Publishing Guide

This checklist keeps the repository usable as both a Python package and a Hugging Face artifact.

Local Checks

python -m pip install -e ".[dev,examples,publish]"
python -m ruff check .
python -m ruff format --check .
python -m pytest
python -m compileall -q src scripts
time-anchor-infer --no-impact
python -m build

Run the plotting example when matplotlib is installed:

time-anchor-sine-test

Hugging Face Model Hub

The current Hugging Face CLI is hf.

hf auth login
hf repo create K-Iwa/time-anchor-modernbert-32m --repo-type model --exist-ok

For a model-only repository, upload the checkpoint files:

hf upload K-Iwa/time-anchor-modernbert-32m models/time-anchor-modernbert-32m . --repo-type model

Users can then load it with:

from time_anchor import TimeAnchorPipeline

pipeline = TimeAnchorPipeline.from_pretrained("K-Iwa/time-anchor-modernbert-32m")

For a combined code/model repository, upload the full project:

hf repo create K-Iwa/time-anchor --repo-type model --exist-ok
hf upload K-Iwa/time-anchor . . --repo-type model --exclude output --exclude dist --exclude .venv

Users should load the nested checkpoint with:

from time_anchor import TimeAnchorPipeline

pipeline = TimeAnchorPipeline.from_pretrained(
    "K-Iwa/time-anchor",
    subfolder="models/time-anchor-modernbert-32m",
)

Use hf upload-large-folder instead of hf upload if the upload is interrupted often or the checkpoint directory grows substantially.

Python Package Distribution

The Python package distribution is intentionally lightweight. MANIFEST.in includes source, documentation, license files, and the small example CSV. Model weights and generated outputs are not part of the package distribution; publish model weights on Hugging Face Hub instead of PyPI.

Check the generated files before uploading:

python -m build
python -m twine check dist/*

GitHub or Source Release

git init
git add .
git commit -m "Prepare Time-Anchor for public release"

Before publishing source-only mirrors, decide whether the checkpoint should be included. If model weights are included, keep Git LFS enabled for *.safetensors.

Release Hygiene

Confirm that output/ only contains generated files and is not committed.
Keep examples/sample_data.csv small.
Update the README examples if the checkpoint folder or default target column changes.
Add benchmark or evaluation notes before advertising quantitative quality claims.