KIDBench (No Cue, Urdu): leaderboard
Metric: Total score (1-5): mean of the five child-safety dimensions, each rated 1-5 by a DeepSeek-V4-Pro judge with the KIDBench rubric and category-specific rules, temperature 0 for responder and judge, on the 500 no-cue KIDBench prompts translated into Urdu by GPT-5.5 (system instruction sets the response language), averaged over the 10 categories, single-turn; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemma 4 31B (IT) | 3.34 |
| 2 | Qwen 3.6 27B | 3.27 |
| 3 | DeepSeek V4 Flash | 3.26 |
| 4 | Gemini 3.1 Flash Lite | 3.25 |
| 5 | Claude Haiku 4.5 | 3.18 |
| 6 | Gemma 3 12B | 3.02 |
| 7 | GPT-5 Mini | 2.9 |
| 8 | Qwen 3.5 4B | 2.7 |
| 9 | Llama 3.3 70B Instruct | 2.63 |
| 10 | Gemma 3 4B | 2.33 |
| 11 | Qwen 3 8B | 2.23 |
| 12 | Llama 3.1 8B Instruct | 1.61 |
Interactive version: theaggregate.ai/benchmark?slug=kidbench-no-cue-urdu · How It Works · Data refreshed daily, snapshot 2026-10-07.