Failing to achieve 2 folds improvement of the peft lora example in week2 lab

Course Q&A Generative AI with Large Language Models
1

Hello,
I am trying to repeat the lab 2 fine-tuning exercise using peft approach on my own Mac mini. I am using flan-T5-small so that the model can fit in the 16G memory. I increased max_step to 10000 and I increased no of epochs to 5. The training finished after 24 hours. However, I only observed 0% - 3% improvement of the rouge scores by the peft model against the original model. Could anybody tell me what I can do to obtain the similar level of improvement as demonstrated in the Lab 2 vedio ?

Here are my peft configurations and training configurations:

lora_config = LoraConfig(r=32, lora_alpha=32, target_modules=[‘q’, ‘v’, ‘k’, ‘o’],
lora_dropout=0.05, bias=‘lora_only’,task_type=TaskType.SEQ_2_SEQ_LM)

peft_training_args=TrainingArguments(
output_dir=output_dir,
learning_rate=1e-4, # Better LR for small model
num_train_epochs=5, # Proper epoch-based training
max_steps=10000,
logging_steps=100,
report_to=“tensorboard”,
)

Thanks a lot!
Quan

When creating a post, please add:

  • Week # must be added in the tags option of the post.
  • Link to the classroom item you are referring to:
  • Description (include relevant info but please do not post solution code or your entire notebook):
1 Like
2

Hello,

Could you please tell me exactly what improvements you are aiming for and why you need them? 3% improvement is normal not perfect, but still decent. If you are looking for significant improvements on your personal Mac, then you would need different models and a larger number of parameters.

1 Like
3

Thanks Igor for the reply! For learning purpose, I am trying to repeat the 2 folds improvements demonstrated in the lab 2 Vedio. Can you please tell me what I should do ? (use larger model? use more training dataset ? )

1 Like