Can I add other decoder models like google/madlad400-3b-mt

[aumcerebral]

Hi ,

Super thanks for the great engine.

Currently, translation is a big issue. So is there any way we could add google/madlad400-3b-mt . Its not an LLM but very powerful model.

[svonava]

Hi @aumcerebral - very interesting model, but quite an old one (2023) - has it been benchmarked vs gemma 4 or qwen 3.6 on your task? Wouldn't those models perform better? Or if the model size is the problem, maybe the qwen3.5 4B could compete with it well?

[aumcerebral]

I have already tried those models . For Translation (of over 90 languages) tasks, gemma4 and qwen3.5 4b are not even near to its accuracy. Also Gemma and qwen family of models are pretrained only on handful of languages. Madlad family is most superior for translation tasks.

[aumcerebral]

Similarly for https://huggingface.co/datalab-to/surya-ocr-2

This model is quite lightweight and accurate across 90+ languages.

[svonava]

Very interesting! Ok I'll add the madlad model to the model-adding-queue (we do a bunch of benchmarking when we tune the configs).

For OCR did you check the models already available? (paddle, miner, docling etc)

[aumcerebral]

Thanks for madlad.

For OCR, I have a full benchmarking across multiple languages.

Apart from lightonocr, glm-ocr, mineru, paddleocr-vl, docling there are more accurate available now that work fabulous both on main stream and low resource languages,
These are chandra, surya, Ovisocr2, qianfan ocr,

You can check this leaderboard. These 16 models are indeed the best among the ocr class.

https://huggingface.co/datasets/allenai/olmOCR-bench

[aumcerebral]

Hello any update on addition of these models?