English

Low-Confidence Gold: Refining Low-Confidence Samples for Efficient Instruction Tuning

Computation and Language 2026-04-09 v7 Artificial Intelligence

Abstract

The effectiveness of instruction fine-tuning for Large Language Models is fundamentally constrained by the quality and efficiency of training datasets. This work introduces Low-Confidence Gold (LCG), a novel filtering framework that employs centroid-based clustering and confidence-guided selection for identifying valuable instruction pairs. Through a semi-supervised approach using a lightweight classifier trained on representative samples, LCG curates high-quality subsets while preserving data diversity. Experimental evaluation demonstrates that models fine-tuned on LCG-filtered subsets of 6K samples achieve superior performance compared to existing methods, with substantial improvements on MT-bench and consistent gains across comprehensive evaluation metrics. The framework's efficacy while maintaining model performance establishes a promising direction for efficient instruction tuning.All open-source assets are publicly available at https://github.com/Lizruletheworld/Low-Confidence_Gold.

Keywords

Cite

@article{arxiv.2502.18978,
  title  = {Low-Confidence Gold: Refining Low-Confidence Samples for Efficient Instruction Tuning},
  author = {Hongyi Cai and Jie Li and Mohammad Mahdinur Rahman and Wenzhen Dong},
  journal= {arXiv preprint arXiv:2502.18978},
  year   = {2026}
}

Comments

Accepted to EMNLP Findings 2025

R2 v1 2026-06-28T21:58:27.391Z