vision-braille-dataset

Chinese ↔ 通用盲文 (Chinese Braille) parallel training data for the Vision-Braille translation models. Three corpora ship together in this repo:

Folder	What it is	Rows	Size
cleaned_2345_v3/	Audited + cleaned braille→Chinese training corpus, stages 2→5, with restored 分词连写 word spacing on stage 2	379,152	426 MB
braille_spaced/	Freshly built passage-level corpus, fully word-spaced, across 通用 / 数学 / 医学 / 中医 / 病理学	80,360	208 MB
cleaned_080126_v1/	Four-stage braille→Chinese training corpus on the word-spaced passages, with a r100→r10 tone-reduction curriculum	645,468	872 MB

Every record is ShareGPT (system / user / assistant) in a messages list, ready for LLaMA-Factory. Both folders keep their original dataset_info.json registration files.

from datasets import load_dataset

ds = load_dataset("Violet-yo/vision-braille-dataset", "c080126_stage2_passage")
ds = load_dataset("Violet-yo/vision-braille-dataset", "spaced_general")

System prompts used throughout:

• braille→Chinese: 你是一个中国盲文翻译助手，请把通用盲文转换成为汉字。
• Chinese→braille: 你是一个中国盲文翻译助手，请把汉字转换成为通用盲文。

Braille is Unicode U+2800–U+28FF. Word division uses U+2800 BRAILLE PATTERN BLANK, not ASCII space. Records carry <BRAILLE_END> / <CHINESE_END> / <TRANSLATION_END> markers.

---

1. cleaned_2345_v3/ — cleaned stage 2→5 corpus

26 datasets, 379,152 rows, cleaned down from 388,416 originals (9,264 rows / 2.39% removed). All braille→Chinese. Tone ratios follow the curriculum convention: r100 = 100% tone marks retained, r10 = 10%.

cleaned_2345_v3/
├── dataset_info.json                              26 LLaMA-Factory entries
├── MANIFEST.json                                  row counts + SHA-256 per file
├── README.md                                      full provenance and caveats
├── stage2_sentence_train_br2zh_r100_{train,val}.jsonl    stage 2, sentence level
└── removed/                                             audit trail, sentence manifests
    └── <dataset>.removed.jsonl

The passage data that used to live here (stage 3 r100, stage 5 r10, and the stage 4
r100→r10 curriculum) has been retired: it was built before 分词连写 word spacing was
available, so its braille is one unbroken cell run. `cleaned_080126_v1/` replaces it
with word-spaced passages. The stage 2 sentence data is unaffected and stays.

What was cleaned

The corpus carried tone-collision corruption: braille tone markers (⠂ tone 2, ⠆ tone 4, ⠔ the final a) collide with digit and punctuation cells, producing rows whose Chinese side cannot be recovered from the braille. Sentence data lost ~0.9% of rows to this; passage data lost 10–16%, which is where the corruption concentrated.

removed/ keeps one record per dropped row, indexed into the original file, so every removal is reviewable and recoverable:

{"line": 16, "reason": "punct_corrupt", "zh_preview": ";坐了几十里马车,哪就累着我了?…"}

Stage 2 carries restored 分词连写 word spacing

Stage-2 sentence braille is word-divided with U+2800; the rest of this folder is not.

Upstream Violet-yo/Chinese-Braille-Dataset-Full-Tone lost its word division in commit 9fa3ec6 ("update the dataset, remove the lables and spaces", 2025-11-26), so every dataset built after that had braille as one unbroken cell run. The revision immediately before it, db9c2013, still has the spacing and was recovered from a local cache, making this a lossless restore from the real source, not heuristic re-segmentation.

before:  ⠕⠄⠓⠊⠄⠓⠻⠁⠍⠢⠆⠗⠄⠃⠴⠄⠎⠪⠆⠙⠖⠆⠐⠕⠄⠃⠴⠄⠚⠴⠂ …
after :  ⠕⠄⠀⠓⠊⠄⠓⠻⠁⠀⠍⠢⠆⠗⠄⠃⠴⠄⠀⠎⠪⠆⠙⠖⠆⠐⠀⠕⠄⠀⠃⠴⠄⠚⠴⠂ …
         我    喜欢     墨尔本      赛道，   我    本人

	before	after
Rows with any word space	1.45%	100.00%
Total U+2800 separators	4,964	6,935,911

The Chinese side is untouched.

Fullwidth digits were repaired

Fullwidth digits (１９３９) never reached the encoder's ASCII-only number branch, so their 数符 (number sign ⠼) was silently dropped:

before:  １９３９年５月  ->  ⠂⠔⠒⠔⠝⠩⠂⠀⠢⠾⠆
after :  1939年5月      ->  ⠼⠁⠊⠉⠊⠝⠩⠂⠀⠼⠑⠾⠆
                            数符+1939  年      数符+5 月

1,839 rows / 3,540 numbers were repaired; 65 numbers (1.8%) had no matching dropped-cell run and were left unchanged and reported rather than guessed. 数符 placement is one ⠼ before the whole number, per GF 0019-2018.

