Discussion about this post

User's avatar
Nan Deng's avatar

The overall board has Datalab at 88.2, DeepSeek-OCR at 77.5, and PaddleOCR-VL at 73.4, and that ranking flips on tables, where DeepSeek-OCR leads at 87.2 against Datalab at 78.9. Handwriting is the other split: Datalab passed 5 of 6 tests and the other two passed 0. A single average would have shipped Datalab to a team whose pages are dense tables. I would keep the page type on each yes/no check, and pick the parser from the failure reasons on that type, not from the column that averages a tax form with a handwritten note.

Ready for more?