Improving LoRA with Variational Learning

Bayesian methods have recently been used to improve LoRA finetuning and, although they improve calibration, their effect on other metrics (such as accuracy) is marginal and can sometimes even be detrimental. Moreover, Bayesian methods also increase computational overheads and require additional tricks for them to work well. Here, we fix these issues by using a recently proposed variational algorithm called IVON. We show that IVON is easy to implement and has similar costs to AdamW, and yet it can also drastically improve many metrics by using a simple posterior pruning technique. We present extensive results on billion-scale LLMs (Llama and Qwen series) going way beyond the scale of existing applications of IVON. For example, we finetune a Llama-3.2-3B model on a set of commonsense reasoning tasks and improve accuracy over AdamW by 1.3% and reduce ECE by 5.4%, outperforming AdamW and other recent Bayesian methods like Laplace-LoRA and BLoB. Overall, our results show that variational learning with IVON can effectively improve LoRA finetuning.

Variational Learning Finds Flatter Solutions at the Edge of Stability

Variational Learning (VL) has recently gained popularity for training deep neural networks and is competitive to standard learning methods. Part of its empirical success can be explained by theories such as PAC-Bayes bounds, minimum description length and marginal likelihood, but there are few tools to unravel the implicit regularization in play. Here, we analyze the implicit regularization of VL through the Edge of Stability (EoS) framework. EoS has previously been used to show that gradient descent can find flat solutions and we extend this result to VL to show that it can find even flatter solutions. This is obtained by controlling the posterior covariance and the number of Monte Carlo samples from the posterior. These results are derived in a similar fashion as the standard EoS literature for deep learning, by first deriving a result for a quadratic problem and then extending it to deep neural networks. We empirically validate these findings on a wide variety of large networks, such as ResNet and ViT, to find that the theoretical results closely match the empirical ones. Ours is the first work to analyze the EoS dynamics in VL.

Rewriting Pre-Training Data Boosts LLM Performance in Math and Code

The performance of large language models (LLMs) in program synthesis and mathematical reasoning is fundamentally limited by the quality of their pre-training corpora. We introduce two openly licensed datasets, released under the Llama 3.3 Community License, that significantly enhance LLM performance by systematically rewriting public data. SwallowCode (approximately 16.1 billion tokens) refines Python snippets from The-Stack-v2 through a novel four-stage pipeline: syntax validation, pylint-based style filtering, and a two-stage LLM rewriting process that enforces style conformity and transforms snippets into self-contained, algorithmically efficient examples. Unlike prior methods that rely on exclusionary filtering or limited transformations, our transform-and-retain approach upgrades low-quality code, maximizing data utility. SwallowMath (approximately 2.3 billion tokens) enhances Finemath-4+ by removing boilerplate, restoring context, and reformatting solutions into concise, step-by-step explanations. Within a fixed 50 billion token training budget, continual pre-training of Llama-3.1-8B with SwallowCode boosts pass@1 by +17.0 on HumanEval and +17.7 on HumanEval+ compared to Stack-Edu, surpassing the baseline model's code generation capabilities. Similarly, substituting SwallowMath yields +12.4 accuracy on GSM8K and +7.6 on MATH. Ablation studies confirm that each pipeline stage contributes incrementally, with rewriting delivering the largest gains. All datasets, prompts, and checkpoints are publicly available, enabling reproducible research and advancing LLM pre-training for specialized domains.

Building Instruction-Tuning Datasets from Human-Written Instructions with Open-Weight Large Language Models

Instruction tuning is crucial for enabling Large Language Models (LLMs) to solve real-world tasks. Prior work has shown the effectiveness of instruction-tuning data synthesized solely from LLMs, raising a fundamental question: Do we still need human-originated signals for instruction tuning? This work answers the question affirmatively: we build state-of-the-art instruction-tuning datasets sourced from human-written instructions, by simply pairing them with LLM-generated responses. LLMs fine-tuned on our datasets consistently outperform those fine-tuned on existing ones. Our data construction approach can be easily adapted to other languages; we build datasets for Japanese and confirm that LLMs tuned with our data reach state-of-the-art performance. Analyses suggest that instruction-tuning in a new language allows LLMs to follow instructions, while the tuned models exhibit a notable lack of culture-specific knowledge in that language. The datasets and fine-tuned models will be publicly available. Our datasets, synthesized with open-weight LLMs, are openly distributed under permissive licenses, allowing for diverse use cases.

