PawChrona-MMQA is an English–Chinese multimodal question-answering dataset for companion-animal health and behavior. It contains 49,247 questions across images, audio, video and text, with separate cat and dog views.

PawChrona-MMQA 是面向宠物健康与行为的中英双语多模态问答数据集，包含 49,247 道题目，涵盖图像、音频、视频和文本，支持分别筛选猫、狗数据。

Dataset / 数据组成

Configuration / 配置	Questions / 题目数	Modality / 模态
`image_vqa`	39,360	Image / 图像
`audio_video_qa`	9,247	Audio and video / 音频与视频
`text_qa`	640	Text / 文本

Species: 34,425 dog · 13,775 cat · 30 mixed · 1,017 unspecified. Breed information is included when reported by the source.

物种分布：狗 34,425 · 猫 13,775 · 混合 30 · 未明确 1,017。来源提供品种信息时一并保留。

Questions / 题型

The primary question formats are 35,864 four-option MCQs, 13,299 open-ended questions and 84 five-option MCQs. Questions, answers and options are paired in English and Chinese. Original open-ended versions are also available through the open_question_* and open_answer_* fields.

主要答题形式为 35,864 道四选一、13,299 道开放题和 84 道五选一。问题、答案及选项均提供中英双语；原始开放题版本保存在 open_question_* 和 open_answer_* 字段中。

Topic / 内容类别	Total / 总数	Train	Dev	Held aside / 单独保留
Image observation and interpretation / 影像观察与判读	10,323	8,851	500	972
Disease categories and veterinary knowledge / 疾病类别与兽医知识	15,046	14,339	500	207
Pet health consultation / 宠物健康咨询	852	333	500	19
Behavior and expression / 行为与表情	20,013	19,202	500	311
Imaging basics and quality / 成像基础与质量	3,013	2,209	500	304

Subsets and splits / 子集与分区

image_vqa, audio_video_qa and text_qa are modality subsets. The five topic categories are stored in broad_category; they are a separate classification axis.

image_vqa、audio_video_qa、text_qa 按模态分组；五大内容类别保存在 broad_category，两者是不同的分类维度。

The train split contains 44,934 questions. The frozen dev split contains 2,500 questions, 500 per topic. Shared sources are allowed; recorded sample/case/subject links, duplicate media, measured near duplicates and concrete question issues determine exclusions. The other 1,813 questions are retained in the related_holdout configuration.

train 共 44,934 题；冻结的 dev 共 2,500 题，每类 500 题。允许同源，按已记录的样本／病例／主体关联、重复媒体、实测近重复及具体题目问题筛选。其余 1,813 题保存在 related_holdout 配置中。

Usage / 使用

from datasets import load_dataset, concatenate_datasets

repo = "PawChrona/PawChrona-MMQA"
names = ["image_vqa", "audio_video_qa", "text_qa"]
train = concatenate_datasets([
    load_dataset(repo, name=name, split="train", streaming=True, token=True)
    for name in names
])
dev = concatenate_datasets([
    load_dataset(repo, name=name, split="dev", streaming=True, token=True)
    for name in names
])
dogs = train.filter(lambda row: row["species"] == "dog")

Use cat, dog, mixed_cat_dog or unknown to filter species. Images, audio and video reuse the existing embedded Parquet media. Split counts are summed across the three modality configurations.

物种筛选值为 cat、dog、mixed_cat_dog 或 unknown；图片、音频与视频复用已有Parquet内嵌媒体。分区总数为三个模态配置之和。

Choose membership with load_dataset(..., split="train") or split="dev". The row-level split column preserves the original shard marker; current membership is defined by the loader configuration and partition manifest.

请通过加载参数 split="train" 或 split="dev" 选择分区；行内 split 列保留原分片标记，当前归属以加载配置及分区清单为准。

Training split / 训练分区 · Development split / 开发分区

See DATA_DICTIONARY.md for the schema and metadata/hf_features.json for Hugging Face feature types.

字段说明见 DATA_DICTIONARY.md，Hugging Face 特征类型见 metadata/hf_features.json。

Sources and access / 来源与访问

The collection draws on published veterinary figures, pet-image datasets, audio/video collections and veterinary QA sources. Source links and attribution are retained in the metadata and third-party notices.

数据来自已发表兽医文献图表、宠物图像数据集、音视频数据及兽医问答来源。来源链接与署名信息保存在元数据和第三方来源说明中。

This version is a research preview. Research access is reviewed individually; commercial use is handled separately. Source materials retain their original terms. See research access, commercial use and rights information.

当前版本为科研预览版，科研访问逐人审核，商业用途单独联系。来源素材保留原有许可条件，详见科研访问、商业使用与权利说明。

Citation / 引用

Author / 作者：PawChrona Team · Version / 版本：2026.10.28-debug.3

@misc{zhao2026pawchrona_mmqa,
  author = {{PawChrona Team}},
  title = {PawChrona-MMQA: Bilingual Companion-Animal Multimodal Question Answering},
  year = {2026},
  publisher = {Hugging Face},
  version = {2026.10.28-debug.3},
  url = {https://huggingface.co/datasets/PawChrona/PawChrona-MMQA},
  note = {Research preview}
}

CITATION.cff · BibTeX

Downloads last month: 62

Total file size:

41.5 GB