Support for the new 450 language translation models from Google T5X "madlad" - apparently Apache-2

[cmp-nct]

Example: https://huggingface.co/jbochi/madlad400-3b-mt/tree/main
In Googles own space: https://huggingface.co/google/madlad400-10b-mt

The guy converted the format of the 3 smallest models (3b,7b,10b) to HF transformers. Given the severe lack in non english output a good translation model would be a gift.
I just tried the CPU demo of the 3B, it produced quite good output, if that gets better with 7B+ it would be a real solution for a huge amount of people.
It could be added as a 2nd stage into llama.cpp

Though the architecture is "T5ForConditionalGeneration" which isn't supported.

So far there was no urgent reason to add those T5 models, they did not stick out as special but the idea to output text in every single language worldwide .. that would be remarkable

[JettScythe]

@cmp-nct can you please link to the PR where this was completed?

[cmp-nct]

@cmp-nct can you please link to the PR where this was completed?

I’ve closed it as not important. Given the many things we need done and the almost zero interest for translation here.
The model from Google was ok but it was quite flawed when testing it in depth.

[vasicvuk]

This would be great if supported +1. @cmp-nct maybe you can reopen the issue so the support for this models is planned

[sorasoras]

That's interesting to me at least.

[easp]

Perhaps relevant: #5763

[MathiasSchindler]

@cmp-nct can you please link to the PR where this was completed?

I’ve closed it as not important. Given the many things we need done and the almost zero interest for translation here. The model from Google was ok but it was quite flawed when testing it in depth.

madlad400 is remarkable in the sense that it has the most permissive license in comparison to meta's nllb200 or seamless communication. The translation quality varies a lot between the 419 listed languages, with some language combinations producing a very decent translation quality. Some language pairs seem to be heavily influenced by the datasets from the European Parliament (with sometimes cute and hilarius translation mistakes coming from the European Parliament debates). I have heard from other users that some languages in the Indic family and Chinese family should not be used at the moment.

I have a huge interest if someone could support it.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.

[Nikola-Milovic]

Surprising lack of interest for translational models, any reasons why?