trickster-2005/cantonese-benchmark-synth

CantoBench-Synth

A synthetic OCR benchmark for written Cantonese (廣東話). 粵語（廣東話）OCR 合成評測資料集。

English · 繁體中文

• Project page: https://trickster-2005.github.io/Cantonese-OCR-Benchmark/
• Code: https://github.com/trickster-2005/Cantonese-OCR-Benchmark
• OCF AI Research Internship 2026

---

English

What this is

4,930 synthetic text-line images testing whether vision-language models can read the characters used in written Cantonese. Everyday writing in Hong Kong uses many characters absent from Standard Written Chinese (嘅, 咗, 冇, 喺, 𨋢), and these barely appear in mainstream OCR training data.

Surveying existing resources turned up three gaps: there are twenty-odd Cantonese datasets on HuggingFace and all are text or audio, with none providing OCR data as images; recent Cantonese benchmarks measure language understanding rather than visual recognition; and some characters are missing from fonts entirely, with 𠮶 (U+20BB6) supported by none of the 33 open-licensed font instances used here.

Structure

Three tiers, with the same character deliberately appearing across them.

Tier	Content	Images	Distinct texts	Length
char	Cantonese-specific characters	1,162	166	1
word	Words containing Cantonese characters	1,781	600	2–4
sentence	Real Cantonese sentences	1,987	667	4–25

The core design. If a model reads 冇 correctly in isolation but returns 有 inside a sentence, the failure is not visual: the language-model prior overrode the glyph. A sentence-only benchmark cannot separate those two cases.

Fifteen minimal pairs are tagged via minimal_pair_with. The matching Standard Chinese characters (有, 既, 左, 但 …) are included as role: counterpart so there is a baseline to compare against.

Character tier by Unicode block: 994 basic, 28 Extension A, 140 Extension B+. Extension blocks are over-represented on purpose: sampling by corpus frequency would fill the benchmark with common characters and never show the rare ones.

Usage

from datasets import load_dataset

ds = load_dataset("trickster-2005/cantonese-benchmark-synth", split="test")
print(ds[0]["text"], ds[0]["tier"], ds[0]["difficulty"])

# Extension-block characters only
ext = ds.filter(lambda r: r["plane"] in ("ExtA", "ExtB+"))

# Minimal pairs, for the character-vs-sentence comparison
pairs = ds.filter(lambda r: r["minimal_pair_with"] != "")

Baseline results

Model	CER ↓	ACC ↑	1−NED ↑	ACC char	ACC sentence	ACC ext. blocks
InternVL3.5 4B	0.3409	29.3%	0.6641	25.6%	20.9%	0.0%
Qwen3-VL 4B	0.7009	17.5%	0.2998	16.7%	5.7%	0.0%

Metrics follow OCR convention: CER (character error rate), ACC (exact sequence match), 1−NED (normalised edit similarity). The prompt deliberately does not mention Traditional Chinese, since "Cantonese characters get corrected into Standard Chinese" is the phenomenon under test.

How it was built

Corpora: Cantonese Wikipedia (CC BY-SA 4.0), HKCanCor (CC BY 4.0), rime-cantonese (CC BY 4.0).

The 154-character Cantonese set is derived by subtracting the mainland, Taiwan and Hong Kong standard character lists plus Japanese-only kanji from corpus character frequencies, then reviewed by hand.

Fonts are SIL OFL 1.1 only (33 instances): Chiron Hei HK, Chiron Sung HK, Noto CJK, Hong Kong Character Set. System Chinese fonts are not used, as their licences do not permit generating a publicly distributed dataset and their extension-block coverage is zero.

Augmentation covers rotation, perspective, elastic distortion, stroke weight, blur, noise and JPEG artefacts across four difficulty levels ({'medium': 1250, 'clean': 1215, 'light': 1709, 'hard': 756}).

Every image passes three checks against tofu characters: the font's cmap is verified before rendering, ink ratio and contrast are verified after, and an independent pass re-checks glyph coverage across the whole dataset.

Reproducible: each image's RNG derives from (seed=20260825, index), independent of worker count.

Full methodology is in the GitHub README.

Licence

CC BY-SA 4.0. The dataset is copyleft because Cantonese Wikipedia is CC BY-SA and that propagates; derivative works must use the same licence.

Please cite this dataset and the three upstream corpora.

Generation code is MIT. Fonts are SIL OFL 1.1 and are not redistributed here.

Limitations

• All images are synthetic; performance on real Cantonese images needs separate validation.
• Horizontal text lines only. Hong Kong signage is frequently vertical.
• Handwriting is approximated by elastic distortion and stroke-weight variation rather than being genuine.
• 𠮶 (U+20BB6) and 𤗲 (U+245F2) are in the character set but produce no images because no available font contains them.

