Research map / Large language models / Instruction and fine-tuning
IDEAL: Data Equilibrium Adaptation for Multi-Capability Language Model Alignment
Chenlin Ming, Chendi Qu, Qizhi Pei, Zhuoshi Pan, Yu Li, Xiaoming Duan, Lijun Wu, Conghui He
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 |
|---|---|---|---|
| use IDEAL gradient-based domain re-weighting instead of Joint SFT uniform mixture | Llama3.1-8B SFT · Overall average, 4 benchmarks, 1 epoch | average score: 54.79 -> 57.87 | Improved |
| use IDEAL re-weighting instead of Joint SFT | Llama3.1-8B SFT · Overall average, 4 benchmarks, 3 epochs | average score: 55.35 -> 59.23 | Improved |
| use IDEAL re-weighting instead of Joint SFT | Llama3.1-8B SFT · HumanEval, 1 epoch | pass@1: 41.26 -> 50.61 | Improved |
| use IDEAL re-weighting on balanced 5-domain mixture | Llama3.1-8B SFT · Extended 8-benchmark average | average score: 54.03 -> 55.20 | Improved |
| use IDEAL re-weighting on balanced 5-domain mixture | Llama3.1-8B SFT · MATH (extended) | accuracy: 30.94 -> 30.20 | Hurt |
| increase joint SFT data from 25k to full 66k (ablation) | Joint SFT · Extended 8-benchmark average | average score: 54.03 -> 53.94 | No clear effect |
| upsample DoReMi data to IDEAL's volume (ablation) | DoReMi D2 · Overall average, 1 epoch | average score: 55.17 -> 55.36 | No clear effect |
| upsample DOGE data to IDEAL's volume (ablation) | DOGE D2 · Overall average, 1 epoch | average score: 53.37 -> 53.81 | No clear effect |
| set scale m to 0.15 rather than 0.1 or 0.3 (ablation) | IDEAL · Sensitivity of m, D1 1 epoch | average score | Improved |
| run second IDEAL iteration (D1 to D2) (ablation) | IDEAL · Overall average, 1 epoch | average score: 56.28 -> 57.87 | Improved |
| train 3 epochs instead of 1 epoch (ablation) | IDEAL D2 · HumanEval | pass@1: 50.61 -> 46.34 | Hurt |
Closest papers by meaning
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.90)
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.89)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.89)
- Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods (ICLR 2026 · similarity 0.88)
- Preference-Oriented Supervised Fine-Tuning: Favoring Target Model over Aligned Large Language Models (AAAI 2025 · similarity 0.88)
- Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.88)
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.
