hessian AI & JAIF on Hugging Face
The central reading surface is the hessian AI & JAIF website, a Protected Static Space under the personal ctauchmann/hessian-recipes namespace. Its pages contain the recipe catalogue, setup guides, measured evidence, research references and domain sketches. The Space also contains runnable source, tests, dependency files and authored fixtures; GitHub is the collaborative source of truth for changes, review and issue tracking. The Space is a generated publication of a specific source commit. Recipe pages link to upstream model/dataset cards and explain local artifacts; this hub does not contain weights, downloaded training data, raw model outputs or private run directories. Model and dataset publication is outside this workflow.
The App is a public research website: a factual overview, an experimental protocol, a searchable methods index, interactive pilot-result comparisons, benchmark proposals and domain sketches. Guide pages share the same layout and add an on-page contents list and copyable code examples. These features work in the Protected Static Space. Share the direct App address above: it needs no login and includes the approved downloadable source files. The Hugging Face repository page and Git clone remain private, as does the development repository on GitHub. Organization profile cards remain deferred.
Result charts are built from the exact committed aggregate evidence included in the source snapshot. They retain sample sizes, inconclusive outcomes and the failed retention gate. They do not read cluster logs or claim live experiment status. The site uses local HTML, CSS, SVG and JavaScript with no runtime external fonts, analytics or model calls; the content remains readable without JavaScript. Typography and purple/lavender colors follow the hessian.AI website; the joint identity includes the official JAIF logo from jaif.ai. Selected official assets are self-hosted with source URLs, checksums and rights scope; the Poppins SIL Open Font License is included. The project-index presentation also draws on FineEnvs: concise scope, methods, implementation links, results and explicit qualification status. The interface validation records desktop/mobile browser checks, accessibility findings and the reviewed fixes.
As of 21 October 2026, the longer NeMo and matched OpenEnv attempts stopped at the GPU token-audit qualification. No longer-run quality result is available; existing pilot results retain their original limitations. See the current programme status.
The Protected App is live, with public recipe/source pages and a separate private upstream-reference collection. Anonymous visitors can read the website and its source downloads; anonymous Hub repository access is denied. The account's private-storage restriction previously blocked authenticated repository resolve downloads; a fresh remote HF clone has not been qualified. Use GitHub for the development checkout. The dated publication evidence records the earlier Private deployment and its access limits; its anonymous-App-denial findings are historical, not the current Protected visibility contract.
Native Hub features and visibility
| Feature | Use here | Visibility boundary |
|---|---|---|
| Static HTML Space | Multi-page recipe and evidence homepage, served without a compute instance. | Protected visibility makes the App and its approved source downloads public while keeping the Space repository private. |
| Collections | Curated lists of related models, datasets, Spaces and papers, with concise item notes. | Personal collections can be private. Private collections do not appear as public profile highlights. Collection visibility does not replace each repository's own access controls. |
| Model cards and dataset cards | Link to upstream ancestry, use, evaluation, limitations and terms; keep this project's local artifact notes in recipe pages. | No model or dataset repository is uploaded in this phase. |
| Organization card | A future institutional navigation page linking the catalogue and release families. | HF displays a card from a public Space named README. A private/protected Space does not display an organization card; organization cards remain deferred. |
| Blog articles | Later public explanations of recipes or research releases. | No public article or profile is edited by this workflow. Keep unpublished working notes in the private GitHub checkout. |
Static Spaces use sdk: static and app_file: index.html. The generated site is self-contained: no external scripts, analytics, model calls or provider keys. Its build manifest records the exact source bytes and generated files. Publication validates the approved personal namespace, target repository and explicit Protected visibility, then verifies the resulting commit and visibility. Private visibility remains available as an explicit deployment mode and the CLI default; neither mode silently changes the existing Space's visibility.
Reusable institutional patterns
These are publishing patterns, not evidence that their model results transfer to this project. Profiles were inspected on 20 October 2026.
- Ai2 / AllenAI groups complete project families
in collections. For example, its AstaBrief collection connects SFT and DPO checkpoints with the corresponding training mixtures; leaderboard Spaces give evaluation its own destination. Reuse the recipe → data → model → evidence links.
- NVIDIA uses a structured organization card with
sections for model families, datasets, frameworks, demonstrations and technical reports. Reuse capability-based entry blocks and links to concrete artifacts, while keeping implementation and qualification boundaries explicit.
- Stanford CRFM provides a research-group
namespace connecting its model work to the institutional project identity. Reuse consistent project naming and provenance, without implying affiliation.
- LeoLM provides a focused German-model project
identity. Reuse a dedicated language-adaptation entry rather than placing every experiment into a generic institutional model list.
For this project, start with recipe pages grouped by user task and references to upstream models, datasets and papers. A private collection can group those links. An institutional public profile and research articles are later publication decisions.
