Document Information Extraction and Grounding on CRC referral dataset (test)
96.1Reading Accuracy (Value Match)Qwen3-VL-8B (FT)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-VL-8B (FT)Regime=Fine-tuned, Access=Open-weight2026.05 | 96.1 | 60.6 | 75.4 | |
| Qwen3-VL-32BRegime=Zero-shot, Access=Open-weight2026.05 | 95.3 | 8 | 60.9 | |
| Claude Opus 4.6Regime=Zero-shot, Access=Closed-source2026.05 | 95.2 | 0.2 | 1 | |
| Gemini 2.5 FlashRegime=Zero-shot, Access=Closed-source2026.05 | 92.6 | 1.2 | 41.4 | |
| Qwen3-VL-4B (FT)Regime=Fine-tuned, Access=Open-weight2026.05 | 90.1 | 25.9 | 46.4 | |
| Qwen3-VL-8BRegime=Zero-shot, Access=Open-weight2026.05 | 88.4 | 6.2 | 37.2 | |
| Qwen3-VL-2B (FT)Regime=Fine-tuned, Access=Open-weight2026.05 | 88.3 | 21 | 32.9 | |
| Qwen3-VL-4BRegime=Zero-shot, Access=Open-weight2026.05 | 87.5 | 0.3 | 33.3 | |
| Qwen3-VL-2BRegime=Zero-shot, Access=Open-weight2026.05 | 62.3 | 0 | — |