castor-labels

Paper: Identically Trained Language Models Break on Different Jailbreaks (arXiv link to follow)

Per-response judge labels for every CASTOR model on 6,226 public benchmark prompts (20 samples and one greedy response each, sample = -1 is the greedy response). Prompts are referenced by benchmark id only; no prompt text and no model generations are included.

208 models. Columns: repo, revision, model_id, source, prompt_id, sample, coherent, coherence_rule, llm_invalid, p_valid, wg_label, wg_partial, wg_vulnerable, wg_harmful, wg_vulnerable_incl, q3g_label, q3g_partial, q3g_vulnerable, q3g_harmful, q3g_vulnerable_incl, refusal_agree.

One response was cut to fit the judges' context: the greedy response (sample = -1) of dd-dolma1_7-no-math-code-1b-large-aux-3-dpo-seed1 to wildjailbreak-00287. WildGuard returned no label for it (JUDGE-ERROR); Qwen3Guard labelled the cut response. Every other label is on the full response.

Safety

The models were safety-tuned on public data, and their bases are public. They can still produce harmful text when prompted, as the released labels show. Model outputs on harmful prompts are not released.

Citation

Chi-Wei Chang, Mu-Tien Kuo, Kuan-Wei Lee, Meng-Hsun Tsai

@misc{castor2026,
  title         = {Identically Trained Language Models Break on Different Jailbreaks},
  author        = {Chang, Chi-Wei and Kuo, Mu-Tien and Lee, Kuan-Wei and Tsai, Meng-Hsun},
  year          = {2026}
}