MOMIJI / README.md
Onely7's picture
Update README.md
cb4e627
|
Raw History Blame
5.54 kB
metadata
language:
  - ja
license: apache-2.0
size_categories:
  - 10M<n<100M
pretty_name: MOMIJI
dataset_info:
  - config_name: default
    features:
      - name: docID
        sequence: string
      - name: url
        dtype: string
      - name: text_list
        dtype: list
      - name: text
        sequence: string
      - name: image_info
        sequence: list

Dataset Card for MOMIJI

Dataset Description

  • GitHub Repository: (Coming soon)
  • Paper: (Coming soon)

MOMIJI (Modern Open Multimodal Japanese filtered Dataset) is a large-scale, carefully curated public dataset of image-text–interleaved web documents. The dataset was extracted from Common Crawl dumps covering February 2024 – January 2025 and contains roughly 56 million Japanese documents, 110 billion characters, and 2.49 billion images. Details of the collection and filtering pipeline will be described in a forthcoming paper.

Image-text–interleaved data is generally used to train large vision-language models (LVLMs) such as LLaVA-OneVision, Idefics 2, NVILA, and Qwen 2.5-VL. Using MOMIJI, we trained our proposed model Heron-NVILA-Lite.

Following OBELICS, we provide an interactive visualization that allows users to explore the contents of MOMIJI. The map shows a subset of 1M of the 56M documents.

MOMIJI Nomic map

Warning and Disclaimer: This content may unintentionally include expressions or information that some may find inappropriate. Please view it at your own discretion and responsibility.

Data Fields

Example (truncated):

{
  "docId": "CC-MAIN-20240518121005-20240518151005-00454_00091",
  "url": "https://tanaka-yuki.com/spring/post-10813/",
  "text_list": [
    "大学の卒業式には、男性ならスーツを選ぶ方がとても多いですが、どんなスーツを着用すればいいのか?と悩みますよね。",
    "...",
    "大学卒業式で男性はスーツが基本",
    "<image1>",
    "...",
    "安定感ならシングルデザイン",
    "<image2>",
    "..."
  ],
  "text": "大学の卒業式には、男性ならスーツを選ぶ方がとても多いですが、どんなスーツを着用すればいいのか?と悩みますよね.\n ...",
  "image_info": [
    {
      "placeholder": "<image1>",
      "url": "https://tanaka-yuki.com/wp-content/uploads/2022/12/collage-1-1.webp",
      "width": 1080,
      "height": 620,
      "original_width": 1080,
      "original_height": 620,
      "exif": "{}",
      "alt": "大学卒業式で男性はスーツが基本"
    },
    "..."
  ]
}

Each sample has five top-level fields:

Field Type Description
docId str Unique ID derived from the source Common Crawl ID
url str Source web URL
text_list list\<str\> Document text split into individual segments
text str Full text with image placeholders
image_info list\<dict\> List of image metadata:
• placeholder (str)
• url (str)
• original_width / original_height (int)
• exif (dict, optional)
• alt (str, optional)

Size and Splits

MOMIJI contains about 56 million Japanese web documents. Because images are provided only as URLs (no binary data), the dataset is roughly 150 GB.

MOMIJI Statistics

Metric Value
Number of documents 56,119,639
Number of images 249,745,953
Total characters 109,980,725,957
Average characters/Doc. 1,959
Average images/Doc. 4.45

Below is a bar chart of document counts by number of images (documents with ≥ 30 images are omitted for readability):

Content Warning

Although an NSFW filter was applied, some links or text samples may still be disturbing. The dataset is intended for scientific or safety analyses by trained researchers.

Disclaimer

MOMIJI does NOT distribute image binaries—only links and metadata. We are not responsible for content accessible via those links. Filtering was performed automatically due to the dataset’s scale.

License Information

Apache License 2.0

Acknowledgements

This model is based on results obtained from a project, JPNP20017, subsidized by the New Energy and Industrial Technology Development Organization (NEDO).