Access to multilingual jailbreak conversations
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains working jailbreak conversations and harmful model replies in English, Hindi, Telugu, Japanese and mixed languages. It is shared for safety research, evaluation and training of safer models only. By requesting access you agree not to use it to attack, harm or deceive people or to make models less safe, and not to redistribute it outside a gated setting.
Log in or Sign Up to review the conditions and access this dataset content.
Tiny Aya multilingual multi-turn jailbreaks
⚠️ Content warning. This dataset contains working jailbreak conversations and harmful model replies, including instructions for illegal or dangerous activity. Access is gated and approved automatically on agreement to the terms.
The attack side of kamikaze-san/tiny-aya-multilingual-safety (Lexsi Labs AI-safety hackathon). The text-free results and labels are public in kamikaze-san/tiny-aya-multilingual-safety-eval.
| Folder | Contents |
|---|---|
attacks/ |
370 held-out STRATA attack conversations plus 12 human-written ones (eval_attacks_combined.jsonl, n = 382), the held-out STRATA set with scores, and a 200-prompt Aya Red-teaming sample (en, hi) |
dpo/ |
Every DPO preference set: attack prompts with refusals (chosen) and compliant replies (rejected), and harmless look-alikes with helpful (chosen) and refusing (rejected) replies. joint_all_v3.jsonl (1,255 pairs) trains the final model. |
replies/<run>/ |
Every evaluated model's replies to the held-out attacks and benign prompts, with Qwen3Guard labels |
stage2_prompts/ |
Stage-2 prompt sets, including the 240 safety-anchor prompts drawn from the stage-1 training attacks |
judge_validation/ |
The 128-row labelling sheet with conversation text |
Attacks were generated by STRATA (an extension of AutoDAN-Turbo) with Qwen3.6-35B-A3B against the official Q4_K_M GGUF of CohereLabs/tiny-aya-global. Contact details and names that appear in generated replies are model output, often placeholders. Licences: our data is CC-BY-NC-4.0; the Aya Red-teaming sample keeps its Apache-2.0 licence.
- Downloads last month
- 134