Example for how I may continue fine tuning the peft model

[therealadityashankar]

Hey!, I apologize if this is a rather generic question

I'm not able to find good examples on how I may continue training with peft over on the peft repository from a stored peft checkpoint,
and since the fine tuning code only shows how to fine tune from scratch, I'd be greatful if I could be given an example on how I may fine tune alpaca from the stored peft checkpoint instead of scratch

thanks, and I really appreciate all the work put into this project!

[tloen]

To load from a saved LoRA adapter, replace

model = get_peft_model(model, config)

with

from peft import PeftModel
model = PeftModel.from_pretrained(model, <PATH>)

To load from a training checkpoint, replace

trainer.train()

with

trainer.train(resume_from_checkpoint=<PATH>)

[zachNA2]

The Path leads to what? Is it the hf_ckpt converted folder? To the previous fine tuned alpaca-lora folder?

[roguh]

It could be a peft_config object: https://github.com/huggingface/peft/blob/main/src/peft/peft_model.py

[bui-thanh-lam]

To load from a training checkpoint, replace

trainer.train()

with

trainer.train(resume_from_checkpoint=<PATH>)

I tried resume training using "resume_from_checkpoint", but all the keys from saved model are mismatched.
Can you confirm that it works properly?

[edwardelric1202]

@bui-thanh-lam I met the same question, have you solved this?

[edwardelric1202]

@tloen Hi, if I use "trainer.train(resume_from_checkpoint=)", should I make any change on "training_args" and the initialization of Trainer?