AI Paper Atlas

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

NeurIPS 2022 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 20 citations (OpenAlex)

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 changedWhereMeasuredEffect
fine-tune on self-generated nontoxic corpus (SGEAT augmented) instead of filtered pre-training data (DAPT)1.3B LM · RealToxicityPrompts fullexpected maximum toxicity: 0.47 -> 0.43Improved
fine-tune on human-annotated nontoxic Jigsaw data1.3B LM · RealToxicityPrompts fullexpected maximum toxicity: 0.57 -> 0.58Hurt
use heuristic single-prompt generation instead of unconditional or augmented prompts (ablation)SGEAT 1.3B · RealToxicityPrompts fullexpected maximum toxicity: 0.43 -> 0.47Hurt
use augmented prompts instead of standard unconditional generation (ablation)SGEAT 1.3B · RealToxicityPrompts fullexpected maximum toxicity: 0.44 -> 0.43Improved
combine SGEAT with DExperts decoding1.3B LM · RealToxicityPrompts fullexpected maximum toxicity: 0.31 -> 0.27Improved
DExperts decoding instead of domain-adaptive training1.3B LM · validation PPL and utilityPPL: 10.18 -> 19.87Hurt
scale model from 126M to 530B (standard LMs)standard LM · RealToxicityPrompts fullexpected maximum toxicity: 0.56 -> 0.57No clear effect
scale model size with SGEAT detoxificationSGEAT (augmented) · RealToxicityPrompts fulltoxicity probability reduction: 27% -> 16% reduction (126M -> 530B)Hurt
train 530B for 5 epochs instead of 3 (ablation)SGEAT (augmented) 530B · validation PPLPPL: 7.86 -> 9.63Hurt
use 100k instead of 50k self-generated samples (ablation)SGEAT (augmented) 530B · RealToxicityPrompts fullexpected maximum toxicity: 0.46 -> 0.45Improved
adapter instead of whole-model adaptation (ablation)SGEAT (augmented) 530B · validation PPLPPL: 7.86 -> 7.22Improved
prefix tuning instead of adapter (ablation)DAPT 1.3B · RealToxicityPrompts fullexpected maximum toxicity: 0.48 -> 0.52Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy