GLaVE-1.2M

GLaVE-1.2M is the training dataset released with GLaVE-Cap: Global-Local Aligned Video Captioning with Vision Expert Integration. It contains fine-grained video captions, intermediate local annotations, scene-level annotations, and multiple-choice question-answer pairs for video understanding.

This repository contains annotations only. It does not redistribute the original video files. The original videos are available from lmms-lab/LLaVA-Video-178K. The data_source and video_source columns identify the corresponding source video.

• Paper: arXiv:2509.11360
• Code: HIT-leaderone/GLaVE-Cap

Dataset Summary

Item	Value
Video-level records	19,702
Raw Parquet shards	20
Videos per shard	1,000, except the final shard with 702
Raw Parquet size	1,105,941,274 bytes (about 1.106 GB / 1.030 GiB)
QA rows in the flattened QA release	1,747,018

The full processing output is provided rather than truncating the release to the number appearing in the dataset name.

Data Files

raw/
├── part-00000.parquet  # 1,000 videos
├── part-00001.parquet  # 1,000 videos
├── ...
├── part-00018.parquet  # 1,000 videos
└── part-00019.parquet  #   702 videos

Each Parquet row represents one video.

Schema

data_source: large_string
video_source: large_string
diff_caption: list<large_string>
detailed_caption: list<large_string>
local_caption: list<large_string>
overview_caption: large_string
scene_list: map<string, Scene>
caption: large_string
general_qa: map<string, QA>

Scene: struct<
  scene_hint: string,
  caption: string,
  QA_pair: map<string, QA>
>

QA: struct<
  Dimension: string,
  Question: string,
  Answer: string,
  Options: list<string>,
  Answer_choices: string
>

The released column names correspond to the generation records as follows:

Released column	Generation field
diff_caption	different
detailed_caption	attention
local_caption	merged

The internal frame_range field was removed from each scene_list entry. Scene IDs, scene hints, scene captions, and scene-level QA pairs are retained.

Loading the Dataset

PyArrow

import pyarrow.parquet as pq

table = pq.read_table(
    "hf://datasets/leaderonehit/GLaVE-1.2M/raw/part-00000.parquet"
)
print(table.schema)
print(table.slice(0, 1).to_pylist()[0])

Download a Snapshot

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="leaderonehit/GLaVE-1.2M",
    repo_type="dataset",
    allow_patterns="raw/*.parquet",
)

scene_list, QA_pair, and general_qa use the Parquet Map logical type. PyArrow returns a Map value as a list of (key, value) pairs; call dict(value) when a regular Python dictionary is preferred.

Source Information

The source columns follow the provenance metadata of LLaVA-Video-178K. Download the original videos from that dataset and use these two columns to locate the corresponding source file. For example:

data_source:  1_2_m_academic_v0_1
video_source: academic_source/Charades/293L4.mp4

License

All data and code are provided strictly for academic and non-commercial use. Please contact the authors if further clarification or permissions are needed. Users are also responsible for complying with the licenses and terms of the corresponding source videos.

Citation

@article{xu2025glavecap,
  title   = {GLaVE-Cap: Global-Local Aligned Video Captioning with Vision Expert Integration},
  author  = {Xu, Wan and Zhu, Feng and Zeng, Yihan and Guo, Yuanfan and Liu, Ming and Xu, Hang and Zuo, Wangmeng},
  journal = {arXiv preprint arXiv:2509.11360},
  year    = {2025}
}