The performance of large language models (LLMs) continues to improve, as reflected in rising scores on standard benchmarks. However, the lack of transparency around training data raises concerns about potential overlap with evaluation sets and the fairness of reported results. Although prior work has proposed methods for detecting data leakage, these approaches primarily focus on identifying outliers and have not been evaluated under controlled simulated leakage conditions. In this work, we compare existing leakage detection techniques, namely permutation and n-gram-based methods, under a continual pretraining setup that simulates real-world leakage scenarios, and additionally explore a lightweight method we call semi-half question. Although semi-half offers a low-cost alternative, our analysis shows that the n-gram method consistently achieves the highest F1-score. We also refine these techniques to support instance-level detection and reduce computational overhead. Leveraging the best-performing method, we create cleaned versions of MMLU and HellaSwag, and re-evaluate several LLMs. Our findings present a practical path toward more reliable and transparent evaluations, and we recommend contamination checks as a standard step before releasing benchmark results.
Common-sense reasoning is a key language model capability because it encapsulates not just specific factual knowledge but rather general language and world understanding. Measuring common-sense reasoning, therefore, is crucial for language models of different sizes and applications. One of the most widely used benchmarks for evaluating such capabilities is HellaSwag; however, in this paper, we show that it has severe construct validity issues. These issues range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. Furthermore, we show that if models are evaluated only on answer texts, or with "Lorem ipsum dolor..." instead of the question, more than 65% of model predictions remain the same, and this cannot be attributed merely to contamination. Since benchmark scores are an essential part of model selection in both research and commercial applications, these validity issues can have severe consequences. In particular, knowing that taking benchmark scores at face value is ubiquitous, inadequate evaluation leads to ill-informed decisions about models. In this paper, we thoroughly investigate critical validity issues posed by HellaSwag and illustrate them with various evaluations using generative language models of different sizes. We argue that this benchmark does not accurately measure common-sense reasoning and, therefore, should not be used for evaluation in its current state. Based on the results of our study, we propose requirements that should be met by future common-sense reasoning benchmarks. In addition, we release GoldenSwag, a corrected subset of HellaSwag, which, to our belief, facilitates acceptable common-sense reasoning evaluation.
Data quality is crucial for training Large Language Models (LLMs). Traditional heuristic filters often miss low-quality text or mistakenly remove valuable content. In this paper, we introduce an LLM-based line-level filtering method to enhance training data quality. We use GPT-4o mini to label a 20,000-document sample from FineWeb at the line level, allowing the model to create descriptive labels for low-quality lines. These labels are grouped into nine main categories, and we train a DeBERTa-v3 classifier to scale the filtering to a 10B-token subset of FineWeb. To test the impact of our filtering, we train GPT-2 models on both the original and the filtered datasets. The results show that models trained on the filtered data achieve higher accuracy on the HellaSwag benchmark and reach their performance targets faster, even with up to 25\% less data. This demonstrates that LLM-based line-level filtering can significantly improve data quality and training efficiency for LLMs. We release our quality-annotated dataset, FinerWeb-10BT, and the codebase to support further work in this area.
Low-rank adaption (LoRA) is a widely used parameter-efficient finetuning method for LLM that reduces memory requirements. However, current LoRA optimizers lack transformation invariance, meaning the actual updates to the weights depends on how the two LoRA factors are scaled or rotated. This deficiency leads to inefficient learning and sub-optimal solutions in practice. This paper introduces LoRA-RITE, a novel adaptive matrix preconditioning method for LoRA optimization, which can achieve transformation invariance and remain computationally efficient. We provide theoretical analysis to demonstrate the benefit of our method and conduct experiments on various LLM tasks with different models including Gemma 2B, 7B, and mT5-XXL. The results demonstrate consistent improvements against existing optimizers. For example, replacing Adam with LoRA-RITE during LoRA fine-tuning of Gemma-2B yielded 4.6\% accuracy gain on Super-Natural Instructions and 3.5\% accuracy gain across other four LLM benchmarks (HellaSwag, ArcChallenge, GSM8K, OpenBookQA).
The LayerNorm (LN) layer in GPT-style transformer models has long been a hindrance to mechanistic interpretability. LN is a crucial component required to stabilize the training of large language models, and LN or the similar RMSNorm have been used in practically all large language models based on the transformer architecture. The non-linear nature of the LN layers is a hindrance for mechanistic interpretability as it hinders interpretation of the residual stream, and makes it difficult to decompose the model into circuits. Some researchers have gone so far as to name "reasons interpretability researchers hate layer norm." In this paper we show that it is possible to remove the LN layers from a pre-trained GPT2-small model by fine-tuning on a fraction (500M tokens) of the training data. We demonstrate that this LN-free model achieves similar performance to the original model on the OpenWebText and ThePile datasets (-0.05 cross-entropy loss), and the Hellaswag benchmark (-0.5% accuracy). We provide our implementation at https://github.com/ApolloResearch/gpt2_noLN, and fine-tuned GPT2-small models at https://huggingface.co/apollo-research/gpt2_noLN. Our work not only provides a simplified model for mechanistic interpretability research, but also provides evidence that the LN layers, at inference time, do not play a crucial role in transformer models.
Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the $\texttt{Open LLM Leaderboard}$ aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from $n > 5000$ LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with $d=28,632$ items in total). From them we distill a sparse benchmark, $\texttt{metabench}$, that has less than $3\%$ of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original $\textit{individual}$ benchmark score with, on average, $1.5\%$ root mean square error (RMSE), (2) reconstruct the original $\textit{total}$ score with $0.8\%$ RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is $r = 0.93$.
We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs). First, during training we apply layer dropout, with low dropout rates for earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. Second, during inference, we show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model. Third, we present a novel self-speculative decoding solution where we exit at early layers and verify and correct with remaining layers of the model. Our proposed self-speculative decoding approach has less memory footprint than other speculative decoding approaches and benefits from shared compute and activations of the draft and verification stages. We run experiments on different Llama model sizes on different types of training: pretraining from scratch, continual pretraining, finetuning on specific data domain, and finetuning on specific task. We implement our inference solution and show speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task. We open source our code and checkpoints at https://github.com/facebookresearch/LayerSkip.
Despite recent advancements showcasing the impressive capabilities of Large Language Models (LLMs) in conversational systems, we show that even state-of-the-art LLMs are morally inconsistent in their generations, questioning their reliability (and trustworthiness in general). Prior works in LLM evaluation focus on developing ground-truth data to measure accuracy on specific tasks. However, for moral scenarios that often lack universally agreed-upon answers, consistency in model responses becomes crucial for their reliability. To address this issue, we propose an information-theoretic measure called Semantic Graph Entropy (SaGE), grounded in the concept of "Rules of Thumb" (RoTs) to measure a model's moral consistency. RoTs are abstract principles learned by a model and can help explain their decision-making strategies effectively. To this extent, we construct the Moral Consistency Corpus (MCC), containing 50K moral questions, responses to them by LLMs, and the RoTs that these models followed. Furthermore, to illustrate the generalizability of SaGE, we use it to investigate LLM consistency on two popular datasets -- TruthfulQA and HellaSwag. Our results reveal that task-accuracy and consistency are independent problems, and there is a dire need to investigate these issues further.
Graphs are commonly used to model complex networks prevalent in modern social media and literacy applications. Our research investigates the vulnerability of these graphs through the application of feature based adversarial attacks, focusing on both decision time attacks and poisoning attacks. In contrast to state of the art models like Net Attack and Meta Attack, which target node attributes and graph structure, our study specifically targets node attributes. For our analysis, we utilized the text dataset Hellaswag and graph datasets Cora and CiteSeer, providing a diverse basis for evaluation. Our findings indicate that decision time attacks using Projected Gradient Descent (PGD) are more potent compared to poisoning attacks that employ Mean Node Embeddings and Graph Contrastive Learning strategies. This provides insights for graph data security, pinpointing where graph-based models are most vulnerable and thereby informing the development of stronger defense mechanisms against such attacks.
The burgeoning interest in Multimodal Large Language Models (MLLMs), such as OpenAI's GPT-4V(ision), has significantly impacted both academic and industrial realms. These models enhance Large Language Models (LLMs) with advanced visual understanding capabilities, facilitating their application in a variety of multimodal tasks. Recently, Google introduced Gemini, a cutting-edge MLLM designed specifically for multimodal integration. Despite its advancements, preliminary benchmarks indicate that Gemini lags behind GPT models in commonsense reasoning tasks. However, this assessment, based on a limited dataset (i.e., HellaSWAG), does not fully capture Gemini's authentic commonsense reasoning potential. To address this gap, our study undertakes a thorough evaluation of Gemini's performance in complex reasoning tasks that necessitate the integration of commonsense knowledge across modalities. We carry out a comprehensive analysis of 12 commonsense reasoning datasets, ranging from general to domain-specific tasks. This includes 11 datasets focused solely on language, as well as one that incorporates multimodal elements. Our experiments across four LLMs and two MLLMs demonstrate Gemini's competitive commonsense reasoning capabilities. Additionally, we identify common challenges faced by current LLMs and MLLMs in addressing commonsense problems, underscoring the need for further advancements in enhancing the commonsense reasoning abilities of these models.
