Nib — licence and credits

Training text

Nib reads two things. English Wikipedia articles as they stood in 2013 or earlier (frequently edited articles come from an October 2011 snapshot, the rest from revisions dated before 2014), written by Wikipedia contributors and licensed under CC BY-SA 3.0. And about 300 public-domain books from Project Gutenberg, all by authors who died before 1928. The exact files and their hashes are on the recipe page.

Weights

The model weights are released under CC BY-SA 4.0, with credit to Wikipedia contributors and Project Gutenberg. Every checkpoint is published with its SHA-256 and can be downloaded from the verify page. The code is in the project repository.

Plain statements

• Unofficial fan project. Not affiliated with, endorsed by, or connected to Rockstar Games or Take-Two Interactive. No marks, art, fonts or characters of theirs are used.
• Not the first baby AI. Others exist; this one is just small, open and verifiable.
• A small model (about 34 million parameters), often wrong, and it may say odd things. Its training data is old encyclopaedia text and old books; it knows nothing after 2013.
• The token has no promised value, no revenue share, no buybacks and no voting. Any holder-only features are a convenience: the weights are open to everyone.
• Training runs on one laptop. When it is offline the site says "napping" and nothing is simulated.
• No date is promised for coherent speech. The checkpoints and their fixed-prompt answers are the only measure.

Software

Trained with MLX; runs in the browser with transformers.js and ONNX Runtime Web; text cleaned with WikiExtractor; hashes posted with the Solana memo program. Fonts: Caveat and Nunito (Google Fonts, OFL). Avatar and mobile drawings are original.