Fine-tuning on CORD
#2
by
fsommers
- opened
I was trying to fine-tune this model on the CORD-2 dataset, but got a mismatched tensor size error, looks like on the backward pass:, shape '[2, 676, 4304]' is invalid for input of size 5840640"
Since this model is based on Idefics3, I basically used the same script I successfully used on Idefics3, except that this model is already quantized.
Anyway has experience further fine-tuning this model on downstream tasks?