VideoChat3-Academic2M

# VideoChat3-Academic2M VideoChat3-Academic2M is the academic video instruction data used by VideoChat3. It re-annotates public academic video datasets for video captioning, video question answering, and fine-grained motion understanding. The dataset follows an evidence-grounded annotation enhancement pipeline. Short answers, option-only labels, and concise captions are rewritten into richer instruction-following responses that mention visible objects, actions, scenes, temporal cues, and supporting evidence. A consistency filtering stage is then used to keep rewritten annotations aligned with the original academic labels. This repository provides JSONL annotation files. The original videos are not duplicated here; users should resolve videos from the original dataset paths listed below. ## Data Sources VideoChat3-Academic2M aggregates academic caption/QA sources including LLaVA-Video, Spoken-MIT, Vript, StarQA, Sports-QA, and Perception-Test. | Source dataset | Original video dataset path | | --- | --- | | LLaVA-Video | https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K | | Spoken-MIT | https://arxiv.org/pdf/2105.04489 | | Vript | https://huggingface.co/datasets/Mutonix/Vript | | StarQA |https://github.com/csbobby/STAR_Benchmark | | Sports-QA | https://huggingface.co/datasets/HopLeeTop/Sports-QA | | Perception-Test | https://github.com/google-deepmind/perception_test | The `acadmic2M.json` file provides the mapping between dataset annotations and video sources. You can refer to it when organizing the dataset structure. ## Citation If you use this data, please cite VideoChat3 and the original video datasets used by the annotations.