Caveats — read before training

1. Val loss is not comparable across the clean/dirty boundary. Passage val sets lost 10–16% of rows, so a cleaned run's val curve cannot be read against an old run's. Checkpoint selection within a single run is unaffected, and improves.

2. stage4_passage_train_br2zh_r10_val is the passage TEST split, not a val split. Its 1,760 rows are identical in content and order to stage3_passage_train_br2zh_r100_test_r10 and share zero rows with stage3_passage_train_br2zh_r100_val — which is why its removal rate is 1.88% while every other val sits at 10–16%. This is inherited from the source configs. It means any stage-5 run does checkpoint selection on test data, so stage-5 test numbers from that setup are optimistically biased.

3. The halfwidth_punct_cjk rule was NOT applied (MANIFEST.json records halfwidth_punct_cjk_rule_applied: false). That rule catches real corruption (? and ! used as Chinese sentence-final punctuation, rendering to cells that collide with vowels the same way tone markers do), but it also removes ~47% of the passage test set. This folder reflects the audited, verified state, so re-running today's cleaner would not reproduce it.

4. Train and val only. No test splits — cleaning changes the denominator, so a cleaned-test score is not comparable to previously published numbers.

Verification performed

• All 26 files present, valid JSONL, every row carries messages
• SHA-256 of every file matches MANIFEST.json
• Counts reconcile: 388,416 original − 9,264 removed = 379,152
• Sources are order-preserving exact subsets of the originals
• No surviving row trips any of the 5 applied removal rules
• Tone-collision fingerprints ;; and :; are at 0

---

2. braille_spaced/ — freshly built spaced corpus

Passage-level Chinese↔braille pairs built from scratch with 分词连写 spacing throughout, to close the no-space gap in the older training data. All source text was downloaded fresh; no prior braille dataset was reused.

Splits are 8:1:1 on source passage id, so they are leak-free. Assembled records are under 2048 characters.

File	Records	Domain
general_shengxie_decode_{train,val}.json	23,405 / 2,925	通用 / 教学
math_shengxie_decode_{train,val}.json	15,753 / 1,969	数学
medical_shengxie_decode_{train,val}.json	15,841 / 1,980	医学 / 西医
tcm_shengxie_decode_{train,val}.json	15,917 / 1,989	中医 / 按摩针灸推拿
pathology_ft/pathology_56_br2zh_{train,val}.json	358 / 6	病理学 LoRA add-on
pathology_ft/pathology_78_br2zh_test.json	70	病理学 held-out class
pathology_{56,78}_{decode,ocr}.json	56 / 56 / 8 / 14	病理学 source pairs

These are the shengxie (省写, strict GF tone-omission) × decode (braille→中文) materialization. dataset_info.json and dataset_card.json in the folder document the full axis space (tone × direction × split) the builder can generate.

Domains and sources

Domain	Source	Field(s)	License
通用/教学	opencsg/chinese-fineweb-edu	text	see source
数学	MU-NLPC/Calc-ape210k	question_chinese + equation	see source
医学	FreedomIntelligence/huatuo_encyclopedia_qa	answers	apache-2.0
中医	SylvanL/Traditional-Chinese-Medicine-Dataset-Pretrain	textbook + clinical text	apache-2.0
病理学	病理学 course exam handouts (PPTX) + real AngelinaReader OCR of braille papers	slide text + OCR	—

Filters applied per domain: 50–400 chars, ≥60% CJK, md5 dedup; medical and TCM additionally cap ASCII letters at 15%; TCM drops classical 文言文 (modern Mandarin only) and boost-filters toward 按摩/推拿/针灸 vocational vocabulary, relaxing to general TCM to reach volume. Math drops power and other notation outside the v1 GB/T 18028-2010 encoder scope.

Assembly kept 87,791 of 90,000 input passages: 1,333 dropped for unsupported notation, 876 for encoding errors.

