Hyperparameters for research

#1
by stepchoi - opened

I am writing a paper on continuous learning that compares different ways to train TinyLllama on the Hermes dataset; continued training, fine tuning (Hermes-2-TinyLllama/TinyHermes), block expansion,...
I would like to cite your work on TinyHermes. Could you please let me know the hyperparameters used (alpha, LR,...)?
Appreciate it.

This is a crap finetune... I just did like 30000 epochs at 1e-05, MLX lora defaults. Thanks for your interest though :D!

Sign up or log in to comment