Jason Phang

Aaron Hurst, Adam Lerer, Adam P. Goucher et al. · openai

GPT-4o is an autoregressive omni model that accepts as input any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs. It's trained end-to-end across text, vision, and audio, meaning all inputs and outputs are processed by the same neural network. GPT-4o can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response time in conversation. It matches GPT-4 Turbo performance on text in English and code, with significant improvement on text in non-English languages, while also being much faster and 50\% cheaper in the API. GPT-4o is especially better at vision and audio understanding compared to existing models. In line with our commitment to building AI safely and consistent with our voluntary commitments to the White House, we are sharing the GPT-4o System Card, which includes our Preparedness Framework evaluations. In this System Card, we provide a detailed look at GPT-4o's capabilities, limitations, and safety evaluations across multiple categories, focusing on speech-to-speech while also evaluating text and image capabilities, and measures we've implemented to ensure the model is safe and aligned. We also include third-party assessments on dangerous capabilities, as well as discussion of potential societal impacts of GPT-4o's text and vision capabilities.

Teven Le Scao, Thomas Wang, Daniel Hesslow et al. · allen-ai, berkeley

The crystallization of modeling methods around the Transformer architecture has been a boon for practitioners. Simple, well-motivated architectural variations can transfer across tasks and scale, increasing the impact of modeling research. However, with the emergence of state-of-the-art 100B+ parameters models, large language models are increasingly expensive to accurately design and train. Notably, it can be difficult to evaluate how modeling decisions may impact emergent capabilities, given that these capabilities arise mainly from sheer scale alone. In the process of building BLOOM--the Big Science Large Open-science Open-access Multilingual language model--our goal is to identify an architecture and training setup that makes the best use of our 1,000,000 A100-GPU-hours budget. Specifically, we perform an ablation study at the billion-parameter scale comparing different modeling practices and their impact on zero-shot generalization. In addition, we study the impact of various popular pre-training corpora on zero-shot generalization. We also study the performance of a multilingual model and how it compares to the English-only one. Finally, we consider the scaling behaviour of Transformers to choose the target model size, shape, and training setup. All our models and code are open-sourced at https://huggingface.co/bigscience .

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Sid Black, Stella Biderman, Eric Hallahan et al.

We introduce GPT-NeoX-20B, a 20 billion parameter autoregressive language model trained on the Pile, whose weights will be made freely and openly available to the public through a permissive license. It is, to the best of our knowledge, the largest dense autoregressive model that has publicly available weights at the time of submission. In this work, we describe \model{}'s architecture and training and evaluate its performance on a range of language-understanding, mathematics, and knowledge-based tasks. We find that GPT-NeoX-20B is a particularly powerful few-shot reasoner and gains far more in performance when evaluated five-shot than similarly sized GPT-3 and FairSeq models. We open-source the training and evaluation code, as well as the model weights, at https://github.com/EleutherAI/gpt-neox.

Yujia Qin, Shengding Hu, Yankai Lin et al. · tsinghua

Humans possess an extraordinary ability to create and utilize tools, allowing them to overcome physical limitations and explore new frontiers. With the advent of foundation models, AI systems have the potential to be equally adept in tool use as humans. This paradigm, i.e., tool learning with foundation models, combines the strengths of specialized tools and foundation models to achieve enhanced accuracy, efficiency, and automation in problem-solving. Despite its immense potential, there is still a lack of a comprehensive understanding of key challenges, opportunities, and future endeavors in this field. To this end, we present a systematic investigation of tool learning in this paper. We first introduce the background of tool learning, including its cognitive origins, the paradigm shift of foundation models, and the complementary roles of tools and models. Then we recapitulate existing tool learning research into tool-augmented and tool-oriented learning. We formulate a general tool learning framework: starting from understanding the user instruction, models should learn to decompose a complex task into several subtasks, dynamically adjust their plan through reasoning, and effectively conquer each sub-task by selecting appropriate tools. We also discuss how to train models for improved tool-use capabilities and facilitate the generalization in tool learning. Considering the lack of a systematic tool learning evaluation in prior works, we experiment with 18 representative tools and show the potential of current foundation models in skillfully utilizing tools. Finally, we discuss several open problems that require further investigation for tool learning. In general, we hope this paper could inspire future research in integrating tools with foundation models.

