MOMIJI Dataset Card

English | 日本語

MOMIJI (Modern Open Multimodal Japanese filtered Dataset) is a large-scale Japanese image-text interleaved dataset built from Common Crawl.

“Interleaved” means that text and images are associated while preserving the order in which they appear within a document. In MOMIJI, image positions are represented by placeholders such as <image1>.

• Dataset: turing-motors/MOMIJI
• Dataset construction code: turingmotors/MOMIJI
• Data generation utility: turingmotors/momiji_generator
• Paper: Coming soon

Please Read Before Use

This repository does not distribute the original Web page text or image files themselves.

Instead, it provides public records that identify documents and images that passed the MOMIJI filtering pipeline. Each record contains information such as the source Web page URL, image URLs, placeholders indicating image positions, image dimensions, and other associated metadata.

The MOMIJI public records alone cannot reproduce the document text or images exactly as they appeared when the dataset was constructed. The text and text_list fields are generated from Web pages that remain accessible at the time of execution by using momiji_generator. Because Web pages may be updated or removed, the generated output may differ from the data available when MOMIJI was originally constructed.

MOMIJI was built from Web data. Although automated filtering, including an NSFW (Not Safe for Work) filter, was applied, the public records and linked resources may still contain offensive, harmful, inaccurate, biased, or otherwise inappropriate content. Before using the data, review the content and handle it at your own discretion and responsibility.

Dataset Overview

MOMIJI was constructed from Common Crawl data collected between February 2024 and January 2025. Common Crawl is a nonprofit project that continuously collects Web pages from around the world and makes the resulting data publicly available.

At the time of construction, the dataset contained approximately 56 million documents, 110 billion characters, and 249 million image references. Here, the number of image references is the total number of image entries recorded in the image_info fields across all documents; it does not guarantee the number of image files that are currently retrievable. The text and images were processed through multiple stages of quality and safety filtering.

MOMIJI can be used to train Large Vision-Language Models (LVLMs), which process both images and text. For example, Heron-NVILA-Lite was trained using MOMIJI.

Approximately one million documents sampled from MOMIJI can be explored in Nomic Atlas. Documents with similar content are placed near one another, making it possible to inspect broad patterns in the dataset visually.

Related Resources and Their Roles

MOMIJI provides three resources that can be used according to your needs.

Resource	Primary role	What is not included or guaranteed
MOMIJI public records	Provide URLs, image placeholders, image dimensions, and other metadata for identifying documents and images that passed the filtering pipeline.	They do not include the full text of the source Web pages, image binaries, or fixed versions of text and text_list.
momiji_generator	Accesses the URLs in the public records and generates text and text_list from Web pages available at the time of execution.	It does not reproduce Web pages exactly as they appeared when MOMIJI was constructed.
MOMIJI construction code	Provides the AWS implementation used to construct MOMIJI and a local implementation for rerunning the main processing stages at a smaller scale.	It does not guarantee that external Web pages or images remain unchanged and accessible from the time of construction.

Quick Start

1. Download the public records

From Files and versions on Hugging Face, download the gzip-compressed JSONL file for the collection period, or Common Crawl release, that you want to process.

For an initial test, we recommend starting with one file and a small number of records rather than processing many files at once.

2. Set up momiji_generator

In an environment where uv is available, run the following commands:

git clone https://github.com/turingmotors/momiji_generator.git
cd momiji_generator
uv sync

3. Generate text and text_list

The following example processes the first 20 records in the input file:

uv run python retrieve_text_from_jsonl.py \
  --input_path /path/to/input.jsonl.gz \
  --output_path /path/to/output.jsonl \
  --limit 20

The --limit option sets the maximum number of records to process. After confirming that the command works as expected, you can omit this option to process all records.

Records processed successfully will include the additional text and text_list fields. Generation may fail when the source Web page is no longer accessible or when its structure has changed.

For Docker-based usage and the latest command-line options, see the momiji_generator README.

Data File Organization

The public records are organized by Common Crawl release.

