minthanthtoo-cs/Burmese-Classics-OCR-RAW

Burmese (Myanmar) Books Dataset – Burmese Classics OCR (Daily Rolling Project)

Overview

A daily rolling dataset of Burmese books, built for AI, OCR, and NLP research.
Each entry includes OCR text with metadata: title, author, page index, and Burmese character ratio.

This project fills a critical gap in Burmese-language resources:

• Scarcity of public-domain Burmese text.
• High technical and financial barriers to corpus building.
• Enables incremental, open access for researchers and developers.

Note: This is a personal, incremental effort — updates may be partial. Source: Derived from the largest freely accessible Burmese book archive, fully open and unrestricted.

⚠️ Error logs and page text-index mismatch all solved up to the current 3rd parquet.

Dataset Format

Data is stored as Parquet shards (~200 MB each) from source JSONL files.

Each Parquet record includes:

• uid – unique identifier, assigned internally by the Burmese Classics Online website
• title – book title
• author – book author
• page – page index (starting from 1)
• burmese_percent – proportion of Burmese characters on the page
• text – OCR text content of the page

1 jsonl line = 1 page (raw OCR output).

features = {
    "uid": datasets.Value("string"),
    "title": datasets.Value("string"),
    "author": datasets.Value("string"),
    "burmese_percent": datasets.Value("float32"),
    "text": datasets.Value("string"),
}

---

Burmese Text Dataset Summary

Dataset Scale (Excluding Predominantly English Content)

• Total Books: ~18,000
  (Currently processed: 13,105)

• Total Pages: ~2,000,000
  (Currently processed: 1,968,300)

• Total Characters: ~3B
  (Currently processed: 2,921,346,487)

• Burmese Characters: ~3B
  (Currently processed: 2,474,165,593)

• Burmese Words: ~290M
  (Currently processed: 227,395,926)

---

Data Processing Status (Parquet-Level)

Shard	Books	Pages	Total Chars	Burmese Chars	Burmese Words	Burmese Tokens
shard_00000	1,421	255,466	362,515,284	305,485,300	29,142,281	649,602,248
shard_00001	1,568	276,365	363,792,509	312,381,022	30,106,015	664,469,926
shard_00002	1,813	267,809	351,985,963	300,219,563	27,244,263	635,428,084
shard_00003	2,162	240,005	350,787,024	299,815,445	26,680,447	633,367,781
shard_00004	1,620	219,715	345,906,476	292,176,302	26,350,542	616,973,180
shard_00005	1,656	215,226	347,054,435	293,146,986	26,366,230	618,963,554
shard_00006	1,485	215,411	348,185,510	294,033,054	26,420,260	621,017,322
shard_00007	1,014	224,962	365,498,334	304,108,340	28,502,450	645,473,176
unknown	366	53,341	85,620,952	72,799,581	6,583,438	153,817,647
Total (current)	13,105	1,968,300	2,921,346,487	2,474,165,593	227,395,926	5,239,112,918

➡ Progress: ~73% of books processed (13,105 / 18,000)
➡ At current rate, full dataset expected to reach ~3B characters and ~290M Burmese words.

---

Progress

Books processed: 13,105 / 18,000
Characters processed: 2.92B / ~3B

Progress: ████████████░░ 73%

---

Tokenization Comparison

Tokenizer / Encoding	Vocab Size	Tokens (Current: 2.92B chars)	Projected (Full ~3B chars)	Notes
GPT-4 cl100k_base	~100k	~5.24B (Burmese tokens)	~5.4–5.5B	Standard GPT-4 encoding
GPT-3 r50k_base	~50k	(est. ~7.5–7.8B)	~8B	Older GPT-3 encoding
WordPiece (mBERT)	119,547	(est. ~1.7B)	~1.8B	Pretrained multilingual vocab
SentencePiece (custom, 8k)	8k (expandable)	(est. ~0.85–0.9B)	~0.9B	Tunable vocab (8k–32k common)
Char-Level	~100–200	2,921,346,487	~3B	Direct mapping, inefficient

---

Efficiency Ratios

• Avg Tokens per Burmese Character: 2.118
• Avg Tokens per Burmese Word: 23.04

---

Notes

• Current processed = 13,105 books (~2.92B chars).
• Full values = projected scaling to 18k books (3B chars).
• Vocabulary size strongly impacts efficiency:
  • Smaller vocab → more tokens, longer sequences, lighter embeddings
  • Larger vocab → fewer tokens, shorter sequences, heavier embeddings
• SentencePiece provides the best trade-off for Burmese efficiency with adjustable vocab sizes.

