Fine-tuning

[alaaj27]

Is it possible to Fine-tune LLaMA for downstream tasks? If so, how can we do that?

Edit: Reading the other opened issues, I realized that neither the training data nor the pre-trained weights were released. How the code is going to be useful anyway?

[archwolf118]

I have same question. Thank you for reply.

[shangdibufashi]

+1

[gsanou]

+1

[tweedmann]

+1

[liaoxingjian]

+1

[aohenuo]

+1

[chembiome]

+1

[comtomru]

+1

[turbo]

There's an RLHF harness here: https://github.com/nebuly-ai/nebullvm/tree/main/apps/accelerate/chatllama

Instruction fine tuning should also be pretty easy.

[mvllwong]

++

[sequoiar]

+1

[kev216]

+1

[ruian0]

• 1

[Chenzongchao]

+1

[secain]

+1

[zanghu]

+1

[wallon-ai]

+1

[hujunchao]

+1

[ruian0]

Found this work from Stanford Alpaca. They read and resave the model(so not dependent on FAIR) and it becomes trainable with the hugging face training framework.

https://crfm.stanford.edu/2023/03/13/alpaca.html
huggingface/transformers#21955

[dneupan1]

+1

[subramen]

We recently shared scripts for finetuning and inference at https://github.com/facebookresearch/llama-recipes