Research map / Large language models / Alignment and preferences
BeaverTails: Towards Improved Safety Alignment of LLM via a Human-Preference Dataset
Jiaming Ji, Mickel Liu, Juntao Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, Yaodong Yang
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 |
|---|---|---|---|
| train separate reward and cost models from decoupled helpfulness/harmlessness rankings with PPO-Lagrangian instead of a single mixed-preference PPO (ablation) | Alpaca-7B · win rate vs Alpaca-7B, harmlessness | win rate: 68.64% -> 82.57% | Improved |
| decoupled preference vs mixed preference (ablation) | Alpaca-7B · win rate vs Alpaca-7B, helpfulness | win rate: 65.07% -> 85.57% | Improved |
| use cost model trained on rankings and sign instead of ensemble of 14 binary classifiers (max) (ablation) | Safe-RLHF · win rate vs Alpaca-7B, harmlessness | win rate: 64.50% -> 82.57% | Improved |
| use cost model rather than classifier ensemble (max) (ablation) | Safe-RLHF · win rate vs Alpaca-7B, helpfulness | win rate: 74.00% -> 85.57% | Improved |
| use max instead of mean of classifier probabilities as cost (ablation) | PPOL-classifier · win rate vs Alpaca-7B, harmlessness | win rate: 59.07% -> 64.50% | Improved |
| train on BeaverTails decoupled data instead of HH-RLHF dataset (ablation) | PPO-based RLHF · win rate vs Alpaca-7B, harmlessness | win rate: 66.21% -> 82.57% | Improved |
| train on BeaverTails decoupled data instead of HH-RLHF dataset (ablation) | PPO-based RLHF · win rate vs Alpaca-7B, helpfulness | win rate: 64.93% -> 85.57% | Improved |
| two-stage annotation (14 harm categories then ranking) instead of single-stage (ablation) | annotation process · annotator-researcher agreement | agreement rate: approximately 15% increase | Improved |
| apply Safe RLHF fine-tuning | Alpaca-7B · static cost model score on red-team prompts | cost | Improved |
Closest papers by meaning
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.89)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.89)
- Does Safety Training of LLMs Generalize to Semantically Related Natural Prompts? (ICLR 2025 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.88)
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.
