README.md

metadata

language:
  - bn
license: cc0-1.0
pretty_name: ApuVerse — Bibhutibhushan Bandyopadhyay Bengali Literary Corpus
task_categories:
  - text-generation
size_categories:
  - n<1K
tags:
  - bengali
  - bangla
  - literature
  - ocr
  - public-domain
  - continued-pretraining
  - style-transfer
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

ApuVerse — Bibhutibhushan Bandyopadhyay Bengali Literary Corpus

Long-form Bengali prose by Bibhutibhushan Bandyopadhyay (1894-1950): the Apu novels (Pather Panchali, Aparajito), Aranyak, Adarsha Hindu Hotel, Ashani Sanket, and seventeen short stories. Extracted from Internet Archive scans of mid-century printed editions, segmented into chapters and stories, and shipped with the OCR confidence scores that produced each record so consumers can filter to the quality level their use case needs.

At a glance

records	104
works	22 (5 long-form, 17 short stories)
characters	2,641,429
words	433,867
tokens (est.)	841,352
language	bn (Beng)
author	বিভূতিভূষণ বন্দ্যোপাধ্যায়
first published	1929–1959
editions scanned	1960–1964
mean OCR confidence	78.6 / 100
licence	cc0-1.0

Each record is one chapter of a novel or one complete short story, so records carry whole narrative arcs rather than fragments.

Text is uncorrected OCR. Rather than silently shipping the errors, every record carries the OCR confidence scores that produced it, so you can filter to whatever quality level your use case needs — see Filtering by quality below.

Contents

Long-form works

work	title (en)	first published	edition	chapters	chars	tokens	mean conf
অপরাজিত	Aparajito	1932	1961	40	931,943	294,890	77.9
পথের পাঁচালী	Pather Panchali	1929	1961	31	490,549	159,850	77.8
আরণ্যক	Aranyak	1939	1964	14	385,928	114,782	84.7
আদর্শ হিন্দু হোটেল	Adarsha Hindu Hotel	1940	1960	1	302,710	101,432	72.1
অশনি সংকেত	Ashani Sanket	1959	1964	1	174,757	55,980	77.0

Short stories

One record each; collection_bn / collection_en name the volume they were printed in.

work	title (en)	first published	edition	records	chars	tokens	mean conf
মেঘ-মল্লার	Megh-Mallar	—	1964	1	35,633	10,858	77.5
মৌরীফুল	Mourifool	—	1964	1	34,250	11,303	76.2
দ্রবময়ীর কাশীবাস	Drabamayeer Kashibas	—	1964	1	32,165	10,891	77.2
তারানাথ তান্ত্রিকের গল্প	Taranath Tantriker Galpo	—	1964	1	30,563	9,563	78.5
কিন্নর দল	Kinnar Dal	—	1964	1	28,059	8,885	79.0
পুঁই মাচা	Pui Macha	—	1964	1	24,812	8,292	75.5
ননুমামা ও আমি	Nonumama O Ami	—	1964	1	22,019	7,137	77.7
ক্যানভাসার কৃষ্ণলাল	Canvasser Krishnalal	—	1964	1	21,641	7,118	76.2
নাস্তিক	Nastik	—	1964	1	20,958	6,651	75.2
ভগ্গুলমামার বাড়ী	Bhagtulmamar Bari	—	1964	1	20,382	6,421	79.8
আহ্বান	Ahoban	—	1964	1	16,476	5,305	77.7
সিঁদুরচরণ	Sindurcharan	—	1964	1	16,154	5,363	76.8
একটি ভ্রমণ-কাহিনী	Ekti Bhraman-Kahini	—	1964	1	14,088	4,441	78.3
কনে দেখা	Kone Dekha	—	1964	1	12,543	3,908	77.9
বিপদ	Bipad	—	1964	1	11,879	3,904	77.7
কুশল পাহাড়ী	Kushal Pahari	—	1964	1	10,311	3,256	75.8
তুচ্ছ	Tuchcha	—	1964	1	3,609	1,122	77.3

Record size distribution

	chars
smallest	3,609
median	18,658
90th percentile	35,633
largest	302,710

Sources

Every record traces back to a scanned printed edition on the Internet Archive.

source PDF	Internet Archive item	pages	works	status
Aaranyak.pdf	in.ernet.dli.2015.453018	200	1	ready
Adarsha-Hindu-hotel.pdf	in.ernet.dli.2015.457479	208	1	ready
Ashani-sanket.pdf	in.ernet.dli.2015.457448	100	1	ready
Aparajita.pdf	in.ernet.dli.2015.455079	818	2	ready
Chotogolpo.pdf	in.ernet.dli.2015.456545	204	17	ready
Arogya-Niketan.pdf	in.ernet.dli.2015.316173	414	0	excluded

Excluded sources

Kept in the manifest rather than deleted, so the reason stays on the record:

• Arogya-Niketan.pdf (excluded) — NOT Bibhutibhushan. This is আরোগ্য নিকেতন (1953) by Tarashankar Bandyopadhyay — the scan's own text names জীবন মশাই, its protagonist. The Internet Archive/DLI record miscredits it to 'Bandyopadhyay, Bibhutibhushan'; the two authors share a surname. Excluded so the corpus stays single-author and the style signal stays clean.

