VideoChat3-LV116K

VideoChat3-LV116K is the long-video instruction data used by VideoChat3. It is designed to complement short academic video data with supervision over longer temporal contexts, where evidence can be sparse, delayed, and distributed across multiple video segments.

The dataset is constructed through a long-video synthesis pipeline. Candidate long videos are filtered for visual quality, semantic content, and temporal coherence. Videos are then split into manageable temporal segments, each segment is annotated and checked for quality, and the validated segment descriptions are assembled into full-video supervision. Based on this evidence ledger, the pipeline synthesizes long-video captions, long-video question answering data, and temporal grounding annotations.

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-LV116K use the following video datasets:

Source dataset	Category	Original video dataset path
CinePile	Movie Summarization & QA	https://huggingface.co/datasets/tomg-group-umd/cinepile
LongVideoDB	Long Video Timelines / Temporal Grounding / Summarization & QA	https://huggingface.co/datasets/LongVideos/LongVideoDB-373K-Videos
SciVideo_Long	SciVideo Summarization & QA	https://huggingface.co/datasets/MCG-NJU/VideoChat3-LV116k/tree/main/SciVideo_Long
SciVideo_Short	SciVideo Summarization & QA	https://huggingface.co/datasets/MCG-NJU/VideoChat3-LV116k/tree/main/SciVideo_Short

The lv116k.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 long-video datasets used by the annotations.