This technical report describes an experiment on autoregressive pre-training of Gemma2 2 billion parameter large language model (LLM) with 10\% on the Lithuanian language component of CulturaX from the point of view of continual learning. We apply elastic weight consolidation (EWC) to the full set of the model's parameters and investigate language understanding benchmarks, consisting of Arc, Belebele, Gsm8K, Hellaswag, MMLU, TruthfulQA, and Winogrande sets (both in English and Lithuanian versions), and perplexity benchmarks. We empirically demonstrate that EWC regularisation allows us not only to mitigate catastrophic forgetting effects but also that it is potentially beneficial for learning of the new task with LLMs.
Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogrande) by providing diverse responses, it also tends to facilitate reward hacking during training. This can lead to a high attack success rate (ASR) when models encounter jailbreaking prompts. The issue is particularly pronounced when employing stronger models like GPT-4o or larger models in the same family to generate chosen responses paired with target model self-generated rejected responses, resulting in dramatically poorer safety outcomes. Furthermore, with respect to safety, using solely self-generated responses (single-model generation) for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models, whether used directly as chosen data or as part of a multi-model response pool. We demonstrate that multi-model preference data exhibits high linear separability between chosen and rejected responses, which allows models to exploit superficial cues rather than internalizing robust safety constraints. Our experiments, conducted on models from the Llama, Mistral, and Qwen families, consistently validate these findings.
In this study, we take a closer look at how Winograd schema challenges can be used to evaluate common sense reasoning in LLMs. Specifically, we evaluate generative models of different sizes on the popular WinoGrande benchmark. We release WinoWhat, a new corpus, in which each instance of the WinoGrande validation set is paraphrased. Additionally, we evaluate the performance on the challenge across five common sense knowledge categories, giving more fine-grained insights on what types of knowledge are more challenging for LLMs. Surprisingly, all models perform significantly worse on WinoWhat, implying that LLM reasoning capabilities are overestimated on WinoGrande. To verify whether this is an effect of benchmark memorization, we match benchmark instances to LLM trainingdata and create two test-suites. We observe that memorization has a minimal effect on model performance on WinoGrande.
Recent copyright agreements between AI companies and content creators underscore the need for fine-grained control over language models' ability to reproduce copyrighted text. Existing defenses-ranging from aggressive unlearning to simplistic output filters-either sacrifice model utility or inadequately address verbatim leakage. We introduce Obliviate, a lightweight post-training method that surgically suppresses exact reproduction of specified sequences while preserving semantic understanding. Obliviate first identifies memorized passages and then, for each target token, minimally adjusts the model's output distribution via a Kullback-Leibler divergence penalty to drive down the probability of exact reproduction. Simultaneously, we enforce a consistency loss on non-target tokens to retain the model's fluency and task performance. We evaluate Obliviate on four popular 6-8B-parameter models (LLaMA-3.1, LLaMA-3.1-Instruct, Qwen-2.5, and Yi-1.5) using synthetic memorization benchmarks and organic copyrighted excerpts (e.g., Moby Dick, Frankenstein, Alice in Wonderland and Les Miserables). Across all settings, Obliviate reduces verbatim recall by two orders of magnitude (e.g., from hundreds of words to fewer than 12) while degrading downstream accuracy by at most 1% on HellaSwag, MMLU, TruthfulQA, and Winogrande. Furthermore, we benchmark Obliviate aganist different unlearning and copyright techniques using the MUSE and CoTaEval benchmarks. These results position Obliviate as a practical, high-fidelity solution for copyright compliance in deployed LLMs.
