OCR with Handwriting is really bad

#1
by dxjheroRon - opened

Why the OCR result for a handwriting image is really bad compared to the test with this model on the Qwen website? Is there some special prompt or setting I should do ?

MLX Community org

Could you elaborate?

I give my handwritten Chinese image to it. On Qwen Website, I choose this 32B Vl model and it gives a very precise OCR result. But when I try it locally with 4bit, even 8 bit version, it will get the first several words and then starts makeing it up with the rest, although you can tell at some point, the hallucination result has a little bit relation with the image, but not much. At some points it even go circles to keep makeing up by itsself.

For this 32B VL, I use LM studio locally with default setting , it gives the hallucination all the times, I try to modified the temperature, but didn't help.

I try locally using a Qwen vl 7B fully directly with the code they offer on github, not with LM studio, it also can give pretty good result, although have a little misktakes, but still pretty good.

So I am not quite sure where is the problem, the quantied version? or the LM studio setting? Naively thinking, the 32B should be better then the 7B version at least. But I do run 7B fully, but 32B with 8bit quant.

Your need to confirm your account before you can post a new comment.

Sign up or log in to comment