# Publish this release as a Hugging Face dataset

Suggested destination: **PureOne/AELYRAITH-III**. The repository title and all
research content come from the supplied package. This kit prepares a dataset
repository containing the manuscript, code, certificates, experiment data,
and AI/expert documentation; it does not publish until you run the uploader.

## Windows

1. Extract the complete ZIP to a normal folder. Do not run files inside the ZIP.
2. Install Python 3.10 or newer if needed.
3. Double-click `UPLOAD_TO_HUGGINGFACE.bat`.
4. If no Hugging Face login or `HF_TOKEN` is available, paste a token at the
   hidden prompt. It must permit writing the target dataset. The uploader
   does not save pasted tokens or add them to Git credentials.

The launcher creates a separate `.upload-venv`, installs the upload dependency,
verifies the release, resolves your account's canonical capitalization, and
uploads only the manifest-listed release files. The research requirements
are not needed for upload. Internet access is needed for dependency installation
and publication.

## Terminal

The verification and upload plan can be inspected completely offline, using
only Python's standard library:

```bash
python upload_to_huggingface.py --dry-run
```

To publish, install the upload client and run:

```bash
python -m pip install -r requirements-upload.txt
python upload_to_huggingface.py
```

The uploader first uses `HF_TOKEN` or your existing cached Hugging Face login.
If neither is available in an interactive terminal, it asks for a token using
a hidden prompt. Do not put a token in commands, source files, or the release.

Optional arguments:

```bash
python upload_to_huggingface.py --repo-id PureOne/AELYRAITH-III --private
python upload_to_huggingface.py --repo-id YourAccount/YourDataset --dry-run
```

`--private` sets the visibility only when creating a new repository. An
existing repository keeps its current visibility. `--repo-id` must belong to
the authenticated account or one of its listed organizations. The Hub still
checks the token's actual write permission. The default never silently switches
publication to another account if a different user is logged in.

Uploads add files and update matching paths. They do not delete unrelated
remote files. Re-running the uploader is the recovery path after an interrupted
upload. Verify the final dataset page after publication; local verification
and mock tests do not establish a successful remote commit.

## Manual alternative

Create a dataset repository on Hugging Face, then upload this extracted
release's manifest-listed files with their relative folder structure intact.
`README.md` belongs at the repository root. Keep `.upload-venv`, local tokens,
caches, and your private work outside uploaded content. The uploader's
`--dry-run` prints the exact planned paths, including `release_manifest.json`.

## Troubleshooting

| Symptom | Action |
|---|---|
| Checksum or verification failure | Extract a fresh release. After intentional edits, review changes and regenerate the release manifest with the release tooling. |
| Namespace mismatch | Log in as the intended owner or pass an explicit permitted `--repo-id`. Canonical account capitalization is preserved. |
| HTTP 401 / 403 | Use a valid token with write access to this dataset and check organization membership where applicable. |
| HTTP 404 | Check the exact namespace and dataset repository type, account access, and token scope. |
| Temporary network or service failure | Retry the same command after connectivity or service availability returns. |

## Implementation references

The implementation follows the official Hugging Face documentation, checked
23 October 2026:

- [HfApi](https://huggingface.co/docs/huggingface_hub/package_reference/hf_api)
- [Uploading files and folders](https://huggingface.co/docs/huggingface_hub/guides/upload)
- [Authentication](https://huggingface.co/docs/huggingface_hub/package_reference/authentication)

`tests/test_upload.py` uses a mocked API to test dataset routing, account
capitalization, exact file allowlisting, and the offline path. These tests make
no Hugging Face requests. Live Windows launcher execution and remote upload
are user-side steps.