Fine tune

[Le-Ngoc-Tu]

Can models fine tuned for a new language run, guy?

[soc8ren]

Is there a training script to finetune for English already? The Readme says:-

⚠️ Quality and voices vary per generation, as the model is not fine-tuned on a specific voice. Use with prefix or fine-tune in order to obtain stable output.

But, I couldn't find any further info on it.

[thomasgauthier]

The readme also says

We provide model checkpoints (1B, 2B) and inference code to accelerate research.

Inference code is indeed all there (🙏) but aside from recommending fine-tuning there are no mentions of training code release. Is a release planned at any point?

[info-wordcab]

@JakubCzarlinski Can you please point us to the fine-tune script?

[JakubCzarlinski]

@JakubCzarlinski Can you please point us to the fine-tune script?

I am not affiliated with Nari Labs, I made contributions because the work is personally of interest to me. Please do not ping me.

Regardless, a fine-tune would be no different from any other pre-training/fine-tune loop for any other model...

[info-wordcab]

@JakubCzarlinski Can you please point us to the fine-tune script?

I am not affiliated with Nari Labs, I made contributions because the work is personally of interest to me. Please do not ping me.

Regardless, a fine-tune would be no different from any other pre-training/fine-tune loop for any other model...

Oops, I just looked at the last contributor and message.

[utkarshshukla2912]

Plus +1000000

[matheusfalcaopinto]

Maybe it uses the same finetune workflow of Dia1?

[TuananhCR]

looking forward too

[TuananhCR]

Maybe it uses the same finetune workflow of Dia1?

So you meant we can use same structure of Dia 1 apply finetuning Dia 2?

[matheusfalcaopinto]

Maybe it uses the same finetune workflow of Dia1?

So you meant we can use same structure of Dia 1 apply finetuning Dia 2?

No, I'm wondering about that but on discord someone sad that will not work, different models etc etc

[Jpisnice]

Fine tune recepie please