llm4lint-vsc

A Static Analysis Tool built by finetuning Qwen2.5 Coder using unsloth. Trained only for python files (For now, You can extend it with your own data if you want).

Features

• Adds 2 additional menu items when you right click a python file
  • Linting of Python source code, highlights problematic parts of the code
  • Lint in interactive mode, talk to the model locally.

Right click on a Python file

Example code

• You can also provide your own data to create a different model (use train.py script): training script and augmentation script
  • Specify your own examples in a csv file with input and output columns. output should be <lineno> - <type>: <issue>. With just 1 example multiple datapoints are created using augmentation.
  • Augmentation of inputs:
    • Variable names are replaced with different names to make sure model does not memorize the code.
    • Additional code is added before and after your input example. (outputs are also adjusted to account for lineno changes).
• Dataset created using this script: https://huggingface.co/datasets/ahmedhus22/python-static-analysis-linting

Requirements

• Download the fine-tuned model from Hugging Face .gguf and Modelfile

• Create ollama Model

ollama create llm4lint7b -f <Modelfile-Path>

For Interactive mode, you need to install llm4lint cli-tool:

pip install llm4lint

Known Issues

• If the output of the linter is not satisfactory, simply rerun it. It will give a different output (most likely, its not deterministic).
• Or prepare your own examples using the training script and augmentation script