Hello Andriy,

I primarily utilize the Intel/orca_dpo_pairs dataset, a true classic in our field! I'm excited to share with you the complete notebook that I employ for training my models. Specifically, for the Krishna model, I leverage LoRA with r = 16, and set the learning rate to 5e-5. This setup has proven to be quite effective for my needs. Looking forward to your feedback!

The collab notebook (I use A100)

https://colab.research.google.com/drive/1NufAApJPWvdFt4d0-lHLVZhJ38DY3A3M?usp=sharing