You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is restricted to authorized defensive-security research. This dataset contains harmful (CBRN) elicitation prompts and successful jailbreak responses. By requesting access you agree to use it solely for evaluating and improving model safety.
Log in or Sign Up to review the conditions and access this dataset content.
JBRL robustness eval
Robustness of defended chat models against our RL-trained red-teamer, on the full ClearHarm held-out set
(179 CBRN prompts). Two configs: sampled (victim temp 1.0, real-world) and greedy (victim temp 0.0).
Each: per-response rows + summary.json (per attacker x victim x judge: ASR / pass@k / attacker_lift).
Judges: HarmBench-13b, StrongREJECT (Qwen2.5-32B), Llama-Guard-3-8B, + ensemble. Attackers: JBRL-v1/v2/v3.
Authorized defensive research only; access-controlled.
- Downloads last month
- 10