JaWildText

JaWildText is a Japanese scene text understanding benchmark for evaluating vision-language models (VLMs) on text-rich real-world images. It is designed to diagnose Japanese OCR, scene text visual question answering, and structured information extraction in practical conditions such as dense signboards, handwritten text, and mobile-captured receipts.

For details on the dataset construction, evaluation protocol, and baseline results, see the paper: JaWildText: A Benchmark for Vision-Language Models on Japanese Scene Text Understanding.

Dataset Summary

Japanese scene text poses challenges that are not fully captured by existing multilingual or document-centric benchmarks, including mixed writing systems, vertical writing, dense layouts, and a large character inventory. JaWildText provides three complementary evaluation tasks:

• Dense STVQA / Board VQA: question answering over text-rich signboards, posters, bulletin boards, and product/package images.
• Receipt KIE: key information extraction from real-world photographs of Japanese receipts.
• Handwriting OCR: page-level transcription of handwritten Japanese text across different writing media and directions.

The dataset contains 3,241 examples across the three task configurations. The default configuration is the union of all task subsets.

Configurations

Configuration	Split	Examples	Description
default	train	3,241	All examples from the three task subsets
board_vqa	train	1,025	Dense scene text visual question answering
handwriting_ocr	train	1,065	Handwritten Japanese OCR
receipt_kie	train	1,151	Receipt key information extraction

Although the dataset is distributed with a train split for compatibility with Hugging Face Datasets, JaWildText is intended as an evaluation benchmark.

Usage

from datasets import load_dataset

board_vqa = load_dataset("llm-jp/jawildtext", "board_vqa")
sample = board_vqa["train"][0]

image = sample["image"]
question = sample["question"]
answer = sample["answer"]

To load all examples:

from datasets import load_dataset

dataset = load_dataset("llm-jp/jawildtext", "default")

Data Fields

All configurations share a common schema. Fields that are not used by a particular task may be null, empty strings, or empty lists.

• subset: subset name, such as board_vqa, handwriting_ocr, or receipt_kie.
• image_id: image identifier.
• filename: original image filename.
• image: input image.
• polygons: text-region annotations.
  • polygon_id: unique text-region identifier within the example.
  • polygon: polygon coordinates in image coordinates.
  • text: annotated text string.
  • direction: writing direction annotation when available.
• question: question text for Dense STVQA / Board VQA examples.
• answer: reference answer for Dense STVQA / Board VQA examples.
• evidence: list of polygon_id values that provide the minimum textual evidence needed to answer the question.
• tool: writing medium/tool metadata for Handwriting OCR examples when available.
• writer_id: anonymized writer identifier for Handwriting OCR examples when available.
• fields: structured receipt fields for Receipt KIE examples.
  • store_name
  • store_address
  • receipt_id
  • date
  • time
  • total_amount
  • tax_amount
  • line_items

Each receipt field contains a value and polygon_ids. The polygon_ids values refer to entries in polygons.

Task Details

Dense STVQA / Board VQA

The board_vqa configuration evaluates whether a model can read and reason over dense Japanese scene text. Each example contains an image, a natural-language question, a reference answer, and evidence region IDs. Questions are designed to require integrating information from one or more text regions rather than simply copying a single visible string.

Receipt KIE

The receipt_kie configuration evaluates structured extraction from Japanese receipt images. The target output is a JSON object containing header fields such as store name, date, time, total amount, and tax amount, as well as line-item information where available.

Handwriting OCR

The handwriting_ocr configuration evaluates page-level transcription of handwritten Japanese text. The subset includes multiple writing media and writing directions, including horizontal and vertical Japanese text.

Evaluation

We follow the evaluation protocol described in the JaWildText paper.

For Dense STVQA / Board VQA, models are prompted to enclose the final answer in \boxed{...}. The extracted answer is evaluated with judge-based accuracy: an LLM verifier compares the model prediction with the reference answer and returns a binary correctness label. Outputs that cannot be parsed receive a score of 0.

For Receipt KIE, models are prompted to output a single JSON object following the predefined schema. Outputs that cannot be parsed as JSON receive a score of 0. We report overall F1 over extracted fields and line items, and field-level accuracy for major header fields.

For Handwriting OCR, models output plain text transcriptions. We compute character-level similarity as max(0, 1 - CER), where CER is the Levenshtein distance between the prediction and reference divided by the reference length. Unicode NFKC normalization is applied before scoring.

The overall score is the unweighted average of Dense STVQA accuracy, Receipt KIE F1, and Handwriting OCR character-level similarity.

Intended Uses

JaWildText is intended for:

• evaluating Japanese scene text understanding in VLMs;
• evaluating Japanese OCR in real-world image conditions;
• evaluating text-centric visual question answering over Japanese scene text;
• evaluating receipt key information extraction and document understanding;
• diagnostic analysis of recognition, reasoning, formatting, and script-specific errors.

Out-of-Scope Uses

JaWildText should not be used for:

• identifying individuals, writers, stores, or customers;
• inferring personal attributes or purchasing behavior;
• surveillance or profiling;
• treating visible trademarks, store names, or product names as endorsement by the rightsholders;
• training or deploying systems in high-stakes settings without additional validation.

Limitations and Ethical Considerations

JaWildText consists of real-world Japanese images and may contain store names, addresses, dates, prices, product names, logos, signs, and other third-party visual information. The dataset is released under Apache-2.0, but the license does not grant trademark rights or imply endorsement by any third-party entities visible in the images.

The dataset reflects images collected in Japan and should not be assumed to represent all Japanese text usage, all receipt formats, or all handwriting styles. Image quality, perspective, lighting, occlusion, and layout complexity vary across examples.

If you identify privacy-sensitive content or other issues in the dataset, please contact the maintainers.

TODO: add contact / takedown address.

License

JaWildText, including both annotations/metadata and images, is released under the Apache License 2.0.

Citation

If you use JaWildText, please cite:

@inproceedings{maeda-etal-2026-jawildtext,
  title     = {{JaWildText}: A Benchmark for Vision-Language Models on Japanese Scene Text Understanding},
  author    = {Maeda, Koki and Okazaki, Naoaki},
  booktitle = {Proceedings of the 20th International Conference on Document Analysis and Recognition (ICDAR)},
  year      = {2026},
  note      = {To appear}
}

References

• Paper: https://arxiv.org/abs/2603.27942
• Dataset: https://huggingface.co/datasets/llm-jp/jawildtext