arXiv cs.CL· Vladimir Necula, David Gvadzabia, Sofia Serrano·· 12 小时前AI 评分24
指令微调早期阶段的数据质量指标实证研究：response ratio 比常用代理指标更能预测训练效用
"High-Quality" Data for Early-Stage Instruction Tuning: An Empirical Study of Data Quality Metrics
AI 导读
一项实证研究在固定微调预算下比较了多种样本级数据质量指标对指令微调训练效用的预测能力，发现由 response ratio 刻画的“答案密集监督”比常用替代代理指标更强、更可迁移。研究按各指标分数将指令-回复样本分为低、中、高三个 token 质量大致相等的分箱，分别微调后以验证困惑度和 LLM-as-judge 评分随累计训练 token 的变化进行比较。
正文
Abstract:"Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which simple example-level metrics, drawn from prior work on data selection, text quality, and instruction tuning, best predict training utility under a fixed finetuning budget. For each metric, we sort instruction-response examples by score, partition them into low, middle, and high bins with roughly equal token mass, finetune separately on each bin, and compare validation perplexity and LLM-as-judge scores for the models' generated answers as a function of cumulative training tokens. Our results suggest that answer-dense supervision, as captured by response ratio, is a stronger and more portable predictor of instruction-tuning utility than commonly used alternative proxies.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.33969 [cs.CL] |
| (or arXiv:2610.33969v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33969 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Sofia Serrano [view email]
[v1]
Sun, 25 Oct 2026 00:49:49 UTC (33,956 KB)
来源：arXiv cs.CL · arxiv.org