IC-818RLHF on general-purpose preference data has negligible net effect on toxicity in Pythia and Llama-7B models

Aaron Jiaxun Li, Satyapriya Krishna, Himabindu Lakkaraju

SourceMore RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness

Using the RealToxicityPrompts benchmark (1.2k toxic and 1.2k nontoxic prompts) with the Perspective API as a toxicity scorer, the authors measure expected maximum toxicity across four prompt settings (toxic/nontoxic user prompts crossed with benign/adversarial system prompts). Toxicity exhibits non-significant fluctuations across RLHF stages: a slight increase after SFT followed by a decrease from PPO or DPO. The net effect is negligible, varies across models, and falls within the error margin. The authors explain this as a cancellation: SFT amplifies toxic content in chosen responses, while PPO/DPO reinforce less toxic outputs because rejected responses are more toxic.

Evidence
correlational
Key metric
net impact on toxicity is negligible; chosen responses have average toxicity 0.13 and high toxicity rate 5.7%, rejected responses 0.18 and 8.6%
Caveat
The net effect is within error bounds and varies across models; the authors note the improvement in the second stage does not guarantee outweighing the negative effect from SFT.
Model
Pythia, LLaMA
Datasets
Anthropic-HH / Anthropic HHH / Anthropic/HH-RLHF [train], RealToxicityPrompts [eval]
Methods
PPO [primary], Direct Preference Optimization / Direct Preference Optimisation / DPO / Rafailov et al. 2023 (Direct Preference Optimization) [primary], Perspective API [eval]
Related findings
IC-815, IC-816, IC-817
Extraction
automatic-extraction