Todor Mihaylov

Hugo Touvron, Louis Martin, Kevin Stone et al. · meta-ai

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.

Susan Zhang, Stephen Roller, Naman Goyal et al. · meta-ai, uw

Large language models, which are often trained for hundreds of thousands of compute days, have shown remarkable capabilities for zero- and few-shot learning. Given their computational cost, these models are difficult to replicate without significant capital. For the few that are available through APIs, no access is granted to the full model weights, making them difficult to study. We present Open Pre-trained Transformers (OPT), a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters, which we aim to fully and responsibly share with interested researchers. We show that OPT-175B is comparable to GPT-3, while requiring only 1/7th the carbon footprint to develop. We are also releasing our logbook detailing the infrastructure challenges we faced, along with code for experimenting with all of the released models.

Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian et al. · meta-ai

Vision-language models trained with contrastive learning on large-scale noisy data are becoming increasingly popular for zero-shot recognition problems. In this paper we improve the following three aspects of the contrastive pre-training pipeline: dataset noise, model initialization and the training objective. First, we propose a straightforward filtering strategy titled Complexity, Action, and Text-spotting (CAT) that significantly reduces dataset size, while achieving improved performance across zero-shot vision-language tasks. Next, we propose an approach titled Concept Distillation to leverage strong unimodal representations for contrastive training that does not increase training complexity while outperforming prior work. Finally, we modify the traditional contrastive alignment objective, and propose an importance-sampling approach to up-sample the importance of hard-negatives without adding additional complexity. On an extensive zero-shot benchmark of 29 tasks, our Distilled and Hard-negative Training (DiHT) approach improves on 20 tasks compared to the baseline. Furthermore, for few-shot linear probing, we propose a novel approach that bridges the gap between zero-shot and few-shot performance, substantially improving over prior work. Models are available at https://github.com/facebookresearch/diht.

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru et al. · berkeley, meta-ai

Recent work has shown that fine-tuning large pre-trained language models on a collection of tasks described via instructions, a.k.a. instruction-tuning, improves their zero and few-shot generalization to unseen tasks. However, there is a limited understanding of the performance trade-offs of different decisions made during the instruction-tuning process. These decisions include the scale and diversity of the instruction-tuning benchmark, different task sampling strategies, fine-tuning with and without demonstrations, training using specialized datasets for reasoning and dialogue, and finally, the fine-tuning objectives themselves. In this paper, we characterize the effect of instruction-tuning decisions on downstream task performance when scaling both model and benchmark sizes. To this end, we create OPT-IML Bench: a large benchmark for Instruction Meta-Learning (IML) of 2000 NLP tasks consolidated into task categories from 8 existing benchmarks, and prepare an evaluation framework to measure three types of model generalizations: to tasks from fully held-out categories, to held-out tasks from seen categories, and to held-out instances from seen tasks. Through the lens of this framework, we first present insights about instruction-tuning decisions as applied to OPT-30B and further exploit these insights to train OPT-IML 30B and 175B, which are instruction-tuned versions of OPT. OPT-IML demonstrates all three generalization abilities at both scales on four different evaluation benchmarks with diverse tasks and input formats -- PromptSource, FLAN, Super-NaturalInstructions, and UnifiedSKG. Not only does it significantly outperform OPT on all benchmarks but is also highly competitive with existing models fine-tuned on each specific benchmark. We release OPT-IML at both scales, together with the OPT-IML Bench evaluation framework.

Xiaochuang Han, Daniel Simig, Todor Mihaylov et al. · berkeley, meta-ai

In-context learning (ICL) improves language models' performance on a variety of NLP tasks by simply demonstrating a handful of examples at inference time. It is not well understood why ICL ability emerges, as the model has never been specifically trained on such demonstrations. Unlike prior work that explores implicit mechanisms behind ICL, we study ICL via investigating the pretraining data. Specifically, we first adapt an iterative, gradient-based approach to find a small subset of pretraining data that supports ICL. We observe that a continued pretraining on this small subset significantly improves the model's ICL ability, by up to 18%. We then compare the supportive subset constrastively with random subsets of pretraining data and discover: (1) The supportive pretraining data to ICL do not have a higher domain relevance to downstream tasks. (2) The supportive pretraining data have a higher mass of rarely occurring, long-tail tokens. (3) The supportive pretraining data are challenging examples where the information gain from long-range context is below average, indicating learning to incorporate difficult long-range context encourages ICL. Our work takes a first step towards understanding ICL via analyzing instance-level pretraining data. Our insights have a potential to enhance the ICL ability of language models by actively guiding the construction of pretraining data in the future.

Momchil Hardalov, Pepa Atanasova, Todor Mihaylov et al. · berkeley

We present bgGLUE(Bulgarian General Language Understanding Evaluation), a benchmark for evaluating language models on Natural Language Understanding (NLU) tasks in Bulgarian. Our benchmark includes NLU tasks targeting a variety of NLP problems (e.g., natural language inference, fact-checking, named entity recognition, sentiment analysis, question answering, etc.) and machine learning tasks (sequence labeling, document-level classification, and regression). We run the first systematic evaluation of pre-trained language models for Bulgarian, comparing and contrasting results across the nine tasks in the benchmark. The evaluation results show strong performance on sequence labeling tasks, but there is a lot of room for improvement for tasks that require more complex reasoning. We make bgGLUE publicly available together with the fine-tuning and the evaluation code, as well as a public leaderboard at https://bgglue.github.io/, and we hope that it will enable further advancements in developing NLU models for Bulgarian.

Mingda Chen, Jingfei Du, Ramakanth Pasunuru et al. · uw

Self-supervised pretraining has made few-shot learning possible for many NLP tasks. But the pretraining objectives are not typically adapted specifically for in-context few-shot learning. In this paper, we propose to use self-supervision in an intermediate training stage between pretraining and downstream few-shot usage with the goal to teach the model to perform in-context few shot learning. We propose and evaluate four self-supervised objectives on two benchmarks. We find that the intermediate self-supervision stage produces models that outperform strong baselines. Ablation study shows that several factors affect the downstream performance, such as the amount of training data and the diversity of the self-supervised objectives. Human-annotated cross-task supervision and self-supervision are complementary. Qualitative analysis suggests that the self-supervised-trained models are better at following task requirements.

Momchil Hardalov, Todor Mihaylov, Dimitrina Zlatkova et al.

We propose EXAMS -- a new benchmark dataset for cross-lingual and multilingual question answering for high school examinations. We collected more than 24,000 high-quality high school exam questions in 16 languages, covering 8 language families and 24 school subjects from Natural Sciences and Social Sciences, among others. EXAMS offers a fine-grained evaluation framework across multiple languages and subjects, which allows precise analysis and comparison of various models. We perform various experiments with existing top-performing multilingual pre-trained models and we show that EXAMS offers multiple challenges that require multilingual knowledge and reasoning in multiple domains. We hope that EXAMS will enable researchers to explore challenging reasoning and knowledge transfer methods and pre-trained models for school question answering in various languages which was not possible before. The data, code, pre-trained models, and evaluation are available at https://github.com/mhardalov/exams-qa.

William Held, Bhargavi Paranjape, Punit Singh Koura et al. · gatech

Large Language Models improve with increasing amounts of high-quality training data. However, leveraging larger datasets requires balancing quality, quantity, and diversity across sources. After evaluating nine baseline methods under both compute- and data-constrained scenarios, we find token-count heuristics outperform manual and learned mixes, indicating that simple approaches accounting for dataset size and diversity are surprisingly effective. Building on this insight, we propose two complementary approaches: UtiliMax, which extends token-based heuristics by incorporating utility estimates from reduced-scale ablations, achieving up to a 10.6x speedup over manual baselines; and Model Estimated Data Utility (MEDU), which leverages LLMs to estimate data utility from small samples, matching ablation-based performance while reducing computational requirements by $\sim$200x. Together, these approaches establish a new framework for automated, compute-efficient data mixing that is robust across training regimes.

Rylan Schaeffer, Punit Singh Koura, Binh Tang et al.

The explosion of high-performing conversational language models (LMs) has spurred a shift from classic natural language processing (NLP) benchmarks to expensive, time-consuming and noisy human evaluations - yet the relationship between these two evaluation strategies remains hazy. In this paper, we conduct a large-scale study of four Chat Llama 2 models, comparing their performance on 160 standard NLP benchmarks (e.g., MMLU, ARC, BIG-Bench Hard) against extensive human preferences on more than 11k single-turn and 2k multi-turn dialogues from over 2k human annotators. Our findings are striking: most NLP benchmarks strongly correlate with human evaluations, suggesting that cheaper, automated metrics can serve as surprisingly reliable predictors of human preferences. Three human evaluations, such as adversarial dishonesty and safety, are anticorrelated with NLP benchmarks, while two are uncorrelated. Moreover, through overparameterized linear regressions, we show that NLP scores can accurately predict human evaluations across different model scales, offering a path to reduce costly human annotation without sacrificing rigor. Overall, our results affirm the continued value of classic benchmarks and illuminate how to harness them to anticipate real-world user satisfaction - pointing to how NLP benchmarks can be leveraged to meet evaluation needs of our new era of conversational AI.

