Research map / Large language models / Instruction and fine-tuning
The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning
Will Hawkins, Kai Rawal, Jonathan Rystrøm, Stratis Tsirtsis, Zihao Fu, Greta Warren, Ryan Brown, Eoin D. Delaney, Sandra Wachter, Brent Mittelstadt, Chris Russell
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 benign English data (alpha/r=8, 1 epoch) | Gemma-3-1B-IT · SORRY-Bench English eval | compliance rate (%): 30.2 -> 65.5 | Hurt |
| fine-tune on benign Hindi data (alpha/r=8) | Gemma-3-1B-IT · SORRY-Bench English eval | compliance rate (%): 44.8 -> 76.8 | Hurt |
| fine-tune on benign Portuguese data | Gemma-3-4B-IT · SORRY-Bench local eval | compliance rate: about 4x increase | Hurt |
| fine-tune on benign data (any language) | Qwen3-0.6B · SORRY-Bench | compliance rate: 83.2 -> 69.8 (EN FT, EN eval) | Improved |
| use more conservative LoRA scaling alpha/r=2 instead of 8 (ablation) | Gemma-3-1B-IT · SORRY-Bench English eval (EN FT) | compliance rate (%): 65.5 (alpha/r=8) vs 44.1 (alpha/r=2) | No clear effect |
| fine-tune in non-English language instead of English | small models · SORRY-Bench local eval | compliance increase vs EN baseline: two-thirds of cases larger than English | Mixed |
| add capability (TinyMMLU) improvement as predictor of compliance change | benign multilingual fine-tuning · TinyMMLU vs SORRY-Bench EN | correlation | No clear effect |
| raise language resource level | Llama-3.2-1B · SORRY-Bench local eval | Spearman r: r = 0.840, p = 0.005 | Improved |
| raise language resource level | Gemma-3-4B · SORRY-Bench local eval | Spearman r: r = 0.225, p = 0.560 | No clear effect |
| larger safety-direction displacement at 80th percentile layer | Llama models · SORRY-Bench EN | compliance change | Mixed |
Closest papers by meaning
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.91)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.91)
- TuCo: Measuring the Contribution of Fine-Tuning to Individual Responses of LLMs (ICML 2025 · similarity 0.91)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.90)
- Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models (ICML 2024 · similarity 0.89)
- On the Role of Attention Heads in Large Language Model Safety (ICLR 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.
