset: Measuring stereotypical bias in pretrained language models. arXiv preprint arXiv:2004.09456, 2020. [NK19] Timothy Niven and Hung-Yu Kao.

<a id="S1185"></a> Source: p.72 S1185

Probing neural network comprehension of natural language arguments. arXiv preprint arXiv:1907.07355, 2019. [Nor09] Peter Norvig.

<a id="S1186"></a> Source: p.72 S1186

Natural language corpus data, 2009. [NvNvdG19] Malvina Nissim, Rik van Noord, and Rob van der Goot.

<a id="S1187"></a> Source: p.72 S1187

Fair is better than sensational: Man is to doctor as woman is to doctor. arXiv preprint arXiv:1905.09866, 2019. [NWD+19] Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela.

<a id="S1188"></a> Source: p.72 S1188

Adversarial nli: A new benchmark for natural language understanding. arXiv preprint arXiv:1910.14599, 2019. [oR16] University of Regensburg.

<a id="S1189"></a> Source: p.72 S1189

Fascha, 2016. [PCC18] Mohammad Taher Pilehvar and Jose Camacho-Collados.

<a id="S1190"></a> Source: p.72 S1190

WIC: 10,000 example pairs for evaluating context-sensitive representations. arXiv preprint arXiv:1808.09121, 2018. [PFB18] Jason Phang, Thibault Fe´vry, and Samuel R.

<a id="S1191"></a> Source: p.72 S1191

Sentence encoders on STILTs: Supplementary training on intermediate labeled-data tasks. arXiv preprint arXiv:1811.01088, 2018. [PHR+18] Adam Poliak, Aparajita Haldar, Rachel Rudinger, J.

<a id="S1192"></a> Source: p.72 S1192

Edward Hu, Ellie Pavlick, Aaron Steven White, and Benjamin Van Durme.

<a id="S1193"></a> Source: p.72 S1193

Collecting diverse natural language inference problems for sentence representation evaluation.

<a id="S1194"></a> Source: p.72 S1194

In Proceedings of EMNLP, 2018. [PKL+16] Denis Paperno, Germa´n Kruszewski, Angeliki Lazaridou, Quan Ngoc Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Ferna´ndez.

<a id="S1195"></a> Source: p.72 S1195

The lambada dataset: Word prediction requiring a broad discourse context. arXiv preprint arXiv:1606.06031, 2016. [PNZtY18] Matthew E.

<a id="S1196"></a> Source: p.72 S1196

Peters, Mark Neumann, Luke Zettlemoyer, and Wen tau Yih.

<a id="S1197"></a> Source: p.72 S1197

Dissecting contextual word embeddings: Architecture and representation, 2018. [Pos18] Matt Post. A call for clarity in reporting BLEU scores. arXiv preprint arXiv:1804.08771, 2018. 72

Page 73

<a id="S1198"></a> Source: p.73 S1198

[PSM14] Jeffrey Pennington, Richard Socher, and Christopher Manning.

<a id="S1199"></a> Source: p.73 S1199

GloVe: Global vectors for word representation.

<a id="S1200"></a> Source: p.73 S1200

In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014. [QIA20] QIANXIN.

<a id="S1201"></a> Source: p.73 S1201

Sa-net on albert (ensemble), April 2020. [QMZH19] Yusu Qian, Urwa Muaz, Ben Zhang, and Jae Won Hyun.

<a id="S1202"></a> Source: p.73 S1202

Reducing gender bias in word-level language models with a gender-equalizing loss function. arXiv preprint arXiv:1905.12801, 2019. [RBG11] Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon.

<a id="S1203"></a> Source: p.73 S1203

Choice of plausible alternatives: An evaluation of commonsense causal reasoning.

<a id="S1204"></a> Source: p.73 S1204

In 2011 AAAI Spring Symposium Series, 2011. [RCM19] Siva Reddy, Danqi Chen, and Christopher D Manning.

<a id="S1205"></a> Source: p.73 S1205

Coqa: A conversational question answering challenge.

