metadata
language:
- code
license: mit
tags:
- code-generation
- bug-fixing
- code-repair
- codet5
- debugging
datasets:
- custom
metrics:
- accuracy
- exact-match
library_name: transformers
pipeline_tag: text2text-generation
brainbug
Model Description
This is a fine-tuned CodeT5 model for automatic bug detection and code repair. The model has been trained to identify and fix various types of programming errors in Python code.
Supported Error Types
- WVAV: Wrong Variable Used in Variable Assignment
- MLAC: Missing Line After Call
- WPFV: Wrong Parameter in Function/Method Call
- And more...
Model Details
- Base Model:
Salesforce/codet5-base - Fine-tuned on: Custom bug-fix dataset
- Task: Code-to-Code generation (bug fixing)
- Language: Python
- Model Size: 220M parameters
Usage
from transformers import T5ForConditionalGeneration, RobertaTokenizer
# Load model and tokenizer
model = T5ForConditionalGeneration.from_pretrained("Sagar123x/brainbug")
tokenizer = RobertaTokenizer.from_pretrained("Sagar123x/brainbug")
# Example: Fix buggy code
faulty_code = """
def check_for_file(self, file_path):
files = self.connection.glob(file_path)
return len(files) == 1
"""
# Prepare input
input_text = f"Fix WVAV: {faulty_code}"
inputs = tokenizer(input_text, return_tensors="pt", max_length=256, truncation=True)
# Generate fix
outputs = model.generate(**inputs, max_length=256, num_beams=5)
fixed_code = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(fixed_code)
Training Details
- Training Epochs: 10
- Batch Size: 1 (with gradient accumulation)
- Learning Rate: 3e-5
- Optimizer: AdamW
- Hardware: NVIDIA RTX 4050 (6GB)
Performance Metrics
- Exact Match Accuracy: 2.60%
- Token-Level Accuracy: 28.52%
- Average Similarity: 76.75%
Limitations
- Trained primarily on Python code
- Best performance on error types seen during training
- May not handle very long code snippets (>256 tokens)
- Requires error type specification for optimal results
Citation
@misc{brainbug-codet5,
author = {Your Name},
title = {BrainBug: CodeT5 for Automatic Bug Repair},
year = {2025},
publisher = {HuggingFace},
howpublished = {\url{https://huggingface.co/Sagar123x/brainbug}}
}
License
MIT License
Contact
For questions or issues, please open an issue on the model repository.