# Third-party and release notices

This model is a full-parameter fine-tune of **SEDD-medium** from Aaron Lou,
Chenlin Meng, and Stefano Ermon:

- Model repository: <https://huggingface.co/louaaron/sedd-medium>
- Pinned revision: `ce71a3c6178b50e899c8be1a1a4c13130308e54f`
- Base weight SHA-256:
  `d93bb0dd1013295a4865848ea546ee3763a5be036cf55ea407e898c0a7a82a33`
- Base config SHA-256:
  `f77926529f6ee5b913981775ebd2f8560c271341ad6dc67a0eab4c56eb032ec5`

The bundled implementation is derived from
**Score-Entropy-Discrete-Diffusion**. Its upstream MIT copyright and license
text are retained in `LICENSE`. See `MODEL_LICENSE.md` for the separate model
weight scope clarification.

The training-data generator was Google's **Mathematics Dataset** at commit
`427f45075f84b8b9774950196ad63867ca20ffb3`; that project is distributed under
Apache License 2.0. The generator source and full training data are not bundled
here. The configured GPT-2 tokenizer is also an external artifact and is not
redistributed by this model repository.

Release-specific loader and experiment code:
<https://github.com/Jrffy666/SEDD/tree/9b26b8980b298ce5cf08013320354ad4141796c1>

This release deliberately includes only the audited online inference weights.
It excludes the raw training checkpoint, optimizer, EMA, RNG state, private
paths and device identifiers, complete training data, and sealed evaluation
examples.