michelangelo-examples

Pip-installable, ready-to-run example models and pipelines for Michelangelo — a lightweight on-ramp for trying Michelangelo without a full monorepo checkout, the heavyweight example extra, or a running sandbox.

pip install michelangelo-examples[<example-name>]

Why this repo

The full example pipelines in michelangelo/python/examples/ are the canonical, full-fidelity reference — they exercise real Cadence/Temporal orchestration, Spark, and a live sandbox, and they aren't going anywhere. But trying one today means checking out the whole monorepo and installing one heavyweight extra that bundles every example's dependencies (torch, transformers, xgboost, pyspark, ray, mlflow, peft, and more) at once, regardless of which example you actually want.

michelangelo-examples is a second, lighter tier: each example is its own pip extra with only the dependencies it needs, plus a plain-Python run_local.py entrypoint that trains/evaluates locally — no Cadence, no Spark, no sandbox. It's the "try it in five minutes" path underneath the full pipeline reference, not a replacement for it.

Examples

Ported so far:

• california_housing — a project (use case) with one pipeline so far, pytorch_train (California Housing price prediction via PyTorch Lightning, migrated from core michelangelo's python/examples/pipelines/california_housing_lightning/). Structuring it as a project with a pipelines/ subfolder leaves room for sibling pipelines against the same use case — e.g. a future xgboost_train pipeline — without another top-level rename.
• bert_cola — fine-tunes BERT for linguistic acceptability classification on the CoLA task, migrated from core michelangelo's python/examples/bert_cola/.
• gpt_oss_20b_finetune — LoRA fine-tuning of GPT-style causal-LM checkpoints (architected for GPT-OSS-20B, tested against GPT-2 as a CPU-feasible proxy), migrated from core michelangelo's python/examples/gpt_oss_20b_finetune/. Requires Python 3.11+ (see that project's own README).
• movielens — trains a tiny Neural Collaborative Filtering model on MovieLens-100k via Ray Train + PyTorch Lightning, migrated from core michelangelo's python/examples/movielens/. Requires Python 3.11+ (see that project's own README).
• nomic_ai — fine-tunes/encodes nomic-ai/nomic-bert-2048, a long-context (2048-token) BERT, over a WikiText-2 sample via PyTorch Lightning, migrated from core michelangelo's python/examples/nomic_ai/. Requires Python 3.11+ (see that project's own README).

v1 candidates (not yet ported): amazon_books_qwen. See the project spec for the full list this repo is drawing from.

Each project lives entirely under src/michelangelo_examples/<project-name>/ — a real subpackage of the one michelangelo_examples package this repo ships, including its pipeline.yaml/README.md (bundled as package data so a plain pip install gets everything, not just the Python code) alongside each pipeline's model/training/task code and __main__.py local runner. Code shared across a project's sibling pipelines (e.g. dataset loading/feature prep) lives in <project-name>/pipelines/libs/.

Installing a project

pip install "michelangelo-examples[california-housing]"
python -m michelangelo_examples.california_housing.pipelines.pytorch_train

Extras are scoped per project, not per pipeline: installing california-housing pulls in the dependencies for every pipeline under that project (today just pytorch_train), since they already share one built image and mostly overlapping dependency sets.

The pip install + python -m commands above run the lightweight local tier only. To run a pipeline's full Cadence-dispatched version against a Michelangelo sandbox (ma project apply → ma pipeline apply → ma pipeline run), see the end-to-end command sequence in that pipeline's own README — e.g. pytorch_train's.

Relationship to other Michelangelo repos

• michelangelo — the core platform; python/examples/ there remains the full pipeline reference (Cadence/Temporal/Spark/sandbox).
• michelangelo-models — pluggable model architecture definitions. Some examples here may import a seed architecture from there instead of defining their own model.

Development

This repo uses uv for dependency management.

uv build
uv run pytest

License

Apache License 2.0