# NOTICE
This repository combines fine-tuned model weights, inference code and example
data from several upstream projects. The components keep their own licenses;
the repository as a whole is **not** relicensed under a single license.
## Model weights
| component | upstream | license |
|---|---|---|
| VRS-HQ fine-tuned checkpoint | [`SitongGong/VRS-HQ`](https://huggingface.co/SitongGong/VRS-HQ) | Apache-2.0 (model card) |
| Chat-UniVi language/vision backbone inside VRS-HQ | [`Chat-UniVi/Chat-UniVi`](https://huggingface.co/Chat-UniVi/Chat-UniVi) | Llama 2 Community License (model card: `llama2`) |
| SAM2 mask branch | [`facebookresearch/sam2`](https://github.com/facebookresearch/sam2) | Apache-2.0 |
| Sa2VA-14B fine-tuned checkpoint | [`ByteDance/Sa2VA-InternVL3-14B`](https://huggingface.co/ByteDance/Sa2VA-InternVL3-14B) | Apache-2.0 |
| InternVL3-14B / InternViT | [`OpenGVLab/InternVL3-14B`](https://huggingface.co/OpenGVLab/InternVL3-14B) | Apache-2.0 |
| Qwen2.5-14B language model | [`Qwen/Qwen2.5-14B`](https://huggingface.co/Qwen/Qwen2.5-14B) | Apache-2.0 |
Because the VRS-HQ checkpoint contains Chat-UniVi parameters, it is
distributed with the Llama 2 Community License obligations. The license text
is included at `LICENSES/Llama-2-Community-License.txt` and is also available
at . **Built with Meta Llama 2.** The
Llama 2 acceptable use policy applies.
Apache-2.0 components are covered by `LICENSES/Apache-2.0.txt`.
## Code
| path | origin | license |
|---|---|---|
| `Sa2VA-14B/checkpoint/*.py` | Sa2VA release (remote code) | Apache-2.0 |
| `Sa2VA-14B/test.py` | this release | Apache-2.0 |
| `VRS-HQ/test.py` | this release | Apache-2.0 |
| `VRS-HQ/code/model/llava`, `code/model/univi` | LLaVA / Chat-UniVi | Apache-2.0 |
| `VRS-HQ/code/model/segment_anything`, `code/sam2` | Segment Anything / SAM2 | Apache-2.0 |
| `VRS-HQ/code/model/VISA_multiseg.py`, `code/dataset`, `code/collate.py` | VRS-HQ / VISA | **no license declared upstream** |
The VRS-HQ GitHub repository () does not
contain a LICENSE file, and neither does the VISA repository it derives from.
Those files are included only so that the published checkpoint can be loaded
and the paper's inference protocol reproduced. They remain the property of
their authors; no license is granted here. Users who need to redistribute
that code should obtain permission from the VRS-HQ authors. The weights
themselves are covered by the upstream model-card license above.
`openai/clip-vit-large-patch14` is downloaded from the Hugging Face Hub at
runtime and is not redistributed in this repository.
## Example data
`examples/cases/` contains six test rows from the public SonoIntent benchmark.
Images and masks are byte-identical to
[`QingboKang/SonoIntent-Annotations`](https://huggingface.co/datasets/QingboKang/SonoIntent-Annotations)
and retain the source licenses:
- **HC18** — CC BY 4.0, Zenodo record 1327317 v2. Cite van den Heuvel et al.,
*PLoS ONE* 13(8):e0200412 (2018), doi:10.1371/journal.pone.0200412.
- **PSFHS (FH-PS-AOP)** — CC BY 4.0, Zenodo record 10969427 V1. Cite Lu et al.,
*Scientific Data* 11:436 (2024), doi:10.1038/s41597-024-03266-4.
## Medical disclaimer
These models and examples are research artifacts. They are not medical
devices, have not been clinically validated, and must not be used for
diagnosis or treatment decisions.