Read the JaSSeT Research Use Terms in LICENSE. Access requests are reviewed manually; dataset files are available only after approval. Research use does not exempt you from applicable source-service restrictions.

JaSSeT: Japanese Synthetic Scene-Text Image Dataset

JaSSeT contains 4,000 synthetic Japanese scene images, paired with requested image text, a question, reference answers, and localized textual evidence. It is designed for research on image-level OCR and visual question answering, including multi-task training on both tasks.

The dataset contains 3,500 training images and 500 development images. Each image is a 1,536 × 1,024 PNG with one question. The files include the image bytes; a separate image download is not required.

Requested text is a generation target, not an exhaustive verified transcript of the final image. The generation process can omit, alter, or add text. Evidence supporting the answer receives additional verification, but the annotations are not guaranteed to be error-free.

Load the dataset

Request access on this page and agree to the JaSSeT Research Use Terms. Requests require manual approval. After your request is approved, sign in locally with hf auth login before loading the dataset. The access form does not request contact details, and sharing of your Hugging Face username and email with the dataset authors is disabled. Hugging Face still manages accounts and access requests.

import json
from datasets import load_dataset

dataset = load_dataset("llm-jp/jasset", token=True)
example = dataset["train"][0]

image = example["image"]              # PIL image
question = example["question"]
answer = example["answer"]
evidence = example["evidence"]
requested_texts = json.loads(example["planned_texts_json"])

The split names are train and dev. For reproducible experiments, pass a dataset commit hash with revision=.

Dataset composition

Primary question category	Value of `qa_category`	Train	Dev	Total
Arithmetic	`arithmetic`	350	50	400
Comparison	`comparison`	525	75	600
Ordering	`order`	525	75	600
Sets	`set`	700	100	800
Mapping	`mapping`	700	100	800
Conditions and exceptions	`condition_exception`	700	100	800
Total		3,500	500	4,000

Questions may combine multiple operations. The primary category is assigned during construction. The splits contain 923 and 134 distinct scene-type names, respectively, with no exact scene-type-name overlap. This does not imply that broader concepts or common phrases are disjoint.

The dataset contains 135,999 requested text strings (mean 34.0 per image; median 32; range 8–56) and 8,737 semantic facts used to derive answers.

Fields

Field	Description
`image_id`	Stable image identifier.
`source_scene_id`	Identifier of the accepted generation design.
`split`	`train` or `dev`.
`image`	Embedded PNG image, decoded as a PIL image by Hugging Face Datasets.
`image_sha256`	SHA-256 of the original PNG bytes.
`width`, `height`	Image dimensions in pixels.
`question`	Japanese question about the image.
`answer`	Reference answer.
`acceptable_answers`	Accepted answer variants.
`qa_category`	Primary reasoning category.
`evidence`	Evidence records containing `region_id`, `fact_id`, `text`, `bboxes_xyxy_px`, and `blind_transcript`.
`planned_texts_json`	JSON object mapping region IDs to text requested during image generation.
`scene_json`	JSON generation plan, including the scene type and requested visual attributes. These are requested attributes, not measurements of the generated image.
`facts_json`	JSON representation of semantic facts and their associated evidence.
`program_json`	JSON expression describing how facts are combined to derive the answer.
`answer_structured_json`	Structured reference answer serialized as JSON.
`counterfactuals_json`	Recorded changes to individual facts that change the derived answer.
`verification_json`	Recorded model-based and programmatic verification results.

Evidence boxes use absolute pixel coordinates [left, top, right, bottom] in the stored image. They are padded evidence crops, can overlap, and are not tight character-level boxes. blind_transcript records a model's transcription of the crop and may include neighboring text or recognition errors.

Requested text preserves repeated strings under separate region IDs. Keep those occurrences when constructing OCR targets. For example:

ocr_target = "\n".join(requested_texts.values())

This gives one serialization of the requested strings; it does not establish a unique visual reading order.

Construction

Sample scene specifications using a taxonomy of 116 compatible combinations of text carrier and content. Assign scene-type-disjoint splits and question-category quotas.
Jointly design image text, questions, answers, semantic facts, and a closed JSON derivation using a language model. Check the derivation programmatically and review the question using Qwen3.5-397B-A17B, with additional checks where needed.
Test whether modifying each fact can change the answer, and whether withholding its planned evidence leaves multiple compatible answers. These checks operate on text references, not image-pixel masking.
Generate the full scene image using GPT Image 2.5 through OpenRouter. The image generator receives the scene and requested text, without the question or reference answer. No source photograph or separately rendered text image is supplied.
Use PP-OCRv5 and vision-language model checks to localize evidence, transcribe crops without showing the expected text, and answer the question from the generated image. Apply bounded repair or regenerate rejected examples.

The archived image-generation model identifier is openai/gpt-image-2.5-sunburst. Some verification and repair steps use alternate readers. The process includes adaptive repairs and does not provide independent human verification of every example.

Requested text lengths use an aggregate histogram from the 745-image JaWildText evaluation cohort, excluding unknown-text regions. This supplies distributional information from the evaluation data. The full-image generation process does not reuse its source images or literal text strings.

Intended use and limitations

Research on Japanese scene-text recognition, visual question answering, and synthetic-data training.
Images, names, signs, and statements are synthetic. They should not be treated as factual information about real people, businesses, locations, products, or services.
Planned non-evidence text may differ from the image. There is no exhaustive final-image transcript or ignore mask.
Model-based verification is fallible, and several checks reuse models. Passing the checks does not guarantee answer correctness, uniqueness, or independence of the checks.
Evidence may also appear outside the stored boxes. Masking those boxes alone need not remove all information relevant to the answer.
Automatic filtering can favor examples that the verification models find easy. Performance on this dataset does not by itself establish performance on real photographs.

License

Use is limited to research purposes under the JaSSeT Research Use Terms. In particular, the dataset must not be used for model development prohibited by the applicable source-service terms. A research purpose does not create an exception to those restrictions.

Authors and accompanying work

Koki Maeda and Naoaki Okazaki.

Accompanying manuscript: Multi-Task Learning of OCR and Visual Question Answering with Synthetic Data for Japanese Scene Text Understanding.

Japanese title: 日本語シーンテキスト理解に向けた合成データを用いたOCRと画像質問応答のマルチタスク学習.

Contact: koki.maeda@nlp.comp.isct.ac.jp

The data files preserve the frozen 4,000-image dataset used in the manuscript.

Downloads last month: 24

Total file size:

10.1 GB