Mikel Artetxe, Shruti Bhosale, Naman Goyal et al.

Mixture of Experts layers (MoEs) enable efficient scaling of language models through conditional computation. This paper presents a detailed empirical study of how autoregressive MoE language models scale in comparison with dense models in a wide range of settings: in- and out-of-domain language modeling, zero- and few-shot priming, and full-shot fine-tuning. With the exception of fine-tuning, we find MoEs to be substantially more compute efficient. At more modest training budgets, MoEs can match the performance of dense models using $\sim$4 times less compute. This gap narrows at scale, but our largest MoE model (1.1T parameters) consistently outperforms a compute-equivalent dense model (6.7B parameters). Overall, this performance gap varies greatly across tasks and domains, suggesting that MoE and dense models generalize differently in ways that are worthy of future study. We make our code and models publicly available for research use.

Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe et al.

Large-scale generative language models such as GPT-3 are competitive few-shot learners. While these models are known to be able to jointly represent many different languages, their training data is dominated by English, potentially limiting their cross-lingual generalization. In this work, we train multilingual generative language models on a corpus covering a diverse set of languages, and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, our model outperforms GPT-3 on 171 out of 182 directions with 32 training examples, while surpassing the official supervised baseline in 45 directions. We conduct an in-depth analysis of different multilingual prompting approaches, showing in particular that strong few-shot learning performance across languages can be achieved via cross-lingual transfer through both templates and demonstration examples. Finally, we evaluate our models in social value tasks such as hate speech detection in five languages and find it has limitations similar to comparable sized GPT-3 models.

Tsvetomila Mihaylova, Pepa Gencheva, Martin Boyanov et al.

We present the system we built for participating in SemEval-2016 Task 3 on Community Question Answering. We achieved the best results on subtask C, and strong results on subtasks A and B, by combining a rich set of various types of features: semantic, lexical, metadata, and user-related. The most important group turned out to be the metadata for the question and for the comment, semantic vectors trained on QatarLiving data and similarities between the question and the comment for subtasks A and C, and between the original and the related question for Subtask B.

Todor Mihaylov, Ivan Koychev, Georgi Georgiev et al.

Recently, Web forums have been invaded by opinion manipulation trolls. Some trolls try to influence the other users driven by their own convictions, while in other cases they can be organized and paid, e.g., by a political party or a PR agency that gives them specific instructions what to write. Finding paid trolls automatically using machine learning is a hard task, as there is no enough training data to train a classifier; yet some test data is possible to obtain, as these trolls are sometimes caught and widely exposed. In this paper, we solve the training data problem by assuming that a user who is called a troll by several different people is likely to be such, and one who has never been called a troll is unlikely to be such. We compare the profiles of (i) paid trolls vs. (ii)"mentioned" trolls vs. (iii) non-trolls, and we further show that a classifier trained to distinguish (ii) from (iii) does quite well also at telling apart (i) from (iii).

Todor Mihaylov, Preslav Nakov

We describe our system for finding good answers in a community forum, as defined in SemEval-2016, Task 3 on Community Question Answering. Our approach relies on several semantic similarity features based on fine-tuned word embeddings and topics similarities. In the main Subtask C, our primary submission was ranked third, with a MAP of 51.68 and accuracy of 69.94. In Subtask A, our primary submission was also third, with MAP of 77.58 and accuracy of 73.39.

Todor Mihaylov, Preslav Nakov

There are different definitions of what a troll is. Certainly, a troll can be somebody who teases people to make them angry, or somebody who offends people, or somebody who wants to dominate any single discussion, or somebody who tries to manipulate people's opinion (sometimes for money), etc. The last definition is the one that dominates the public discourse in Bulgaria and Eastern Europe, and this is our focus in this paper. In our work, we examine two types of opinion manipulation trolls: paid trolls that have been revealed from leaked reputation management contracts and mentioned trolls that have been called such by several different people. We show that these definitions are sensible: we build two classifiers that can distinguish a post by such a paid troll from one by a non-troll with 81-82% accuracy; the same classifier achieves 81-82% accuracy on so called mentioned troll vs. non-troll posts.

