CodeSim_cpp

The following codes are the solutions (1st place, private score: 0.9911) for the dacon competition.

1. Environmental settings

1.1 Clone this repository

git clone https://github.com/GNOEYHEAT/CodeSim_cpp.git
cd CodeSim_cpp

1.2 Install packages

pip install -r requirements.txt 

Directory Structure

├── Dataset
│   ├── train_code
│   │   ├── problem001
│   │   ├── ...
│   │   └── problem500
│   ├── sample_submission.csv
│   ├── sample_train.csv
│   └── test.csv
├── Utils
│   ├── CodeLM_utils.py
│   └── Preprocessing_utils.py
├── Preprocess.py
└── CodeLM.py

2. Data Proprocessing

python Preprocess.py 

3. Train Model

python CodeLM.py

4. Experiment Results

The final submission is GraphCodeBERT+UniXcoder.

Preprocessed Datasets

Datasets

• The hyperparameters are as follows:
  • truncation_side='left', bm25='bm25plus'

Pre-trained Models

Models

• The hyperparameters are as follows:
  • truncation_side='left', optimizer='adamw', learning_rate=0.00003

Results

Preds

index	CodeBERT Model	frac	text_len	Pr Acc	Pl Acc	Val Acc
exp_30	GraphCodeBERT	0.01	512	0.98859	0.98831	0.99641
exp_31	GraphCodeBERT	0.02	512	0.98909	0.98892	0.99794
exp_32	UniXcoder	0.01	1024	0.98942	0.98911	0.99606
exp_31+32	GraphCodeBERT+UniXcoder	-	-	0.99111	0.99084	-