DimV-Ai/kuzushiji-character-dataset-v1

Filename components:

• U+81EA → Unicode code point of the character
• 100241706 → Book / manuscript ID
• 00004_2 → Page image identifier
• X1852_Y1736 → Character position on the page (top-left corner)

---

📑 Metadata (metadata.csv)

The metadata.csv file provides structured annotations for each character instance and serves as the main index linking text labels to image data.

CSV Columns

Column name	Description
source	Origin of the annotation. Values include codh (Center for Open Data in the Humanities) and ogihan (Ogihan annotation pipeline).
book_id	Identifier of the manuscript or book in which the character appears. Corresponds to CODH / Ogihan book IDs.
page_image	Identifier of the page image within the book. Used to reconstruct page-level context.
unicode	Unicode code point of the character (e.g. U+81EA). This is the canonical label for the character.
x	X-coordinate (in pixels) of the character’s top-left corner on the original page image.
y	Y-coordinate (in pixels) of the character’s top-left corner on the original page image.
char_image_path	Relative path to the character image inside the tar archives. This path can be used to locate the image within the corresponding images-*.tar.gz file.

Together, these columns allow:

• character-level classification
• reconstruction of page layout
• filtering by manuscript, character, or source
• alignment between visual and textual representations

---

🧠 Intended Use

This dataset is suitable for:

• Kuzushiji OCR and recognition
• Character-level image classification
• Vision–language / multimodal models
• Contrastive or few-shot learning
• Digital Humanities research on historical Japanese texts

The tar-sharded format is particularly well suited for:

• large-scale training
• streaming data loaders (e.g. WebDataset)
• cloud and HPC environments

---

🛠️ Example: Loading Metadata

import pandas as pd

df = pd.read_csv("metadata.csv")
df.head()