JMangaBench_Read v1.3

Crop images for JMangaBench_Read, a Japanese manga OCR benchmark. The main set contains the text a reader reads and a translation pipeline translates (speech, narration, asides): 7,789 text boxes from 1,847 pages by 1,847 artists, all pages first posted between 2025-10-01 and 2026-10-01. Sound effects are kept in a separate SFX set (380 boxes, unchanged since v1.2).

v1.3 adds 4,660 boxes from 1,116 new pages and removes 13 v1.2 boxes. Details: https://github.com/muscgab/JMangaBench_Read

Found an error? Please open an issue at https://github.com/muscgab/JMangaBench_Read/issues.

• Leaderboard, labels, LLM readings, model predictions and scoring code: https://github.com/muscgab/JMangaBench_Read
• Interactive leaderboard: https://muscgab.github.io/JMangaBench_Read/

Files

• main/<key>.png (split test): main-set crops at original resolution, 7,789.
• sfx/<key>.png (split sfx): SFX-set crops, 380.
• main/metadata.jsonl, sfx/metadata.jsonl: one row per crop with file_name and the label fields (text, text_rin, gold, gold_hard, printed, non_printed, font, role, orientation, page_id, box, rating, origin, review, ...). Same content as data/gt.jsonl and data/gt_sfx.jsonl in the GitHub repository. Score with eval/score.py there.

Usage

hf download muscgab/JMangaBench_Read --repo-type dataset --local-dir JMangaBench_Read

For evaluation only. Do not use for training.

License

CC BY-NC-SA 4.0. Copyright of the original images remains with their respective artists.