BART fine-tuning

• train.py : train and valid
• inference.py : eval and make summarization on test dataset
• summarizer folder
  • dataloader.py : dataloader of xsum and cnn/dm dataset
  • default.py : summarization model
  • scheduler.py : learning rate scheduler
  • utils.py : logger function

Following directories should be created

• ./outputs : store model checkpoints

• XSum using https://github.com/EdinburghNLP/XSum dataset or download preprocessed data

I run codes on NVIDIA GeForce RTX 3090. It takes 10 ~ 13 hours for 1 epoch.

python train.py --dataset-path [xsum or cnn/dm dataset] --output-dir [output dir path] --epochs 10 --max-learning-rate 2e-3 --batch-size 4 --valid-batch-size 8

python inference.py --pretrained-ckpt-path [fine-tuned model path] --dataset-path [xsum or cnn/dm dataset] --output-path [saved result path]