# Prepare a Hugging Face research release

Hugging Face is the main reading and discovery surface for this collection. A
private personal Static Space contains recipe pages, source snapshots, setup
guides and evidence; GitHub retains development history, tests and issue tracking.
See the [Hub layout and institutional examples](../examples/huggingface/README.md).
The current Hub scope is recipe documentation, executable source and aggregate
evidence, with links to upstream models and datasets. Trained weights, downloaded
data, raw model outputs and run artifacts remain on the cluster. Model/dataset
publication is outside the current delivery.
See the official [organization card documentation](https://huggingface.co/docs/hub/organizations-cards)
and [model card documentation](https://huggingface.co/docs/hub/model-cards).

No institutional organization is created automatically. The institutional
namespace, maintainers, access roles, and publication authority still need to be
established. A draft organization profile is in `templates/huggingface/organization.md`;
it can be adapted for an approved organization's **public Space named `README`**.
That public organization-card mechanism is separate from the current private
personal recipe Space.

Each research release should link its model or adapter, original dataset card and
data permissions, pinned training recipe, dependency/container identity, evaluation
specification and report, hardware profile, and maintainer. Do not place private
training samples, patient information, tokens, or raw agent logs into a public card.
Evaluation reports include benchmark/reference snapshots and full generated outputs;
treat them as private evidence when the task data is confidential. The local bundle
does not redact them.
Review code, base-model weights, training data, teacher-generated data, and benchmark
licenses separately.

## Create a local bundle

Fill every TODO in a copy of `templates/huggingface/model-card.md`, then select the
files explicitly:

```bash
hessian-recipes bundle \
  --model-card /absolute/completed-model-card.md \
  --report runs/candidate/report.json \
  --training-manifest /absolute/training-manifest.json \
  --artifact /absolute/adapter_config.json \
  --artifact /absolute/adapter_model.safetensors \
  --output releases/reasoning-adapter-candidate
```

The builder creates a new directory atomically, copies only named files, hashes
each copy, and writes `bundle.json`. It verifies evaluation evidence and refuses
to overwrite an existing bundle. It does not establish artifact compatibility or
scientific validity: card completion and a successful copy are insufficient for
that. All bundles remain candidates with `release_eligible: false`; review is an
explicit future step. Synthetic fixture results remain development evidence.

For the personal private recipe hub, the explicit publication command is
documented in the [Hub guide](../examples/huggingface/README.md). Model and dataset
bundles remain local artifacts; building a bundle does not upload it. This project
does not upload model weights or datasets to Hugging Face in the current scope.
An institutional release additionally requires its owners to establish the
namespace, access, licenses and publication terms.