CC-MAIN-2024-10/
CC-MAIN-2024-18/
CC-MAIN-2024-22/
CC-MAIN-2024-26/
CC-MAIN-2024-30/
CC-MAIN-2024-33/
CC-MAIN-2024-42/
CC-MAIN-2024-46/
CC-MAIN-2024-51/
CC-MAIN-2025-05/

The repository contains approximately 56 million records, with a total file size of about 14.7 GB. Because it primarily stores URLs and metadata rather than the full text of source Web pages or image binaries, it is considerably smaller than the complete corpus available at the time of construction.

Data Format

Top-level fields in the public records

Each JSONL record corresponds to one document that passed the final MOMIJI filtering pipeline.

Field	Type	Description
docId	str	Document ID derived from the original Common Crawl record
url	str	URL of the source Web page
image_info	list[dict]	Image placeholders, URLs, dimensions, and other metadata for images in the document

Fields in image_info

Each element of image_info may contain the following fields:

Field	Type	Description
placeholder	str	A string such as <image1> representing the image position within the document
url	str	URL of the source image
width / height	int	Image width and height recorded during processing
original_width / original_height	int	Original image width and height, when available
exif	str or object	EXIF-related information, such as camera and shooting-condition metadata, when available
alt	str	Alt text, or alternative image description, recorded from the source page when available

The following example uses fictional values for illustration:

{
  "docId": "CC-MAIN-YYYYMMDDHHMMSS-EXAMPLE_00001",
  "url": "https://example.com/article",
  "image_info": [
    {
      "placeholder": "<image1>",
      "url": "https://example.com/images/example.jpg",
      "width": 1024,
      "height": 768,
      "original_width": 1024,
      "original_height": 768,
      "exif": "{}",
      "alt": "Example image description"
    }
  ]
}

Fields added by momiji_generator

The public records do not contain fixed versions of text and text_list. When momiji_generator is run, it generates the following fields from Web pages that can be retrieved successfully:

Field	Type	Description
text	str	Full extracted document text, including image placeholders
text_list	list[str]	Text segments and image placeholders arranged in their original order within the document

The following example uses fictional values for illustration:

{
  "docId": "CC-MAIN-YYYYMMDDHHMMSS-EXAMPLE_00001",
  "url": "https://example.com/article",
  "text_list": [
    "This is the first text segment.",
    "<image1>",
    "This is the second text segment."
  ],
  "text": "This is the first text segment.\n<image1>\nThis is the second text segment.",
  "image_info": [
    {
      "placeholder": "<image1>",
      "url": "https://example.com/images/example.jpg",
      "width": 1024,
      "height": 768,
      "original_width": 1024,
      "original_height": 768,
      "exif": "{}",
      "alt": "Example image description"
    }
  ]
}

The contents of text and text_list depend on the Web pages available when momiji_generator is executed. They therefore may not match the snapshot used when MOMIJI was constructed.

MOMIJI Statistics

The following values were calculated from the data available when MOMIJI was constructed.

Metric	Value
Number of documents	56,119,639
Number of image references	249,745,953
Total number of characters	109,980,725,957
Average characters per document	1,959
Average image references per document	4.45

Data generated later with momiji_generator may have different character and image counts because source Web pages may have been updated or removed.

The bar chart below shows the number of documents for each number of image references per document. For readability, documents containing 31 or more image references are omitted.

Construction Pipeline and Released Code

The MOMIJI construction code is available in the MOMIJI GitHub repository. The repository contains two implementations:

• build_momiji_on_aws: The implementation used to construct the full MOMIJI dataset with services such as Amazon S3 for cloud storage, AWS Lambda for code execution, and AWS Step Functions for workflow management
• build_momiji: A local implementation that replaces AWS-specific storage and distributed processing with local files and execution scripts

The local implementation corresponds to the main filtering rules, processing order, thresholds, image-placeholder handling, and output fields used by the AWS implementation. However, it does not reproduce the large-scale distributed execution environment itself.

What Can and Cannot Be Reproduced

Web data changes over time. When considering the reproducibility of MOMIJI, it is important to distinguish among using the public records, generating data from currently available Web pages, and rerunning the construction pipeline.

Using the public records

By downloading gzip-compressed JSONL files from Hugging Face, you can use the document and image references that passed the final filtering stage without rerunning the entire construction pipeline.

Generating currently available interleaved data

Running momiji_generator on the public records generates text and text_list from Web pages that are currently accessible.

