Research map / Large language models / Instruction and fine-tuning
From Macro to Micro: Probing Dataset Diversity in Language Model Fine-Tuning
Haoyu Li, Xuhong Li, Yiming Dong, Kun Liu
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 |
|---|---|---|---|
| increase token-level (micro) diversity of responses | SFT data, Llama-2-7B · AlpacaEval 2.0 + Arena Hard (10K/20K/30K) | score slope: slope 5.35 / 6.68 / 10.06 (x10^-2) | Improved |
| increase token-level (micro) diversity of instructions | SFT data, Llama-2-7B · AlpacaEval 2.0 + Arena Hard (10K/20K/30K) | score slope: slope -5.45 / -2.83 / -1.80 (x10^-2) | Hurt |
| increase tag-level (meso) diversity | SFT data, Llama-2-7B · AlpacaEval 2.0 + Arena Hard | score slope: slope 4.93 vs macro 2.37 on instructions (10K) | Improved |
| increase semantic (macro) diversity of instructions | SFT data, Llama-2-7B · AlpacaEval 2.0 + Arena Hard | score slope: slope 2.37 (10K) | Improved |
| increase dataset size from 10K to 30K | SFT data, Llama-2-7B · AlpacaEval 2.0 + Arena Hard | score | Improved |
| apply micro response diversity to Llama-2-70B and Llama-3-8B (ablation) | micro response strategy · 10K, AlpacaEval 2.0 + Arena Hard | score | Improved |
| control response length to about 500 tokens (ablation) | micro response strategy · 10K, AlpacaEval 2.0 + Arena Hard | score vs diversity relation | No clear effect |
| use word segmentation instead of model tokenizer for micro diversity (ablation) | micro response strategy · 10K, AlpacaEval 2.0 + Arena Hard | score | Hurt |
| use information entropy of response tokens as diversity metric | diversity metrics · 10K datasets | Pearson correlation with score: 0.58 | Improved |
Closest papers by meaning
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.91)
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.90)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.89)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.89)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.88)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (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.
