/PEFT-Chinese-Fact-Verification

Parameter-Efficient Fine-Tuning for Chinese Fact Checking

Primary LanguagePython

PEFT-Chinese-Fact-Verification

Claim Verification

  • P-Tuning
  • P-Tuning v2
  • LoRA (rank=4 is the best)

Template Engineering

  • We only define 10 soft tokens for peft

Preprocess

python preprocess.py

--dataset choose the dataset path you want to preprocess (default='datasets/unpreprocess/train.json')
--save_file save the preprocess file (default='datasets/preprocessed/train.json')

Basic Usage

python main.py

Arguments

--plm bert, bert-large, roberta, ernie, roberta-large (Use large model is the best)
--type ptuningv1, ptuningv2, lora
--train Remember to set this if you want to train.
--eval Remember to set this if you want to evaluate.
--train_file (default: 'datasets/preprocessed/train.json')
--valid_file (default: 'datasets/preprocessed/dev.json')
--test_file (default: 'datasets/preprocessed/test.json')

Checkpoint