Datasets:
| verified 251,408 (ambiguous dropped 37,615) -> 250,842 after dedup | |
| train 196,297 | |
| eval_held 6,061 | |
| eval_unseen_phrasing 8,723 | |
| eval_unseen_dialect 39,761 | |
| eval_hard 728 | |
| held-out families in train: none | |
| held-out variety in train : none | |
| --- audit --- | |
| class mix: {'safe': '17.8%', 'safe_sensitive': '11.8%', 'privacy_pii': '8.2%', 'specialized_advice': '8.0%', 'self_harm': '7.9%', 'prompt_injection': '7.6%', 'jailbreak': '7.5%', 'sexual_content': '7.4%', 'hate_harassment': '7.0%', 'violent_weapons': '6.0%', 'misinformation': '5.8%', 'nonviolent_crime': '5.0%'} | |
| length shortcut: majority 0.1785 | best threshold 0.1798 | leak +0.0013 | |
| variety mix: {'عربي مختلط بالإنجليزية': '17.1%', 'اللهجة المصرية': '16.8%', 'اللهجة الشامية': '16.7%', 'اللهجة الخليجية': '16.7%', 'اللهجة المغاربية': '16.4%', 'العربية الفصحى': '16.3%'} | |
| TOTAL: 250,842 | |