Jason Phang, Herbie Bradley, Leo Gao et al. · cambridge

Over the past two years, EleutherAI has established itself as a radically novel initiative aimed at both promoting open-source research and conducting research in a transparent, openly accessible and collaborative manner. EleutherAI's approach to research goes beyond transparency: by doing research entirely in public, anyone in the world can observe and contribute at every stage. Our work has been received positively and has resulted in several high-impact projects in Natural Language Processing and other fields. In this paper, we describe our experience doing public-facing machine learning research, the benefits we believe this approach brings, and the pitfalls we have encountered.

Jason Phang, Yi Mao, Pengcheng He et al. · microsoft-research

Fine-tuning large language models for different tasks can be costly and inefficient, and even methods that reduce the number of tuned parameters still require full gradient-based optimization. We propose HyperTuning, a novel approach to model adaptation that uses a hypermodel to generate task-specific parameters for a fixed downstream model. We demonstrate a simple setup for hypertuning with HyperT5, a T5-based hypermodel that produces soft prefixes or LoRA parameters for a frozen T5 model from few-shot examples. We train HyperT5 in two stages: first, hyperpretraining with a modified conditional language modeling objective that trains a hypermodel to generate parameters; second, multi-task fine-tuning (MTF) on a large number of diverse language tasks. We evaluate HyperT5 on P3, MetaICL and Super-NaturalInstructions datasets, and show that it can effectively generate parameters for unseen tasks. Moreover, we show that using hypermodel-generated parameters as initializations for further parameter-efficient fine-tuning improves performance. HyperTuning can thus be a flexible and efficient way to leverage large language models for diverse downstream applications.

Xiangru Tang, Yiming Zong, Jason Phang et al.

Despite the remarkable capabilities of Large Language Models (LLMs) like GPT-4, producing complex, structured tabular data remains challenging. Our study assesses LLMs' proficiency in structuring tables and introduces a novel fine-tuning method, cognizant of data structures, to bolster their performance. We unveil Struc-Bench, a comprehensive benchmark featuring prominent LLMs (GPT-NeoX-20B, GPT-3.5, GPT-4, and Vicuna), which spans text tables, HTML, and LaTeX formats. Our proposed FormatCoT aids in crafting format-specific instructions from the intended outputs to populate this benchmark. Addressing the gap in task-centered evaluation, we propose two innovative metrics, P-Score (Prompting Score) and H-Score (Heuristical Score), to more accurately gauge LLM performance. Our experiments show that applying our structure-aware fine-tuning to LLaMA-7B leads to substantial performance gains, outshining its LLM counterparts across most measures. In-depth error analysis and creating an ability map across six dimensions -- coverage, formatting, reasoning, comprehension, pragmatics, and hallucination -- highlight areas for future enhancements and suggest forthcoming research trajectories. Our code and models can be found at https://github.com/gersteinlab/Struc-Bench.

Tomasz Korbak, Kejian Shi, Angelica Chen et al.

Language models (LMs) are pretrained to imitate internet text, including content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, and more. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the trade-off between alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores given by a reward model. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.

Alex Wang, Richard Yuanzhe Pang, Angelica Chen et al.

Summarization datasets are often assembled either by scraping naturally occurring public-domain summaries -- which are nearly always in difficult-to-work-with technical domains -- or by using approximate heuristics to extract them from everyday text -- which frequently yields unfaithful summaries. In this work, we turn to a slower but more straightforward approach to developing summarization benchmark data: We hire highly-qualified contractors to read stories and write original summaries from scratch. To amortize reading time, we collect five summaries per document, with the first giving an overview and the subsequent four addressing specific questions. We use this protocol to collect SQuALITY, a dataset of question-focused summaries built on the same public-domain short stories as the multiple-choice dataset QuALITY (Pang et al., 2021). Experiments with state-of-the-art summarization systems show that our dataset is challenging and that existing automatic evaluation metrics are weak indicators of quality.

