Allen AI CoCoNot scores whether a chat model withholds answers on 1,001 contextual noncompliance prompts, plus a 379-item contrast set for over-refusal.
unassessed
| Category | safety |
|---|---|
| Subcategory | contextual noncompliance and over-refusal, beyond unsafe queries |
| Page status | active |
| Metric | compliance_rate (percent; original: lower is better; contrast: higher is better) |
| Direction | lower_is_better |
| Unit | % |
| Dataset size | 1001 |
| Dataset licence | ODC-By (LICENSE.md); card also links Allen AI IMPACT-LR, which 404'd |
| Publisher | Allen Institute for AI (with University of Washington, Microsoft Research, Samaya AI) |
CoCoNot (Contextually, Comply Not) tests when a chat model should not comply. The original test set is 1,001 English prompts across five taxonomy categories: incomplete, unsupported, indeterminate, safety, and humanizing requests. A judge labels the reply ACCEPTABLE (noncompliant) or UNACCEPTABLE (compliant). A 379-prompt contrast set asks the reverse: the model should answer, and the judge labels COMPLIANCE versus NONCOMPLIANCE. It is not jailbreak success and not [xstest](xstest.md), which only probes over-refusal on superficially unsafe wording.
Single-turn English prompt in, free-text reply out. inspect_evals defaults to subset=original, no system prompt, temperature 0, max_tokens 256, one epoch. Optional Tulu-style system prompt. Default judge openai/gpt-3.5-turbo.
No model card in ModelSpec reports this benchmark yet.