---
license: mit
base_model: microsoft/phi-3-mini-4k-instruct
tags:
- llm
- code-generation
- bug-fixing
- lora
- peft
- python
datasets:
- mbpp
metrics:
- exact_match
- similarity
---

# DebugGPT LoRA Adapter for Phi-3 Mini

A lightweight LoRA adapter fine-tuned on synthetic Python bug-fixing tasks using the MBPP dataset. This model enhances the ability of Phi-3 Mini to detect and correct common Python syntax errors while preserving general language capabilities.

---

## Model Description

- **Base Model:** microsoft/phi-3-mini-4k-instruct
- **Fine-Tuning Method:** QLoRA (Low-Rank Adaptation with 4-bit quantization)
- **Task:** Automated Python bug fixing

The model takes buggy Python code as input and generates the corrected version.

---

## Intended Use

This model is designed for:

- Python debugging assistance
- Educational coding tools
- AI-assisted code correction
- Research experiments in code repair

### Out-of-Scope Use

- Production-critical systems
- Security-sensitive applications
- Complex multi-file debugging

---

## Dataset

We use the **MBPP (Mostly Basic Python Problems)** dataset. Since MBPP contains correct code, we generate a bug-fixing dataset by injecting synthetic bugs.

### Data Format

Each example follows an instruction-tuning format:

```json
{
  "instruction": "Fix the bug in the following Python code",
  "input": "<buggy code>",
  "output": "<correct code>"
}
```

### Bug Injection Strategy

We introduce controlled bugs such as:

- Operator replacement (`+` → `-`)
- Comparison changes (`>` → `<`)
- Removal of return statements

### Dataset Size

| Split      | Samples |
|------------|---------|
| Train      | ~374    |
| Validation | ~90     |
| Test       | ~500    |

---

## Training Procedure

### Method: QLoRA

To enable efficient training on limited hardware:

- Base model loaded in 4-bit precision (NF4)
- Base weights frozen
- Only LoRA adapters trained

### LoRA Configuration

| Parameter       | Value                              |
|-----------------|------------------------------------|
| Rank (r)        | 16                                 |
| Alpha           | 32                                 |
| Dropout         | 0.05                               |
| Target Modules  | q_proj, k_proj, v_proj, o_proj     |

### Training Configuration

| Parameter              | Value   |
|------------------------|---------|
| Epochs                 | 3       |
| Learning Rate          | 2e-4    |
| Batch Size             | 1       |
| Gradient Accumulation  | 8       |
| Precision              | FP16    |
| Optimizer              | AdamW   |

---

## Hardware & Frameworks

- **GPU:** NVIDIA Tesla T4
- **Frameworks:** Hugging Face Transformers, PEFT (LoRA), TRL (SFTTrainer), Weights & Biases

---

## Evaluation Results

### Performance Summary

| Metric                  | Base Model    | Fine-Tuned Model   |
|-------------------------|---------------|--------------------|
| Syntax Fix Accuracy     | Low           | Noticeably Higher  |
| Indentation Correction  | Inconsistent  | Reliable           |
| Variable Error Fixing   | Occasional    | Improved           |
| Complex Logic Bugs      | Limited       | Limited (unchanged)|
| Instruction Adherence   | Moderate      | High               |

> **Note:** Quantitative metrics (e.g., exact match accuracy, CodeBLEU) were not computed due to dataset and tooling constraints.

---

## Example

### Input — Buggy Code

```python
for i in range(5)
    print(i)
```

### Output — Fixed Code

```python
for i in range(5):
    print(i)
```

---

## Limitations

- Small dataset size limits generalization
- Focused primarily on syntax-level bugs
- Limited performance on complex logical errors
- Not evaluated on large-scale real-world codebases

---

## Discussion

### What Worked Well

- QLoRA enabled efficient fine-tuning on limited hardware
- Significant improvement in syntax correction tasks
- Strong adherence to instruction format

### Challenges

- Limited dataset size
- Lack of quantitative evaluation metrics
- Difficulty handling complex multi-line logic bugs

### Ethical Considerations

- The model may generate incorrect fixes for complex bugs
- Should be used as an assistive tool, not a final authority
- Users should validate outputs before deployment

---

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/phi-3-mini-4k-instruct"
)

tokenizer = AutoTokenizer.from_pretrained(
    "microsoft/phi-3-mini-4k-instruct"
)

model = PeftModel.from_pretrained(
    base_model,
    "Sud1212/phi3-debug-llm-lora"
)

prompt = "Fix the bug:\nfor i in range(5)\n    print(i)"

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## Resources

- **GitHub Repository:** [Phi3-debugLLM-LoRA](https://github.com/suddhumaddi/Phi3-debugLLM-LoRA)
- **Weights & Biases Dashboard:** [W&B Project](https://wandb.ai/suddhumaddi-woxsen-university/huggingface)
- **Dataset (MBPP):** [Hugging Face Datasets](https://huggingface.co/datasets/mbpp)

---

## Author

**Sudarshan Maddi**
Woxsen University

---

## License

MIT License