TimeLens2-93K
TimeLens2-93K is a large-scale, long-video temporal grounding dataset. This release contains 23,793 videos and 93,232 text–temporal interval pairs, including 12,091 multi-span pairs. The videos range from short clips to nearly 100 minutes and cover broad web domains such as entertainment, education, sports, news, science and technology, gaming, travel, vehicles, music, and daily life.
TimeLens2-93K offers a rare combination of scale, long-context coverage, multi-span supervision, and explicit label verification. We hope it serves as a strong foundation for training generalist video models that can search long timelines, return auditable evidence, and transfer across grounding instructions instead of merely memorizing one timestamp format.
Paper
TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs
Highlights
- Long and diverse videos: from short clips to nearly 100-minute videos across a wide range of real-world domains.
- Multi-span by design: annotations can contain a variable number of temporal intervals when evidence recurs at disjoint moments.
- High-confidence labels: cross-agent agreement, semantic verification, and boundary-focused refinement form complementary quality-control stages.
- Training-ready formats: use clean raw annotations or diverse multi-turn SFT conversations with varied instructions, answer syntax, and timestamp representations.
Files
| Path | Description |
|---|---|
videos/videos-xxxxx-of-00018.tar | Video archives split into 18 shards |
TimeLens2-93K_preview.jsonl | Viewer-friendly join with video ID, annotations list, and messages list |
TimeLens2-93K_raw_annotations.jsonl | Raw text queries and their temporal intervals |
TimeLens2-93K_conversations.jsonl | Multi-turn SFT conversations with diverse query and answer formats |
Annotation formats
Raw annotations
Each line contains one video and all of its text–timestamp annotations:
{"video_id":"-IA08sFr96U","annotations":[{"text":"Two men alternate singing on stage while holding microphones","timestamps":[[28,190]]}]}
Conversations
Each line contains one video and a multi-turn conversation. Questions and answers use diverse natural-language and timestamp formats:
{"video_id":"-IA08sFr96U","messages":[{"role":"user","content":"Retrieve the video segment timestamps matching: 'Two men alternate singing on stage while holding microphones'.\nReturn the result as an array of [start, end] pairs in seconds."},{"role":"assistant","content":"[[28.0, 190.0]]"}]}
Loading annotations
from datasets import load_dataset
preview = load_dataset("MCG-NJU/TimeLens2-93K", split="train")
raw = load_dataset(
"MCG-NJU/TimeLens2-93K", "raw_annotations", split="train"
)
conversations = load_dataset(
"MCG-NJU/TimeLens2-93K", "conversations", split="train"
)
To download the video archives:
hf download MCG-NJU/TimeLens2-93K \
--repo-type dataset \
--include "videos/*.tar" \
--local-dir TimeLens2-93K
License
This dataset is released under the Apache License 2.0.
Citation
@misc{zhu2026timelens2,
title={TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs},
author={Yuhan Zhu and Changlian Ma and Xiangyu Zeng and Xinhao Li and Zhiqiu Zhang and Songze Li and Jun Zhang and Tianxiang Jiang and Yuandong Yang and Ziang Yan and Zikang Wang and Xinyu Chen and Haoran Chen and Shaowei Zhang and Limin Wang},
year={2026},
eprint={2607.17423},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.17423},
}