Schema

column	type	meaning
text	string	the training field: NFC Bengali prose, paragraphs split by \n\n
id	string	stable record id, <dataset>/<work>/<unit>
work_bn / work_en / work_id	string	the work this record belongs to
collection_bn / collection_en	string	set for stories in a collection, else null
unit_type	string	chapter or story
unit_index	int32	position within the work
n_chars / n_words / n_tokens_est	int32	size; n_tokens_est is null if no tokenizer was configured
first_published / edition_year	int32	year of the work vs. of the scanned printing
ia_identifier / source_url / source_pdf	string	provenance
ocr_engine / ocr_params / ocr_source	string	how the text was produced
conf_mean	float32	mean OCR word confidence (0-100), null if unknown
conf_p10	float32	10th-percentile page confidence — catches local damage a mean hides
low_conf_word_frac	float32	fraction of words scored below 70
pages	list<struct>	per source page: page_index, printed_page, char_start, char_end, conf_mean, low_conf_word_frac

conf_* columns are null, never 0.0, when confidence is genuinely unknown — a fabricated zero would read as worst-possible quality and silently drop good text.

Filtering by quality

from datasets import load_dataset

ds = load_dataset("<repo>", split="train")

# 1. drop whole records that are mostly noise
good = ds.filter(lambda r: (r["conf_mean"] or 0) >= 75.0)

# 2. or excise only the bad pages, keeping the rest of the chapter
def drop_bad_pages(rec, floor=70.0):
    keep = [rec["text"][p["char_start"]:p["char_end"]]
            for p in rec["pages"] if (p["conf_mean"] or 0) >= floor]
    return {"text": "".join(keep)}

clean = ds.map(drop_bad_pages)

The pages spans are contiguous and tile text exactly, so slicing and rejoining them never loses or duplicates characters.

Intended use

Continued pretraining and LoRA/PEFT style tuning on next-token prediction — the records are raw prose, not instruction pairs. Text is not pre-chunked to a token window: packing to 2048/4096 with an EOS at document boundaries is a training-time decision, and pre-chunking would sever sentences and take that choice away. n_tokens_est is provided so you can plan packing without tokenizing first.

There is a single train split. For style evaluation, hold out a work yourself (filtering on work_id) so the held-out text is genuinely unseen.

# hold out one novel for evaluation
train = ds.filter(lambda r: r["work_id"] != "aranyak")
eval_ = ds.filter(lambda r: r["work_id"] == "aranyak")

How the text was produced

1. Extract — text and per-word confidence are read from the Internet Archive hOCR derivative of each scan. The PDF's own invisible text layer carries the same characters but no confidence, so hOCR is preferred wherever an item exists.
2. Normalise — Unicode NFC. Zero-width joiners are preserved: in Bengali they carry the reph / ya-phala distinction, so stripping them would change words.
3. Strip page furniture — running headers and page numbers are removed. The same running title is OCR'd differently on nearly every page, so headers are detected by fuzzy-clustering candidate lines rather than by exact match.
4. Reflow — printed line breaks are rejoined into paragraphs, including across page boundaries. Paragraphs are separated by \n\n.
5. Segment — split into works by verified page ranges, then into chapters by printed headings.

Limitations

• OCR is uncorrected (tesseract 5.0.0-alpha-20201231-10-g1236). Mean record confidence is 78.6/100, and 2 of 104 records fall below 75. Expect wrong conjuncts, stray Latin characters and mangled punctuation. The text is good enough to learn style and syntax from, and not a reliable edition to quote from.
• Front matter, editorial prefaces and critical appendices were excluded by page range, but header and page-number removal is heuristic and a few survive mid-text.
• Paragraph boundaries are reconstructed from printed line breaks and are approximate; dialogue segments more reliably than continuous narration.
• Chapter splits come from printed headings. Books whose headings did not survive OCR appear as a single long record rather than being split arbitrarily.
• Scans are of mid-century reprints, not first editions, so spelling reflects the printing house's conventions of the day.
• Single author and single genre: this is literary prose from one writer, and is not a general-purpose Bengali corpus.

Provenance and rights

Scans and OCR come from the Internet Archive; see ia_identifier and source_url on each record, and the Sources table above.

Cataloguing note: Internet Archive/DLI author fields for this collection are not always reliable. Attribution here was checked against each scan's own title page, not taken from the metadata record.

The author died in 1950, so these works are public domain in India (life + 60, expired 2011). US status is not automatic: works published 1929 or later by a foreign author may have had US copyright restored under the URAA, which would cover Aranyak (1939) and Adarsha Hindu Hotel (1940). Confirm before publishing this dataset publicly.