README.md

Data format

For local files, the finetuning script supports only json as input file format. The input file structure should be the same as standard QA datasets like SQuAD v2.0.

To view sample input files, see the files here.

Training & Evaluation

To see list of all available options, do python run_question_answering.py -h. There are three ways to provide input data files to the script:

• with flag --dataset_dir <path> where <path> points to the directory containing files with prefix train, validation and test.
• with flags --train_file <path> / --train_file <path> / --validation_file <path> / --test_file <path>.
• with a dataset from Huggingface Datasets Library, usng the keys --dataset_name <name> and dataset_config_name <name> (optional)

For the following commands, we are going to use the --dataset_dir <path> to provide input files.

Finetuning

For finetuning and inference on the test set using the best model during validation (on single GPU), a minimal example is as follows:

$ python ./run_question_answering.py \
    --model_name_or_path "csebuetnlp/banglat5" \
    --dataset_dir "sample_inputs/" \
    --output_dir "outputs/" \
    --learning_rate=5e-4 \
    --warmup_ratio 0.1 \
    --label_smoothing_factor 0.1 \
    --gradient_accumulation_steps 4 \
    --weight_decay 0.1 \
    --lr_scheduler_type "linear"  \
    --per_device_train_batch_size=8 \
    --per_device_eval_batch_size=8 \
    --max_source_length 512 \
    --max_target_length 30 \
    --logging_strategy "epoch" \
    --save_strategy "epoch" \
    --evaluation_strategy "epoch" \
    --greater_is_better true --load_best_model_at_end \
    --metric_for_best_model f1 \
    --num_train_epochs=5 \
    --do_train --do_eval --do_predict \
    --predict_with_generate

Evaluation

• To calculate metrics on test set / inference on raw data, use the following snippet:

$ python ./run_question_answering.py \
    --model_name_or_path <path/to/trained/model> \
    --dataset_dir "sample_inputs/" \
    --output_dir "outputs/" \
    --per_device_eval_batch_size=8 \
    --overwrite_output_dir \
    --do_predict --predict_with_generate