SenseNova-U1-8B-SI800K-SFT
sensenova/SenseNova-U1-8B-MoT-SFT
further supervised-fine-tuned on SenseNova-SI-800K for spatial understanding.
This is an understanding-focused SFT: only the language/understanding branch is optimized on the spatial corpus. The image-generation branch of the unified NEO-unify architecture is inherited unchanged from the base model (not further trained here).
Details
| Base | sensenova/SenseNova-U1-8B-MoT-SFT (NEO-unify, NEOChatModel) |
| Data | SenseNova-SI-800K — 832K image+text spatial-intelligence samples (depth / distance / size / relative position / multi-view spatial relations) |
| Objective | understanding-only SFT (generation branch untouched) |
| Steps | ≈183K (8×H100; lr 2e-4 constant; seq len 20480) |
| Precision | bfloat16 |
Built with the SenseNova-U1 training stack; recipe summarized above.
Usage
The model uses custom code (NEOChatModel); load with trust_remote_code=True.
See the base model card
for inference instructions — they apply unchanged.
License
Apache-2.0, inherited from the base model.