混合训练可避免数学精调中的灾难性遗忘，保持通用能力。
Mitigating Catastrophic Forgetting in Mathematical Reasoning Finetuning through Mixed Training
- 训练时交替穿插数学与通用任务数据，缓解遗忘问题。
- 1:1混合比率达12.0%数学准确率，同时保留86.2%通用推理能力。
- 少量通用数据（6.2%）即可有效防止遗忘，适合多任务场景。
在对大型语言模型进行数学推理精调时，模型会出现灾难性遗忘，丧失原有能力。我们以Flan-T5-Base（250M参数）为基础，在DeepMind Mathematics数据集上训练，并在MultiNLI上测量遗忘情况。仅用数学数据训练使数学准确率从3.1%提升至12.0%，但导致NLI准确率从81.0%暴跌至16.5%（下降64.5个百分点），且在前1000步内即发生。我们提出混合训练策略，将数学与NLI样本交错训练。结果表明，混合训练完全消除遗忘，同时保持同等数学性能：1:1比例达12.0%数学准确率（与纯数学训练相当），且维持86.2%的NLI准确率。系统评估了1:1至15:1的混合比例，发现即使仅6.2%的NLI数据暴露也具显著正则化效果。研究说明专业化训练无需牺牲通用能力，对更大模型的扩展具有重要启示。
原文摘要 · Abstract (English)
When finetuning large language models for specialized tasks such as mathematical reasoning, models exhibit catastrophic forgetting, losing previously learned capabilities. We investigate this by finetuning Flan-T5-Base (250M parameters) on the DeepMind Mathematics dataset and measuring forgetting on MultiNLI. Math-only training improves mathematical accuracy from 3.1\% to 12.0\% but causes NLI accuracy to collapse from 81.0\% to 16.5\%--a 64.5 percentage point drop occurring within the first 1,000 training steps. We propose mixed training strategies that interleave mathematical and NLI examples during training. Our results demonstrate that mixed training completely eliminates catastrophic forgetting while maintaining equivalent mathematical performance: the balanced 1:1 ratio achieves 12.0\% math accuracy (matching math-only) while preserving 86.2\% NLI accuracy. We systematically explore mixing ratios from 1:1 to 15:1, finding that even minimal NLI exposure (6.2\%) provides effective regularization. These findings demonstrate that specialization need not require forgetting general capabilities, with implications for scaling to larger models where mixed training may confer additional benefits beyond forgetting prevention.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。