Weight provenance: is lingbot-map initialized from / distilled from VGGT's checkpoint, or trained from scratch?

[di-vad01]

Thank you for the great work and for releasing the code under Apache-2.0.

The README/model card lists VGGT and DINOv2 under 'builds upon several excellent open-source projects,' but it isn't clear from the README, model card, or LICENSE.txt whether this refers to:

(a) architectural inspiration only (LingBot-Map trained from scratch, unrelated model weights), or
(b) the released lingbot-map / lingbot-map-stage1 checkpoints being initialized from, fine-tuned from, or distilled using Meta's VGGT pretrained weights (which are released under a non-commercial research license) or DINOv2 weights, or
(c) any other dataset (ScanNet, Habitat, BlendedMVS, Co3D, etc.) that carries research-only/non-commercial terms.

Could you confirm:

1. Were the released checkpoints initialized from, fine-tuned from, or distilled from VGGT's or DINOv2's pretrained weights, or trained entirely from scratch?
2. What dataset(s) were used to train the released checkpoints, and do any of them carry non-commercial or research-only license terms?
3. Does the Apache-2.0 LICENSE.txt cover the released model weights as well as the code, or only the code?

This matters for evaluating commercial use of the released checkpoints. Happy to help update the README/model card with a short provenance note if useful. Thanks!

[LinZhuoChen]

Thanks for the question. The released checkpoints were not initialized from, fine-tuned from, or distilled from VGGT weights; we follow the VGGT architecture, and only the ViT backbone is initialized from DINOv2 (Apache-2.0). Training details and the full list of datasets are in our paper (Sec. 4.1–4.3): https://arxiv.org/abs/2604.14141. The Apache-2.0 license in LICENSE.txt covers both the code and the released model weights.