This repository contains scripts and tools to collect, preprocess, and prepare Go code/test file pairs for fine-tuning the Phi-3 language model for unit test generation. Phi-3 mini was selected due size and performance on tasks related to code-generation.
This project was imlemented from scratch using bash, GitHub API, and Unsloth library as a part of MSc project. Initial training was performed on Kaggle environment. The goal of the project was to evaluate the possibility of generating unit tests using a small fine-tuned model and compare results with state-of-the-art models.
| File | Description |
|---|---|
fetch-repositories-urls.py |
Python script to fetch repository URLs from GitHub API using batching. |
clone-repositories-optimised.sh |
Efficiently clones the repositories fetched earlier using shallow clones and other tricks. |
collect-files-in-folder.sh |
Filters and copies Go files into a single folder for further processing. |
preprocess-files.sh |
Cleans and filters the collected files to prepare them for dataset generation. |
generate-training-data.sh |
Converts preprocessed files into structured input/output pairs suitable for model fine-tuning. |
fine-tuning-and-inference-project-go-phi.ipynb |
A notebook to fine-tune the Phi-3 model on the generated dataset and perform inference on the trained model. |
- Repository Fetching: Automatically gather Go repository URLs according to specified criteria.
- Preprocessing Pipeline: Clean, filter, and normalize code and corresponding tests.
- Training Dataset Generation: Generate a dataset that contains Go code and test pairs.
- Training pipeline using Unsloth library: Jupyter notebook that contains the training and evaluation pipeline.
- Git installed
- GitHub API key needs to be generated
- Environment to run the notebook
- Bash
This project is licensed under the MIT license.
- Upload fine-tuned model to Hugging Face
- Upload generated dataset to Hugging Face
- Add instructions on how to run the code