Repository evidence
Hugging Face model API, exact repository file listing, model card, config files and repository license files. These are the default sources for factual model fields.
SOURCE POLICY · EVIDENCE STANDARD
Every published field should be traceable to a source, an explicit classification rule, or a clearly labeled derivation. Unknown values stay unknown.
EVIDENCE HIERARCHY
The registry prefers direct repository evidence, then publisher documentation, then transparent derivation.
Hugging Face model API, exact repository file listing, model card, config files and repository license files. These are the default sources for factual model fields.
Official model documentation, legal terms and organization pages are used when a field is defined outside the source repository.
Calculations made from verified inputs, such as weight-only memory estimates from parameter counts. Derived values are never presented as publisher claims.
FIELD → SOURCE
This mapping is the practical contract behind “field-verified”.
| Registry field | Primary evidence | Rule / limitation |
|---|---|---|
| Weight availability | Hugging Face API + exact repository file list | Recognized weight artifacts must be present; repository existence alone is not enough. |
| Weight files | Exact repository file list | Exact filenames are stored rather than inferred from a model name. |
| Formats / precision | Repository artifacts + API/config dtype evidence | Observed artifacts and dtype signals only; third-party conversions are not silently attributed to the source repository. |
| License | Model-card license metadata + repository LICENSE/licence file | The declared terms are recorded with their evidence URL when available. |
| Commercial use | Checked license terms | A structured classification of the checked terms; not legal advice. |
| Context window | config.json + explicit model-card statement | Structured config is preferred; explicit publisher text is used when config does not expose the value. |
| Parameters | Hugging Face metadata/config + existing publisher declaration | Used for display and derived memory calculations when sufficiently supported. |
| Base model / lineage | Declared model-card metadata | No parent model is invented from naming similarity. |
| Training recipe | Publisher model card + repository files | Presence/disclosure signal, not an independent reconstruction of training. |
| Training data | Publisher/model-card disclosure | “Not disclosed” means not found in the checked standard evidence, not that no disclosure exists anywhere. |
| Runtime support | Publisher tags, model card and repository artifacts | Source-derived compatibility signals; Open Model Weights does not imply that every runtime was independently executed. |
| Downloads / likes | Hugging Face model API | Popularity metadata; useful for discovery, not a quality score. |
| Repository created / updated | Hugging Face model API | Repository creation is labeled as a release-date proxy where no separate structured release date exists. |
| Hardware memory | Derived from verified parameter count | Weight-only estimate. KV cache, activations, runtime overhead and sharding are excluded. |
VERIFICATION SEMANTICS
A published record has passed the verifier for its listed source repository. Verification is field-specific: one field can be verified while another remains “not disclosed”. The label does not mean every fact about a model has been independently reproduced.
Not disclosed means the checked standard repository/API/config/model-card evidence did not expose the value. It does not assert that the information exists nowhere else.
Latest registry verification run represented on this page.
Only candidates that pass the publication verifier count toward the public target.
FRESHNESS
The daily pipeline reads the current source-repository revision/SHA and fresh API metadata.
Prior field evidence is retained. The new revision check is recorded separately from the last full field verification.
The pipeline re-fetches the repository evidence needed for a full field-by-field verification.
Meaningful repository or field changes are added to the rolling change history.
At registry scale, repeatedly downloading unchanged model cards, configs and license files adds latency and rate-limit pressure without adding evidence. Revision-aware verification preserves freshness while keeping full verification meaningful.
Open change history →BOUNDARIES
Unknown or missing values remain explicit instead of being filled from naming conventions or assumptions.
A community conversion is not treated as the publisher’s own weight release unless the record explicitly identifies that repository as its source.
Hardware estimates are weight-only calculations, not claims that a model will run within that amount of VRAM/RAM.
Runtime support is source-derived unless Open Model Weights explicitly labels a runtime as independently tested.
License and commercial-use labels summarize checked terms for comparison; users remain responsible for legal review.
Downloads and likes are discovery signals only. They do not become a model-quality or safety score.
REPRODUCIBILITY
The public registry, versioned JSON layer, model schema, OpenAPI description and verification methodology are exposed so researchers and developers can inspect the record structure directly.
REPOSITORY IDENTITY
At large scale, some indexed repositories are maintained directly by model publishers and others can be community-maintained repositories. Open Model Weights therefore describes the indexed Hugging Face URL as the source repository unless publisher ownership is separately established. Organization location, when shown on developer pages, is likewise an organization fact and not a nationality assigned to a model.