Research map / Large language models / Instruction and fine-tuning
Data Difficulty and the Generalization–Extrapolation Tradeoff in LLM Fine-Tuning
Siyuan Liu, Tinghong Chen, Xinghan Li, Yifei Wang, Jingzhao Zhang
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 |
|---|---|---|---|
| remove the longest 10% (hardest) examples from the SFT data (ablation) | SFT on OpenR1-Math-94k · Math500, small data sizes (2k-5k) | accuracy change vs full SFT | Improved |
| remove the shortest 10% (easiest) examples from the SFT data (ablation) | SFT on OpenR1-Math-94k · Math500, large data sizes (20k-50k) | accuracy change vs full SFT | Improved |
| remove the longest 10% examples from the SFT data at large data size (ablation) | SFT on OpenR1-Math-94k · Math500, large data sizes | accuracy change vs full SFT | Hurt |
| use harder (longer CoT) training data at fixed data size (ablation) | SFT on OpenMath · avg of Math500, AIME24, Minerva Math | accuracy gain over base | Mixed |
| increase SFT data size at fixed difficulty (ablation) | SFT on OpenMath · avg of Math500, AIME24, Minerva Math | accuracy gain over base | Improved |
| use a stronger base model (7B instead of 1.5B) (ablation) | SFT on OpenMath · avg of Math500, AIME24, Minerva Math | accuracy gain over base | Improved |
| increase data size on iGSM synthetic data (ablation) | SFT on iGSM, base Ops[2-8]2k · iGSM test (ops 2-20) | SFT improvement over base | Improved |
| fine-tune on very hard iGSM data with small data budget (ablation) | SFT on iGSM, base Ops[2-8]2k · iGSM test (ops 2-20) | accuracy | Hurt |
| fine-tune on easy iGSM data (ablation) | SFT on iGSM, base Ops[2-8]2k · iGSM test (ops 2-20) | accuracy | Mixed |
| replace SFT with DFT token reweighting when data are too hard for the data size (ablation) | SFT on iGSM, base Ops[2-8]2k · iGSM, limited data and hard difficulty | improvement over base | Improved |
| replace SFT with DFT when data size is large (ablation) | SFT on iGSM, base Ops[2-8]2k · iGSM, large data sizes | improvement over base | Hurt |
Closest papers by meaning
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.92)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.91)
- Two-stage LLM Fine-tuning with Less Specialization and More Generalization (ICLR 2024 · similarity 0.90)
- When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.90)
- One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.90)
- FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain (ICML 2025 · similarity 0.89)
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.
