Week2: training args for offline models

Hi all,

I was wondering if in the week2 lab, the same training parameters was used to train the offline models (full fine tune and peft). If not would like to ask the organizers to share the training aArgs used for the offline models. Thanks in advance!

Hi @r_j , most probably the hyperparameters were the same, unless one specific training would have required a different value in any hyperparameter. Can you share more where this question is coming from? because I think that the main purpose of the topic was to present both training methods and how they produce so similar results even though the scope of the training is so different.

Hi @Juan_Olano
The question is, What were the parameter for training the model which we loaded using aws cli.

In the studio env, it was only 1 epoch and max_steps=1. How do we reproduce the same instruct model in full finetuning ?

Ah, got it. I don’t have that information, but if I were to run the whole process (which according to the instructor it took several hours) I would probably start with some 20 epochs and based on metrics I would go up or down. I’ll see if I can find the params used during the full training, but most probably whatever you do with your own data will need something different and you will have to experiment with different values until you get your ideal result.

I agree @Juan_Olano , have been experimenting with it to reproduce the same results. Increasing the learning rate to 5e-5 is blunder (overfitting badly).

Training for 5 epochs makes more sense qualitatively but gain in rouge is not large as shown during Lab (aws model).

PEFT LORA tuning had large margins in gains (as shown in lab) with epoch=3 (keeping other hyperparams constant).

Thanks for sharing these results @Abhijeet_Kumar1 !!! And yes, in this world it is about experimentation.