Jason Phang, Yao Zhao, Peter J. Liu

While large pretrained Transformer models have proven highly capable at tackling natural language tasks, handling long sequence inputs continues to be a significant challenge. One such task is long input summarization, where inputs are longer than the maximum input context of most pretrained models. Through an extensive set of experiments, we investigate what model architectural changes and pretraining paradigms can most efficiently adapt a pretrained Transformer for long input summarization. We find that a staggered, block-local Transformer with global encoder tokens strikes a good balance of performance and efficiency, and that an additional pretraining phase on long sequences meaningfully improves downstream summarization performance. Based on our findings, we introduce PEGASUS-X, an extension of the PEGASUS model with additional long input pretraining to handle inputs of up to 16K tokens. PEGASUS-X achieves strong performance on long input summarization tasks comparable with much larger models while adding few additional parameters and not requiring model parallelism to train.

Julian Michael, Ari Holtzman, Alicia Parrish et al.

We present the results of the NLP Community Metasurvey. Run from May to June 2022, the survey elicited opinions on controversial issues, including industry influence in the field, concerns about AGI, and ethics. Our results put concrete numbers to several controversies: For example, respondents are split almost exactly in half on questions about the importance of artificial general intelligence, whether language models understand language, and the necessity of linguistic structure and inductive bias for solving NLP problems. In addition, the survey posed meta-questions, asking respondents to predict the distribution of survey responses. This allows us not only to gain insight on the spectrum of beliefs held by NLP researchers, but also to uncover false sociological beliefs where the community's predictions don't match reality. We find such mismatches on a wide range of issues. Among other results, the community greatly overestimates its own belief in the usefulness of benchmarks and the potential for scaling to solve real-world problems, while underestimating its own belief in the importance of linguistic structure, inductive bias, and interdisciplinary science.

Alicia Parrish, Harsh Trivedi, Ethan Perez et al.

Current QA systems can generate reasonable-sounding yet false answers without explanation or evidence for the generated answer, which is especially problematic when humans cannot readily check the model's answers. This presents a challenge for building trust in machine learning systems. We take inspiration from real-world situations where difficult questions are answered by considering opposing sides (see Irving et al., 2018). For multiple-choice QA examples, we build a dataset of single arguments for both a correct and incorrect answer option in a debate-style set-up as an initial step in training models to produce explanations for two candidate answers. We use long contexts -- humans familiar with the context write convincing explanations for pre-selected correct and incorrect answers, and we test if those explanations allow humans who have not read the full context to more accurately determine the correct answer. We do not find that explanations in our set-up improve human accuracy, but a baseline condition shows that providing human-selected text snippets does improve accuracy. We use these findings to suggest ways of improving the debate set up for future data collection efforts.

Alicia Parrish, Harsh Trivedi, Nikita Nangia et al.

The use of language-model-based question-answering systems to aid humans in completing difficult tasks is limited, in part, by the unreliability of the text these systems generate. Using hard multiple-choice reading comprehension questions as a testbed, we assess whether presenting humans with arguments for two competing answer options, where one is correct and the other is incorrect, allows human judges to perform more accurately, even when one of the arguments is unreliable and deceptive. If this is helpful, we may be able to increase our justified trust in language-model-based systems by asking them to produce these arguments where needed. Previous research has shown that just a single turn of arguments in this format is not helpful to humans. However, as debate settings are characterized by a back-and-forth dialogue, we follow up on previous results to test whether adding a second round of counter-arguments is helpful to humans. We find that, regardless of whether they have access to arguments or not, humans perform similarly on our task. These findings suggest that, in the case of answering reading comprehension questions, debate is not a helpful format.

Stella Biderman, Hailey Schoelkopf, Lintang Sutawika et al. · cmu

Effective evaluation of language models remains an open challenge in NLP. Researchers and engineers face methodological issues such as the sensitivity of models to evaluation setup, difficulty of proper comparisons across methods, and the lack of reproducibility and transparency. In this paper we draw on three years of experience in evaluating large language models to provide guidance and lessons for researchers. First, we provide an overview of common challenges faced in language model evaluation. Second, we delineate best practices for addressing or lessening the impact of these challenges on research. Third, we present the Language Model Evaluation Harness (lm-eval): an open source library for independent, reproducible, and extensible evaluation of language models that seeks to address these issues. We describe the features of the library as well as case studies in which the library has been used to alleviate these methodological concerns.

Betty Li Hou, Kejian Shi, Jason Phang et al.

