ADR-0003: marola-ml is MIT again, and the model is marola-gpt¶
| Status | Proposed (2026-10-14), by the owner's order |
| Supersedes | the licence of ADR 0002 (its decision 3), and the name marolaGPT |
Context¶
The owner, 14 October 2026: "change license of marola-ml to MIT, so you can distribute with no problem. Dont do any minting for now, change the name to marola-gpt not marolagpt, name this prepare for 0.0.1 release on hugging face, update README.md license".
Decision¶
- MIT.
LICENSEis the MIT License, Copyright (c) 2026 marola.dev, byte for byte the one v0.1.0 was released with. It covers the code and the model it trains.sea/nanogpt/keeps nanoGPT's own MIT License. - The name is marola-gpt, everywhere: the recipes (
just marola-gptand its steps), the tags (marola-gpt-vX.Y.Z), the release directory (sea/release/marola-gpt-<version>/), the Hugging Face repository (h0ffmann/marola-gpt) and the model card. - The first release is 0.0.1.
just sea-releasecounts from 0.0.0, so its first patch ismarola-gpt-v0.0.1. - Public on Hugging Face.
sea/publish_hf.pycreates the repository public;--privateis the exception. - No DOI for now. marola-ml#56 waits.
Consequences¶
- The restricted licence was on
mainfrom #53 to this change, and no release was made under it: the only release is v0.1.0, MIT. The question of relicensing contributors' code is gone. - ADR 0002's chain (the recipes, the release directory, the publish workflow) stands, under the new name; its licence and its "no distribution outside marola" consequences do not.
marola-gpt-publish.ymlstill keeps no workflow artifact; the reason now is only that a release belongs on Hugging Face, not in a run's storage.