Data contamination in model evaluation has become increasingly prevalent with the growing popularity of large language models. It allows models to "cheat" via memorisation instead of displaying true capabilities. Therefore, contamination analysis has become an crucial part of reliable model evaluation to validate results. However, existing contamination analysis is usually conducted internally by large language model developers and often lacks transparency and completeness. This paper presents an extensive data contamination report for over 15 popular large language models across six popular multiple-choice QA benchmarks. We also introduce an open-source pipeline that enables the community to perform contamination analysis on customised data and models. Our experiments reveal varying contamination levels ranging from 1\% to 45\% across benchmarks, with the contamination degree increasing rapidly over time. Performance analysis of large language models indicates that data contamination does not necessarily lead to increased model metrics: while significant accuracy boosts of up to 14\% and 7\% are observed on contaminated C-Eval and Hellaswag benchmarks, only a minimal increase is noted on contaminated MMLU. We also find larger models seem able to gain more advantages than smaller models on contaminated test sets.
Despite their impressive performance in a wide range of NLP tasks, Large Language Models (LLMs) have been reported to encode worrying-levels of gender biases. Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally costly. Moreover, one might not even have access to the model parameters for performing debiasing such as in the case of closed LLMs such as GPT-4. To address this challenge, we propose bias suppression that prevents biased generations of LLMs by simply providing textual preambles constructed from manually designed templates and real-world statistics, without accessing to model parameters. We show that, using CrowsPairs dataset, our textual preambles covering counterfactual statements can suppress gender biases in English LLMs such as LLaMA2. Moreover, we find that gender-neutral descriptions of gender-biased objects can also suppress their gender biases. Moreover, we show that bias suppression has acceptable adverse effect on downstream task performance with HellaSwag and COPA.
Beyond the success story of adversarial training (AT) in the recent text domain on top of pre-trained language models (PLMs), our empirical study showcases the inconsistent gains from AT on some tasks, e.g. commonsense reasoning, named entity recognition. This paper investigates AT from the perspective of the contextualized language representation outputted by PLM encoders. We find the current AT attacks lean to generate sub-optimal adversarial examples that can fool the decoder part but have a minor effect on the encoder. However, we find it necessary to effectively deviate the latter one to allow AT to gain. Based on the observation, we propose simple yet effective \textit{Contextualized representation-Adversarial Training} (CreAT), in which the attack is explicitly optimized to deviate the contextualized representation of the encoder. It allows a global optimization of adversarial examples that can fool the entire model. We also find CreAT gives rise to a better direction to optimize the adversarial examples, to let them less sensitive to hyperparameters. Compared to AT, CreAT produces consistent performance gains on a wider range of tasks and is proven to be more effective for language pre-training where only the encoder part is kept for downstream tasks. We achieve the new state-of-the-art performances on a series of challenging benchmarks, e.g. AdvGLUE (59.1 $ \rightarrow $ 61.1), HellaSWAG (93.0 $ \rightarrow $ 94.9), ANLI (68.1 $ \rightarrow $ 69.3).
Data Augmentation (DA) is known to improve the generalizability of deep neural networks. Most existing DA techniques naively add a certain number of augmented samples without considering the quality and the added computational cost of these samples. To tackle this problem, a common strategy, adopted by several state-of-the-art DA methods, is to adaptively generate or re-weight augmented samples with respect to the task objective during training. However, these adaptive DA methods: (1) are computationally expensive and not sample-efficient, and (2) are designed merely for a specific setting. In this work, we present a universal DA technique, called Glitter, to overcome both issues. Glitter can be plugged into any DA method, making training sample-efficient without sacrificing performance. From a pre-generated pool of augmented samples, Glitter adaptively selects a subset of worst-case samples with maximal loss, analogous to adversarial DA. Without altering the training strategy, the task objective can be optimized on the selected subset. Our thorough experiments on the GLUE benchmark, SQuAD, and HellaSwag in three widely used training setups including consistency training, self-distillation and knowledge distillation reveal that Glitter is substantially faster to train and achieves a competitive performance, compared to strong baselines.
Commonsense AI has long been seen as a near impossible goal -- until recently. Now, research interest has sharply increased with an influx of new benchmarks and models. We propose two new ways to evaluate commonsense models, emphasizing their generality on new tasks and building on diverse, recently introduced benchmarks. First, we propose a new multitask benchmark, RAINBOW, to promote research on commonsense models that generalize well over multiple tasks and datasets. Second, we propose a novel evaluation, the cost equivalent curve, that sheds new insight on how the choice of source datasets, pretrained language models, and transfer learning methods impacts performance and data efficiency. We perform extensive experiments -- over 200 experiments encompassing 4800 models -- and report multiple valuable and sometimes surprising findings, e.g., that transfer almost always leads to better or equivalent performance if following a particular recipe, that QA-based commonsense datasets transfer well with each other, while commonsense knowledge graphs do not, and that perhaps counter-intuitively, larger models benefit more from transfer than smaller ones. Last but not least, we introduce a new universal commonsense reasoning model, UNICORN, that establishes new state-of-the-art performance across 8 popular commonsense benchmarks, aNLI (87.3%), CosmosQA (91.8%), HellaSWAG (93.9%), PIQA (90.1%), SocialIQa (83.2%), WinoGrande (86.6%), CycIC (94.0%) and CommonsenseQA (79.3%).