Resources

• Code: https://github.com/Hui-design/TSPO
• Paper: https://arxiv.org/pdf/2508.04369

Dataset Information

TSPO-10K: This repository contains the JSON annotations for our TSPO-10K dataset.

Video Data Source

The video data used in our project is from LLaVA-Video-178K, which can be downloaded via: https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K. You don't need to download the llava_hound video inside it.

Video Needle-in-a-Haystack Dataset

As described in our paper, the Video Needle-in-a-Haystack dataset contains synthesized long videos (10-60 minutes). These long videos are composite results obtained by stitching together target videos and irrelevant videos. Implementation Note: The video synthesis is implicitly implemented through code rather than being an explicit video dataset. The relevant code can be found at: https://github.com/Hui-design/TSPO/tree/main/src/open_tspo/trainer/tspo_trainer.py#L469~#L480