Grounded OCR on OCR-IDL, TabMe++, and PubMed-OCR (10.5K held-out pages)
0.202CER (text)GutenOCR-7B
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GutenOCR-7BParams=7B, Backbone=Qwen2.5-VL, Training Stage=3a, Coord. boxes=Yes2026.01 | 0.202 | 0.28 | 0.147 | 0.129 | 0.787 | 0.882 | 0.819 | |
| GutenOCR-3BParams=3B, Backbone=Qwen2.5-VL, Training Stage=3a, Coord. boxes=Yes2026.01 | 0.224 | 0.293 | 0.185 | 0.109 | 0.798 | 0.877 | 0.811 | |
| Qwen3-VL-8BParams=8B, Backbone=N/A, Coord. boxes=Yes2026.01 | 0.285 | 0.491 | 0.306 | 0.689 | 0.049 | 0.025 | 0.384 | |
| Infinity-Parser-7BParams=7B, Backbone=Qwen2.5-VL-7B, Coord. boxes=No2026.01 | 0.31 | 0.586 | 0.386 | 0.613 | 0.098 | 0.112 | 0.386 | |
| Qwen2.5-VL-7BParams=7B, Backbone=N/A, Coord. boxes=Yes2026.01 | 0.333 | 0.522 | 0.633 | 0.53 | 0.111 | 0.285 | 0.396 | |
| olmOCR 2Params=7B, Backbone=Qwen2.5-VL-7B, Coord. boxes=No2026.01 | 0.365 | 0.53 | 0.515 | 0.745 | 0.039 | 0.024 | 0.318 | |
| Nanonets OCR2-3BParams=3B, Backbone=Qwen2.5-VL-3B, Coord. boxes=No2026.01 | 0.369 | 0.518 | 0.645 | 0.807 | 0.001 | 0.001 | 0.277 | |
| Chandra OCRParams=8B, Backbone=Qwen3-VL-8B, Coord. boxes=Yes2026.01 | 0.39 | 0.593 | 0.867 | 0.721 | 0.012 | 0.005 | 0.241 | |
| PaddleOCR-VLParams=0.9B, Backbone=PaddleOCR-VL-0.9B, Coord. boxes=No2026.01 | 0.399 | 0.539 | 1 | 0.936 | 0.001 | 0.002 | 0.188 | |
| Qwen2.5-VL-3BParams=3B, Backbone=N/A, Coord. boxes=Yes2026.01 | 0.508 | 0.584 | 0.379 | 0.699 | 0.135 | 0.121 | 0.348 | |
| DeepSeek-OCRParams=3B, Backbone=DeepSeek-3B-MoE, Coord. boxes=Yes2026.01 | 0.702 | 0.809 | 0.998 | 0.94 | 0.01 | 0.003 | 0.094 |