Large Language Models (LLMs) have shown remarkable performance across various tasks, yet significant disparities remain for non-English languages, and especially native African languages. This paper addresses these disparities by creating approximately 1 million human-translated words of new benchmark data in 8 low-resource African languages, covering a population of over 160 million speakers of: Amharic, Bambara, Igbo, Sepedi (Northern Sotho), Shona, Sesotho (Southern Sotho), Setswana, and Tsonga. Our benchmarks are translations of Winogrande and three sections of MMLU: college medicine, clinical knowledge, and virology. Using the translated benchmarks, we report previously unknown performance gaps between state-of-the-art (SOTA) LLMs in English and African languages. Finally, using results from over 400 fine-tuned models, we explore several methods to reduce the LLM performance gap, including high-quality dataset fine-tuning (using an LLM-as-an-Annotator), cross-lingual transfer, and cultural appropriateness adjustments. Key findings include average mono-lingual improvements of 5.6% with fine-tuning (with 5.4% average mono-lingual improvements when using high-quality data over low-quality data), 2.9% average gains from cross-lingual transfer, and a 3.0% out-of-the-box performance boost on culturally appropriate questions. The publicly available benchmarks, translations, and code from this study support further research and development aimed at creating more inclusive and effective language technologies.
As large language models (LLMs) increasingly shape the AI landscape, fine-tuning pretrained models has become more popular than in the pre-LLM era for achieving optimal performance in domain-specific tasks. However, pretrained LLMs such as ChatGPT are periodically evolved, i.e., model parameters are frequently updated), making it challenging for downstream users with limited resources to keep up with fine-tuning the newest LLMs for their domain application. Even though fine-tuning costs have nowadays been reduced thanks to the innovations of parameter-efficient fine-tuning such as LoRA, not all downstream users have adequate computing for frequent personalization. Moreover, access to fine-tuning datasets, particularly in sensitive domains such as healthcare, could be time-restrictive, making it crucial to retain the knowledge encoded in earlier fine-tuned rounds for future adaptation. In this paper, we present PortLLM, a training-free framework that (i) creates an initial lightweight model update patch to capture domain-specific knowledge, and (ii) allows a subsequent seamless plugging for the continual personalization of evolved LLM at minimal cost. Our extensive experiments cover seven representative datasets, from easier question-answering tasks {BoolQ, SST2} to harder reasoning tasks {WinoGrande, GSM8K}, and models including {Mistral-7B, Llama2, Llama3.1, and Gemma2}, validating the portability of our designed model patches and showcasing the effectiveness of our proposed framework. For instance, PortLLM achieves comparable performance to LoRA fine-tuning with reductions of up to 12.2x in GPU memory usage. Finally, we provide theoretical justifications to understand the portability of our model update patches, which offers new insights into the theoretical dimension of LLMs' personalization.
This paper proposes a novel prompt engineering technique called Judgment of Thought (JoT) that is specifically tailored for binary logical reasoning tasks. JoT employs three roles$\unicode{x2014}$lawyer, prosecutor, and judge$\unicode{x2014}$to facilitate more reliable and accurate reasoning by the model. In this framework, the judge utilizes a high$\unicode{x2010}$level model, while the lawyer and prosecutor utilize low$\unicode{x2010}$level models. This structure helps the judge better understand the responses from both the lawyer and prosecutor, enabling a more accurate judgment. Experimental results on large language model (LLM) benchmark datasets, such as BigBenchHard and Winogrande, demonstrate that JoT outperforms existing methods, including Chain of Thought (CoT) and Self$\unicode{x2010}$Consistency (SC), in binary logical reasoning tasks. Additionally, in real$\unicode{x2010}$world tasks, such as Fake News Detection and SMS Spam Detection, JoT shows comparable or improved performance compared to existing techniques. JoT significantly enhances the accuracy and reliability of models in binary reasoning tasks and show potential for practical applicability across various domains. Future research should aim to further broaden the applicability of JoT and optimize its implementation for real$\unicode{x2010}$world problem$\unicode{x2010}$solving.
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 $r = 0.93$.
