Llama 2: Open Foundation and Fine-Tuned Chat Models

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

Weisfeiler and Lehman Go Measurement Modeling: Probing the Validity of the WL Test

The expressive power of graph neural networks is usually measured by comparing how many pairs of graphs or nodes an architecture can possibly distinguish as non-isomorphic to those distinguishable by the $k$-dimensional Weisfeiler-Lehman ($k$-WL) test. In this paper, we uncover misalignments between practitioners' conceptualizations of expressive power and $k$-WL through a systematic analysis of the reliability and validity of $k$-WL. We further conduct a survey ($n = 18$) of practitioners to surface their conceptualizations of expressive power and their assumptions about $k$-WL. In contrast to practitioners' opinions, our analysis (which draws from graph theory and benchmark auditing) reveals that $k$-WL does not guarantee isometry, can be irrelevant to real-world graph tasks, and may not promote generalization or trustworthiness. We argue for extensional definitions and measurement of expressive power based on benchmarks; we further contribute guiding questions for constructing such benchmarks, which is critical for progress in graph machine learning.

Call for Papers -- The BabyLM Challenge: Sample-efficient pretraining on a developmentally plausible corpus

We present the call for papers for the BabyLM Challenge: Sample-efficient pretraining on a developmentally plausible corpus. This shared task is intended for participants with an interest in small scale language modeling, human language acquisition, low-resource NLP, and cognitive modeling. In partnership with CoNLL and CMCL, we provide a platform for approaches to pretraining with a limited-size corpus sourced from data inspired by the input to children. The task has three tracks, two of which restrict the training data to pre-released datasets of 10M and 100M words and are dedicated to explorations of approaches such as architectural variations, self-supervised objectives, or curriculum learning. The final track only restricts the amount of text used, allowing innovation in the choice of the data, its domain, and even its modality (i.e., data from sources other than text is welcome). We will release a shared evaluation pipeline which scores models on a variety of benchmarks and tasks, including targeted syntactic evaluations and natural language understanding.

Language model acceptability judgements are not always robust to context

Targeted syntactic evaluations of language models ask whether models show stable preferences for syntactically acceptable content over minimal-pair unacceptable inputs. Most targeted syntactic evaluation datasets ask models to make these judgements with just a single context-free sentence as input. This does not match language models' training regime, in which input sentences are always highly contextualized by the surrounding corpus. This mismatch raises an important question: how robust are models' syntactic judgements in different contexts? In this paper, we investigate the stability of language models' performance on targeted syntactic evaluations as we vary properties of the input context: the length of the context, the types of syntactic phenomena it contains, and whether or not there are violations of grammaticality. We find that model judgements are generally robust when placed in randomly sampled linguistic contexts. However, they are substantially unstable for contexts containing syntactic structures matching those in the critical test content. Among all tested models (GPT-2 and five variants of OPT), we significantly improve models' judgements by providing contexts with matching syntactic structures, and conversely significantly worsen them using unacceptable contexts with matching but violated syntactic structures. This effect is amplified by the length of the context, except for unrelated inputs. We show that these changes in model performance are not explainable by simple features matching the context and the test inputs, such as lexical overlap and dependency overlap. This sensitivity to highly specific syntactic features of the context can only be explained by the models' implicit in-context learning abilities.

The Curious Case of Absolute Position Embeddings

Transformer language models encode the notion of word order using positional information. Most commonly, this positional information is represented by absolute position embeddings (APEs), that are learned from the pretraining data. However, in natural language, it is not absolute position that matters, but relative position, and the extent to which APEs can capture this type of information has not been investigated. In this work, we observe that models trained with APE over-rely on positional information to the point that they break-down when subjected to sentences with shifted position information. Specifically, when models are subjected to sentences starting from a non-zero position (excluding the effect of priming), they exhibit noticeably degraded performance on zero to full-shot tasks, across a range of model families and model sizes. Our findings raise questions about the efficacy of APEs to model the relativity of position information, and invite further introspection on the sentence and word order processing strategies employed by these models.

