DocVQA (val)
Multimodal · 2020-07-01
DocVQA asks extractive questions about scanned industry documents; the validation split has 5,349 questions on 1,286 page images. Answers are scored by ANLS (normalized Levenshtein similarity, zeroed below 0.5) averaged over questions and stored here on a 0-100 scale (the official 0-1 value times 100).
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-4B | 97.1 |
| Gemma 4 12B IT | 97.0 |
| LFM3-VL-2B | 96.7 |
| Gemini 3.5 Flash-Lite | 96.2 |
| Qwen3.5-2B | 95.6 |
| GPT-5.6 Luna | 94.8 |
| LFM2.5-VL-3B | 91.1 |
| Gemma 4 E2B IT | 90.4 |