pr-assist

LoRA fine-tuned Qwen2.5-1.5B-Instruct for pull request code review.

Default adapter: toastcoder/pr-review-qwen-lora

Project layout

pr-assist/
├── src/
│   ├── inference.py   # CLI and library API for review generation
│   ├── train.py       # LoRA fine-tuning script
│   ├── prompts.py     # PR prompt templates
│   └── chunking.py    # Split large diffs by file
├── scripts/
│   ├── create_subset.py
│   └── verify_adapter.py
└── notebooks/         # Exploratory training notebooks

Legacy experiments and local checkpoints live under archive/.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run inference

Interactive mode:

python -m src.inference

Review a pull request from flags:

python -m src.inference \
  --title "Add JWT auth middleware" \
  --description "Protect user routes with JWT tokens." \
  --diff-file path/to/changes.diff

Review a large diff in file chunks:

python -m src.inference \
  --title "Refactor auth module" \
  --description "Split auth into middleware and helpers." \
  --diff-file path/to/large.diff \
  --chunk

Pipe a full prompt on stdin:

cat prompt.txt | python -m src.inference

Training

Create a small local dataset:

python scripts/create_subset.py --size 1000 --output data/processed/train_small.jsonl

Fine-tune LoRA:

python -m src.train \
  --data-files data/processed/train_small.jsonl \
  --output-dir models/pr-assist-qwen

Smoke test

python scripts/verify_adapter.py