patch-sft 0.1.0

GitHub issue-to-diff dataset pipeline for supervised fine-tuning

patch

patch is a Python pipeline for building supervised fine-tuning (SFT) datasets from real GitHub engineering history.

It collects closed GitHub issues and linked merged pull requests, extracts unified diffs, applies quality filters, formats examples into messages[] chat records, and publishes train/test splits to Hugging Face.

What this project does

• Collects issue -> merged PR pairs from curated open-source repositories
• Fetches review counts, changed files, and unified diffs
• Applies quality filters to keep focused, learnable examples
• Formats output for ChatML-style instruction fine-tuning
• Merges and splits into deterministic train.jsonl / test.jsonl
• Pushes a DatasetDict to Hugging Face Hub

Project layout

patch/
├── pyproject.toml
├── .env.example
├── Makefile
├── config/
│   └── repos.py
├── src/
│   └── patch/
│       ├── collect.py
│       ├── process.py
│       ├── merge.py
│       ├── push.py
│       ├── peek.py
│       ├── manifest.py
│       └── filters.py
├── data/
│   ├── raw/
│   ├── processed/
│   └── hf_upload/
└── scripts/
    ├── run_collect.py
    ├── run_process.py
    ├── run_merge.py
    ├── run_push.py
    └── run_peek.py

Requirements

• Python 3.11+
• uv
• GitHub token with API access
• Hugging Face account/token (or CLI login)

Setup

make sync
make init-env

Then edit .env with your values:

• GITHUB_TOKEN
• HF_REPO_ID
• HF_TOKEN (optional if using hf auth login)
• DATA_DIR (optional, defaults to ./data)

Run the pipeline

All repos

make collect
make process
make merge
make push

Single repo

make collect-repo REPO=apache/arrow
make process-repo REPO=apache/arrow

One-command full run

make pipeline

Peek mode (GraphQL preview)

Use peek when you want to inspect GraphQL request structure and one sample issue/PR result without writing files.

make peek

• Hardcoded target: pola-rs/polars
• Prints a request example plus a sample normalized record to stdout
• Does not write dataset data to disk
• Includes full issue body, full PR body, changed files, and unified diff text in the sample

Output formats

Raw record (data/raw/*.jsonl)

Each line is a JSON object with fields such as:

• repo, language, domain, license, collected_at
• issue_number, issue_title, issue_body, issue_labels, issue_created_at
• pr_number, pr_title, pr_body, pr_merged_at, base_branch, base_sha, merge_sha, closing_pr_confidence
• diff, review_count, changed_files, additions, deletions, has_tests, test_files_changed

Processed record (data/processed/*.jsonl)

Each line is formatted for instruction SFT:

• messages: system, user, assistant
• metadata: repo/language/domain identifiers plus issue/pr and provenance fields

HF upload files (data/hf_upload/)

• train.jsonl
• test.jsonl

Quality filters

Records are kept only if they pass all checks:

• issue body length >= 100 chars
• changed diff lines between 5 and 500
• review count >= 1
• changed files <= 10

Resumability and invariants

• data/raw/ is append-only
• collection is safe to re-run and deduplicates by issue number
• data/manifest.json is the source of truth for per-repo progress
• manifest writes are flushed during collection for crash safety
• each stage runs independently from files on disk

Authentication notes

GitHub

• GITHUB_TOKEN is required for collection and peek commands

Hugging Face

Push supports either auth path:

1. Set HF_TOKEN in .env, or
2. Run hf auth login and use cached CLI credentials

Helpful targets

make help
make check

• make help: list all available commands
• make check: compile Python files for quick syntax validation