PIJ - Criminal Profiling: leaderboard
F1 for criminal profiling on the PIJ human-comparison subset of 50 samples per task, including Unknown labels. These are the human-comparison results rather than the full 2,500-case headline evaluation.
Metric: F1 (%, 50-sample human evaluation, include Unknown). Source: arxiv.org. Saturation forecast: Around 2033. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro (Non-reasoning) | 51.38 | #112 (DeepSeek V4 Pro) |
| 2 | Gemini 3 Flash | 48.5 | #103 |
| 3 | GPT-5.4 Mini | 47.63 | #228 |
| 4 | Qwen 3 Max | 47.2 | #213 |
| 5 | Llama 3.3 70B | 45.46 | #578 |
| 6 | GPT-5.4 | 44.72 | #83 |
| 7 | Claude Sonnet 4.6 | 42.39 | #94 |
| 8 | DeepSeek V4 Pro (Reasoning) | 42.19 | #112 (DeepSeek V4 Pro) |
| 9 | Qwen 3 14B | 40.9 | #542 |
| 10 | Qwen 3 8B | 37.3 | #678 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pij-criminal-profiling · How It Works · Data refreshed daily, snapshot 2026-11-05.