Large Language Models (LLMs) are able to provide assistance on a wide range of information-seeking tasks. However, model outputs may be misleading, whether unintentionally or in cases of intentional deception. We investigate the ability of LLMs to be deceptive in the context of providing assistance on a reading comprehension task, using LLMs as proxies for human users. We compare outcomes of (1) when the model is prompted to provide truthful assistance, (2) when it is prompted to be subtly misleading, and (3) when it is prompted to argue for an incorrect answer. Our experiments show that GPT-4 can effectively mislead both GPT-3.5-Turbo and GPT-4, with deceptive assistants resulting in up to a 23% drop in accuracy on the task compared to when a truthful assistant is used. We also find that providing the user model with additional context from the passage partially mitigates the influence of the deceptive model. This work highlights the ability of LLMs to produce misleading information and the effects this may have in real-world situations.

Yada Pruksachatkun, Phil Yeres, Haokun Liu et al.

We introduce jiant, an open source toolkit for conducting multitask and transfer learning experiments on English NLU tasks. jiant enables modular and configuration-driven experimentation with state-of-the-art models and implements a broad set of tasks for probing, transfer learning, and multitask training experiments. jiant implements over 50 NLU tasks, including all GLUE and SuperGLUE benchmark tasks. We demonstrate that jiant reproduces published performance on a variety of tasks and models, including BERT and RoBERTa. jiant is available at https://jiant.info.

Yiqiu Shen, Nan Wu, Jason Phang et al.

Medical images differ from natural images in significantly higher resolutions and smaller regions of interest. Because of these differences, neural network architectures that work well for natural images might not be applicable to medical image analysis. In this work, we extend the globally-aware multiple instance classifier, a framework we proposed to address these unique properties of medical images. This model first uses a low-capacity, yet memory-efficient, network on the whole image to identify the most informative regions. It then applies another higher-capacity network to collect details from chosen regions. Finally, it employs a fusion module that aggregates global and local information to make a final prediction. While existing methods often require lesion segmentation during training, our model is trained with only image-level labels and can generate pixel-level saliency maps indicating possible malignant findings. We apply the model to screening mammography interpretation: predicting the presence or absence of benign and malignant lesions. On the NYU Breast Cancer Screening Dataset, consisting of more than one million images, our model achieves an AUC of 0.93 in classifying breasts with malignant findings, outperforming ResNet-34 and Faster R-CNN. Compared to ResNet-34, our model is 4.1x faster for inference while using 78.4% less GPU memory. Furthermore, we demonstrate, in a reader study, that our model surpasses radiologist-level AUC by a margin of 0.11. The proposed model is available online: https://github.com/nyukat/GMIC.

Jason Phang, Michael Lampe, Lama Ahmad et al.

As AI chatbots see increased adoption and integration into everyday life, questions have been raised about the potential impact of human-like or anthropomorphic AI on users. In this work, we investigate the extent to which interactions with ChatGPT (with a focus on Advanced Voice Mode) may impact users' emotional well-being, behaviors and experiences through two parallel studies. To study the affective use of AI chatbots, we perform large-scale automated analysis of ChatGPT platform usage in a privacy-preserving manner, analyzing over 3 million conversations for affective cues and surveying over 4,000 users on their perceptions of ChatGPT. To investigate whether there is a relationship between model usage and emotional well-being, we conduct an Institutional Review Board (IRB)-approved randomized controlled trial (RCT) on close to 1,000 participants over 28 days, examining changes in their emotional well-being as they interact with ChatGPT under different experimental settings. In both on-platform data analysis and the RCT, we observe that very high usage correlates with increased self-reported indicators of dependence. From our RCT, we find that the impact of voice-based interactions on emotional well-being to be highly nuanced, and influenced by factors such as the user's initial emotional state and total usage duration. Overall, our analysis reveals that a small number of users are responsible for a disproportionate share of the most affective cues.

Jason Phang

Gisting (Mu et al., 2023) is a simple method for training models to compress information into fewer token representations using a modified attention mask, and can serve as an economical approach to training Transformer-based hypernetworks. We introduce HyperLlama, a set of Gisting-based hypernetworks built on Llama-2 models that generates task-specific soft prefixes based on few-shot inputs. In experiments across P3, Super-NaturalInstructions and Symbol Tuning datasets, we show that HyperLlama models can effectively compress information from few-shot examples into soft prefixes. However, they still underperform multi-task fine-tuned language models with full attention over few-shot in-context examples. We also show that HyperLlama-generated soft prefixes can serve as better initializations for further prefix tuning. Overall, Gisting-based hypernetworks are economical and easy to implement, but have mixed empirical performance.

