patch-sft

patch-sft 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.

Using the published dataset

If you just want the pre-built dataset, no GitHub token or local pipeline needed:

import patch

ds = patch.load("your-hf-username/patch-sft")
print(ds)
# DatasetDict({
#     train: Dataset({...}),
#     test:  Dataset({...}),
# })

df = ds["train"].to_pandas()

Or install and use it as a dependency in another project:

# pyproject.toml
dependencies = ["patch-sft>=0.1.0"]

import patch

# Download from HuggingFace Hub
ds = patch.load("your-hf-username/patch-sft")

# Train split only
train_ds = patch.load("your-hf-username/patch-sft", split="train")

Running your own collection

To collect from GitHub and build your own dataset:

Requirements

• Python 3.11+
• uv
• GitHub personal access token
• Hugging Face account/token (for pushing)

Setup

git clone https://github.com/your-username/patch-sft
cd patch-sft
make sync
make init-env

Edit .env with your values:

GITHUB_TOKEN=ghp_...        # required for collection
HF_REPO_ID=yourname/patch-sft
HF_TOKEN=hf_...             # optional if using `huggingface-cli login`
DATA_DIR=./data             # optional, defaults to ./data
PYPI_API_TOKEN=pypi-...     # only needed for `make publish`

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

# Full pipeline in one shot
make pipeline

Programmatic collection

import patch
from patch import RepoConfig

patch.collect(
    repos=[
        RepoConfig("apache", "arrow", "python", "data-engineering"),
        RepoConfig("fastapi", "fastapi", "python", "web"),
    ],
    token="ghp_...",      # or set GITHUB_TOKEN env var
    data_dir="./data",
)

CLI

After installation, the patch-sft command is available:

patch-sft collect [--repo owner/repo]
patch-sft process [--repo owner/repo]
patch-sft merge
patch-sft push
patch-sft peek

Peek mode

Preview GraphQL request structure and one sample record without writing any files:

make peek
# or
patch-sft peek

Hardcoded target is pola-rs/polars. Prints a request example and a normalized sample record to stdout.

Project layout

patch-sft/
├── pyproject.toml
├── .env.example
├── Makefile
├── src/
│   └── patch/
│       ├── __init__.py     # public API: collect(), load(), RepoConfig, REPOS
│       ├── cli.py          # patch-sft CLI entry point
│       ├── repos.py        # curated repo list and RepoConfig
│       ├── collect.py      # GitHub GraphQL + REST collection
│       ├── process.py      # quality filters and SFT formatting
│       ├── merge.py        # train/test split → Parquet
│       ├── push.py         # HuggingFace Hub upload
│       ├── peek.py         # single-sample GraphQL preview
│       ├── manifest.py     # per-repo progress tracking
│       └── filters.py      # record quality filters
└── data/
    ├── raw/                # per-repo JSONL (append-only)
    ├── processed/          # filtered and formatted JSONL
    └── hf_upload/          # train.parquet / test.parquet

Output formats

Raw record (data/raw/*.jsonl)

Each line is a JSON object:

• 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

Failed diff fetches are written to data/raw/*.errors.jsonl and retried automatically on the next run.

Processed record (data/processed/*.jsonl)

Formatted for instruction SFT:

• messages: [{role: system}, {role: user}, {role: assistant}]
• metadata: repo/language/domain identifiers plus issue/PR provenance fields

HF upload (data/hf_upload/)

• train.parquet
• test.parquet

Quality filters

Records are kept only if they pass all checks:

• Issue body length ≥ 100 characters
• Changed diff lines between 5 and 500
• At least 1 PR review
• Changed files ≤ 10

Resumability

• data/raw/ is append-only; collection deduplicates by issue number
• data/manifest.json tracks per-repo progress and is flushed during collection
• Each pipeline stage is independent and reads from files on disk
• Re-running collection is safe — already-collected issues are skipped

Publishing to PyPI

make publish

Requires ~/.pypirc with a valid PyPI token (username __token__, password pypi-...), or PYPI_API_TOKEN set in your environment.