VisionJEV
VisionJEV 1.0.0 is a 27B multimodal decision model fine-tuned from Qwen3.8-27B. It answers typed questions over text and multiple images, returning probabilities in a single forward pass per question:
- Choice: select among supplied options.
- Noul: estimate the probability of a yes/no proposition.
- Score: return a probability distribution and expected value over ordered levels.
Training
VisionJEV is trained on curated, publicly available text and multimodal data. Data preparation includes checks for overlap with evaluation samples, quality filtering, and conversion into a unified decision format. The model is then trained with full-parameter supervised fine-tuning.
Public Decision Index 0.3
| Panel | VisionJEV |
|---|---|
| Text | 66.46 |
| Vision | 76.77 |
Usage
Inference code is included in inference/. See the installation and usage instructions for the Python API, HTTP service and Decision Index evaluation.
License
- Downloads last month
- 117
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for taoszhang/VisionJEV
Base model
Qwen/Qwen3.8-27B