Dataset Viewer
Auto-converted to Parquet
image
imagewidth (px)
224
224
text
stringlengths
2
6
label
stringclasses
2 values
is_real_word
bool
2 classes
pair_id
stringlengths
9
9
category
stringclasses
4 values
char_count
int32
2
6
囚人
positive
true
pair_0000
K-K
2
人囚
negative
false
pair_0000
K-K
2
戦艦
positive
true
pair_0001
K-K
2
艦戦
negative
false
pair_0001
K-K
2
付随
positive
true
pair_0002
K-K
2
随付
negative
false
pair_0002
K-K
2
癌腫
positive
true
pair_0003
K-K
2
腫癌
negative
false
pair_0003
K-K
2
外題
positive
true
pair_0004
K-K
2
題外
negative
false
pair_0004
K-K
2
果報
positive
true
pair_0005
K-K
2
果報
negative
false
pair_0005
K-K
2
試料
positive
true
pair_0006
K-K
2
料試
negative
false
pair_0006
K-K
2
遊惰
positive
true
pair_0007
K-K
2
惰遊
negative
false
pair_0007
K-K
2
壮語
positive
true
pair_0008
K-K
2
語壮
negative
false
pair_0008
K-K
2
意志
positive
true
pair_0009
K-K
2
志意
negative
false
pair_0009
K-K
2
曲節
positive
true
pair_0010
K-K
2
節曲
negative
false
pair_0010
K-K
2
長期
positive
true
pair_0011
K-K
2
期長
negative
false
pair_0011
K-K
2
善徳
positive
true
pair_0012
K-K
2
徳善
negative
false
pair_0012
K-K
2
倒立
positive
true
pair_0013
K-K
2
立倒
negative
false
pair_0013
K-K
2
抽斗
positive
true
pair_0014
K-K
2
斗抽
negative
false
pair_0014
K-K
2
出費
positive
true
pair_0015
K-K
2
費出
negative
false
pair_0015
K-K
2
註釈
positive
true
pair_0016
K-K
2
釈註
negative
false
pair_0016
K-K
2
翠色
positive
true
pair_0017
K-K
2
色翠
negative
false
pair_0017
K-K
2
要言
positive
true
pair_0018
K-K
2
言要
negative
false
pair_0018
K-K
2
科目
positive
true
pair_0019
K-K
2
目科
negative
false
pair_0019
K-K
2
疾走
positive
true
pair_0020
K-K
2
走疾
negative
false
pair_0020
K-K
2
忠魂
positive
true
pair_0021
K-K
2
魂忠
negative
false
pair_0021
K-K
2
製靴
positive
true
pair_0022
K-K
2
靴製
negative
false
pair_0022
K-K
2
宿六
positive
true
pair_0023
K-K
2
六宿
negative
false
pair_0023
K-K
2
時期
positive
true
pair_0024
K-K
2
期時
negative
false
pair_0024
K-K
2
畜殺
positive
true
pair_0025
K-K
2
殺畜
negative
false
pair_0025
K-K
2
加味
positive
true
pair_0026
K-K
2
味加
negative
false
pair_0026
K-K
2
勉強
positive
true
pair_0027
K-K
2
強勉
negative
false
pair_0027
K-K
2
有情
positive
true
pair_0028
K-K
2
情有
negative
false
pair_0028
K-K
2
鞏膜
positive
true
pair_0029
K-K
2
膜鞏
negative
false
pair_0029
K-K
2
糞便
positive
true
pair_0030
K-K
2
便糞
negative
false
pair_0030
K-K
2
矢印
positive
true
pair_0031
K-K
2
印矢
negative
false
pair_0031
K-K
2
全盛
positive
true
pair_0032
K-K
2
盛全
negative
false
pair_0032
K-K
2
梟雄
positive
true
pair_0033
K-K
2
雄梟
negative
false
pair_0033
K-K
2
武臣
positive
true
pair_0034
K-K
2
臣武
negative
false
pair_0034
K-K
2
歳次
positive
true
pair_0035
K-K
2
次歳
negative
false
pair_0035
K-K
2
注入
positive
true
pair_0036
K-K
2
入注
negative
false
pair_0036
K-K
2
爆燃
positive
true
pair_0037
K-K
2
燃爆
negative
false
pair_0037
K-K
2
猛威
positive
true
pair_0038
K-K
2
威猛
negative
false
pair_0038
K-K
2
心肝
positive
true
pair_0039
K-K
2
肝心
negative
false
pair_0039
K-K
2
悲劇
positive
true
pair_0040
K-K
2
劇悲
negative
false
pair_0040
K-K
2
嫡男
positive
true
pair_0041
K-K
2
男嫡
negative
false
pair_0041
K-K
2
内臓
positive
true
pair_0042
K-K
2
臓内
negative
false
pair_0042
K-K
2
騒動
positive
true
pair_0043
K-K
2
動騒
negative
false
pair_0043
K-K
2
鉄蹄
positive
true
pair_0044
K-K
2
蹄鉄
negative
false
pair_0044
K-K
2
雲母
positive
true
pair_0045
K-K
2
母雲
negative
false
pair_0045
K-K
2
引換
positive
true
pair_0046
K-K
2
換引
negative
false
pair_0046
K-K
2
苦悶
positive
true
pair_0047
K-K
2
悶苦
negative
false
pair_0047
K-K
2
火刑
positive
true
pair_0048
K-K
2
刑火
negative
false
pair_0048
K-K
2
後生
positive
true
pair_0049
K-K
2
生後
negative
false
pair_0049
K-K
2
End of preview. Expand in Data Studio

Japanese Multi-Character Recognition Dataset

Dataset Description

This dataset is designed to evaluate Vision-Language Models (VLMs) on their ability to distinguish between meaningful Japanese words and meaningless character sequences. It contains 2,000 images of Japanese text (1,000 positive/negative pairs).

Dataset Summary

  • Total Samples: 2,000 (1,000 pairs)
  • Categories:
    • K-K: 2-character Kanji compounds (250 pairs)
    • K+H: Kanji + Hiragana combinations (250 pairs)
    • H-only: Pure Hiragana words (250 pairs)
    • T-only: Pure Katakana words (250 pairs)
  • Task: Binary classification (real word vs. shuffled non-word)
  • Image Size: 224x224 pixels
  • Font: Japanese Gothic (NotoSansCJK or similar)

Dataset Structure

Each example contains:

  • image: The rendered text image (PIL Image)
  • text: The text shown in the image
  • label: "positive" (real word) or "negative" (shuffled)
  • is_real_word: Boolean indicating if it's a real word
  • pair_id: Links positive/negative pairs
  • category: One of ["K-K", "K+H", "H-only", "T-only"]
  • char_count: Number of characters in the text

Usage

from datasets import load_dataset

dataset = load_dataset("Silviase/JPMultiCharRecog")

# Example
print(dataset['train'][0])
# {
#   'image': <PIL.Image>,
#   'text': '学校',
#   'label': 'positive',
#   'is_real_word': True,
#   'pair_id': 'pair_0001',
#   'category': 'K-K',
#   'char_count': 2
# }

Evaluation

The dataset is designed for OCR evaluation of Japanese text recognition capabilities in Vision-Language Models.

License

This dataset is released under the MIT License.

Downloads last month
57