dhad-vision / NOTICE.md
Alshami's picture
dhad-1-vision-preview: stage-1 connector + unfinished stage-2 step675 snapshot (Apache-2.0)
e3f7919 verified
|
Raw History Blame Contribute Delete
1.83 kB
# NOTICE — licences and third-party components
This repository ships **only** Dhad's own connector weights (`connector.pt`), its
architecture/inference code (`code/`), and the unfinished stage-2 snapshot. It does
**not** re-upload the base models, the frozen encoder, or any training data.
## Shipped here (Dhad) — Apache License 2.0
| File | Copyright | Licence |
|---|---|---|
| `connector.pt` (connector + markers) | Dhad / dhadlab | Apache-2.0 (see `LICENSE`) |
| `code/connector.py`, `code/graft.py`, `code/infer.py`, `code/encoder.py` | Dhad / dhadlab | Apache-2.0 |
| `stage2-step675/*` (unfinished) | Dhad / dhadlab | Apache-2.0 |
## Referenced, not shipped
| Component | Use | Licence |
|---|---|---|
| `inception42/Jais-3-0.9B` | frozen language trunk | Apache-2.0 (declared in its model card; the repo has no LICENSE/NOTICE file) |
| `google/siglip2-base-patch16-naflex` | frozen vision encoder | Apache-2.0; cite [arXiv:2502.14786](https://huggingface.co/papers/2502.14786) |
| `FreedomIntelligence/ALLaVA-4V` | training images (not shipped) | CC-BY-NC-4.0 (card); images are LAION web crawl, rights not owned by the uploaders, per-image licence unknown |
| `FreedomIntelligence/ALLaVA-4V-Arabic` | training captions (not shipped) | Apache-2.0 (card); machine translation of GPT-4V captions |
## Notes
- We redistribute only a trained connector. No images, captions, or SigLIP2 feature
files are included.
- The upstream training-data terms are stated above for provenance. If you
redistribute this model or build on it, review those upstream terms yourself.
- The unfinished stage-2 LoRA targets the Jais trunk; it is published only as an
explicitly-labelled unfinished preview.
The project's full licence audit (`docs/LICENSES_AUDIT.md`) is not published here.
Nothing in this notice is legal advice.