MCG-NJU/VideoJev ? reverse-engineered prompt

Build me a Python project called VideoJev that can understand videos and answer questions about them.

I want it to take a video file and a few prompts, then do two things. First, find the time range where a described moment happens in the video. Second, pick the right answer from a small set of choices about what is happening in the video. It should return clean JSON with the result, including the time interval for grounding and the chosen answer for multiple choice.

Please include a simple demo I can run locally, plus the code to load a pretrained model from Hugging Face, run inference on a video, and print results. Also add training and evaluation scripts, basic config files, and tests so I can verify everything works.

If you need to check current library details or model loading examples, look up the latest docs online. Keep the project practical and easy to run with clear setup instructions.