Research map / Large language models / Instruction and fine-tuning
Task-Aware Data Selection via Proxy-Label Enhanced Distribution Matching for LLM Finetuning
Hao Cheng, Rui Zhang, Ling Li, Na Di, Jiaheng Wei, Zhaowei Zhu, Bo Han
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 |
|---|---|---|---|
| add OOD filtering before incremental sampling (ablation) | Align pipeline · TruthfulQA | accuracy: 3.5 -> 42.6 | Improved |
| add incremental sampling after OOD filtering (ablation) | Align pipeline · TruthfulQA | accuracy: 42.1 -> 42.6 | Improved |
| add incremental sampling after OOD filtering (ablation) | Align pipeline · BBH | accuracy: 61.9 -> 62.5 | Improved |
| add incremental sampling after OOD filtering (ablation) | Align pipeline · TyDiQA | accuracy: 56.4 -> 59.1 | Improved |
| add LLM proxy labels with distribution alignment instead of inserting labels into RDS+ (ablation) | RDS+ · TruthfulQA | accuracy: 5.4 -> 46.4 | Improved |
| insert proxy label information into RDS+ (ablation) | RDS+ · MMLU | accuracy: 63.6 -> 63.7 | No clear effect |
| add LLM annotation (joint P(X,Y) alignment) versus X only (ablation) | Our pipeline · TruthfulQA | accuracy: 40.8 -> 46.4 | Improved |
| use LLM-based OOD filtering instead of embedding similarity (ablation) | Our pipeline · TruthfulQA | accuracy: 28.3 -> 46.4 | Improved |
| raise min score threshold from 6 to 7 | Align_topic · TruthfulQA | accuracy: 24.7 -> 46.9 | Improved |
| raise min score threshold from 6 to 7 | Align_topic · MMLU | accuracy: 65.0 -> 64.5 | Hurt |
| increase selection size from 1K to 5K | Align_style · TruthfulQA | accuracy: 6.8 -> 32.4 | Improved |
| 20% synthetic label noise | Our pipeline · MMLU | accuracy: 64.3 -> 64.1 | No clear effect |
Closest papers by meaning
- TSDS: Data Selection for Task-Specific Model Finetuning (NeurIPS 2024 · similarity 0.89)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.88)
- Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations (NeurIPS 2025 · similarity 0.88)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.88)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.88)
- MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · 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.
