A 帮助 B 而 B 伤害 A：指令微调混合中的定向迁移
A helps B while B hurts A: directed transfer in instruction-tuning mixture
- University of Cologne（科隆大学）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本研究提出迁移图，通过有符号估计源任务对目标任务的正负影响，在指令微调混合中筛选任务，显著提升推理目标准确率。
AI中文摘要：
将语言模型适应于专门语料库意味着在固定预算下选择要训练的指令微调任务，而测试一种选择需要一次微调运行。常见的启发式方法会添加更多源任务或选择与目标相似的源。前者假设迁移永远不会是负面的；后者假设迁移是对称的。我们表明这两个假设都失败：任务 A 可以帮助任务 B，而 B 伤害 A，因此帮助性是有序源-目标对的有符号属性。我们引入了迁移图，一种对每个源帮助或伤害每个保留目标程度的有符号估计。我们在 Qwen3 和 Mistral 模型（参数规模从 0.6B 到 32B）上进行了数百次微调运行来拟合该图，所有源均来自一个语料库，且不包含目标的训练示例。该图预测了未见混合上保留目标的准确率：在那些运行之前记录，其预测误差不到混合不可知基线的一半。该图特定于其目标和语料库，但可跨模型规模迁移：在一个规模上预先选择的混合在测试的每个其他规模上都优于训练所有源任务。因此，迁移是数据的一种属性。该图选择有帮助的任务并丢弃干扰的任务：在推理目标（因果解释、多跳问题和方法论批评）上的准确率比训练所有源任务高出最多 14 个百分点。
英文摘要：
Adapting a language model to a specialized corpus means choosing which instruction-tuning tasks to train on under a fixed budget, and testing one choice costs a fine-tuning run. Common heuristics add more source tasks or pick sources similar to the target. The first assumes transfer is never negative; the second, that it is symmetric. We show that both assumptions fail: task $A$ can help task $B$ while $B$ hurts $A$, so helpfulness is a signed property of ordered source--target pairs. We introduce the transfer map, a signed estimate of how much each source helps or hurts each held-out target. We fit the map in hundreds of fine-tuning runs on Qwen3 and Mistral models from 0.6B to 32B parameters, with all sources drawn from one corpus and no training examples from the target. The map predicts a held-out target's accuracy on unseen mixtures: recorded before those runs, its predictions have less than half the error of a mixture-agnostic baseline. The map is specific to its target and corpus but transfers across model scale: a mixture selected in advance at one size beats training on all source tasks at every other size we tested. Transfer is thus a property of the data. The map selects the tasks that help and drops the one that interferes: accuracy on the reasoning targets (causal explanation, multi-hop questions and methodological critique) rises by up to 14 percentage points over training on all source tasks.