Angelica Chen, Jason Phang, Alicia Parrish et al.

Large language models (LLMs) have achieved widespread success on a variety of in-context few-shot tasks, but this success is typically evaluated via correctness rather than consistency. We argue that self-consistency is an important criteria for valid multi-step reasoning in tasks where the solution is composed of the answers to multiple sub-steps. We propose two types of self-consistency that are particularly important for multi-step reasoning -- hypothetical consistency (a model's ability to predict what its output would be in a hypothetical other context) and compositional consistency (consistency of a model's final outputs when intermediate sub-steps are replaced with the model's outputs for those steps). We demonstrate that multiple variants of the GPT-3/-4 models exhibit poor consistency rates across both types of consistency on a variety of tasks.

Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi et al.

To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%).

Jason Phang, Angelica Chen, William Huang et al.

More capable language models increasingly saturate existing task benchmarks, in some cases outperforming humans. This has left little headroom with which to measure further progress. Adversarial dataset creation has been proposed as a strategy to construct more challenging datasets, and two common approaches are: (1) filtering out easy examples and (2) model-in-the-loop data collection. In this work, we study the impact of applying each approach to create more challenging evaluation datasets. We adapt the AFLite algorithm to filter evaluation data, and run experiments against 18 different adversary models. We find that AFLite indeed selects more challenging examples, lowering the performance of evaluated models more as stronger adversary models are used. However, the resulting ranking of models can also be unstable and highly sensitive to the choice of adversary model used. Moreover, AFLite oversamples examples with low annotator agreement, meaning that model comparisons hinge on the most contentiously labeled examples. Smaller-scale experiments on the adversarially collected datasets ANLI and AdversarialQA show similar findings, broadly lowering performance with stronger adversaries while disproportionately affecting the adversary model.

Alicia Parrish, Angelica Chen, Nikita Nangia et al.

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluates model responses at two levels: (i) given an under-informative context, we test how strongly responses reflect social biases, and (ii) given an adequately informative context, we test whether the model's biases override a correct answer choice. We find that models often rely on stereotypes when the context is under-informative, meaning the model's outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conflicts, with this difference widening to over 5 points on examples targeting gender for most models tested.

Jason Phang, Haokun Liu, Samuel R. Bowman

Despite the success of fine-tuning pretrained language encoders like BERT for downstream natural language understanding (NLU) tasks, it is still poorly understood how neural networks change after fine-tuning. In this work, we use centered kernel alignment (CKA), a method for comparing learned representations, to measure the similarity of representations in task-tuned models across layers. In experiments across twelve NLU tasks, we discover a consistent block diagonal structure in the similarity of representations within fine-tuned RoBERTa and ALBERT models, with strong similarity within clusters of earlier and later layers, but not between them. The similarity of later layer representations implies that later layers only marginally contribute to task performance, and we verify in experiments that the top few layers of fine-tuned Transformers can be discarded without hurting performance, even with no further tuning.

Clara Vania, Phu Mon Htut, William Huang et al.

Recent years have seen numerous NLP datasets introduced to evaluate the performance of fine-tuned models on natural language understanding tasks. Recent results from large pretrained models, though, show that many of these datasets are largely saturated and unlikely to be able to detect further progress. What kind of datasets are still effective at discriminating among strong models, and what kind of datasets should we expect to be able to detect future improvements? To measure this uniformly across datasets, we draw on Item Response Theory and evaluate 29 datasets using predictions from 18 pretrained Transformer models on individual test examples. We find that Quoref, HellaSwag, and MC-TACO are best suited for distinguishing among state-of-the-art models, while SNLI, MNLI, and CommitmentBank seem to be saturated for current strong models. We also observe span selection task format, which is used for QA datasets like QAMR or SQuAD2.0, is effective in differentiating between strong and weak models.

Leo Gao, Stella Biderman, Sid Black et al.

Recent work has demonstrated that increased training dataset diversity improves general cross-domain knowledge and downstream generalization capability for large-scale language models. With this in mind, we present \textit{the Pile}: an 825 GiB English text corpus targeted at training large-scale language models. The Pile is constructed from 22 diverse high-quality subsets -- both existing and newly constructed -- many of which derive from academic or professional sources. Our evaluation of the untuned performance of GPT-2 and GPT-3 on the Pile shows that these models struggle on many of its components, such as academic writing. Conversely, models trained on the Pile improve significantly over both Raw CC and CC-100 on all components of the Pile, while improving performance on downstream evaluations. Through an in-depth exploratory analysis, we document potentially concerning aspects of the data for prospective users. We make publicly available the code used in its construction.

