RFD 1141: Publishing artifacts

Problem

Code goes to GitHub, and the goal manifest places every repository on a side of the hexagon. Artifacts had no such rule. A trained adapter, a rendered corpus and a set of vertex weights are not code: too large for a source repository, regenerated rather than edited, and carrying the licence of whatever they derive from as well as our own. Left on a desk they go with the desk. Pushed into a source repository they bloat a history nobody can shrink again.

A checkpoint also lies about its own size: the trainer writes the whole model, so a 19.5 MiB adapter arrives as 14.8 GiB of mostly somebody else’s Apache-2.0 base weights.

Decision

Artifacts go to Hugging Face. Code stays on GitHub, and each names the other. A dataset repository holds corpora; a model repository holds weights, in safetensors, carrying only the tensors that were trained.

Every artifact repository carries a CITATION.cff naming its title, its licence, and every source it derives from. An artifact without one cannot answer what it is made of, which is condition 1.

The name matches the source repository, and the README names the side it sits on, so a reader who finds the artifact first reaches the code.

References

  • SKILL.md is the procedure. DETAILS.md says what is not published.

Details

The measurements and the retractions