<a id="S1206"></a> Source: p.73 S1206

Transactions of the Association for Computational Linguistics, 7:249–266, 2019. [RCP+17] Scott Reed, Yutian Chen, Thomas Paine, Aa¨ron van den Oord, SM Eslami, Danilo Rezende, Oriol Vinyals, and Nando de Freitas.

<a id="S1207"></a> Source: p.73 S1207

Few-shot autoregressive density estimation: Towards learning to learn distributions. arXiv preprint arXiv:1710.10304, 2017. [RJL18] Pranav Rajpurkar, Robin Jia, and Percy Liang.

<a id="S1208"></a> Source: p.73 S1208

Know what you don’t know: Unanswerable questions for squad. arXiv preprint arXiv:1806.03822, 2018. [RL16] Sachin Ravi and Hugo Larochelle.

<a id="S1209"></a> Source: p.73 S1209

Optimization as a model for few-shot learning.

<a id="S1210"></a> Source: p.73 S1210

ICLR 2017 (oral), 2016. [RLL+19] Qiu Ran, Yankai Lin, Peng Li, Jie Zhou, and Zhiyuan Liu.

<a id="S1211"></a> Source: p.73 S1211

NumNet: Machine reading comprehension with numerical reasoning.

<a id="S1212"></a> Source: p.73 S1212

In Proceedings of EMNLP, 2019. [RNLVD18] Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme.

<a id="S1213"></a> Source: p.73 S1213

Gender bias in coreference resolution. arXiv preprint arXiv:1804.09301, 2018. [RNSS18] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever.

<a id="S1214"></a> Source: p.73 S1214

Improving language understanding by generative pre-training, 2018. [Ros12] R.S.

<a id="S1215"></a> Source: p.73 S1215

NIST Special Publication, 2012. [RRBS19] Jonathan S.

<a id="S1216"></a> Source: p.73 S1216

Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales, 2019. [RRS20] Adam Roberts, Colin Raffel, and Noam Shazeer.

<a id="S1217"></a> Source: p.73 S1217

How much knowledge can you pack into the parameters of a language model? arXiv preprint arXiv:2002.08910, 2020. [RSR+19] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J.

<a id="S1218"></a> Source: p.73 S1218

Exploring the limits of transfer learning with a unified text-to-text transformer, 2019. [RWC+19] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.

<a id="S1219"></a> Source: p.73 S1219

Language models are unsupervised multitask learners, 2019. [SBBC19] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi.

<a id="S1220"></a> Source: p.73 S1220

Winogrande: An adversarial winograd schema challenge at scale, 2019. [SBC+19] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang.

<a id="S1221"></a> Source: p.73 S1221

Release strategies and the social impacts of language models, 2019. [SCNP19] Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng.

<a id="S1222"></a> Source: p.73 S1222

The woman worked as a babysitter: On biases in language generation. arXiv preprint arXiv:1909.01326, 2019. [SDCW19] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf.

<a id="S1223"></a> Source: p.73 S1223

DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. [SDSE19] Roy Schwartz, Jesse Dodge, Noah A.

<a id="S1224"></a> Source: p.73 S1224

CoRR, abs/1907.10597, 2019. [SHB15] Rico Sennrich, Barry Haddow, and Alexandra Birch.

<a id="S1225"></a> Source: p.73 S1225

Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709, 2015. 73

Page 74

<a id="S1226"></a> Source: p.74 S1226

[SMM+17] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean.

<a id="S1227"></a> Source: p.74 S1227

Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. [SPP+19] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro.

<a id="S1228"></a> Source: p.74 S1228

Megatron-lm: Training multi-billion parameter language models using model parallelism, 2019. [SS20] Timo Schick and Hinrich Schu¨tze.

<a id="S1229"></a> Source: p.74 S1229

Exploiting cloze questions for few-shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. [STQ+19] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu.