On the Relationship Between Double Descent of CNNs and Shape/Texture Bias Under Learning Process

The double descent phenomenon, which deviates from the traditional bias-variance trade-off theory, attracts considerable research attention; however, the mechanism of its occurrence is not fully understood. On the other hand, in the study of convolutional neural networks (CNNs) for image recognition, methods are proposed to quantify the bias on shape features versus texture features in images, determining which features the CNN focuses on more. In this work, we hypothesize that there is a relationship between the shape/texture bias in the learning process of CNNs and epoch-wise double descent, and we conduct verification. As a result, we discover double descent/ascent of shape/texture bias synchronized with double descent of test error under conditions where epoch-wise double descent is observed. Quantitative evaluations confirm this correlation between the test errors and the bias values from the initial decrease to the full increase in test error. Interestingly, double descent/ascent of shape/texture bias is observed in some cases even in conditions without label noise, where double descent is thought not to occur. These experimental results are considered to contribute to the understanding of the mechanisms behind the double descent phenomenon and the learning process of CNNs in image recognition.

Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization

The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more. As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs. All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.

Why We Build Local Large Language Models: An Observational Analysis from 35 Japanese and Multilingual LLMs

Why do we build local large language models (LLMs)? What should a local LLM learn from the target language? Which abilities can be transferred from other languages? Do language-specific scaling laws exist? To explore these research questions, we evaluated 35 Japanese, English, and multilingual LLMs on 19 evaluation benchmarks for Japanese and English, taking Japanese as a local language. Adopting an observational approach, we analyzed correlations of benchmark scores, and conducted principal component analysis (PCA) on the scores to derive \textit{ability factors} of local LLMs. We found that training on English text can improve the scores of academic subjects in Japanese (JMMLU). In addition, it is unnecessary to specifically train on Japanese text to enhance abilities for solving Japanese code generation, arithmetic reasoning, commonsense, and reading comprehension tasks. In contrast, training on Japanese text could improve question-answering tasks about Japanese knowledge and English-Japanese translation, which indicates that abilities for solving these two tasks can be regarded as \textit{Japanese abilities} for LLMs. Furthermore, we confirmed that the Japanese abilities scale with the computational budget for Japanese text.

Lion Cub: Minimizing Communication Overhead in Distributed Lion

Communication overhead is a key challenge in distributed deep learning, especially on slower Ethernet interconnects, and given current hardware trends, communication is likely to become a major bottleneck. While gradient compression techniques have been explored for SGD and Adam, the Lion optimizer has the distinct advantage that its update vectors are the output of a sign operation, enabling straightforward quantization. However, simply compressing updates for communication and using techniques like majority voting fails to lead to end-to-end speedups due to inefficient communication algorithms and reduced convergence. We analyze three factors critical to distributed learning with Lion: optimizing communication methods, identifying effective quantization methods, and assessing the necessity of momentum synchronization. Our findings show that quantization techniques adapted to Lion and selective momentum synchronization can significantly reduce communication costs while maintaining convergence. We combine these into Lion Cub, which enables up to 5x speedups in end-to-end training compared to Lion. This highlights Lion's potential as a communication-efficient solution for distributed training.

Balancing Speed and Stability: The Trade-offs of FP8 vs. BF16 Training in LLMs

Large Language Models (LLMs) have attracted significant attention due to their human-like language understanding and generation capabilities, as well as their applicability across various domains. These models, characterized by their massive scale and extensive training data, continue to push the boundaries of what is possible in natural language processing. The Llama 3 series, for instance, exemplifies this trend with its flagship model boasting 405 billion parameters trained on 15.6 trillion tokens. The immense computational demands associated with training such models have spurred ongoing research into optimizing the efficiency of the training process, particularly through the use of lower-precision formats. NVIDIA's H100 GPU, which introduces support for FP8 in addition to the more conventional FP16 and BF16 formats, has emerged as a focal point in this optimization effort. Preliminary studies suggest that FP8 could offer substantial reductions in training time without sacrificing model performance when compared to BF16, making it a promising candidate for large-scale model training. However, the broader implications of adopting FP8, particularly in terms of training stability and downstream task performance, have yet to be fully understood. In this study, we delve into the practical trade-offs involved in adopting FP8 over BF16 for training LLMs.

Variational Low-Rank Adaptation Using IVON

We show that variational learning can significantly improve the accuracy and calibration of Low-Rank Adaptation (LoRA) without a substantial increase in the cost. We replace AdamW by the Improved Variational Online Newton (IVON) algorithm to finetune large language models. For Llama-2 with 7 billion parameters, IVON improves the accuracy over AdamW by 2.8% and expected calibration error by 4.6%. The accuracy is also better than the other Bayesian alternatives, yet the cost is lower and the implementation is easier. Our work provides additional evidence for the effectiveness of IVON for large language models. The code is available at https://github.com/team-approx-bayes/ivon-lora.