The official repository for paper: Better Language Models of Code Through Self Improvement

Findings of The 61st Annual Meeting of the Association for Computaional Linguistics (ACL 2023)

In this paper, we propose a simple augmentation technique to improve performance of code language models on sequence generation tasks. In particular, after fine-tuning a pre-trained language model on a specific sequence generation task, we

• Use this fine-tuned model to create an augmented version of training data and then
• Continue to fine-tune this model on the augmented dataset.

The overall training pipeline as well as process of data augmentation are depicted by the below figures.

Please refer to our paper for detailed results. In summary, our experiments showed that this method, when applied to popular pre-trained code models (CodeBERT, CodeT5, and UniXCoder), significantly improves performance on code summarization and code generation tasks.

conda create -n code-self-improve -y
conda activate code-self-improve
conda install pip -y
pip install -r requirements.txt

Our method was demonstrated on code-summarization and code-generation data from CodeXGLUE benchmark. To download data dependencies, run

cd data/
./download.sh

We give example script for reproducing our paper reported results. We take UniXCoder on code summarization as our demonstration.

First, navigate to the UniXCoder directory by

cd UniXCoder/code-summarization

python generate_augmented_data.py \
    --model_name_or_path microsoft/unixcoder-base \
    --train_filename ./data/codesearchnet/python/train.jsonl \
    --output_dir ./augmeted_data \
    --max_soucre_length 256 \
    --max_target_length 128 \
    --beam_size 10 \
    --eval_batch_size 64

The augmented data will be saved into ./augmented_data/python/pseudo_data.jsonl by running this script

Passing the path to the augmented data derived by the above script, i.e ./augmented_data/python/pseudo_data.jsonl to the flag --train_filename

python run.py \
    --do_train \
    --do_eval \
    --do_test \
    --model_name_or_path microsoft/unixcoder-base \
    --train_filename ./augmented_data/python/pseudo_data.jsonl \
    --dev_filename ./data/codesearchnet/python/valid.jsonl \
    --test_filename ./data/codesearcnet/python/test.jsonl \
    --output_dir ./output_dir \
    --max_source_length 256 \
    --max_target_length 128 \
    --beam_size 10 \
    --train_batch_size 64 \
    --eval_batch_size 64 \
    --learning_rate 5e-6 \
    --gradient_accumulation_step 1 \
    --num_train_epoch 10

Final checkpoint will be saved as ./output_dir/python/checkpoint-best-bleu/pytorch_model.bin

Our code inherits the following repositories:

• CodeBERT
• UniXCoder
• CodeT5

We thank all the researchers who have made their code publicly available to facilitate the research community as well as our work.

@misc{to2023better,
      title={Better Language Models of Code through Self-Improvement}, 
      author={Hung Quoc To and Nghi D. Q. Bui and Jin Guo and Tien N. Nguyen},
      year={2023},
      eprint={2304.01228},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}