Research map / Large language models / Instruction and fine-tuning
Difficulty–Diversity Collaborative Filtering for Data-Efficient LLM Fine-Tuning
Long P. Hoang, Wenxuan Zhang, Wei Lu
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 DDCF selection instead of Random 1K | Qwen2.5-Math-7B · AIME24 | pass@32: 38.6 -> 49.0 | Improved |
| use DDCF instead of Perplexity selection | Qwen2.5-Math-7B · ID avg | accuracy: 69.3 -> 70.2 | Improved |
| use DDCF instead of Random selection | Qwen3-8B-Base · ID avg | accuracy: 83.6 -> 85.0 | Improved |
| use DDCF 1K subset instead of full dataset | Qwen3-8B-Base · OOD avg | accuracy: 80.4 -> 80.5 | No clear effect |
| use DDCF 1K subset instead of full dataset | Qwen2.5-Math-7B · ID avg | accuracy: 75.8 -> 70.2 | Hurt |
| increase difficulty weight lambda from 0 to 0.2 (ablation) | DDCF · AMC23 | pass@32: 81.1 -> 83.3 | Improved |
| increase difficulty weight lambda from 0.2 to 1.0 (ablation) | DDCF · SVAMP | accuracy: about 4 points decline | Hurt |
| fine-tune on 1K DDCF subset | Qwen2.5-Math-1.5B · ID avg | accuracy: 54.3 -> 45.4 | Hurt |
| increase DDCF data size from 1K to 8K+ (ablation) | Qwen2.5-Math-1.5B · ID avg | accuracy: 45.4 -> 56.9 (k=16,000) | Improved |
| use DDCF on MMLU STEM instead of DiSF | Qwen2.5-7B · MMLU STEM | accuracy: 74.7 -> 81.1 | Improved |
| add more models to the correctness predictor (1 to 23) (ablation) | correctness predictor · held-out accuracy | accuracy: 81.5 -> 82.7 | Improved |
| increase seed questions from 1K to 17.5K (ablation) | correctness predictor · held-out accuracy | accuracy: 80.0 -> 82.7 | Improved |
Closest papers by meaning
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.90)
- Towards Scalable Data Diversification for Language Model Pretraining via Leverage Score Sampling (NeurIPS 2026 · similarity 0.90)
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.90)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.90)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.89)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 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.
