Research map / Large language models / Instruction and fine-tuning
Importance-Aware Data Selection for Efficient LLM Instruction Tuning
Tingyu Jiang, Shen Li, Yiyao Song, Lan Zhang, Hualei Zhu, Yuan Zhao, Xiaohang Xu, Kenjiro Taura, Hao Henry Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select top-MIWV samples instead of training on full dataset (1% of Alpaca) | LLaMA2-7B instruction tuning · AlpacaEval | win rate: 27.75 -> 39.50 | Improved |
| select top-MIWV samples instead of full data (1% of Alpaca) | LLaMA2-13B instruction tuning · AlpacaEval | win rate: 35.00 -> 41.30 | Improved |
| select 1% of WizardLM data by MIWV instead of full data | LLaMA2-13B · Open LLM Leaderboard average | average score: 60.23 -> 60.30 | No clear effect |
| select data with random sampling instead of MIWV (ablation) | LLaMA-7B on Alpaca · pairwise win rate vs official Alpaca | win rate | Hurt |
| select data with high prompt loss instead of high MIWV (ablation) | LLaMA-7B on Alpaca · pairwise win rate vs official Alpaca | win rate | Hurt |
| select low MIWV instead of high MIWV samples (ablation) | LLaMA-7B on Alpaca · pairwise win rate vs official Alpaca | win rate | Hurt |
| swap embedding model for retrieval (Bge-en-large vs Gte-base-en-v1.5) (ablation) | MIWV on LLaMA-7B Alpaca · pairwise win rate | win rate | No clear effect |
| add ICL one-shot retrieval to IFD Score (ablation) | IFD Score on LLaMA-7B Alpaca 1% · five test sets win rate | win rate: 0.939 -> 1.017 | Improved |
| use MIWV scoring instead of ICL + IFD Score (ablation) | Alpaca 1% LLaMA-7B · five test sets win rate | win rate: 1.017 -> 1.140 | Improved |
| MIWV selection vs other selection methods at 1% scale | LLaMA2-7B Alpaca · WizardLM test set | win rate vs full data: 1.119 vs 1.092 (RECOST) | Improved |
| increase fraction of selected data toward the full set (ablation) | MIWV subsets · pairwise win rate | win rate | Hurt |
| apply MIWV to Qwen2.5-7B/14B | Qwen2.5 instruction tuning · LIMA test set | win rate | Improved |
Closest papers by meaning
- NICE Data Selection for Instruction Tuning in LLMs with Non-differentiable Evaluation Metric (ICML 2025 · similarity 0.93)
- PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization (ICLR 2024 · similarity 0.92)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.91)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.91)
- Instruction Tuning With Loss Over Instructions (NeurIPS 2024 · similarity 0.91)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
