TandA for answer sentence selection on covidQA

The pre-trained Ansq-RoBERTa Large model (transfer model) provided by wqa_tanda repo. The transformers directory and run_glue.py are adaptation and modification from wqa_tanda repo, the rest is author's code.

Description

This work is for advanced machine learnine course's final assignment. You can read all detail work's description (in bahasa) at technical-report

Requirement

• Python 3.7
• datasets 1.8.0
• boto3 1.17.97
• sacremoses 0.0.45
• sentencepiece-0.1.95
• nltk 3.2.5

Reproduce Step

• Clone This Repo
• Download ASNQ Transfer's Model From RoBERTa-Large ASNQ
• Training Adapt

!python3 run_glue.py \
    --model_type roberta \
    --model_name_or_path  [model-name-or-path] \
    --task_name ASNQ \
    --do_train \
    --do_eval \
    --evaluate_during_training \
    --do_lower_case \
    --data_dir [dataset_dir] \
    --per_gpu_train_batch_size 15 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir [ouput_dir] \
    --tensorboard_log_dir_loc=[tensor_dir] \

• Eval Model's Performance

!python3 run_glue.py \
    --model_type roberta \
    --model_name_or_path  [model-name-or-path] \
    --task_name ASNQ \
    --do_test \
    --do_eval \
    --evaluate_during_training \
    --do_lower_case \
    --data_dir [dataset_dir] \
    --per_gpu_train_batch_size 15 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir [ouput_dir] \
    --tensorboard_log_dir_loc=[tensor_dir] \

Note: --do_test is argument for using test data in evaluation process.

Alternatif reproduce way

In this repo also provided sample notebook (Example Notebook) to run the program. You can change the parameters as you wish

Help

You can submit a GitHub issue for asking a question or help. Or you can contact me directly at [email protected] as well