Research question
Do safety-tuned models over-refuse benign prompts?
Key findings
- 250 safe prompts that superficially resemble unsafe ones; Llama 2 refused up to 38% - Exaggerated safety trades off against helpfulness
XSTest
Introduced systematic measurement of over-refusal, establishing that safety tuning can make models refuse harmless requests.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.