Research map / Large language models / Language, safety and interpretability
Exploring the Limits of Domain-Adaptive Training for Detoxifying Large-Scale Language Models
Boxin Wang, Wei Ping, Chaowei Xiao, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, Bo Li, Anima Anandkumar, Bryan Catanzaro
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 self-generated nontoxic corpus (SGEAT augmented) instead of filtered pre-training data (DAPT) | 1.3B LM · RealToxicityPrompts full | expected maximum toxicity: 0.47 -> 0.43 | Improved |
| fine-tune on human-annotated nontoxic Jigsaw data | 1.3B LM · RealToxicityPrompts full | expected maximum toxicity: 0.57 -> 0.58 | Hurt |
| use heuristic single-prompt generation instead of unconditional or augmented prompts (ablation) | SGEAT 1.3B · RealToxicityPrompts full | expected maximum toxicity: 0.43 -> 0.47 | Hurt |
| use augmented prompts instead of standard unconditional generation (ablation) | SGEAT 1.3B · RealToxicityPrompts full | expected maximum toxicity: 0.44 -> 0.43 | Improved |
| combine SGEAT with DExperts decoding | 1.3B LM · RealToxicityPrompts full | expected maximum toxicity: 0.31 -> 0.27 | Improved |
| DExperts decoding instead of domain-adaptive training | 1.3B LM · validation PPL and utility | PPL: 10.18 -> 19.87 | Hurt |
| scale model from 126M to 530B (standard LMs) | standard LM · RealToxicityPrompts full | expected maximum toxicity: 0.56 -> 0.57 | No clear effect |
| scale model size with SGEAT detoxification | SGEAT (augmented) · RealToxicityPrompts full | toxicity probability reduction: 27% -> 16% reduction (126M -> 530B) | Hurt |
| train 530B for 5 epochs instead of 3 (ablation) | SGEAT (augmented) 530B · validation PPL | PPL: 7.86 -> 9.63 | Hurt |
| use 100k instead of 50k self-generated samples (ablation) | SGEAT (augmented) 530B · RealToxicityPrompts full | expected maximum toxicity: 0.46 -> 0.45 | Improved |
| adapter instead of whole-model adaptation (ablation) | SGEAT (augmented) 530B · validation PPL | PPL: 7.86 -> 7.22 | Improved |
| prefix tuning instead of adapter (ablation) | DAPT 1.3B · RealToxicityPrompts full | expected maximum toxicity: 0.48 -> 0.52 | Hurt |
Closest papers by meaning
- Large Language Models can Become Strong Self-Detoxifiers (ICLR 2025 · similarity 0.92)
- UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation (ICLR 2025 · similarity 0.91)
- Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing (NeurIPS 2025 · similarity 0.90)
- When Bad Data Leads to Good Models (ICML 2025 · similarity 0.89)
- Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models (ICML 2024 · similarity 0.88)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 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.
