Main Script: QuestionAnswering.py
The script uses HuggingFace library for managing the datasets, importing/exporting models and training the models.
There are various variables at the start of the script.
train: Training a new model
PEFT: Whether to use PEFT during training
tf32/fp16: Mixed precision training choice
trained_model: Name of trained model (to be pushed to HF Hub)
train_checkpoint: Checkpoint of training (None by default)
squad_shift: Whether to include extra data (squadshift)
base_tokenizer: Tokenizer of base model
base_model: Pre-trained model
test: Testing a model
tokenizer_list/model_list/question_list: Which tokenizer, model and questions to be tested.
CUDA is enabled if applicable. Require user to login into HuggingFace Hub (via command line token or through script) if training. Alternative is to not push to hub, a local repository will be created.
Huggingface repositories created (models created)
- botcon/XLNET_squad_finetuned_large
- botcon/XLNET_squadshift_finetuned_large
- botcon/LUKE_squad_finetuned_large
- botcon/LUKE_squadshift_finetuned_large
- botcon/LUKE_squad_what