Encoding standards

• GF 0019-2018 国家通用盲文方案 for all Chinese prose, including 省写 tone omission, 简写 contractions, and 分词连写 word division
• GB/T 18028-2010 for mathematical notation in the 数学 domain

Sample records

decode direction, shengxie tone, braille truncated with …:

general

[user]  ⠬⠂⠇⠮⠀⠵⠼⠁⠌⠂⠀⠋⠦⠆⠇⠳⠂⠀⠱⠀⠓⠭⠀⠛⠲⠁⠛⠲⠀⠱⠺⠄⠞⠊⠄…<BRAILLE_END>
[asst]  鱼类增殖放流是向公共水体人工投放水生生物苗种,主动增殖资源的活动…<TRANSLATION_END>

math (Chinese GF 0019-2018 + math GB/T 18028-2010)

[user]  ⠥⠄⠝⠩⠛⠊⠂⠀⠞⠲⠓⠾⠂⠀⠉⠧⠁⠛⠫⠁⠀⠊⠆⠥⠀⠛⠯⠁⠱⠥⠁…⠰⠄⠅⠡⠄⠹⠀⠇⠭⠄⠌⠲⠄…<BRAILLE_END>
[asst]  五年级同学参加义务捐书活动，五1班捐了500本，五2班捐的本数是五1班80%…<TRANSLATION_END>

tcm (按摩/针灸/推拿 vocational)

[user]  ⠝⠬⠄⠐⠀⠼⠑⠓⠀⠎⠺⠐⠀⠼⠃⠚⠁⠚⠀⠝⠩⠀⠼⠉⠀⠾⠀⠼⠃⠁⠀⠚⠆…<BRAILLE_END>
[asst]  女，58岁，2010年3月21日初诊。患者自述3周前因着凉而出现腰两侧冷痛重着…<TRANSLATION_END>

---

Intended use and limitations

Built for training and evaluating Chinese braille ↔ Chinese machine translation, primarily Qwen-family models fine-tuned with LLaMA-Factory. Tone markers matter for pronunciation, and the tone-ratio curriculum (r100 → r10) exists to teach the model to recover Chinese from increasingly tone-sparse braille, which is what real-world braille looks like.

Limitations to be aware of:

• The medical, TCM, and pathology content is training data, not clinical guidance. It comes from encyclopedia text and course material and has not been reviewed for medical accuracy.
• Braille in braille_spaced/ is machine-encoded from Chinese text via the GF 0019-2018 encoder, not transcribed by human braille readers, except the pathology OCR pairs. The encoder is validated cell-for-cell against the standard's worked examples, but it is not a substitute for human transcription review. Three known GF 0019-2018 rule gaps remain unimplemented.
• Word segmentation for 分词连写 comes from jieba plus GF 0019-2018 rules, so segmentation disagreements with a human transcriber are expected at the margins.
• cleaned_2345_v3/ outside stage 2 is not word-spaced. Do not mix spaced and unspaced data into one stage without intending to.

Licensing

license: other, because this is a derivative of several upstream corpora under different terms. Check the per-domain source table above and honor each upstream license. The 医学 and 中医 sources are apache-2.0; the 通用 and 数学 sources carry their own terms on the HF Hub. The braille encoding, cleaning, spacing restoration, and assembly work in this repo is released for research use.

Citation

@misc{vision_braille_dataset,
  title  = {vision-braille-dataset: Chinese--Braille parallel corpora with 分词连写 word spacing},
  author = {Wu, Alan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Violet-yo/vision-braille-dataset}
}

---

3. cleaned_080126_v1/ — four-stage word-spaced training corpus

The corpus behind the cleaned_080126_v1 training pipeline. Braille→Chinese throughout, word-divided with U+2800, <BRAILLE_END> / <TRANSLATION_END> markers on every record. Replaces the retired cleaned_2345_v3 passage data.

Stage	What it trains	Train	Val
1	Sentence, 100% tone	298,639	37,344
2	Passage, 100% tone	70,916	8,863
3	Passage curriculum, r100 → r10	70,916 (10 chunks)	8,863
4	Passage, 10% tone	70,916	8,863

cleaned_080126_v1/
├── dataset_info.json                                    27 LLaMA-Factory entries
├── stage1_sentence_spaced_br2zh_r100_{train,val}.jsonl
├── stage2_passage_spaced_br2zh_r100_{train,val,test}.jsonl
├── stage3_curriculum_mix_separate/
│   └── stage3_curriculum_spaced_br2zh_mix_{r100..r10}_{train,val}.jsonl   20 files
└── stage4_passage_spaced_br2zh_r10_{train,val}.jsonl

Stage 1 is the cleaned_2345_v3 sentence corpus rewritten into the passage record format (Braille content: header, end markers), so the prompt shape does not change under the model between stages.

Stages 2–4 are built from the same source passages as braille_spaced/ (通用 / 数学 / 医学 / 中医), re-encoded at full tone. Stage 3 splits the corpus into ten disjoint chunks and thins tone marks to 100%, 90%, … 10% respectively; stage 4 applies 10% to the whole corpus. Realised retention runs above target (r10 lands near 15.7%) because tone cells immediately following a punctuation cell are always kept.

Chinese targets use fullwidth punctuation throughout, matching what the braille encodes. The source passages carry ASCII , . : ; ? ! ( ), which the GF 0019 encoder maps to fullwidth before rendering; the same span-aware mapping is applied to the Chinese side so the two agree. Decimal points and thousands separators inside math spans are left alone.