⚠️ These totals exclude categories likely to contain mostly English content: Technical, Newspaper | Journal, and Language.

---

Accessing Data

The dataset can be accessed programmatically from Hugging Face using the datasets library:

from datasets import load_dataset

dataset = load_dataset("minthanthtoo-cs/Burmese-Classics-OCR-RAW")
print(dataset)

# Access the 7th entry (index 6)
sample = dataset['train'][6]

# Print metadata and 1.2k text snippet
print(f"Title       : {sample['title']}")
print(f"Author      : {sample['author']}")
print(f"UID         : {sample['uid']}")
print(f"Burmese %   : {sample['burmese_percent']:.2f}%")
print("Text snippet:", sample['text'][:1200].replace('\n', ' '), "...")

• Each shard is automatically loaded as a dataset split.
• The Parquet format ensures fast, columnar access and efficient filtering.

---

Processing Notes

• Each book contains ~3k–100k Burmese characters (~2k–100k tokens).
• Only books meeting a minimum Burmese content threshold are included.
• Dataset is rolling: new books are added daily.

---

Visualizations (Preview)

(These charts highlight which PDFs provide the most content relative to their file size in MB)

Distribution of Pages per Book

File Size vs Pages (MB/Page)

---

Categories Summary

⚠️ Note: The following stats are estimated from 24k parsed PDF metatdata only (out of 36k total).
Numbers may shift slightly once all files are fully processed.

Category	Books	Total Size (MB)	Total Pages	Avg Size/Page	Est Tokens
Technical	1,891	26,909.73	626,402	0.048	212,976,680
Language	4,165	33,951.70	721,935	0.090	259,896,600
Novel	3,770	32,987.31	152,693	0.107	369,517,060
Religion	2,725	40,064.28	441,158	0.107	114,701,080
Knowledge	1,586	19,082.50	113,560	0.117	31,796,800
Business	176	2,165.26	9,976	0.117	2,294,480
Translation	758	11,792.22	65,170	0.122	16,292,500
Martial Art	1,651	25,345.99	146,292	0.126	32,184,240
Mystery	486	6,523.44	18,631	0.130	4,285,130
History	407	6,248.73	32,805	0.148	8,857,350
Health	263	3,830.64	20,831	0.151	4,686,975
Rare	71	1,182.29	3,198	0.150	639,600
Poem	241	2,082.35	10,070	0.161	906,300
Detective	256	3,305.23	10,310	0.202	2,371,300
Textbooks	131	2,281.23	11,332	0.199	3,512,920
Other	408	4,759.20	21,994	0.219	4,398,800
Magazine	364	8,689.53	21,011	0.215	4,202,200
Agricultural	85	447.88	3,237	0.222	647,400
Comic	978	22,719.21	39,699	0.234	27,789,300
Thriller	1,146	10,426.67	47,389	0.103	114,681,380
Lyrics	325	719.00	10	0.800	2,400
Newspaper / Journal	14,472	127,541.61	371,311	0.290	408,442,100
TOTALS	24,240	276,077.47	2,903,645	—	1,628,594,035

Dataset Stats

• Total categories: 23
• Total books: 24,240
• Total size: 276,077.47 MB
• Total pages: 2,903,645
• Pages per MB (global): 10.52
• Estimated tokens: ~1.63B

---

License

This dataset is open-source under the Apache 2.0 License, intended to support Burmese AI research and development.

• Allows use, modification, and redistribution
• Attribution required

Apache 2.0 License

---

Credits

• The dataset was parsed from Burmese Classics Online.
• Thanks to the original site for providing access to Burmese literary works.

---

Disclaimer

This is a personal project. Data may be incomplete, inconsistent, or partial due to the rolling update process. Use for research purposes only.

---