libo200404/VQA-UM

UMBench

English and Chinese multiple-choice visual question-answering data for urban-management imagery. The two language configurations share the same images and aligned question identifiers.

面向城市管理图像的中英文多项选择视觉问答数据。两种语言配置共用图片，并通过一致的题目编号逐题对齐。

Release contents / 本版本内容

This release provides 581 image files and 3305 image-backed questions per language. The full supplied annotations are also preserved: they contain 2,761 questions per language referencing 588 image filenames. Seven referenced images are unavailable in this release, so their 33 questions are excluded from the loadable en and zh configurations and recorded separately. No original annotation records are deleted.

本版本提供 581 个图片文件，每种语言各 3305 道有对应图片的题目。同时保留完整原始标注：每种语言各 2,761 道题，引用 588 个图片文件名。其中 7 张图片本版本未提供，对应的 33 道题未纳入可加载的 en、zh 配置，而是单独记录；原始标注记录全部保留。

Item / 项目	Count / 数量
Image files included / 提供的图片文件	581
Distinct image contents by binary comparison / 按文件内容去重后的图片数	549
Questions in each loadable language configuration / 每种可加载语言配置的题目数	2,728
Rows whose answer belongs to the available options / 每种配置中答案属于该题选项的题目数	2,723
Questions per language in the original annotations / 每种语言的原始标注题目数	2,761
Image filenames referenced by the original annotations / 原始标注引用的图片文件名数	588
Missing referenced images / 本版本缺失的被引用图片	7
Questions excluded from each configuration because of missing images / 每种配置因缺图排除的题目数	33

English and Chinese are two language versions of the same questions, not additional independent samples. Each configuration has one test split; no training or validation partition is defined.

中英文是同一组题目的两种语言版本，不应重复计为额外独立样本。每种配置均使用单一 test 划分，未另设训练集或验证集。

Files / 文件组织

Path / 路径	Contents / 内容
metadata_en.jsonl	English records with available images, one question per line / 有对应图片的英文标注，每行一道题
metadata_zh.jsonl	Chinese records with available images, one question per line / 有对应图片的中文标注，每行一道题
images/	Shared image files / 中英文共用图片
original/annotations_en.json	Complete original English annotations / 完整原始英文标注
original/annotations_zh.json	Complete original Chinese annotations / 完整原始中文标注
quality/missing_image_questions.jsonl	The seven missing image filenames and their 33 affected questions / 7 个缺失图片文件名及其对应的 33 道题

The original JSON files retain their supplied structure. Explicit configuration paths keep them separate from the loadable JSONL records. Image filename 51 was normalized to 051 to match its annotation reference; the image content was not changed.

原始 JSON 保留收到时的结构。加载配置显式指定 JSONL 路径，避免与原始 JSON 重复读取。图片文件名中的 51 已规范为 051，以匹配标注引用；图片内容未改动。

Record fields / 字段说明

Field	Type	Meaning / 含义
question_id	string	Six-digit identifier aligned across languages, e.g. 000001 / 中英文共用的六位题目编号
image_id	string	Original image filename stem, preserving leading zeros, e.g. 001 / 原图片文件名去掉扩展名，保留前导零
file_name	string	Image path relative to the metadata file, e.g. images/001.png / 相对于标注文件的图片路径
question	string	Question text in the selected language / 当前语言的题干
choices	string	Original options retained as a single string / 原始选项文本，保留为单个字符串
answer	string	Original answer annotation / 原始答案标注
task_ids	list of strings	Original task labels; a question may have multiple labels / 原始任务标签，一题可对应多个标签
answer_label	string	Single uppercase A–Z label after whitespace/period normalization; empty if invalid / 仅规范空白和句点后的单个大写答案字母；格式无效时为空字符串
answer_valid	bool	Whether the normalized answer matches an option label for this question / 规范后的答案是否属于该题选项
image_available	bool	Whether the referenced image is included; true for every row in the loadable configurations / 引用图片是否已提供；可加载配置中均为 true

question_id values are retained when missing-image records are excluded, so identifiers may have gaps. Join the language versions by question_id. ImageFolder converts file_name to an image feature during loading. answer_valid checks whether the normalized answer is one of the question’s option labels; it does not establish that the answer is visually correct.

排除缺图记录时保留原 question_id，因此编号可能不连续。中英文应通过 question_id 对齐。ImageFolder 加载时会将 file_name 转换为 image 特征。answer_valid 检查规范后的答案是否属于该题的选项，不代表答案已通过图像事实核验。

Data quality and evaluation / 数据质量与评测口径

The release preserves the supplied answers and task labels. It does not replace uncertain annotations with guessed answers. The following records remain available for inspection; locations use image filename stem + 1-based question index within that image, rather than the global question_id:

本版本保留原始答案和任务标签，不以猜测补写不确定标注。以下记录仍保留，便于核查；位置采用图片文件名去掉扩展名 + 图片内从 1 开始的题目序号，并非全局 question_id：

• Answers outside the available option labels / 答案不属于该题选项： 065#3, 093#3, 192#5, 529#1, 573#3. All five rows have answer_valid=false. Four retain D even though their options end at C; the 3. answer in 573#3 has an empty answer_label / 五条记录均为 answer_valid=false。其中四条保留了超出 A–C 选项范围的 D；573#3 的原始答案为 3.，其 answer_label 为空。
• Empty task labels / 任务标签为空： 104#1.
• Duplicate question entry / 重复题目： 476#3 is retained / 保留该条记录。
• Repeated image content / 图片内容重复： 29 groups contain binary-identical image files. Original image identifiers are retained, giving 581 files with 549 distinct contents / 共有 29 组图片文件内容完全相同；保留原图片标识，因此 581 个文件对应 549 种不同内容。

For answer-label accuracy on the supplied image-backed data, filtering with answer_valid leaves 2,723 rows per language. Report whether this filter is applied and how missing labels or duplicates are handled. Task IDs are multi-label, so a question may contribute to more than one task-specific count. Counts from the full original annotations and the image-backed configurations have different scopes and should not be mixed.

在本版本有图片的条目上计算答案字母准确率时，按 answer_valid 筛选后，每种语言剩余 2,723 条记录。报告结果时应说明是否采用此筛选，以及缺失标签、重复项的处理方式。任务 ID 为多标签，一题可能计入多个任务的统计。完整原始标注与可加载配置的统计范围不同，应分别报告。

Load with Hugging Face Datasets / 使用方式

The configuration uses metadata_filenames, supported by datasets 5.1.0. Local loading was tested with that version.

本配置使用 metadata_filenames，已通过 datasets 5.1.0 的本地加载验证。

pip install "datasets[vision]>=5.1.0"

from datasets import load_dataset

english = load_dataset("libo200404/VQA-UM", "en", split="test")
chinese = load_dataset("libo200404/VQA-UM", "zh", split="test")

assert list(english["question_id"]) == list(chinese["question_id"])
print(english["question"][0])
print(chinese["question"][0])

# Optional: retain rows with answers that match an available option label for scoring.
evaluation = english.filter(lambda row: row["answer_valid"])
assert len(evaluation) == 2723

sample = evaluation[0]
image = sample["image"]
prompt = sample["question"] + "\n" + sample["choices"]
gold_label = sample["answer_label"]

Use the selected language's question and choices together. The answer and answer_label fields are scoring targets, not part of the input prompt.

构造输入时应搭配使用同一种语言的 question 与 choices。answer 和 answer_label 用作评分目标，不应加入模型输入提示。

License / 许可

The dataset card retains the repository's existing MIT license declaration.

本数据卡沿用仓库已有的 MIT 许可声明。