Build and review locally
From the repository root, in the dependency-free core environment:
.venv/bin/python -m hessian_recipes.hf_recipe_hub \
--output runs/huggingface/recipe-hub-review --visibility protected
python -m http.server 8090 --directory runs/huggingface/recipe-hub-reviewOpen the local page, review hub-manifest.json, and use a fresh build directory for each snapshot. An explicit allowlist selects tracked source, dependencies, scripts, tests, documentation, cards and small authored development fixtures. Unrelated repository files, .env files, tokens, ignored run data and weights are excluded. Executable-source links resolve to the included snapshot. Start at Run from Hugging Face for the snapshot layout and authenticated repository-clone route without a GitHub dependency.
Publication uses the optional installed huggingface_hub client and an already authenticated personal account. The approved target is only ctauchmann/hessian-recipes with repository type space and explicit Protected visibility. There is no implicit collection, organization, blog or model publication.
Publish after GitHub checks
Deploy Hugging Face site publishes the generated site after the checks workflow succeeds for a push to main in this repository. Pull-request and fork runs cannot publish. It checks out the exact tested commit, builds the allowlisted snapshot and validates local links in a job without Hugging Face authentication. A separate publication job uses GitHub OIDC with huggingface_hub==1.33.0 to obtain short-lived credentials for spaces/ctauchmann/hessian-recipes.
CI explicitly uses publish --visibility protected --verification repository. It independently rebuilds the source bundle, checks the approved Protected target, and publishes with the existing repository head as a compare-and-swap condition. It verifies the exact remote repository files, then the publicly served manifest, assets and source files. Before and after publication, it requires public App access and denial of anonymous Hub raw-file access. In Protected mode its receipt records served_manifest_matches: true and app_verified_by_ci: true; no App JWT is needed to read the public site.
During the earlier Private deployment on 21 October 2026, the repository-scoped OIDC publisher could upload and inspect the private Space, but its request to mint an App-access JWT returned HTTP 401. That historical CI receipt recorded served_manifest_matches: null and app_verified_by_ci: false. Protected deployment verifies the public App directly. The separate App check can also run from JAIF using local credentials outside the checkout; no personal token is stored in GitHub. The publisher's default verification mode remains strict; Private mode retains authenticated App verification or an explicitly chosen repository-only check.
For Private mode, the authenticated check obtains a five-minute Space JWT from the first-party /api/spaces/{namespace}/{repo}/jwt endpoint and verifies the manifest using __sign. The endpoint is described in Hugging Face's OpenAPI schema. Tokens stay in memory; signed URLs are never logged or saved in receipts.
The Space must have a Hugging Face Trusted Publisher registration with these exact claims: repository ChTauchmann/hessian-recipes, branch main, and workflow deploy-huggingface.yml. The build job cannot mint OIDC tokens; only the publication job has id-token: write, and HF_OIDC_RESOURCE is scoped to its publication step. No stored HF_TOKEN repository secret is required. Do not place a user token in source files, workflow inputs or clone URLs.
HF attributes these repository-scoped writes to its synthetic OIDC publisher; this is separate from interactive publication with the owning personal account. Both paths restrict the target to the approved Space and require its selected visibility to match the explicit build/publication mode.
The workflow serializes publications and does not cancel an upload in progress. Before uploading, it verifies that anonymous requests can read the existing Protected website but cannot read raw files from the Hub repository; it repeats these checks after publication. A repository visibility flag alone does not pass the checks. The website includes the approved source downloads even though Hub Git access remains private. This workflow builds website files only and never starts cluster training or evaluation jobs. It rechecks that the candidate is the current main commit and has successful push checks before publishing, so an older queued run cannot replace a newer source snapshot. After upload it verifies remote repository file identities and the public served manifest/assets/source files and denial of anonymous Hub raw-file access, then writes a deployment receipt and build manifest to a seven-day Actions artifact. Protected App verification is part of successful CI deployment. The static bundle passed between jobs is retained for one day. It does not publish models, datasets, collections or profile cards.
For a retry after fixing a transient hosting error or configuring the publisher, open Actions → Deploy Hugging Face site → Run workflow and select main. Manual runs require successful push checks for that same current commit. A deployment failure does not change source history; inspect that workflow's log and receipt before retrying. Automated source/build checks do not replace the recorded browser and model-quality qualification.
The earlier Private-mode companion-collection helper follows the official collection API. It requires a verified Private Space receipt and creates only a private list of links: this Space, upstream Qwen3-1.7B, MathOctopus and MGSM. Its notes distinguish training data from evaluation and record that the earlier MGSM64 subset was already inspected. It copies no upstream files and changes no public profile. The existing private collection is separate from the Protected website deployment.
