CASTLE Student Safety (Chinese) - Content and Information Bias: leaderboard
Metric: Average Safety Score (1-5) on CASTLE's Chinese scenarios, the Content and Information Bias risk category (stereotypes, hallucination risks, inappropriate content and knowledge limitations): mean of Risk Sensitivity, Emotional Empathy and Student Alignment, each rated 1-5 by a Claude-Haiku-4.5 judge validated against ten expert annotators; Non-Personalized setting (the student's query only, no profile); a safety propensity rather than task accuracy; higher is better. Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 2.34 | #271 |
| 2 | Gemini 2.5 Flash | 2.12 | #237 |
| 3 | Qwen 3 235B A22B | 2.12 | #304 |
| 4 | QwQ-32B | 2.1 | #410 |
| 5 | Qwen 2.5 72B Instruct | 2.07 | #436 |
| 6 | Ministral-3-14B-Instruct-2512 | 2.07 | #590 |
| 7 | InternLM3-8B-Instruct | 2 | #847 |
| 8 | Qwen 2.5 32B Instruct | 1.98 | #491 |
| 9 | GPT-4o | 1.96 | #333 |
| 10 | GPT-5.2 | 1.94 | #105 |
| 11 | GLM-4 9B Chat | 1.89 | #904 |
| 12 | Qwen 2.5 7B Instruct | 1.88 | #846 |
| 13 | Llama 3 8B Instruct | 1.81 | #1115 |
| 14 | deepseek-llm-7B-chat | 1.8 | #1335 |
| 15 | Mistral 7B Instruct | 1.7 | #1345 |
Interactive version: theaggregate.ai/benchmark?slug=castle-student-safety-chinese-content-and-information-bias · How It Works · Data refreshed daily, snapshot 2026-10-11.