Jos Rozen

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.

Germán Kruszewski, Jos Rozen, Marc Dymetman

Pre-trained language models and other generative models have revolutionized NLP and beyond. However, these models tend to reproduce undesirable biases present in their training data. Also, they may overlook patterns that are important but challenging to capture. To address these limitations, researchers have introduced distributional control techniques. These techniques, not limited to language, allow controlling the prevalence (i.e., expectations) of any features of interest in the model's outputs. Despite their potential, the widespread adoption of these techniques has been hindered by the difficulty in adapting complex, disconnected code. Here, we present disco, an open-source Python library that brings these techniques to the broader public.

Dongyoung Go, Tomasz Korbak, Germán Kruszewski et al.

Aligning language models with preferences can be posed as approximating a target distribution representing some desired behavior. Existing approaches differ both in the functional form of the target distribution and the algorithm used to approximate it. For instance, Reinforcement Learning from Human Feedback (RLHF) corresponds to minimizing a reverse KL from an implicit target distribution arising from a KL penalty in the objective. On the other hand, Generative Distributional Control (GDC) has an explicit target distribution and minimizes a forward KL from it using the Distributional Policy Gradient (DPG) algorithm. In this paper, we propose a new approach, f-DPG, which allows the use of any f-divergence to approximate any target distribution that can be evaluated. f-DPG unifies both frameworks (RLHF, GDC) and the approximation methods (DPG, RL with KL penalties). We show the practical benefits of various choices of divergence objectives and demonstrate that there is no universally optimal objective but that different divergences present different alignment and diversity trade-offs. We show that Jensen-Shannon divergence strikes a good balance between these objectives, and frequently outperforms forward KL divergence by a wide margin, leading to significant improvements over prior work. These distinguishing characteristics between divergences persist as the model size increases, highlighting the importance of selecting appropriate divergence objectives.

Dongyoung Go, Tomasz Korbak, Germán Kruszewski et al.

As language models (LMs) become more capable, it is increasingly important to align them with human preferences. However, the dominant paradigm for training Preference Models (PMs) for that purpose suffers from fundamental limitations, such as lack of transparency and scalability, along with susceptibility to overfitting the preference dataset. We propose Compositional Preference Models (CPMs), a novel PM framework that decomposes one global preference assessment into several interpretable features, obtains scalar scores for these features from a prompted LM, and aggregates these scores using a logistic regression classifier. Through these simple steps, CPMs allow to control which properties of the preference data are used to train the preference model and to build it based on features that are believed to underlie the human preference judgment. Our experiments show that CPMs not only improve generalization and are more robust to overoptimization than standard PMs, but also that best-of-n samples obtained using CPMs tend to be preferred over samples obtained using conventional PMs. Overall, our approach demonstrates the benefits of endowing PMs with priors about which features determine human preferences while relying on LM capabilities to extract those features in a scalable and robust way.

Nadezhda Chirkova, Germán Kruszewski, Jos Rozen et al.

Autoregressive language models (LMs) map token sequences to probabilities. The usual practice for computing the probability of any character string (e.g. English sentences) is to first transform it into a sequence of tokens that is scored by the model. However, there are exponentially many token sequences that represent any given string. To truly compute the probability of a string one should marginalize over all tokenizations, which is typically intractable. Here, we analyze whether the practice of ignoring the marginalization is justified. To this end, we devise an importance-sampling-based algorithm that allows us to compute estimates of the marginal probabilities and compare them to the default procedure in a range of state-of-the-art models and datasets. Our results show that the gap in log-likelihood is no larger than 0.5% in most cases, but that it becomes more pronounced for data with long complex words.

Victor Sanh, Albert Webson, Colin Raffel et al.

Large language models have recently been shown to attain reasonable zero-shot generalization on a diverse set of tasks (Brown et al., 2020). It has been hypothesized that this is a consequence of implicit multitask learning in language models' pretraining (Radford et al., 2019). Can zero-shot generalization instead be directly induced by explicit multitask learning? To test this question at scale, we develop a system for easily mapping any natural language tasks into a human-readable prompted form. We convert a large set of supervised datasets, each with multiple prompts with diverse wording. These prompted datasets allow for benchmarking the ability of a model to perform completely held-out tasks. We fine-tune a pretrained encoder-decoder model (Raffel et al., 2020; Lester et al., 2021) on this multitask mixture covering a wide variety of tasks. The model attains strong zero-shot performance on several standard datasets, often outperforming models up to 16x its size. Further, our approach attains strong performance on a subset of tasks from the BIG-bench benchmark, outperforming models up to 6x its size. All trained models are available at https://github.com/bigscience-workshop/t-zero and all prompts are available at https://github.com/bigscience-workshop/promptsource.

