Research map / Large language models / Language, safety and interpretability
Scaling Towards the Information Boundary of Instructions through Data Synthesizing
Li Du, Hanyu Zhao, Yiming Ju, Tengfei Pan
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 |
|---|---|---|---|
| fine-tune on InfInstruct-Sub instead of official instruct model | Llama-3-8B · AlpacaEval 2.0 | win rate: 22.9 -> 36.2 | Improved |
| fine-tune on InfInstruct-Sub instead of official instruct model | Llama-3-8B · Arena-Hard-v0.1 | score: 20.6 -> 35.3 | Improved |
| fine-tune on InfInstruct-Sub instead of official instruct model | Mistral-7B-v0.2 · AlpacaEval 2.0 | win rate: 17.1 -> 41.2 | Improved |
| fine-tune on InfInstruct-Sub instead of official instruct model | Qwen-2-7B · Arena-Hard-v0.1 | score: 19.6 -> 27.7 | Improved |
| use InfInstruct-Sub instead of Magpie instruction data | Llama-3-8B · Arena-Hard-v0.1 | score: 14.9 -> 35.3 | Improved |
| enlarge instruction set size without targeting depth (GenQA 6.47M vs smaller sets) | Llama-3-8B · Arena-Hard-v0.1 | score: GenQA 6.47M: 3.0 vs WildChat 652k: 8.7 | Hurt |
| select informative seed instructions rather than random samples (ablation) | Llama-3-8B (50k) · AlpacaEval 2.0 | win rate: 4.5 -> 6.4 | Improved |
| select informative seed instructions rather than random samples (ablation) | Llama-3-8B (50k) · Arena-Hard-v0.1 | score: 3.5 -> 3.9 | Improved |
| increase depth and coverage of 20k-instruction subsets (ablation) | Llama-3-8B · AlpacaEval and Arena-Hard | win rate / score | Improved |
Closest papers by meaning
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.89)
- Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning (ICLR 2025 · similarity 0.88)
- Beyond IID: Optimizing Instruction Finetuning from the Perspective of Instruction Interaction and Dependency (AAAI 2025 · similarity 0.88)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.88)
- #InsTag: Instruction Tagging for Analyzing Supervised Fine-tuning of Large Language Models (ICLR 2024 · similarity 0.88)
- MUFFIN: Curating Multi-Faceted Instructions for Improving Instruction Following (ICLR 2024 · similarity 0.87)
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.
