Instruction Data Selection via Answer Divergence
For practitioners of instruction tuning, ADG offers a practical and effective signal for selecting high-quality instruction data without requiring human annotation or external models.
The paper proposes Answer Divergence-Guided Selection (ADG), a method for selecting instruction data based on the geometric structure of multiple high-temperature generations per instruction. Fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors across two backbones and three public instruction pools on six benchmarks.
Instruction tuning relies on large instruction-response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.