This is the implement of the paper, CASR: Generating Complicated Sequences with Autoregressive Self-Boost Refinement.
The overview logic is in utils.cascade_trainer.CascadeSeq2SeqTrainer.train_all
build and run the docker image with Dockerfile_sing for fine-tuning and Dockerfile_ada for adapter-tuning
set the MASTER_PORT and TASK environment variable:
export MASTER_PORT=12345
export TASK=webqsp (or mtop, kvret)
To train finetuning+sepenc+continue, run:
bash script/ft-t5-con/run_sing.sh
To train continue on Sudoku, run:
BART_SIZE=base bash script/trainer_sudoku.sh
To train CASR-Large, run:
bash script/trainer_large.sh
bash baselines/inat/train.sh # INAT
bash baselines/levenshtein/train.sh # Levenshtein
bash baselines/bidirectional/train.sh # Bidirectional
python -m baselines.progressive.prepare_vocab && bash baselines/progressive/train.sh # Progressive
To try CASR on ChatGPT, run:
TASK=<your task> python empirical/chatgpt-test.py
Scripts of empirical studies are in this directory empirical
More cases are in this directory cases