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

• URL: http://arxiv.org/abs/2505.02881v3
• Date: Fri, 04 Jul 2025 06:10:53 GMT
• Title: Rewriting Pre-Training Data Boosts LLM Performance in Math and Code
• Abstract summary: We introduce two datasets, released under the Llama 3.3 Community License, that significantly enhance large language models (LLMs) performance.<n> SwallowCode refines Python snippets through a novel four-stage pipeline: syntax validation, pylint-based style filtering, and a two-stage rewriting process.<n>We show that 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.
• Score: 21.587613905318705
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• FLeX: Fourier-based Low-rank EXpansion for multilingual transfer [0.0]
  Cross-lingual code generation is critical in enterprise environments where multiple programming languages coexist.<n>This paper investigates whether fine-tuning methods and parameter enhancements can improve cross-lingual transfer from Python to languages like Java.
  arXiv  Detail & Related papers  (2026-04-06T19:26:13Z)
• Enhancing LLM-Based Code Generation with Complexity Metrics: A Feedback-Driven Approach [6.289275189295223]
  We investigate the relationship between code complexity and the success of Large Language Models generated code.<n>We propose an iterative feedback method, where LLMs are prompted to generate correct code based on complexity metrics from previous failed outputs.<n>Experiment results show that our approach makes notable improvements, particularly with a smaller LLM.
  arXiv  Detail & Related papers  (2025-05-29T19:06:14Z)
• S$^2$R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning [51.84977135926156]
  We introduce S$2$R, an efficient framework that enhances LLM reasoning by teaching models to self-verify and self-correct during inference.<n>Our results demonstrate that Qwen2.5-math-7B achieves an accuracy improvement from 51.0% to 81.6%, outperforming models trained on an equivalent amount of long-CoT distilled data.
  arXiv  Detail & Related papers  (2025-02-18T13:40:22Z)
• Large Language Models as Code Executors: An Exploratory Study [29.545321608864295]
  This paper pioneers the exploration of Large Language Models (LLMs) as code executors. We are the first to examine this feasibility across various LLMs, including OpenAI's o1, GPT-4o, GPT-3.5, DeepSeek, and Qwen-Coder. We introduce an Iterative Instruction Prompting (IIP) technique that processes code snippets line by line, enhancing the accuracy of weaker models by an average of 7.22%.
  arXiv  Detail & Related papers  (2024-10-09T08:23:22Z)
• LLM Surgery: Efficient Knowledge Unlearning and Editing in Large Language Models [16.67999382790238]
  Large language models (LLMs) have revolutionized various domains, yet their utility comes with challenges related to outdated or problematic knowledge embedded during pretraining. This paper addresses the challenge of modifying LLMs to unlearn problematic and outdated information while efficiently integrating new knowledge without retraining from scratch. Using Llama2-7B, we demonstrate that LLM Surgery can achieve significant forgetting on the unlearn set, a 20% increase in accuracy on the update set, and maintain performance on the retain set.
  arXiv  Detail & Related papers  (2024-09-19T19:07:01Z)
• LeDex: Training LLMs to Better Self-Debug and Explain Code [36.604898865514365]
  LeDex is a training framework that significantly improves the self-ging capability of LLMs.<n>We propose an automated pipeline to collect a high-quality dataset for code explanation and refinement.<n>We perform supervised fine-tuning (SFT) and further reinforcement learning (RL) on both success and failure trajectories with a novel reward design.
  arXiv  Detail & Related papers  (2024-05-28T23:20:24Z)
• JiuZhang3.0: Efficiently Improving Mathematical Reasoning by Training Small Data Synthesis Models [110.45794710162241]
  Existing work either collects large-scale math-related texts for pre-training, or relies on stronger LLMs to synthesize massive math problems. We propose an efficient way that trains a small LLM for math problem synthesis, to efficiently generate sufficient high-quality pre-training data. We leverage it to synthesize 6 million math problems for pre-training our JiuZhang3.0 model, which only needs to invoke GPT-4 API 9.3k times and pre-train on 4.6B data.
  arXiv  Detail & Related papers  (2024-05-23T09:43:19Z)
• Natural Is The Best: Model-Agnostic Code Simplification for Pre-trained Large Language Models [6.646510073473929]
  We propose SlimCode, a model-agnostic code simplification solution for Large Language Models. SlimCode can improve the state-of-the-art technique by 9.46% and 5.15% in terms of MRR and BLEU score on code search and summarization.
  arXiv  Detail & Related papers  (2024-05-18T06:15:52Z)
• LLM-Assisted Code Cleaning For Training Accurate Code Generators [53.087019724256606]
  We investigate data quality for code and find that making the code more structured and readable leads to improved code generation performance of the system. We build a novel data-cleaning pipeline that uses these principles to transform existing programs. We evaluate our approach on two challenging algorithmic code generation benchmarks and find that fine-tuning CodeLLaMa-7B improves the performance by up to 30% compared to fine-tuning on the original dataset.
  arXiv  Detail & Related papers  (2023-11-25T02:45:50Z)
• Large Language Model-Aware In-Context Learning for Code Generation [75.68709482932903]
  Large language models (LLMs) have shown impressive in-context learning (ICL) ability in code generation. We propose a novel learning-based selection approach named LAIL (LLM-Aware In-context Learning) for code generation.
  arXiv  Detail & Related papers  (2023-10-15T06:12:58Z)
• MLLM-DataEngine: An Iterative Refinement Approach for MLLM [62.30753425449056]
  We propose a novel closed-loop system that bridges data generation, model training, and evaluation. Within each loop, the MLLM-DataEngine first analyze the weakness of the model based on the evaluation results. For targeting, we propose an Adaptive Bad-case Sampling module, which adjusts the ratio of different types of data. For quality, we resort to GPT-4 to generate high-quality data with each given data type.
  arXiv  Detail & Related papers  (2023-08-25T01:41:04Z)
• Getting More Juice Out of Your Data: Hard Pair Refinement Enhances Visual-Language Models Without Extra Data [122.282521548393]
  Contrastive Language-Image Pre-training (CLIP) has become the standard for cross-modal image-text representation learning.<n>We introduce HELIP, a cost-effective strategy that improves CLIP models by exploiting challenging text-image pairs within existing datasets in continuous training.
  arXiv  Detail & Related papers  (2023-05-09T07:00:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.