mcg-nju/videojev is the public repository for VideoJev, a system model for video understanding and interaction. FACTS records it as created on October 8, 2026, and there is no publicly available latest release. The focus is therefore not on the changes of a specific version but on the model and approach presented in the repository. Instead of generating answer sentences, VideoJev scores candidates directly to handle video questions.
This approach applies to temporal grounding and multiple-choice video question answering. Temporal grounding is the task of finding the segment of a video that corresponds to a question or description. Multiple-choice question answering means picking the answer among the given options. The README introduces VideoJev-4B and VideoJev-35B-A3B and explains data preparation and training, as well as inference and evaluation methods. Handling both problems through candidate scoring without any text generation is the central feature of this repository. The repository presents the related results tables together with the paper and includes a papers with code badge. The project is written in Python and uses the Apache-2.0 license.
That said, the performance numbers are not results reproduced and confirmed directly in the repository. The results tables are provided as images in the paper, and run verification was not confirmed either. The public model configuration and evaluation materials therefore give clues about the approach, but it is difficult to judge actual performance independently from this information alone.