Research map / Large language models / Instruction and fine-tuning
Token-level Data Selection for Safe LLM Fine-tuning
Yanping Li, Zhening Liu, Zijian Li, Zehong Lin, Jun Zhang
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 progressive refinement (TOSS-Pro) on top of TOSS (ablation) | TOSS, Llama-3-8B-Instruct · HEx-PHI | win rate: 87.54 -> 93.79 | Improved |
| use metric-driven instead of random selection of samples for refining the safety-degraded model (ablation) | TOSS-Pro · HEx-PHI | win rate: 87.37 -> 93.79 | Improved |
| use metric-driven instead of random selection (ablation) | TOSS-Pro · SlimOrca test | win rate: 63.12 -> 68.85 | Improved |
| add a second progressive iteration (ablation) | TOSS-Pro · HEx-PHI | win rate: 91.13 -> 93.79 | Improved |
| add a second progressive iteration (ablation) | TOSS-Pro · ANTHROPIC HH test | win rate: 88.54 -> 88.85 | No clear effect |
| token-level TOSS instead of sample-level SEAL | Llama-3-8B-Instruct · HEx-PHI | win rate: 68.83 -> 87.54 | Improved |
| token-level TOSS instead of sample-level SEAL | Llama-3-8B-Instruct · SlimOrca test | win rate: 57.41 -> 68.37 | Improved |
| token-level TOSS instead of SEAL at 30% harmful data | fine-tuning mixture · HEx-PHI | win rate: 63.37 -> 69.71 | Improved |
| token-level TOSS instead of SEAL at 70% harmful data | fine-tuning mixture · SlimOrca test | win rate: 51.30 -> 64.48 | Improved |
| token-level TOSS instead of SEAL on benign data | fine-tuning on benign data · HEx-PHI | win rate: 55.57 -> 55.75 | No clear effect |
| token-level TOSS instead of SEAL on GSM8K fine-tuning | Llama-3-8B-Instruct · HEx-PHI | win rate: 53.04 -> 93.18 | Improved |
| token-level TOSS instead of SEAL on GSM8K fine-tuning | Llama-3-8B-Instruct · GSM8K | win rate: 60.62 -> 62.97 | Improved |
Closest papers by meaning
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (ICML 2025 · similarity 0.91)
- Token Cleaning: Fine-Grained Data Selection for LLM Supervised Fine-Tuning (ICML 2025 · similarity 0.91)
- Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024 · similarity 0.91)
- Learning and Forgetting Unsafe Examples in Large Language Models (ICML 2024 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- What Makes and Breaks Safety Fine-tuning? A Mechanistic Study (NeurIPS 2024 · similarity 0.90)
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.
