marimo-batch-mlflow
Installation
SKILL.md
marimo-batch-mlflow
Opinionated fork of marimo-team/skills/marimo-batch that:
- Uses Tyro for CLI parsing (works with
dataclass,pydantic.BaseModel, orattrs) instead ofmo.cli_args()+ manual help-table rendering. - Uses MLflow + mlflow-widgets for experiment tracking instead of Weights and Biases.
- Keeps the dual-mode pattern (
mo.app_meta().mode == "script") so a singlenotebook.pyis both the UI for iteration and the entry point foruv run notebook.py --epochs 50batch jobs.
When to use this vs upstream marimo-batch
| Concern | upstream marimo-batch |
this skill |
|---|---|---|
| CLI parsing | mo.cli_args() + hand-rolled rich.Table for --help |
tyro.cli(ModelParams) — auto --help, type coercion, validation |
| Params model | Pydantic BaseModel |
dataclass (primary) or pydantic.BaseModel (alternative) |
| Tracking backend | Weights and Biases (wandb) |
MLflow (mlflow + optional mlflow-widgets for live charts) |
| Live training UI | none — W&B web dashboard only | mlflow_widgets.MlflowChart cell, gated off in script mode |
| Grid launcher | HF Jobs + WANDB_API_KEY secret |
HF Jobs + MLFLOW_TRACKING_URI (+ optional MLFLOW_TRACKING_TOKEN) |
Pick this skill when the user has a self-hosted MLflow server (or local ./mlruns is fine) and prefers strongly-typed CLIs. Pick upstream when the user is already on W&B.