Thibaut Thonet, Jos Rozen, Laurent Besacier

Research on Large Language Models (LLMs) has recently witnessed an increasing interest in extending the models' context size to better capture dependencies within long documents. While benchmarks have been proposed to assess long-range abilities, existing efforts primarily considered generic tasks that are not necessarily aligned with real-world applications. In contrast, we propose a new benchmark for long-context LLMs focused on a practical meeting assistant scenario in which the long contexts consist of transcripts obtained by automatic speech recognition, presenting unique challenges for LLMs due to the inherent noisiness and oral nature of such data. Our benchmark, ELITR-Bench, augments the existing ELITR corpus by adding 271 manually crafted questions with their ground-truth answers, as well as noisy versions of meeting transcripts altered to target different Word Error Rate levels. Our experiments with 12 long-context LLMs on ELITR-Bench confirm the progress made across successive generations of both proprietary and open models, and point out their discrepancies in terms of robustness to transcript noise. We also provide a thorough analysis of our GPT-4-based evaluation, including insights from a crowdsourcing study. Our findings indicate that while GPT-4's scores align with human judges, its ability to distinguish beyond three score levels may be limited.

Thibaut Thonet, Germán Kruszewski, Jos Rozen et al.

LLM-powered conversational assistants are often deployed in a one-size-fits-all manner, which fails to accommodate individual user preferences. Recently, LLM personalization -- tailoring models to align with specific user preferences -- has gained increasing attention as a way to bridge this gap. In this work, we specifically focus on a practical yet challenging setting where only a small set of preference annotations can be collected per user -- a problem we define as Personalized Preference Alignment with Limited Data (PPALLI). To support research in this area, we introduce two datasets -- DnD and ELIP -- and benchmark a variety of alignment techniques on them. We further propose FaST, a highly parameter-efficient approach that leverages high-level features automatically discovered from the data, achieving the best overall performance.

Germán Kruszewski, Pierre Erbacher, Jos Rozen et al.

Reinforcement Learning (RL) has become the de facto standard for tuning LLMs to solve tasks involving reasoning. However, growing evidence shows that models trained in such way often suffer from a significant loss in diversity. We argue that this arises because RL implicitly optimizes the "mode-seeking" or "zero-forcing" Reverse KL to a target distribution causing the model to concentrate mass on certain high-probability regions of the target while neglecting others. In this work, we instead begin from an explicit target distribution, obtained by filtering out incorrect answers while preserving the relative probabilities of correct ones. Starting from a pre-trained LLM, we approximate this target distribution using the $α$-divergence family, which unifies prior approaches and enables direct control of the precision-diversity trade-off by interpolating between mode-seeking and mass-covering divergences. On a Lean theorem-proving benchmark, our method achieves state-of-the-art performance along the coverage-precision Pareto frontier, outperforming all prior methods on the coverage axis.

Matthias Gallé, Jos Rozen, Germán Kruszewski et al.

The power of natural language generation models has provoked a flurry of interest in automatic methods to detect if a piece of text is human or machine-authored. The problem so far has been framed in a standard supervised way and consists in training a classifier on annotated data to predict the origin of one given new document. In this paper, we frame the problem in an unsupervised and distributional way: we assume that we have access to a large collection of unannotated documents, a big fraction of which is machine-generated. We propose a method to detect those machine-generated documents leveraging repeated higher-order n-grams, which we show over-appear in machine-generated text as compared to human ones. That weak signal is the starting point of a self-training setting where pseudo-labelled documents are used to train an ensemble of classifiers. Our experiments show that leveraging that signal allows us to rank suspicious documents accurately. Precision at 5000 is over 90% for top-k sampling strategies, and over 80% for nucleus sampling for the largest model we used (GPT2-large). The drop with increased size of model is small, which could indicate that the results hold for other current and future large language models.

Hady Elsahar, Maximin Coavoux, Matthias Gallé et al.

We address the problem of unsupervised abstractive summarization of collections of user generated reviews with self-supervision and control. We propose a self-supervised setup that considers an individual document as a target summary for a set of similar documents. This setting makes training simpler than previous approaches by relying only on standard log-likelihood loss. We address the problem of hallucinations through the use of control codes, to steer the generation towards more coherent and relevant summaries.Finally, we extend the Transformer architecture to allow for multiple reviews as input. Our benchmarks on two datasets against graph-based and recent neural abstractive unsupervised models show that our proposed method generates summaries with a superior quality and relevance.This is confirmed in our human evaluation which focuses explicitly on the faithfulness of generated summaries We also provide an ablation study, which shows the importance of the control setup in controlling hallucinations and achieve high sentiment and topic alignment of the summaries with the input reviews.