Code Reviewer — Fine-Tuned Llama 3.2 3B

📖 About

This project fine-tunes Llama 3.2 3B using QLoRA to produce structured Python code reviews. The model identifies bugs, style issues, and improvements, and formats every response with severity levels: Critical, Warning, or Info.

Training was performed on a free Google Colab T4 GPU in 2.5 minutes using Unsloth. The final LoRA adapter is only 88 MB — a 70x reduction compared to the full base model.

📊 Training Metrics

500

Training examples

2.5 min

Training time (T4)

2.40 → 0.11

Loss (start → end)

24M

Trainable params (0.75%)

88 MB

LoRA adapter size

⚡ Benchmark: Base vs. Fine-Tuned

The fine-tuned model produces concise, structured reviews with severity levels, while the base Llama 3.2 outputs generic, verbose explanations.

Test case: Off-by-one error

def calculate_average(numbers):
    total = 0
    for i in range(len(numbers) + 1):
        total += numbers[i]
    return total / len(numbers)

❌ Base Llama 3.2 3B

#### Critical
The function `calculate_average` has a bug. It does not handle the case when the input list `numbers` is empty. In such a case, the function will raise a `ZeroDivisionError`...

#### Warning
The variable name `total` is not very descriptive...

#### Warning
The function uses a for loop with an index that goes up to `len(numbers) + 1`...

✅ Fine-Tuned Code Reviewer

**Critical: Off-by-one error**
- `range(len(numbers) + 1)` goes one past the end, causing `IndexError`.
- **Fix:** Use `range(len(numbers))` or iterate directly.

🛠️ Tech Stack

Llama 3.2 3B QLoRA Unsloth PEFT TRL HuggingFace Google Colab FastAPI

🔗 Links

🤗 LoRA Adapter on HF Hub 📦 GitHub Repo 📓 Colab Notebook

⚠️ This is a static showcase page. The model runs on Google Colab or locally via the FastAPI service in the GitHub repo.

🔍 Code Reviewer