VennLicense audit notes#
These are maintainer working notes supporting the publication decision for the venn-1 head. They are not legal advice. User-facing attribution is in licenses.
Head training sources#
Decision, November 7, 2026: the project owner confirmed that the published
venn-1 head can be redistributed. The public model card and license page now
record that permission. The source review below remains as historical context;
it is not an unresolved release blocker.
The venn-1 generic head was trained on CLINC, MASSIVE, Banking77, DBpedia, Yahoo Answers and AG News. The current source cards declare:
| Dataset | Declared license | Notes |
|---|---|---|
| CLINC | CC BY 3.0 | |
| MASSIVE | CC BY 4.0 | |
| Banking77 | CC BY 4.0 | |
| DBpedia | CC BY-SA 3.0 | |
| Yahoo Answers | Unknown | No license metadata on the source card |
| AG News | Unknown | The card describes research/noncommercial use; its formal licensing section is unspecified |
Data revisions and license-card revisions are pinned separately. The converted parquet commits used for training omit source cards, so these card findings do not by themselves establish the permissions in force at training time, and they do not clear any derived head.
Reasoning#
Missing dataset metadata is not an established prohibition on distributing Venn heads, and dataset licenses do not automatically become weight licenses. This audit has found no explicit ban on distributing these heads. Retraining is not required solely because a card omits a license; assess the actual applicable terms before a weight publication decision. The AG News research/noncommercial wording is a specific training/use question to assess, not proof of a blanket derived-weight ban.
The source-only geometric benchmark needs no Venn head weights, so its publication does not depend on this review.
Other corpora#
SLURP text is CC BY 4.0, but its audio is CC BY-NC 4.0 according to upstream. ESC-50 is noncommercial. Other media and adapter corpora have individual terms. Their corpora, embeddings and fitted weights are excluded from the public release. Yelp's custom terms have not been cleared for distributing the ordinal evaluation head.