dots ocr, result of some pages are wrong, some are correct.

[nissansz]

Example of wrong results,
left is result by chatllm.py, right is by dots ocr runner / llama-cpp.
Don't know why there is such problem, and how to resolve?

[foldl]

You can just use the one that gives the correct answer.

[foldl]

@nissansz can you share some pages that trigger wrong results?

[nissansz]

I think you can try any jpg page. Then compare with original model result.

Did you convert model for dotsocr1.5?
Maybe can try new model if you have.

[foldl]

fix a issue with chat template (5581533).

Tested dots.ocr-1.5 with demo_image1.jpg, the bbox results looks good.

[nissansz]

How to convert dots.ocr-1.5 model to bin or gguf file? Where to download?

[foldl]

dots.ocr-1.5 can be converted just like the elder version. Here is the quantized models.

Note, I am using --max-proj-tokens 8000 to make my GPU happy, but the image is resized.

[nissansz]

Thank you.
Where is the script to convert the model to bin?
Do you know how to convert it to gguf for llama-cpp?

[foldl]

Converting to gguf for llama.cpp is not supported.

[nissansz]

Ok. Where is script to convert to bin?

[foldl]

Here is the script: https://github.com/foldl/chatllm.cpp/blob/master/convert.py
Or, you can download quantized dots.ocr-1.5 from here.

[nissansz]

Thank you. Will try later.

[nissansz]

In llama-cpp, there is a model filr mmproj-Dots.Ocr-F16.gguf

Do you know use of this file?

[nissansz]

Do you know any free model for chatgpt3.5?

[foldl]

In llama-cpp, there is a model filr mmproj-Dots.Ocr-F16.gguf

Do you know use of this file?

You can ask for help from the community of llama.cpp.

[nissansz]

In llama-cpp, there is a model filr mmproj-Dots.Ocr-F16.gguf
Do you know use of this file?

You can ask for help from the community of llama.cpp.

dotsocr is not officially supported by llama-cpp. No help is available.