Jason Phang, Jungkyu Park, Krzysztof J. Geras

Saliency maps that identify the most informative regions of an image for a classifier are valuable for model interpretability. A common approach to creating saliency maps involves generating input masks that mask out portions of an image to maximally deteriorate classification performance, or mask in an image to preserve classification performance. Many variants of this approach have been proposed in the literature, such as counterfactual generation and optimizing over a Gumbel-Softmax distribution. Using a general formulation of masking-based saliency methods, we conduct an extensive evaluation study of a number of recently proposed variants to understand which elements of these methods meaningfully improve performance. Surprisingly, we find that a well-tuned, relatively simple formulation of a masking-based saliency model outperforms many more complex approaches. We find that the most important ingredients for high quality saliency map generation are (1) using both masked-in and masked-out objectives and (2) training the classifier alongside the masking model. Strikingly, we show that a masking model can be trained with as few as 10 examples per class and still generate saliency maps with only a 0.7-point increase in localization error.

Nan Wu, Zhe Huang, Yiqiu Shen et al.

Breast cancer is the most common cancer in women, and hundreds of thousands of unnecessary biopsies are done around the world at a tremendous cost. It is crucial to reduce the rate of biopsies that turn out to be benign tissue. In this study, we build deep neural networks (DNNs) to classify biopsied lesions as being either malignant or benign, with the goal of using these networks as second readers serving radiologists to further reduce the number of false positive findings. We enhance the performance of DNNs that are trained to learn from small image patches by integrating global context provided in the form of saliency maps learned from the entire image into their reasoning, similar to how radiologists consider global context when evaluating areas of interest. Our experiments are conducted on a dataset of 229,426 screening mammography exams from 141,473 patients. We achieve an AUC of 0.8 on a test set consisting of 464 benign and 136 malignant lesions.

Jason Phang, Iacer Calixto, Phu Mon Htut et al.

Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tasks and moderate improvements on question-answering target tasks. MNLI, SQuAD and HellaSwag achieve the best overall results as intermediate tasks, while multi-task intermediate offers small additional improvements. Using our best intermediate-task models for each target task, we obtain a 5.4 point improvement over XLM-R Large on the XTREME benchmark, setting the state of the art as of June 2020. We also investigate continuing multilingual MLM during intermediate-task training and using machine-translated intermediate-task data, but neither consistently outperforms simply performing English intermediate-task training.

Yada Pruksachatkun, Jason Phang, Haokun Liu et al.

While pretrained models such as BERT have shown large gains across natural language understanding tasks, their performance can be improved by further training the model on a data-rich intermediate task, before fine-tuning it on a target task. However, it is still poorly understood when and why intermediate-task training is beneficial for a given target task. To investigate this, we perform a large-scale study on the pretrained RoBERTa model with 110 intermediate-target task combinations. We further evaluate all trained models with 25 probing tasks meant to reveal the specific skills that drive transfer. We observe that intermediate tasks requiring high-level inference and reasoning abilities tend to work best. We also observe that target task performance is strongly correlated with higher-level abilities such as coreference resolution. However, we fail to observe more granular correlations between probing and target task performance, highlighting the need for further work on broad-coverage probing benchmarks. We also observe evidence that the forgetting of knowledge learned during pretraining may limit our analysis, highlighting the need for further work on transfer learning methods in these settings.

Phu Mon Htut, Jason Phang, Shikha Bordia et al.

We investigate the extent to which individual attention heads in pretrained transformer language models, such as BERT and RoBERTa, implicitly capture syntactic dependency relations. We employ two methods---taking the maximum attention weight and computing the maximum spanning tree---to extract implicit dependency relations from the attention weights of each layer/head, and compare them to the ground-truth Universal Dependency (UD) trees. We show that, for some UD relation types, there exist heads that can recover the dependency type significantly better than baselines on parsed English text, suggesting that some self-attention heads act as a proxy for syntactic structure. We also analyze BERT fine-tuned on two datasets---the syntax-oriented CoLA and the semantics-oriented MNLI---to investigate whether fine-tuning affects the patterns of their self-attention, but we do not observe substantial differences in the overall dependency relations extracted using our methods. Our results suggest that these models have some specialist attention heads that track individual dependency types, but no generalist head that performs holistic parsing significantly better than a trivial baseline, and that analyzing attention weights directly may not reveal much of the syntactic knowledge that BERT-style models are known to learn.

