Research map / Large language models / Instruction and fine-tuning
Be Careful When Fine-tuning On Open-Source LLMs: Your Fine-tuning Data Could Be Secretly Stolen!
Zhexin Zhang, Yuhao Sun, Junxiao Yang, Shiyao Cui, yuanchao zhang, Hongning Wang, Minlie Huang
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 SFT backdoor training with extraction instruction | Qwen2.5-7B · Dolly extraction | mean match ratio (%): 13.4 -> 29.8 | Improved |
| use GRPO instead of SFT for backdoor training | Qwen2.5-7B · Dolly extraction | mean match ratio (%): 29.8 -> 33.2 | Improved |
| use GRPO instead of SFT for backdoor training | Llama3.1-8B · Dolly extraction | mean match ratio (%): 43.3 -> 38.5 | Hurt |
| increase model size from 7B to 32B | Qwen2.5 SFT backdoor · Dolly extraction | mean match ratio (%): 29.8 -> 49.2 | Improved |
| add backdoor training | Qwen2.5-7B · AlpacaEval 2 / MMLU | win rate / accuracy: 28.0 -> 33.0; MMLU 71.3 -> 71.3 | No clear effect |
| test with paraphrased extraction instruction instead of training instruction (ablation) | SFT backdoor · Dolly extraction | mean match ratio (%): 29.8 -> 17.6 | Hurt |
| train backdoor with fabricated trigger instruction (ablation) | SFT backdoor · Dolly extraction detection by paraphrase probing | mean match ratio (%): paraphrase-probe 17.6 -> 11.2 | Improved |
| add DP-SGD during downstream fine-tuning (ablation) | Llama3.1-8B SFT backdoor · MATH extraction | mean match ratio (%): 50.9 -> 0.9 (eps=4.0) | Hurt |
| add DP-SGD during downstream fine-tuning (ablation) | Llama3.1-8B · MATH500 | accuracy (%): 14.0 -> 1.2 (eps=4.0) | Hurt |
| increase downstream fine-tuning epochs (ablation) | Qwen2.5-7B SFT backdoor · Dolly extraction | match ratio | Improved |
Closest papers by meaning
- Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safety (ICML 2025 · similarity 0.92)
- Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024 · similarity 0.88)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.88)
- SafetyPrompts: A Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety (AAAI 2025 · similarity 0.88)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.88)
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (ICML 2025 · 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.