The modern autoregressive Large Language Models (LLMs) have achieved outstanding performance on NLP benchmarks, and they are deployed in the real world. However, they still suffer from limitations of the autoregressive training paradigm. For example, autoregressive token generation is notably slow and can be prone to \textit{exposure bias}. The diffusion-based language models were proposed as an alternative to autoregressive generation to address some of these limitations. We evaluate the recently proposed Score Entropy Discrete Diffusion (SEDD) approach and show it is a promising alternative to autoregressive generation but it has some short-comings too. We empirically demonstrate the advantages and challenges of SEDD, and observe that SEDD generally matches autoregressive models in perplexity and on benchmarks such as HellaSwag, Arc or WinoGrande. Additionally, we show that in terms of inference latency, SEDD can be up to 4.5$\times$ more efficient than GPT-2. While SEDD allows conditioning on tokens at abitrary positions, SEDD appears slightly weaker than GPT-2 for conditional generation given short prompts. Finally, we reproduced the main results from the original SEDD paper.
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.
The machine learning community has witnessed impressive advancements since large language models (LLMs) first appeared. Yet, their massive memory consumption has become a significant roadblock to large-scale training. For instance, a 7B model typically requires at least 60 GB of GPU memory with full parameter training, which presents challenges for researchers without access to high-resource environments. Parameter Efficient Fine-Tuning techniques such as Low-Rank Adaptation (LoRA) have been proposed to alleviate this problem. However, in most large-scale fine-tuning settings, their performance does not reach the level of full parameter training because they confine the parameter search to a low-rank subspace. Attempting to complement this deficiency, we investigate the layerwise properties of LoRA on fine-tuning tasks and observe an unexpected but consistent skewness of weight norms across different layers. Utilizing this key observation, a surprisingly simple training strategy is discovered, which outperforms both LoRA and full parameter training in a wide range of settings with memory costs as low as LoRA. We name it Layerwise Importance Sampled AdamW (LISA), a promising alternative for LoRA, which applies the idea of importance sampling to different layers in LLMs and randomly freezes most middle layers during optimization. Experimental results show that with similar or less GPU memory consumption, LISA surpasses LoRA or even full parameter tuning in downstream fine-tuning tasks, where LISA consistently outperforms LoRA by over 10%-35% in terms of MT-Bench score while achieving on-par or better performance in MMLU, AGIEval and WinoGrande. On large models, specifically LLaMA-2-70B, LISA surpasses LoRA on MT-Bench, GSM8K, and PubMedQA, demonstrating its effectiveness across different domains.
Large language models (LLMs) are trained on massive internet corpora that often contain copyrighted content. This poses legal and ethical challenges for the developers and users of these models, as well as the original authors and publishers. In this paper, we propose a novel technique for unlearning a subset of the training data from a LLM, without having to retrain it from scratch. We evaluate our technique on the task of unlearning the Harry Potter books from the Llama2-7b model (a generative language model recently open-sourced by Meta). While the model took over 184K GPU-hours to pretrain, we show that in about 1 GPU hour of finetuning, we effectively erase the model's ability to generate or recall Harry Potter-related content, while its performance on common benchmarks (such as Winogrande, Hellaswag, arc, boolq and piqa) remains almost unaffected. We make our fine-tuned model publicly available on HuggingFace for community evaluation. To the best of our knowledge, this is the first paper to present an effective technique for unlearning in generative language models. Our technique consists of three main components: First, we use a reinforced model that is further trained on the target data to identify the tokens that are most related to the unlearning target, by comparing its logits with those of a baseline model. Second, we replace idiosyncratic expressions in the target data with generic counterparts, and leverage the model's own predictions to generate alternative labels for every token. These labels aim to approximate the next-token predictions of a model that has not been trained on the target data. Third, we finetune the model on these alternative labels, which effectively erases the original text from the model's memory whenever it is prompted with its context.
