| license: other | |
| language: [en] | |
| tags: [red-teaming, jailbreak, robustness, defensive-security, CBRN] | |
| configs: | |
| - config_name: sampled | |
| data_files: sampled/responses.jsonl | |
| - config_name: greedy | |
| data_files: greedy/responses.jsonl | |
| extra_gated_prompt: >- | |
| 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. | |
| # 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–v5 | |
| (v1–v4 single-turn, v5 multi-turn) — identical 7-victim panel across all versions for comparability. | |
| **Authorized defensive research only; access-controlled.** | |