Update README.md

84237cb verified 7 months ago

3.33 kB

---
license: mit
tags:
  - codellama
  - linux
  - bugfix
  - lora
  - qlora
  - git-diff
base_model: codellama/CodeLLaMA-7b-Instruct-hf
model_type: LlamaForCausalLM
library_name: peft
pipeline_tag: text-generation
---

# CodeLLaMA-Linux-BugFix

A fine-tuned version of `CodeLLaMA-7B-Instruct`, designed specifically for Linux kernel bug fixing using QLoRA (Quantized Low-Rank Adaptation). The model learns to generate Git diff patches based on buggy C code and commit messages.

---

## 🎯 Overview

This project targets automated Linux kernel bug fixing by:

- Mining real commit data from kernel Git history
- Training a QLoRA model to generate Git-style fixes
- Evaluating performance using BLEU and ROUGE
- Supporting integration into code review pipelines

---

## 📊 Performance Results

**BLEU Score**: 33.87

**ROUGE Scores**:
- ROUGE-1: P=0.3775, R=0.7306, F1=0.4355
- ROUGE-2: P=0.2898, R=0.6096, F1=0.3457
- ROUGE-L: P=0.3023, R=0.6333, F1=0.3612

These results show that the model generates high-quality diffs with good semantic similarity to ground-truth patches.

---

## 🧠 Model Configuration

- **Base model**: `CodeLLaMA-7B-Instruct`
- **Fine-tuning**: QLoRA (LoRA r=64, α=16, dropout=0.1)
- **Quantization**: 4-bit NF4
- **Training**: 3 epochs, batch size 64, LR 2e-4
- **Precision**: bfloat16 with gradient checkpointing
- **Hardware**: 1× NVIDIA H200 (144 GB VRAM)

---

## 🗃️ Dataset

- 100,000 samples from Linux kernel Git commits
- Format: JSONL with `"prompt"` and `"completion"` fields
- Content: C code segments + commit messages → Git diffs
- Source: Bug-fix commits filtered by keywords like `fix`, `null`, `race`, `panic`

---

## 🚀 Usage

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

model = AutoModelForCausalLM.from_pretrained("codellama/CodeLLaMA-7b-Instruct-hf")
model = PeftModel.from_pretrained(model, "train/output/qlora-codellama-bugfix")
tokenizer = AutoTokenizer.from_pretrained("codellama/CodeLLaMA-7b-Instruct-hf")

prompt = '''
Given the following original C code:
```c
if (!file->filter)
    return;

Instruction: Fix the null pointer dereference

Return the diff that fixes it: '''

inputs = tokenizer(prompt, return_tensors="pt") outputs = model.generate(**inputs, max_length=512, temperature=0.1) fix = tokenizer.decode(outputs[0], skip_special_tokens=True) print(fix)


---

## 📁 Structure

CodeLLaMA-Linux-BugFix/ ├── dataset/ # Raw and processed JSONL files ├── dataset_builder/ # Scripts for mining & formatting commits ├── train/ # Training scripts & checkpoints ├── evaluate/ # Evaluation scripts & results └── requirements.txt # Dependencies


---

## 📈 Metrics

| Metric   | Score  |
|----------|--------|
| BLEU     | 33.87  |
| ROUGE-1  | 0.4355 |
| ROUGE-2  | 0.3457 |
| ROUGE-L  | 0.3612 |

---

## 🔬 Use Cases

- Kernel patch suggestion tools
- Code review assistants
- Bug localization + repair research
- APR benchmarks for kernel code

---

## 📄 License

MIT License

---

## 📚 References

- [CodeLLaMA](https://arxiv.org/abs/2308.12950)
- [QLoRA](https://arxiv.org/abs/2305.14314)
- [LoRA](https://arxiv.org/abs/2106.09685)