Google released T5X checkpoints for MADLAD-400 a couple of months ago, but nobody could figure out how to run them. Turns out the vocabulary was wrong, but they uploaded the correct one last week.
I’ve converted the models to the safetensors format, and I created this space if you want to try the smaller model.
I also published quantized GGUF weights you can use with candle. It decodes at ~15tokens/s on a M2 Mac.
It seems that NLLB is the most popular machine translation model right now, but the license only allows non commercial usage. MADLAD-400 is CC BY 4.0.
Thanks a lot for converting and quantizing these. I have a couple of questions.
How does it compare to ALMA? (13B)
Is it capable of translating more than 1 sentence at a time?
Is there a way to specify source language or does it always detect it on its own?
Thanks!
- I’m not familiar with ALMA, but it seems to be similar to MADLAD-400. Both are smaller than NLLB-54B, but competitive with it. Because ALMA is a LLM and not a seq2seq model with cross-encoding, I’d guess it’s faster.
- You can translate up to 128 tokens at the time.
- You can only specify the target language, not the source language.