docker pull ghcr.io/iq2i/ai-code-review:latest

# Review your codebase
docker run --rm -v $(pwd):/workspace ghcr.io/iq2i/ai-code-review:latest /workspace/src

llama.cpp

bash

# Download the model
wget https://huggingface.co/iq2i/ai-code-review/resolve/main/model-Q4_K_M.gguf

# Run inference
./llama-cli -m model-Q4_K_M.gguf -p "Review this code: ..."

Python (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama(model_path="model-Q4_K_M.gguf")
output = llm("Review this code: ...", max_tokens=512)
print(output)

Output Format

The model outputs concise text-based code reviews:

**SQL injection vulnerability**

User input is concatenated directly into a raw SQL query without parameterization or escaping.

Impact: An attacker can execute arbitrary SQL commands, potentially dumping the entire database, deleting data, or escalating privileges. For example: keyword=' OR '1'='1' -- would return all products.

Suggestion:
Use parameter binding: DB::select("SELECT * FROM products WHERE name LIKE ?", ['%' . $keyword . '%']) or better, use Eloquent: Product::where('name', 'like', '%' . $keyword . '%')->get()

Training

Training examples: 100+ real-world code issues
Format: ChatML conversation format with concise reviews
Framework: MLX for Apple Silicon acceleration
Method: LoRA adapters (r=4, alpha=8)
Iterations: 625

For training details, see the GitHub repository.

Limitations

Should be used as a supplementary tool, not a replacement for human review
May not catch all edge cases or security vulnerabilities
Best results on common programming patterns and frameworks

License

Apache 2.0

Citation

bibtex

@software{ai_code_review,
  title = {AI Code Review Model},
  author = {IQ2i Team},
  year = {2025},
  url = {https://github.com/iq2i/ai-code-review}
}