zzlynxSG/SEA-Omni-Video-v1

SEA-Omni Video Benchmark (SEA-Video-v1)

Southeast-Asia-oriented video benchmark released with our EMNLP 2026 paper (see Citation). It captures culturally grounded multimodal cues and intra-sentential code-switching across English, Chinese, Malay, Indonesian and Thai; tasks cover video summarisation, audio QA, visual QA and multiple-choice QA.

3,142 annotated video clips across 4 tasks, 5 languages (en, zh, id, ms, th). Source videos are YouTube videos licensed under CC BY 3.0 and are used for evaluation only; this benchmark is intended for evaluation only.

Data Format

Parquet file (data.parquet) with columns:

Column	Description
question	Input prompt / task instruction
answer	Ground-truth reference answer
image	Empty for this benchmark
audio	Relative path to audio file (media/xxx.mp3)
video	Relative path to video file (media/xxx.mp4)
certainty	Answer certainty flag
task_type	One of: video_summary, video_aqa, video_vqa, video_mcqa

metadata_test.json contains YouTube URLs, durations, and language IDs for each video segment.

Task Distribution

Task	Count	Description
Video Summary	675	Summarize the video + audio content
Audio QA	641	Answer questions based on audio track
Visual QA	668	Answer questions based on visual content
MCQA	1,158	Multiple-choice QA (options embedded in question)

Loading

import pyarrow.parquet as pq
import json

# Load annotations
table = pq.read_table("data.parquet")
df = table.to_pandas()

# Load metadata
with open("metadata_test.json") as f:
    metadata = json.load(f)

# Access first row
row = df.iloc[0]
question, answer = row["question"], row["answer"]
video_path = row["video"]  # e.g., "media/segment_1_9dad688c.mp4"
audio_path = row["audio"]  # e.g., "media/segment_1_747af689.mp3"

# Filter by task
summary = df[df["task_type"] == "video_summary"]
mcqa = df[df["task_type"] == "video_mcqa"]

Citation

If you use this dataset, please cite the following papers:

@inproceedings{zhang2026culturally,
      title={Culturally-Aware Alignment for Multimodal Models via Group Reward Optimization for Southeast Asia},
      author={Longyin Zhang and Shuo Sun and Yingxu He and Cheng Yi Lewis Won and Bowei Zou and Ai Ti Aw},
      booktitle={EMNLP 2026 Award Nominee},
      year={2026},
}

@misc{zhang2026unlockingcognitivecapabilitiesanalyzing,
      title={Unlocking Cognitive Capabilities and Analyzing the Perception-Logic Trade-off},
      author={Longyin Zhang and Shuo Sun and Yingxu He and Won Cheng Yi Lewis and Muhammad Huzaifah Bin Md Shahrin and Hardik Bhupendra Sailor and Heng Meng Jeremy Wong and Tarun Kumar Vangani and Yi Ma and Qiongqiong Wang and Minh Duc Pham and Ridong Jiang and Jingtao Li and Jingyi Liao and Zhuohan Liu and Yanfeng Lu and Manas Gupta and Ai Ti Aw},
      year={2026},
      eprint={2602.23730},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.23730},
}