The pasta from ruc-datalab

pasta's People

Contributors

Stargazers

Watchers

pasta's Issues

a problem about loading pasta checkpoints

Hello Author，
Thank you very much for your excellent work. But we encountered a problem when loading the pasta checkpoint as described in the readme. The problem always seems to be in the following sections：

PASTA/src/utils/dataset.py

Line 16 in 2f5e910

self.tokenizer = DebertaV2Tokenizer.from_pretrained(data_args.tokenizer_path)

Looking forward to your recovery！

Fine-tuning on semtabfacts

Hello! I tried to fine-tune on the tabfact and semtabfacts based on the pretrained PASTA. The tabfact result can be reproduced successfully, while I met some problems on semtabfacts.

I tried to load the checkpoint you provide (https://drive.google.com/file/d/1cUq1G9Mx_gNwmNsptBEnRaVor8c70sE2/view?usp=sharing), but got an error "AttributeError: Can't get attribute 'ContextPooler' on <module 'main' from 'src/run_finetune.py'>
".
Then I tried to use fine-tuned PASTA to initialize the training, but the accuracy on semtabfacts test set is only 0.77 or so. Initializing with trained model on the TabFact I've finetuned got similar result. The training parameters are same with those in your script files.

So how to reproduce the result on semtabfacts? Could you give me some suggestions?

Recommend Projects