baaivision / CapsFusion

[CVPR 2024] CapsFusion: Rethinking Image-Text Data at Scale
193 stars 5 forks source link

training costs for fine-tuning LLaMA (CapsFus-LLaMA) #6

Open YoojLee opened 6 months ago

YoojLee commented 6 months ago

Hi, thanks for such a great work! I would like to ask you of training costs for fine-tuning LLaMA2-13B on the caption fusion task. If possible, please let me know which gpu you have used and how many days (or hours) it costs!

yqy2001 commented 6 months ago

Thank you for your interest. The finetuning cost is about 1-2 days with 8 A800-80G gpus based on Alpaca's codebase, as only 2M samples are enough (2 epochs).

YoojLee commented 6 months ago

Thanks for quick reply!