However, the resulting data reflects the Web pages available at the time of execution. It does not provide a fixed reconstruction of the pages as they appeared when MOMIJI was constructed.

Rerunning the construction pipeline

By providing the same Common Crawl WARC files, or a subset of them, to build_momiji or build_momiji_on_aws, you can rerun text extraction and the main filtering stages.

Reprocessing the full MOMIJI dataset requires substantial storage and computational resources. Even when full-scale reprocessing is impractical, processing a small number of WARC files with the local implementation allows you to inspect the following:

• Filtering rules
• Order in which filters are applied
• Main threshold values
• Intermediate and final outputs

Factors that make exact reproduction difficult

• Changes to source Web pages: Pages may have been updated, removed, relocated, access-restricted, or structurally modified.
• Changes to external images: Some images are retrieved from external URLs and may have been replaced or removed, making it impossible to retrieve the same images available at construction time.
• Errors in automated processing: Large-scale automated filtering may leave false positives or false negatives, inaccurate metadata, or inaccessible URLs.
• Differences in execution environments: Even with the same WARC files, differences in dependencies or settings may lead to different processing results.

Intended Uses

MOMIJI is primarily intended for information-analysis tasks such as:

• Training, evaluating, and auditing machine learning models
• Research on Large Language Models (LLMs) and Large Vision-Language Models (LVLMs)
• Safety and bias analysis
• Text mining and data mining
• Statistical analysis

In Japan, some uses for information-analysis purposes may fall within the scope of Article 30-4 of the Copyright Act. For more information, see the materials on AI and copyright published by the Agency for Cultural Affairs, Japan.

The inclusion of a record, URL, or generated field in MOMIJI does not mean that the source content may be used for every purpose. Users should verify the following according to their intended use and jurisdiction:

• The legal basis or permission applicable to the intended use
• Applicable laws and regulations
• The rights of copyright holders and other rights holders of the source content
• The terms of use and access restrictions of the source Web sites
• Whether third-party content may be redistributed or publicly transmitted

This dataset card provides general information and does not constitute legal advice.

License and Rights

This repository contains both materials created by the publishers and information derived from third parties.

Materials created by the publishers

The Creative Commons Attribution 4.0 International License applies to materials for which the publishers hold the necessary rights, including this dataset card, publisher-authored documentation, statistical tables, and descriptions of the data format.

Information derived from third parties

Rights to the text, images, alt text, URLs, and other information derived from source Web pages remain with their respective rights holders. Unless the relevant rights holder separately provides such content under CC BY 4.0, the CC BY 4.0 license does not apply to this third-party information.

MOMIJI does not redistribute the full text of source Web pages or the binaries of source images as a fixed corpus. The public records provide reference information for identifying Web documents and images that passed the filtering pipeline and for supporting information analysis and data generation.

Software

The source code for momiji_generator, build_momiji_on_aws, and build_momiji is provided separately under the MIT License. The MIT License applies to the software and its accompanying documentation. It does not apply to third-party text or images retrieved or processed by the software.

Content and Quality Limitations

MOMIJI was constructed by automatically processing large-scale Web data. Although an NSFW filter was applied, the following types of content or issues may remain:

• Offensive, explicit, violent, discriminatory, or otherwise inappropriate content
• Inaccurate, biased, or outdated information
• Filtering errors
• Inaccurate or missing metadata
• URLs that are no longer accessible
• Web pages or images that have changed since MOMIJI was constructed

Before using the data, we recommend performing additional quality, safety, and duplicate checks appropriate to your intended purpose.

Disclaimer

The publishers do not control the current content, availability, safety, or terms of use of external Web sites linked from the public records.

The following are not guaranteed:

• Continued accessibility of all URLs
• Consistency between current Web pages or images and their versions at the time of construction
• Accuracy of all records and metadata
• Removal of all inappropriate content by the automated filters
• Suitability of MOMIJI for any particular purpose

Before using MOMIJI or resources available through external links, review their content and applicable conditions and use them at your own responsibility.

Acknowledgments

This dataset is based on results obtained from project JPNP20017, commissioned by the New Energy and Industrial Technology Development Organization (NEDO), Japan.