Recent work on explainable NLP has shown that few-shot prompting can enable large pretrained language models (LLMs) to generate grammatical and factual natural language explanations for data labels. In this work, we study the connection between explainability and sample hardness by investigating the following research question - "Are LLMs and humans equally good at explaining data labels for both easy and hard samples?" We answer this question by first collecting human-written explanations in the form of generalizable commonsense rules on the task of Winograd Schema Challenge (Winogrande dataset). We compare these explanations with those generated by GPT-3 while varying the hardness of the test samples as well as the in-context samples. We observe that (1) GPT-3 explanations are as grammatical as human explanations regardless of the hardness of the test samples, (2) for easy examples, GPT-3 generates highly supportive explanations but human explanations are more generalizable, and (3) for hard examples, human explanations are significantly better than GPT-3 explanations both in terms of label-supportiveness and generalizability judgements. We also find that hardness of the in-context examples impacts the quality of GPT-3 explanations. Finally, we show that the supportiveness and generalizability aspects of human explanations are also impacted by sample hardness, although by a much smaller margin than models. Supporting code and data are available at https://github.com/swarnaHub/ExplanationHardness
Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).
Language models built using semi-supervised machine learning on large corpora of natural language have very quickly enveloped the fields of natural language generation and understanding. In this paper we apply a zero-shot approach independently developed by a number of researchers now gaining recognition as a significant alternative to fine-tuning for evaluation on common sense tasks. A language model with relatively few parameters and training steps compared to a more recent language model (T5) can outperform it on a recent large data set (TimeDial), while displaying robustness in its performance across a similar class of language tasks. Surprisingly, this result is achieved by using a hyperparameter-free zero-shot method with the smaller model, compared to fine-tuning to the larger model. We argue that robustness of the smaller model ought to be understood in terms of compositionality, in a sense that we draw from recent literature on a class of similar models. We identify a practical cost for our method and model: high GPU-time for natural language evaluation. The zero-shot measurement technique that produces remarkable stability, both for ALBERT and other BERT variants, is an application of pseudo-log-likelihoods to masked language models for the relative measurement of probability for substitution alternatives in forced choice language tasks such as the Winograd Schema Challenge, Winogrande, and others. One contribution of this paper is to bring together a number of similar, but independent strands of research. We produce some absolute state-of-the-art results for common sense reasoning in binary choice tasks, performing better than any published result in the literature, including fine-tuned efforts. We show a remarkable consistency of the model's performance under adversarial settings, which we argue is best explained by the model's compositionality of representations.
We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-domain (TLD). Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we translate and adapt the WinoGrande dataset for co-reference resolution. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.
Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.
Training a model to provide natural language explanations (NLEs) for its predictions usually requires the acquisition of task-specific NLEs, which is time- and resource-consuming. A potential solution is the few-shot out-of-domain transfer of NLEs from a parent task with many NLEs to a child task. In this work, we examine the setup in which the child task has few NLEs but abundant labels. We establish four few-shot transfer learning methods that cover the possible fine-tuning combinations of the labels and NLEs for the parent and child tasks. We transfer explainability from a large natural language inference dataset (e-SNLI) separately to two child tasks: (1) hard cases of pronoun resolution, where we introduce the small-e-WinoGrande dataset of NLEs on top of the WinoGrande dataset, and (2)~commonsense validation (ComVE). Our results demonstrate that the parent task helps with NLE generation and we establish the best methods for this setup.
In this work, we propose Masked Noun-Phrase Prediction (MNPP), a pre-training strategy to tackle pronoun resolution in a fully unsupervised setting. Firstly, We evaluate our pre-trained model on various pronoun resolution datasets without any finetuning. Our method outperforms all previous unsupervised methods on all datasets by large margins. Secondly, we proceed to a few-shot setting where we finetune our pre-trained model on WinoGrande-S and XS separately. Our method outperforms RoBERTa-large baseline with large margins, meanwhile, achieving a higher AUC score after further finetuning on the remaining three official splits of WinoGrande.
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%).