Local PDF extraction model improves on messy tables and charts
A model has been improved for difficult PDFs with multi-column layouts, merged table cells, charts, and partly scanned pages. The base VLM scored about 46% on , then reached 91.1% after LoRA training and several structure changes. On , the score moved from 46% to 79%, described as second-place level.
Tables were the main improvement area because they are often where PDF breaks down. The model runs on a user’s own hardware, so documents do not need to be sent to an outside API. It is described as strong on charts and tables, with near-zero , meaning it does not add rows or numbers that are not in the source.
The project is now looking for hard PDFs to test where it fails.
Key points
- A messy-PDF model is claimed to improve from 46% to 91.1% on .
- On , the claimed score rises from 46% to 79%.
- The biggest improvement is table , especially avoiding made-up rows or numbers.
- The model runs on local hardware instead of sending documents to an outside API.
- The project is seeking difficult PDFs to test its weak spots.