Learning Transductions to Test Systematic Compositionality

Recombining known primitive concepts into larger novel combinations is a quintessentially human cognitive capability. Whether large neural models in NLP acquire this ability while learning from data is an open question. In this paper, we look at this problem from the perspective of formal languages. We use deterministic finite-state transducers to make an unbounded number of datasets with controllable properties governing compositionality. By randomly sampling over many transducers, we explore which of their properties (number of states, alphabet size, number of transitions etc.) contribute to learnability of a compositional relation by a neural network. In general, we find that the models either learn the relations completely or not at all. The key is transition coverage, setting a soft learnability limit at 400 examples per transition.

DataPerf: Benchmarks for Data-Centric AI Development

Machine learning (ML) research has generally focused on models, while the most prominent datasets have been employed for everyday ML tasks without regard for the breadth, difficulty, and faithfulness of these datasets to the underlying problem. Neglecting the fundamental importance of datasets has caused major problems involving data cascades in real-world applications and saturation of dataset-driven criteria for model quality, hindering research growth. To solve this problem, we present DataPerf, a benchmark package for evaluating ML datasets and dataset-working algorithms. We intend it to enable the "data ratchet," in which training sets will aid in evaluating test sets on the same problems, and vice versa. Such a feedback-driven strategy will generate a virtuous loop that will accelerate development of data-centric AI. The MLCommons Association will maintain DataPerf.

Perturbation Augmentation for Fairer NLP

Unwanted and often harmful social biases are becoming ever more salient in NLP research, affecting both models and datasets. In this work, we ask: does training on demographically perturbed data lead to more fair language models? We collect a large dataset of human annotated text perturbations and train an automatic perturber on it, which we show to outperform heuristic alternatives. We find: (i) Language models (LMs) pre-trained on demographically perturbed corpora are more fair, at least, according to our current best metrics for measuring model fairness, and (ii) LMs finetuned on perturbed GLUE datasets exhibit less demographic bias on downstream tasks. We find that improved fairness does not come at the expense of accuracy. Although our findings appear promising, there are still some limitations, as well as outstanding questions about how best to evaluate the (un)fairness of large language models. We hope that this initial exploration of neural demographic perturbation will help drive more improvement towards fairer NLP.

"I'm sorry to hear that": finding bias in language models with a holistic descriptor dataset

As language models grow in popularity, their biases across all possible markers of demographic identity should be measured and addressed in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes, and are commonly used with preset bias tests that presuppose which types of biases the models exhibit. In this work, we present a new, more inclusive dataset, HOLISTICBIAS, which consists of nearly 600 descriptor terms across 13 different demographic axes. HOLISTICBIAS was assembled in conversation with experts and community members with lived experience through a participatory process. We use these descriptors combinatorially in a set of bias measurement templates to produce over 450,000 unique sentence prompts, and we use these prompts to explore, identify, and reduce novel forms of bias in several generative models. We demonstrate that our dataset is highly efficacious for measuring previously unmeasurable biases in token likelihoods and generations from language models, as well as in an offensiveness classifier. We will invite additions and amendments to the dataset, and we hope it will help serve as a basis for easy-to-use and more standardized methods for evaluating bias in NLP models.

Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality

We present a novel task and dataset for evaluating the ability of vision and language models to conduct visio-linguistic compositional reasoning, which we call Winoground. Given two images and two captions, the goal is to match them correctly - but crucially, both captions contain a completely identical set of words, only in a different order. The dataset was carefully hand-curated by expert annotators and is labeled with a rich set of fine-grained tags to assist in analyzing model performance. We probe a diverse range of state-of-the-art vision and language models and find that, surprisingly, none of them do much better than chance. Evidently, these models are not as skilled at visio-linguistic compositional reasoning as we might have hoped. We perform an extensive analysis to obtain insights into how future work might try to mitigate these models' shortcomings. We aim for Winoground to serve as a useful evaluation set for advancing the state of the art and driving further progress in the field. The dataset is available at https://huggingface.co/datasets/facebook/winoground.