Scholarlink

Instruction-tuning data quality has emerged as a key determinant of large language model (LLM) performance. In this work, we propose a novel data selection framework designed to evaluate and improve instruction data quality. The approach begins by extracting and clustering semantic units from the outputs generated by multiple LLMs. Then, we design an enhanced semantic entropy metric that integrates both the salience of semantic units and their divergence from cluster centroids to assess the clarity and consistency of the question in an instruction sample. Finally, mutual information is employed to evaluate the informativeness and alignment of the answer in an instruction sample. Experimental results demonstrate that our method significantly reduces evaluation overhead while achieving superior fine-tuning performance compared to state-of-the-art data quality evaluation methods.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Instruction-Tuning Data Quality Evaluation Method Based on Semantic Entropy and Mutual Information

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Huijun Xuan,
  • Yanfei Lin,
  • Cong Liu,
  • Xueyu Li,
  • Xiaoli Zheng,
  • Xuening Sun,
  • Mukai Chen,
  • Xiao Liu,
  • Muqi Luo,
  • Zhilin Du,
  • Enxiao Liu

摘要

Instruction-tuning data quality has emerged as a key determinant of large language model (LLM) performance. In this work, we propose a novel data selection framework designed to evaluate and improve instruction data quality. The approach begins by extracting and clustering semantic units from the outputs generated by multiple LLMs. Then, we design an enhanced semantic entropy metric that integrates both the salience of semantic units and their divergence from cluster centroids to assess the clarity and consistency of the question in an instruction sample. Finally, mutual information is employed to evaluate the informativeness and alignment of the answer in an instruction sample. Experimental results demonstrate that our method significantly reduces evaluation overhead while achieving superior fine-tuning performance compared to state-of-the-art data quality evaluation methods.