Todor Mihaylov, Anette Frank

In this work, we propose to use linguistic annotations as a basis for a \textit{Discourse-Aware Semantic Self-Attention} encoder that we employ for reading comprehension on long narrative texts. We extract relations between discourse units, events and their arguments as well as coreferring mentions, using available annotation tools. Our empirical evaluation shows that the investigated structures improve the overall performance, especially intra-sentential and cross-sentential discourse relations, sentence-internal semantic role relations, and long-distance coreference relations. We show that dedicating self-attention heads to intra-sentential relations and relations connecting neighboring sentences is beneficial for finding answers to questions in longer contexts. Our findings encourage the use of discourse-semantic annotations to enhance the generalization capacity of self-attention models for reading comprehension.

Todor Mihaylov, Peter Clark, Tushar Khot et al.

We present a new kind of question answering dataset, OpenBookQA, modeled after open book exams for assessing human understanding of a subject. The open book that comes with our questions is a set of 1329 elementary level science facts. Roughly 6000 questions probe an understanding of these facts and their application to novel situations. This requires combining an open book fact (e.g., metals conduct electricity) with broad common knowledge (e.g., a suit of armor is made of metal) obtained from other sources. While existing QA datasets over documents or knowledge bases, being generally self-contained, focus on linguistic understanding, OpenBookQA probes a deeper understanding of both the topic---in the context of common knowledge---and the language it is expressed in. Human performance on OpenBookQA is close to 92%, but many state-of-the-art pre-trained QA methods perform surprisingly poorly, worse than several simple neural baselines we develop. Our oracle experiments designed to circumvent the knowledge retrieval bottleneck demonstrate the value of both the open book and additional facts. We leave it as a challenge to solve the retrieval problem in this multi-hop setting and to close the large gap to human performance.

Todor Mihaylov, Anette Frank

We introduce a neural reading comprehension model that integrates external commonsense knowledge, encoded as a key-value memory, in a cloze-style setting. Instead of relying only on document-to-question interaction or discrete features as in prior work, our model attends to relevant external knowledge and combines this knowledge with the context representation before inferring the answer. This allows the model to attract and imply knowledge from an external knowledge source that is not explicitly stated in the text, but that is relevant for inferring the answer. Our model improves results over a very strong baseline on a hard Common Nouns dataset, making it a strong competitor of much more complex models. By including knowledge explicitly, our model can also provide evidence about the background knowledge used in the RC process.

Todor Mihaylov, Zornitsa Kozareva, Anette Frank

Reading comprehension is a challenging task in natural language processing and requires a set of skills to be solved. While current approaches focus on solving the task as a whole, in this paper, we propose to use a neural network `skill' transfer approach. We transfer knowledge from several lower-level language tasks (skills) including textual entailment, named entity recognition, paraphrase detection and question type classification into the reading comprehension model. We conduct an empirical evaluation and show that transferring language skill knowledge leads to significant improvements for the task with much fewer steps compared to the baseline model. We also show that the skill transfer approach is effective even with small amounts of training data. Another finding of this work is that using token-wise deep label supervision for text classification improves the performance of transfer learning.

Todor Mihaylov, Daniel Belchev, Yasen Kiprov et al.

We transfer a key idea from the field of sentiment analysis to a new domain: community question answering (cQA). The cQA task we are interested in is the following: given a question and a thread of comments, we want to re-rank the comments so that the ones that are good answers to the question would be ranked higher than the bad ones. We notice that good vs. bad comments use specific vocabulary and that one can often predict the goodness/badness of a comment even ignoring the question, based on the comment contents only. This leads us to the idea to build a good/bad polarity lexicon as an analogy to the positive/negative sentiment polarity lexicons, commonly used in sentiment analysis. In particular, we use pointwise mutual information in order to build large-scale goodness polarity lexicons in a semi-supervised manner starting with a small number of initial seeds. The evaluation results show an improvement of 0.7 MAP points absolute over a very strong baseline and state-of-the art performance on SemEval-2016 Task 3.

Todor Mihaylov, Anette Frank

This paper describes two supervised baseline systems for the Story Cloze Test Shared Task (Mostafazadeh et al., 2016a). We first build a classifier using features based on word embeddings and semantic similarity computation. We further implement a neural LSTM system with different encoding strategies that try to model the relation between the story and the provided endings. Our experiments show that a model using representation features based on average word embedding vectors over the given story words and the candidate ending sentences words, joint with similarity features between the story and candidate ending representations performed better than the neural models. Our best model achieves an accuracy of 72.42, ranking 3rd in the official evaluation.