This project uses GraphCodeBERT, fine-tuned on the CodeXGLUE / Devign Defect Detection dataset, to classify C/C++ functions as:

• Clean (0) – Non-defective
• Defective (1) – Likely vulnerable or risky

The goal is to build a practical, developer-friendly tool that can detect potential code defects using state-of-the-art transformer models.

• ✔ Fine-tuned GraphCodeBERT-base model
• ✔ Local inference using Python
• ✔ Batch prediction (clean + defective examples)
• ✔ Terminal-friendly output
• ✔ Ready for dataset evaluation, integration into CI, or further fine-tuning

👉 MODEL DOWNLOAD LINK:
https://drive.google.com/file/d/1Q3_x5eaYQ-jlntAgGuo9sau5Kic-P5r8/view?usp=sharing

Place the downloaded model folder inside your project like this:

DefectPrediction/
│
├── final_graphcodebert_balanced_best/
│   ├── config.json
│   ├── model.safetensors
│   ├── vocab.json
│   ├── tokenizer.json
│   ├── merges.txt
│   └── ...
│
├── run_inference.py
├── requirements.txt
└── README.md

git clone https://github.com/SreehariU/DefectPrediction
cd DefectPrediction

python3 -m venv env
source env/bin/activate

pip install -r requirements.txt

Modify or use run_inference.py:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
import torch

MODEL_DIR = "./final_graphcodebert_balanced_best"

tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_DIR)

pipe = pipeline(
    "text-classification",
    model=model,
    tokenizer=tokenizer,
    top_k=None,
    device=-1
)

sample_code = """ 
void swap(int *a, int *b){ int t=*a; *a=*b; *b=t; }
"""

out = pipe(sample_code)[0]
clean_prob = out[0]["score"]
defect_prob = out[1]["score"]
prediction = "defective" if defect_prob > clean_prob else "clean"

print("Prediction:", prediction)
print("Defect probability:", defect_prob)

Run:

python run_inference.py

run_inference.py also includes batch testing of 10 clean + 10 defective samples.

Trained on:

• CodeXGLUE – C/C++ Defect Detection (Devign)

Each sample contains:

• func → raw function code
• target → 0 (clean) or 1 (defective)

PRs welcome!
Ask for:

• Gradio UI
• FastAPI server
• Evaluation tools
• More datasets