Receipt Information Extraction on ReceiptBench 1.0 (test)
79.5Overall F1Qwen3-VL-8B (SFT+GRPO)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Qwen3-VL-8B (SFT+GRPO)Category=Ours (Fine-tuned), Size=8B, Fine-tuning protocol=SFT+GRPO2026.05 | 79.5 | 84.88 | 94.16 | 85.47 | 63.73 | |
| Qwen3-VL-4B (SFT+GRPO)Category=Ours (Fine-tuned), Size=4B, Fine-tuning protocol=SFT+GRPO2026.05 | 77.88 | 82.26 | 92.98 | 84.17 | 62.15 | |
| Qwen3-VL-8B (SFT)Category=Ours (Fine-tuned), Size=8B, Fine-tuning protocol=SFT2026.05 | 77.36 | 81.55 | 91.8 | 82.73 | 64.62 | |
| Qwen3-VL-4B (SFT)Category=Ours (Fine-tuned), Size=4B, Fine-tuning protocol=SFT2026.05 | 77.23 | 80.03 | 89.64 | 81.84 | 64.78 | |
| Gemini-3-ProCategory=Proprietary & API-based2026.05 | 73.73 | 73.6 | 90.86 | 87.14 | 57.81 | |
| Qwen3-VL-PlusCategory=Proprietary & API-based2026.05 | 72.1 | 73.06 | 90 | 87.87 | 54.84 | |
| GPT-5Category=Proprietary & API-based2026.05 | 70.76 | 73.04 | 87.43 | 87.06 | 48.93 | |
| Qwen3-VL-32BCategory=General Open, Size=32B2026.05 | 67.51 | 66.45 | 83.1 | 86.9 | 48.52 | |
| InternVL3.5-241BCategory=Proprietary & API-based, Size=241B2026.05 | 67.42 | 68.53 | 87.91 | 80.24 | 51.12 | |
| Qwen3-VL-8BCategory=General Open, Size=8B2026.05 | 65.45 | 66.64 | 85.77 | 79.36 | 47.92 | |
| InternVL3-2B (SFT)Category=Ours (Fine-tuned), Size=2B, Fine-tuning protocol=SFT2026.05 | 64.96 | 70.77 | 85.14 | 73.14 | 45.09 | |
| InternVL3-78BCategory=General Open, Size=78B2026.05 | 64.43 | 64.86 | 84.44 | 78.74 | 46.3 | |
| PaddleOCR-VLCategory=Specialized, Size=0.9B2026.05 | 63.44 | 61.33 | 81.25 | 80.67 | 47.46 | |
| Qwen3-VL-4BCategory=General Open, Size=4B2026.05 | 62.61 | 64.07 | 75.03 | 75.88 | 49.09 | |
| olmOCR-7BCategory=Specialized, Size=7B2026.05 | 62.28 | 57.36 | 80.3 | 78.54 | 48.49 | |
| InternVL3-8BCategory=General Open, Size=8B2026.05 | 57.72 | 56.96 | 76.61 | 73.71 | 46.09 | |
| DianJin-OCR-R1Category=Specialized, Size=7B2026.05 | 49.79 | 50.38 | 71.26 | 59.92 | 29.82 | |
| InternVL3-2B (SFT+GRPO)Category=Ours (Fine-tuned), Size=2B, Fine-tuning protocol=SFT+GRPO2026.05 | 44.66 | 62.79 | 46.96 | 39.54 | 29.95 | |
| DeepSeek-OCR-smallCategory=Specialized, Size=3B2026.05 | 39.59 | 36.88 | 50.81 | 55.17 | 36.37 | |
| InternVL3-2BCategory=General Open, Size=2B2026.05 | 38.07 | 38.51 | 58.21 | 46.37 | 30.77 |