Acknowledgement

Produced with assistance from Claude (Anthropic). All figures, methods and claims were reviewed and verified by the author.

---

繁體中文

這是什麼

4,930 張合成文字行影像，用來測視覺語言模型認不認得粵語書寫特有的字。 香港的日常書寫大量使用標準中文沒有的字（嘅、咗、冇、喺、𨋢）， 而這些字在主流 OCR 的訓練資料中極少出現。

盤點現有資源後發現三件事：HuggingFace 上的粵語資料集有二十多個，全部是純文字或語音， 未找到以影像形式提供的；近年的粵語基準測試評估的是語言理解而非視覺辨識； 部分字連字型都缺，𠮶（U+20BB6）在本專案使用的 33 個開源字型實例中支援數為 0。

結構

三層結構，同一個字刻意跨層出現。

層級	內容	張數	相異文字	字數
char	粵語特有字	1,162	166	1
word	含粵語字的詞	1,781	600	2–4
sentence	真實粵語句子	1,987	667	4–25

核心設計。 如果模型在單字層把 冇 認對、在句子層卻讀成 有， 那就不是視覺問題，而是語言模型的先驗蓋過了字形。 只測句子的評測分不出這兩種情況。

15 組最小對立對以 minimal_pair_with 標記。書面語對照字（有、既、左、但等） 以 role: counterpart 收錄，作為比較的基準線。

單字層 Unicode 分區：基本區 994、 擴展 A 區 28、擴展 B 區以上 140。 擴展區刻意超額配置：按語料頻率抽樣的話，常用字會塞滿評測集，罕見字一次都不會出現。

使用方式

from datasets import load_dataset

ds = load_dataset("trickster-2005/cantonese-benchmark-synth", split="test")
print(ds[0]["text"], ds[0]["tier"], ds[0]["difficulty"])

# 只取擴展區字元
ext = ds.filter(lambda r: r["plane"] in ("ExtA", "ExtB+"))

# 只取最小對立對，做單字層與句子層的對照
pairs = ds.filter(lambda r: r["minimal_pair_with"] != "")

基線跑分

模型	CER ↓	ACC ↑	1−NED ↑	單字層 ACC	短句層 ACC	擴展區 ACC
InternVL3.5 4B	0.3409	29.3%	0.6641	25.6%	20.9%	0.0%
Qwen3-VL 4B	0.7009	17.5%	0.2998	16.7%	5.7%	0.0%

指標採 OCR 慣例：CER（字元錯誤率）、ACC（整句完全正確率）、1−NED（正規化編輯相似度）。 prompt 刻意不提「繁體中文」，因為「粵語字被糾正成書面中文」正是要測的現象。

製作方式

語料：粵語維基百科 （CC BY-SA 4.0）、HKCanCor（CC BY 4.0）、 rime-cantonese（CC BY 4.0）。

154 個粵語特有字的字表，是從語料字頻扣掉兩岸港的標準字表與日文專用漢字 推導而來，再經人工複核。

字型只用 SIL OFL 1.1（33 個實例）：昭源黑體、昭源宋體、 Noto CJK、香港民間字集。不使用系統中文字型，因為其授權不允許用於產生公開散布的資料集， 且擴展區覆蓋率為 0。

增強包含旋轉、透視、彈性形變、筆畫粗細、模糊、雜訊、JPEG 壓縮痕跡， 分四個難度等級（{'medium': 1250, 'clean': 1215, 'light': 1709, 'hard': 756}）。

每張影像都通過三道防豆腐字檢查：渲染前驗證字型 cmap、渲染後驗證墨水比例與對比、 最後再對整份資料集獨立複查字形覆蓋。

可重現：每張圖的亂數由 (seed=20260825, index) 推導，與 worker 數量無關。

完整方法論見 GitHub README。

授權

CC BY-SA 4.0。 因粵語維基百科是 CC BY-SA 且會傳染，本資料集受 copyleft 約束， 衍生作品須以相同授權釋出。

請同時引用本資料集與三個上游語料來源。

生成程式碼為 MIT。字型為 SIL OFL 1.1，不隨資料集散布。

限制

• 全部是合成影像，真實粵語影像上的表現需另外驗證。
• 只有橫排文字行。香港招牌大量使用直排。
• 手寫只以彈性形變與筆畫粗細變化近似，並非真實手寫。
• 𠮶（U+20BB6）與 𤗲（U+245F2）在字表中但產不出影像，因為沒有可用字型收錄。

聲明

由 Claude（Anthropic）輔助產出，所有數據、方法與敘述均經作者人工審核確認。