InternVL icon indicating copy to clipboard operation
InternVL copied to clipboard
Published 3 months ago • verified publisher icon OpenGVLab

An issue with the extraction front

Open Dineshkumar-Anandan-ZS0367 opened this issue 4 months ago • 2 comments

Motivation

I’m extracting data using the internvl3 14B parameter model. I provided a prompt along with an image (300 DPI).

In one case, the image contains an ID key with the value 143J44363, but the model skips the alphabet ‘J’ and returns 143344363 instead.

How can we address this issue? Does anyone have suggestions or ideas?

Dineshkumar-Anandan-ZS0367 avatar Sep 01 '25 17:09 Dineshkumar-Anandan-ZS0367

Our config sets the max_dynamic_patch to 12 by default. You can set it to 24 if you want to improve its OCR ability.

Weiyun1025 avatar Sep 01 '25 17:09 Weiyun1025

  1. I’m using lmdeploy with OpenAI Chat Completions. If I change a parameter in the config.json file, will it take effect? In Transformers it does, but how about in lmdeploy?

  2. How can I improve extraction accuracy on faxed documents? Are there any other parameters I should be focusing on?

Dineshkumar-Anandan-ZS0367 avatar Sep 02 '25 12:09 Dineshkumar-Anandan-ZS0367