Alex Warstadt, Yu Cao, Ioana Grosu et al.

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing in English, as a case study for our experiments. NPIs like "any" are grammatical only if they appear in a licensing environment like negation ("Sue doesn't have any cats" vs. "Sue has any cats"). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model's grammatical knowledge in a given domain.

Thibault Févry, Jason Phang, Nan Wu et al.

We trained and evaluated a localization-based deep CNN for breast cancer screening exam classification on over 200,000 exams (over 1,000,000 images). Our model achieves an AUC of 0.919 in predicting malignancy in patients undergoing breast cancer screening, reducing the error rate of the baseline (Wu et al., 2019a) by 23%. In addition, the models generates bounding boxes for benign and malignant findings, providing interpretable predictions.

Jungkyu Park, Jason Phang, Yiqiu Shen et al.

Radiologists typically compare a patient's most recent breast cancer screening exam to their previous ones in making informed diagnoses. To reflect this practice, we propose new neural network models that compare pairs of screening mammograms from the same patient. We train and evaluate our proposed models on over 665,000 pairs of images (over 166,000 pairs of exams). Our best model achieves an AUC of 0.866 in predicting malignancy in patients who underwent breast cancer screening, reducing the error rate of the corresponding baseline.

Yiqiu Shen, Nan Wu, Jason Phang et al.

Deep learning models designed for visual classification tasks on natural images have become prevalent in medical image analysis. However, medical images differ from typical natural images in many ways, such as significantly higher resolutions and smaller regions of interest. Moreover, both the global structure and local details play important roles in medical image analysis tasks. To address these unique properties of medical images, we propose a neural network that is able to classify breast cancer lesions utilizing information from both a global saliency map and multiple local patches. The proposed model outperforms the ResNet-based baseline and achieves radiologist-level performance in the interpretation of screening mammography. Although our model is trained only with image-level labels, it is able to generate pixel-level saliency maps that provide localization of possible malignant findings.

Nan Wu, Jason Phang, Jungkyu Park et al.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 images). Our network achieves an AUC of 0.895 in predicting whether there is a cancer in the breast, when tested on the screening population. We attribute the high accuracy of our model to a two-stage training procedure, which allows us to use a very high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and find our model to be as accurate as experienced radiologists when presented with the same data. Finally, we show that a hybrid model, averaging probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To better understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, model design, training procedure, errors, and properties of its internal representations.

Jason Phang, Thibault Févry, Samuel R. Bowman

Pretraining sentence encoders with language modeling and related unsupervised tasks has recently been shown to be very effective for language understanding tasks. By supplementing language model-style pretraining with further training on data-rich supervised tasks, such as natural language inference, we obtain additional performance improvements on the GLUE benchmark. Applying supplementary training on BERT (Devlin et al., 2018), we attain a GLUE score of 81.8---the state of the art (as of 02/24/2019) and a 1.4 point improvement over BERT. We also observe reduced variance across random restarts in this setting. Our approach yields similar improvements when applied to ELMo (Peters et al., 2018a) and Radford et al. (2018)'s model. In addition, the benefits of supplementary training are particularly pronounced in data-constrained regimes, as we show in experiments with artificially limited training data.

Thibault Févry, Jason Phang

In sentence compression, the task of shortening sentences while retaining the original meaning, models tend to be trained on large corpora containing pairs of verbose and compressed sentences. To remove the need for paired corpora, we emulate a summarization task and add noise to extend sentences and train a denoising auto-encoder to recover the original, constructing an end-to-end training regime without the need for any examples of compressed sentences. We conduct a human evaluation of our model on a standard text summarization dataset and show that it performs comparably to a supervised baseline based on grammatical correctness and retention of meaning. Despite being exposed to no target data, our unsupervised models learn to generate imperfect but reasonably readable sentence summaries. Although we underperform supervised models based on ROUGE scores, our models are competitive with a supervised baseline based on human evaluation for grammatical correctness and retention of meaning.