sherpa-ml 0.1.3

Sherpa-ML: production-grade ML repo scaffolder with presets (vision/tabular/nlp/minimal), Hydra, MLflow, Docker, CI.

Sherpa-ML

Sherpa-ML is a scaffolding tool that generates production-ready ML repositories with opinionated presets (vision / tabular / NLP / minimal). It standardizes testing, typing, linting, configs, experiment tracking, Docker, optional serving, and CI — so a single command gives you a repo that installs, trains, logs, tests, and (optionally) serves out of the box.

• Consistent skeletons across domains
• Batteries included: Hydra, MLflow, pre-commit, Makefile, CI, Docker
• Velocity: sherpa-ml new my-repo --preset vision runs end-to-end in minutes

---

Quickstart

python -m pip install -U pip
pip install sherpa-ml
sherpa-ml --help

Generate a repo (vision, PyTorch, Hydra, MLflow, Docker slim, CI, extras):

sherpa-ml new demo-vision \
  --preset vision \
  --framework pytorch \
  --config hydra \
  --tracking mlflow \
  --serving false \
  --docker slim \
  --ci true \
  -e pre-commit -e "ruff+black+mypy" -e "pytest+coverage" -e makefile -e devcontainer \
  --render

Inside the generated repo:

cd demo-vision
pip install -e .[dev]
pytest -q
python -m demo_vision.train
python -m demo_vision.eval

If MLflow is enabled:

mlflow ui --backend-store-uri mlruns --port 5000

---

CLI Overview

sherpa-ml new <repo_name> [options]

Options
  --preset        vision | tabular | nlp | minimal
  --framework     pytorch | sklearn | transformers
  --tracking      mlflow | none                 (default: mlflow)
  --config        hydra | plain-yaml            (default: hydra)
  --serving/--no-serving                        (default: no-serving)
  --docker        slim | cuda | none            (default: slim)
  --ci/--no-ci                                  (default: ci)
  --license       MIT | Apache-2.0              (default: MIT)
  --extra/-e      dvc, pre-commit, ruff+black+mypy, pytest+coverage, makefile, devcontainer, minio (multi)
  --pkg           Override Python package name (defaults to repo_name slug)
  --plan          Dry-run plan table (no writes)
  --render        Write files to disk
  --force         Allow rendering into non-empty folder

Other commands
  sherpa-ml version
  sherpa-ml telemetry [--enable/--disable | --show]

---

Presets

• Vision (PyTorch): CIFAR-10 loaders, ResNet18/MobileNetV2, AMP training, checkpoints, ONNX export, tiny ORT bench.
• Tabular (scikit-learn): CSV loader, pipelines, RF/GBM baselines, metrics (acc/F1/AUC), calibration plot.
• NLP (HF Transformers): Tokenization pipeline, Trainer fine-tune on SST-2/AG News (small subsets), inference CLI.
• Minimal: Clean skeleton with tests/CI/config; ideal starting point for custom pipelines.

---

Quality Bar (Generated Repos)

• Typing: mypy strict
• Style: ruff + black (pre-commit)
• Tests: pytest smoke ≤ 30s
• Repro: seed control, env pinned
• Security: optional pip-audit / Trivy workflows
• CI: GitHub Actions matrix (3.10–3.12) if enabled

---

Developing Sherpa-ML (this repo)

python -m venv .venv && source .venv/bin/activate  # Windows: .\.venv\Scripts\Activate.ps1
pip install -e .[dev]
pre-commit install
pytest -q
sherpa-ml --help

Local end-to-end check

sherpa-ml new work-vision --preset vision --framework pytorch --config hydra --tracking none --docker none --no-ci --render
cd work-vision && pip install -e .[dev] && pytest -q && python -m work_vision.train

---

Telemetry (Opt-in)

On first run you’ll be asked to share anonymous template usage (no code/data/PII). You can disable at any time:

sherpa-ml telemetry --disable
# or set env: SHERPA_TELEMETRY=0

---

License

This project is licensed under MIT.