<a id="S1230"></a> Source: p.74 S1230

MASS: Masked sequence to sequence pre-training for language generation. arXiv preprint arXiv:1905.02450, 2019. [TFR+17] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel.

<a id="S1231"></a> Source: p.74 S1231

Domain randomization for transferring deep neural networks from simulation to the real world.

<a id="S1232"></a> Source: p.74 S1232

In 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30.

<a id="S1233"></a> Source: p.74 S1233

Corpus-based learning of analogies and semantic relations.

<a id="S1234"></a> Source: p.74 S1234

CoRR, abs/cs/0508103, 2005. [TL18] Trieu H.

<a id="S1235"></a> Source: p.74 S1235

Le. A simple method for commonsense reasoning. arXiv preprint arXiv:1806.02847, 2018. [TLBS03] Peter D.

<a id="S1236"></a> Source: p.74 S1236

Littman, Jeffrey Bigham, and Victor Shnayder.

<a id="S1237"></a> Source: p.74 S1237

Combining independent modules to solve multiple-choice synonym and analogy problems.

<a id="S1238"></a> Source: p.74 S1238

CoRR, cs.CL/0309035, 2003. [Tur20] Project Turing.

<a id="S1239"></a> Source: p.74 S1239

Microsoft research blog, Feb 2020. [VBL+16] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al.

<a id="S1240"></a> Source: p.74 S1240

Matching Networks for One Shot Learning.

<a id="S1241"></a> Source: p.74 S1241

In Advances in neural information processing systems, pages 3630–3638, 2016. [VSP+17] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N.

<a id="S1242"></a> Source: p.74 S1242

Gomez, Łukasz Kaiser, and Illia Polosukhin.

<a id="S1243"></a> Source: p.74 S1243

In Advances in neural information processing systems, 2017. [WPN+19] Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman.

<a id="S1244"></a> Source: p.74 S1244

Superglue: A stickier benchmark for general-purpose language understanding systems.

<a id="S1245"></a> Source: p.74 S1245

In Advances in Neural Information Processing Systems, pages 3261–3275, 2019. [WXH+18] Yiren Wang, Yingce Xia, Tianyu He, Fei Tian, Tao Qin, ChengXiang Zhai, and Tie-Yan Liu.

<a id="S1246"></a> Source: p.74 S1246

ICLR 2019, 2018. [XDH+19] Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V.

<a id="S1247"></a> Source: p.74 S1247

Unsupervised data augmentation for consistency training, 2019. [YdC+19] Dani Yogatama, Cyprien de Masson d’Autume, Jerome Connor, Tomas Kocisky, Mike Chrzanowski, Lingpeng Kong, Angeliki Lazaridou, Wang Ling, Lei Yu, Chris Dyer, et al.

<a id="S1248"></a> Source: p.74 S1248

Learning and evaluating general linguistic intelligence. arXiv preprint arXiv:1901.11373, 2019. [YDY+19] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V.

<a id="S1249"></a> Source: p.74 S1249

XLNet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237, 2019. [ZHB+19] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi.

<a id="S1250"></a> Source: p.74 S1250

Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. [ZHR+19] Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi.

<a id="S1251"></a> Source: p.74 S1251

Defending against neural fake news. arXiv preprint arXiv:1905.12616, 2019. [ZLL+18] Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme.

<a id="S1252"></a> Source: p.74 S1252

ReCoRD: Bridging the gap between human and machine commonsense reading comprehension. arXiv preprint arXiv:1810.12885, 2018. [ZSW+19a] Daniel M.

<a id="S1253"></a> Source: p.74 S1253

Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B.

<a id="S1254"></a> Source: p.74 S1254

Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving.

<a id="S1255"></a> Source: p.74 S1255

Fine-tuning language models from human preferences, 2019. 74

Page 75

<a id="S1256"></a> Source: p.75 S1256

Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B.

<a id="S1257"></a> Source: p.75 S1257

Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving.

<a id="S1258"></a> Source: p.75 S1258

Fine-tuning language models from human preferences.