CIGAR: Contrastive learning for GHA Recommendation

This repository contains the code and the dataset of our paper.

• CIGAR contains the source code of our model
• CIGAR_variants contains the variants of our model (i.e. RoBERTa w/ or w/o finetune)
• data folder contains our dataset, along with the notebook for building the dataset.
• plots folder contains plots for data visualization.

To train and save our model, run the run.py file with the following commands:

python run.py \
    --output_dir=./saved_models/CIGAR_t50 \
    --model_type=roberta \
    --tokenizer_name=roberta-base \
    --model_name_or_path=roberta-base \
    --do_train \
    --train_data_file=../../data/train.csv.gz  \
    --eval_data_file=../../data/valid.csv.gz  \
    --test_data_file=../../data/test.csv.gz  \
    --epoch 10 \
    --block_size 128 \
    --train_batch_size 16 \
    --eval_batch_size 16 \
    --learning_rate 5e-5 \
    --max_grad_norm 1.0 \
    --evaluate_during_training \
    --seed 123456 2>&1| tee train_t50.log

• results folder contains the predictions provided by our model with various $\tau$ selected. The structure of the results is as follows:

  name	actual_action	suggested_actions
  upload source artifact	0	[0, 22, 13, 51, 88]
  ...	...	...
  select latest xcode	99	[99, 36, 37, 76, 14]