Safety IndexEvaluationsWildGuardTest
WildGuardTest
Refusal calibration · Harmful requestsCounts in the indexPlanned
Whether the model refuses the harmful prompts and answers the harmless ones in WildGuardTest, a set of 1,725 ordinary and adversarial prompts with human harm labels.
- Metric
- % refused (harmful half). Higher is better.
- Dataset
- huggingface.co/datasets/allenai/wildguardmix · 1,725 · ODC-BY (gated; check card)
- Grader
- WildGuard-7B refusal head (paper's own openifier, run locally)
- Status
- Planned
- Replicability
- public
- Notes
- Dataset access is being arranged before the first run.
Results
Planned; no measurements yet.