fine-tuning on docvqa ,anls only 40%

clovaai / donut

Official Implementation of OCR-free Document Understanding Transformer (Donut) and Synthetic Document Generator (SynthDoG), ECCV 2022

https://arxiv.org/abs/2111.15664

MIT License

5.74k stars 466 forks source link

fine-tuning on docvqa ,anls only 40% #265

Open ShuoZhang2003 opened 11 months ago

ShuoZhang2003 commented 11 months ago

I tried fine-tuning donut on docvqa, but got anls result of only 40%. I followed the code provided by the author exactly. Are there any problems that I haven't noticed? I would be very grateful if you could answer my question!

bswethav commented 8 months ago

@ShuoZhang2003 , i want to finetune donut on docvqa as well. is your usecase single question multiple answers ( OR single question single answer, kindly help