Python ML

Python machine learning projects (PyTorch, Hugging Face, scikit-learn) with raw data, notebooks, checkpoints and experiment tracking.

Set this profile

ccpp cpo profile set python-ml

Inside Claude Code, prefix the command with ! to run it in the session. Downloading a profile needs a Builder or Team plan; listing them does not.

6 policies, grouped by what they guard.

Requests, downloads and tool output are validated before they reach SQL, a shell, eval or a deserializer.

no-trust-remote-code
No new `trust_remote_code=True`: it runs arbitrary code from the model hub
no-unsafe-deserialization
No new unsafe loading of untrusted pickles (`weights_only=False`, `allow_pickle=True`)

Deletions, overwrites and irreversible commands wait for the user or a scoped target.

raw-data-immutable
Raw datasets are read-only; derived data goes to data/interim or data/processed

Routing, server and client boundaries, hooks and plugins the way the framework expects them.

notebooks-via-notebook-tool
Jupyter notebooks are edited cell by cell with NotebookEdit, not rewritten as raw JSON

Packages change through the package manager, lockfiles stay consistent and the project's tools are used.

dependencies-in-manifest
No ad-hoc `pip install` of packages; dependencies are declared in the project manifest

Build outputs, generated code and tool-owned files are regenerated, never edited by hand.

no-handwritten-artifacts
Model weights, checkpoints and binary datasets come from code, never from a file tool