| name | train-smoke |
| description | Run a minimal train/import smoke for ML crates (Rust or Python HF). Use before claiming a training path works. |
| metadata | {"author":"tz-forge","version":"1.0","domain":"ml"} |
| allowed-tools | Bash(cargo:*) Bash(uv:*) Bash(python:*) Bash(pytest:*) |
Train smoke
When to use
- After scaffolding
ml-rust-crate or ml-python-hf
- Before long GPU runs — validate imports and unit tests first
- CI local gate extension for ML kinds
Rust path
./scripts/check.sh --quick
cargo test --all-features
Python HF path
uv sync --all-groups 2>/dev/null || uv sync
uv run python -m {{project_snake}}.train_smoke
uv run pytest -q
Rules
- Prefer CPU / no-download smoke unless user explicitly requests a model pull.
- Do not commit
wandb/, checkpoints/, *.safetensors, large .bin/.pt.
- Record command + pass/fail in the PR or kickoff note.
Output
Train smoke
===========
kind: ml-rust-crate | ml-python-hf
imports: ✅/❌
tests: ✅/❌
notes: ...