PEFT fine-tuning on Flan-t5-base model does not change inference results

[malsadi]

Hello All,

I followed the same guidance on how to fine-tune a base model for summarisation task using LoRA technique. I applied the same code but the the final results have not changed. I’m not sure why this is happened, given that the model, dataset and LoRA configurations are the same. The only difference is that I have all steps executed on my local environment. I have not used the files from checkpoints.
The only thing which I don’t feel correct is that after training the peft_model on the tokenized assets the logs shows learning_rate is 0.

#load the foundation model from disk
model_local_path = 'C:\\Users\\ms\\.cache\\huggingface\\hub\\models--google--flan-t5-base\\snapshots\\7bcac572ce56db69c1ea7c8af255c5d7c9672fc2\\'
foundation_model = AutoModelForSeq2SeqLM.from_pretrained(model_local_path, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_local_path)

[gent.spah]

I am not sure if there are other files needed to run the labs on your machine but may I say that these to functions/methods:

foundation_model = AutoModelForSeq2SeqLM.from_pretrained(model_local_path, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_local_path)

have actually more parameters that you might need to set.

Search about the in google and try to see if any of them has a learning rate parameter!

[malsadi]

Many thanks for your reply,
I’m not sure what exactly what you are referring to. However, the model ( all the files listed in the attached picture ) was successfully uploaded to my disk and then I was able to load it and start interacting with it ( asking questions and getting answers )