arXiv：cs.CL· Zile Yang, Ling Li, Na Di, Jinlong Pang, Yao Zhou, Hao Cheng, Bo Han, Jiaheng Wei·· 6 小时前
用规则引导的 LLM 重构回收低质量指令数据
Reclaiming Low-Quality Instruction Data with Rule-Guided LLM Reconstruction
AI 导读
研究提出规则引导的 LLM 重构框架，从 123,786 条低分样本中构建出 54,888 条训练样本，用于监督微调 SFT 数据筛选。在三个 7B/8B 基础 LLM 上，重构数据的下游平均表现高于低分源数据集及所有评估过的数据选择基线，且样本量远少于约 300K 的全量数据集却取得更高平均分。逐步消融显示规则引导重构是相较直接融合的最大提升来源，人工审查则发现源整合与答案忠实度仍有局限。
正文
Abstract:Data selection for supervised fine-tuning (SFT) often discards low-scoring instruction examples that may retain useful supervision. We propose a rule-guided LLM reconstruction framework to reclaim this information. The framework identifies low-scoring samples through LLM-based rating and score correction, selects semantically representative sources, and reconstructs instruction-response pairs through strategy-guided generation, criteria-based validation, and failure-guided repair. From 123,786 low-scoring samples, it constructs 54,888 training examples. Across three 7B/8B base LLMs, the reconstructed data achieves higher average downstream performance than the low-scoring source set and all evaluated data-selection baselines. Despite using substantially fewer examples than the full source dataset (approximately 300K), it achieves higher reported average scores than full-data training across all three base models. Stepwise ablations identify rule-guided reconstruction as the largest source of observed improvements over direct fusion, while a manual audit reveals remaining limitations in source integration and answer faithfulness. These findings support reconstruction as a complementary approach to selection-based SFT data curation.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2510.23160 [cs.CL] |
| (or arXiv:2510.23160v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2510.23160 arXiv-issued DOI via DataCite |
Submission history
From: Zile Yang [view email]
[v1]
Mon, 27 Oct 2025 09:39:22 UTC (4,852 KB)
[v2]
Wed, 14 Oct 2026 14:53:40 UTC (6,594 KB)
来源：arXiv：cs.CL · arxiv.org