Question about fine-tuning recipe?

#3
by phanhoang - opened

Hi, thanks for your great model!

I have a question about fine-tuning process.
What are the values of the two parameters, lora_rank and lora_alpha, when fine-tuning this model?
And did you adjust the freeze_vision_tower = false parameter during fine-tuning?

phanhoang changed discussion status to closed
phanhoang changed discussion status to open
EraX JS Company org
edited Oct 9

freeze_vision_tower = false is courageous. Needs to trial & errors.

Large datasets likes 2-3 millions, then you should open vision, better quality.

LoRA rank wise, 128 should be a good one, it improves reasoning.

Cheers,
Nguyên

@thusinh1969 thank you for your response!

Sign up or log in to comment