--- license: cc-by-nc-sa-4.0 language: - en - zh base_model: Qwen/Qwen3.5-4B-Base tags: - image-classification - decision-model - multimodal - lora - custom-code --- # JEVision Give JEVision an image, a question, and possible answers. It returns a probability for each answer. - **Image + text input.** Designed for fast visual decisions. - **About 0.4 seconds:** median 356 ms on our 100-request local test, using an RTX 5060 Ti 16 GB. Loading, network time, and queues are excluded. - **135 MB download is the adapter and decision head.** The setup below also downloads the required Qwen3.5-4B base model. ## Run it Requirements: Linux or Windows WSL2, Python 3.10+, Docker with NVIDIA GPU support, and an NVIDIA GPU. Tested on an RTX 5060 Ti 16 GB. First setup downloads several GB and takes time. Run these commands in a Linux/WSL terminal: ```bash python3 -m venv .venv source .venv/bin/activate python -m pip install --upgrade huggingface_hub hf download heack/JEVision --local-dir jevision cd jevision docker build -t jevision-local . docker run --rm -v "$PWD:/release" -w /release jevision-local download_base.py --root /release/data docker run --rm --gpus all -v "$PWD:/release" -w /release jevision-local example.py --root /release/data --checkpoint /release ``` The example creates three colored images and checks all nine predictions. A successful run exits without an error and saves `verification.json`. To test your own picture, copy it into the `jevision` folder as `photo.png`, then run: ```bash docker run --rm --gpus all -v "$PWD:/release" -w /release jevision-local example.py --root /release/data --checkpoint /release --image photo.png --question "Which animal is visible?" --choices cat,dog,bird ``` Choose answers that match your question. The output contains each answer's probability. Keep the `data` folder to avoid downloading the base again. ## Results Internal evaluation: **94.14% accuracy over 2,423 questions** from 900 labeled states. These are internal results, not an independent benchmark. The Docker runtime and image examples passed on the tested GPU. [Evaluation](evaluation.json) · [Usage verification](verification-clean.json) · [Technical details](DETAILS.md) ## Online API No hosted endpoint is available yet. Hugging Face rejected this account's ZeroGPU deployment request with a PRO subscription requirement. The commands above run the model locally. ## License Noncommercial use: [CC BY-NC-SA 4.0](LICENSE). Commercial use: **US$1,000 once for a perpetual license to this model's purchased version**. Each model is licensed separately. Hosting and API usage are not included. Upstream licenses and prior valid rights remain in effect.

Select USD and include `heack/JEVision`, your Hugging Face username, and the licensee's name with your payment. [Read the commercial terms](COMMERCIAL-LICENSE.md).