Research map / Large language models / Instruction and fine-tuning
Selection of LLM Fine-Tuning Data Based on Orthogonal Rules
Xiaomin Li, Mingye Gao, Zhiwei Zhang, Chang Yue, Hong Hu
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 |
|---|---|---|---|
| select data with DPP-chosen 10 orthogonal rules instead of No Rule scoring | data selection for fine-tuning · IMDB sentiment (Pythia-1B) | accuracy: 51.1 -> 53.5 | Improved |
| select data with DPP rules instead of No Rule scoring | data selection · MMLU medical avg (Pythia-1B) | accuracy: 29.2 -> 31.6 | Improved |
| select data with DPP rules instead of No Rule scoring | data selection · MMLU math avg (Llama3-8B) | accuracy: 38.9 -> 40.8 | Improved |
| select data with DPP rules instead of No Rule scoring | data selection · Code avg (Llama3-8B) | score: 45.9 -> 48.6 | Improved |
| fine-tune on 20K DPP-selected samples versus the backbone | Llama3-8B LoRA · Code avg | score: 45.4 -> 48.6 | Improved |
| use QuRating four fixed rules instead of No Rule | data selection · IMDB sentiment (Pythia-1B) | accuracy: 51.1 -> 47.7 | Hurt |
| use QuRating four fixed rules instead of No Rule | data selection · Code avg (Llama3-8B) | score: 45.9 -> 43.1 | Hurt |
| use GPT-generated uncorrelated rules instead of DPP rules | data selection · MMLU math avg (Llama3-8B) | accuracy: 40.8 -> 37.3 | Hurt |
| fine-tune on all 1M SlimPajama samples versus 20K DPP-selected | fine-tuning · Code avg (Llama3-8B) | score: 48.6 -> 45.8 | Hurt |
| uniform sampling 20K versus backbone | fine-tuning · Code avg (Llama3-8B) | score: 45.4 -> 38.7 | Hurt |
| DSIR selection versus DPP rules | data selection · MMLU medical avg (Pythia-1B) | accuracy: 29.2 -> 23.9 | Hurt |
| use DPP-selected rules instead of random 10 rules or all 50 rules (ablation) | rule selection · downstream fine-tuning | benchmark performance | Improved |
Closest papers by meaning
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.92)
- One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.91)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · similarity 0.90)
- LLM Data Selection and Utilization via Dynamic Bi-level Optimization (ICML 2025 · similarity 0.90)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.90)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 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.
