Reproducible install
After publication under DataFog/datafog-pii, record the immutable 40-character Hub commit SHA. Do not deploy an unpinned main revision. The release tag will be v0.1.0.
hf download DataFog/datafog-pii --revision <40-character-commit-sha> --local-dir ./datafog-pii-release
python3 ./datafog-pii-release/tools/install_bundle.py --snapshot ./datafog-pii-release --destination ./datafog-pii-bundle
The installer verifies the complete snapshot manifest, checks host OS/architecture, refuses an existing destination, safely extracts the versioned archive, and restores executable mode. Run datafog-pii-bundle/datafog-pii datafog-pii-bundle and send one JSON object with a text string per stdin line. Successful inference never requires Python or network access.
Policy for PR #32:
version = 1
[model]
bundle_directory = "/absolute/path/to/datafog-pii-bundle"
Reproduce actual MCP verification with a clean checkout:
git clone https://github.com/DataFog/datafog-mcp.git
git -C datafog-mcp checkout c2c45ec62728d85b7472131a3df9ccb67afd4a02
cd datafog-mcp
uv sync --frozen --group dev --python 3.12
DATAFOG_TEST_MODEL_BUNDLE=/absolute/path/to/datafog-pii-bundle uv run --frozen pytest tests/test_model_runtime.py tests/test_csv_processing.py --no-cov
Use the frozen lock, not the GLiNER comparison requirements. Fresh archive transport tests can be run locally before publication; only a real Hub download establishes Hub delivery. The release report distinguishes these statuses.