Toolster/ebook-format-samples

E-book Format Samples

Four public-domain books, each converted into up to 23 e-book and document formats: 88 files in total. For every book there is also the XHTML source it was converted from and a plain-text ground truth, so you always know what a file is supposed to contain.

Why this dataset

Legal sample files of old e-book formats are hard to find. Search for a LIT, LRF or TCR file and you mostly get pirated books or files of unknown origin, with no idea what text is inside. Here the same known text exists in every format, which makes the set useful for:

• Testing format recognition. File type detectors, MIME sniffing, ML classifiers of file types. Every row of metadata.csv gives the extension, MIME type, first 16 bytes and, for Palm database based files, the type/creator code. Several formats share a container (EPUB, KEPUB, DOCX, HTMLZ, TXTZ and PMLZ are all ZIP; MOBI, AZW3 and three PDB flavours are all Palm databases), which is exactly where detectors get confused.
• Developing converters and readers for rare formats. LIT, LRF, TCR, SNB, RB, PMLZ, eReader, PalmDOC and zTXT files with a text you can check against. Good for writing a new parser, fixing an old one or adding a format to an existing reader.
• Measuring text extraction. Compare what your parser gets out of a file with ground_truth/<book>.txt. Useful before feeding e-books into a search index or an LLM pipeline, where silent text loss is hard to notice.
• Regression tests for reading apps. Illustrations, a footnote, block quotes, scene breaks, a table of contents, Cyrillic text, accented Spanish and a book made of several separate works.

Books

id	Language	Text	Features
alice_en	English	Lewis Carroll, Alice's Adventures in Wonderland, chapters I-II (1865)	2 illustrations by John Tenniel, italics, scene breaks
art_of_war_en	English	Sun Tzu, The Art of War, chapter I, translated by Lionel Giles (1910)	footnote, commentary as block quotes, numbered paragraphs
quijote_es	Spanish	Miguel de Cervantes, Don Quijote de la Mancha, chapters I-II (1605)	accented letters, ¡¿ punctuation, long chapter titles
chekhov_ru	Russian	Anton Chekhov, three short stories (1883-1884)	Cyrillic, dialogue dashes, non-breaking spaces, three works in one book

Formats

variant	Format	Extension
epub2, epub3	EPUB 2 and EPUB 3	.epub
kepub	Kobo EPUB	.kepub
azw3	Kindle KF8	.azw3
mobi7	Mobipocket 6 (old MOBI)	.mobi
mobi_joint	MOBI + KF8 in one file	.mobi
fb2	FictionBook 2	.fb2
lit	Microsoft Reader	.lit
lrf	Sony BroadBand eBook	.lrf
pdb_ereader, pdb_palmdoc, pdb_ztxt	Palm database: eReader, PalmDOC, zTXT	.pdb
rb	RocketBook	.rb
tcr	Psion TCR	.tcr
snb	Shanda Bambook	.snb
pmlz	eReader PML in ZIP	.pmlz
htmlz	HTML in ZIP (Calibre)	.htmlz
txtz	text in ZIP (Calibre)	.txtz
html_zip	HTML in plain ZIP	.zip
txt	plain text, UTF-8	.txt
rtf	Rich Text Format	.rtf
docx	Word	.docx
pdf	PDF	.pdf

Missing on purpose: rb for the Spanish and Russian books, and pdb_ereader and pmlz for the Russian book (see below).

Layout

metadata.csv                         one row per file
files/<book>/<book>.<variant>.<ext>  the samples
ground_truth/<book>.txt              expected text of each book
source/<book>/<book>.html            XHTML source (+ images) all formats were made from

metadata.csv columns: file_name, book_id, language, format, variant, extension, mime, text_encoding (for 8-bit and text formats), size_bytes, sha256, magic_hex (first 16 bytes), pdb_type_creator (bytes 60-67 of Palm database files, such as BOOKMOBI, TEXtREAd, PNRdPPrs), ground_truth, converter, options, roundtrip_word_recall.

MIME types follow common practice. Formats without any widely used MIME type (TCR, SNB, PMLZ, HTMLZ, TXTZ) are listed as application/octet-stream.

How the files were made

All files were produced from the XHTML sources with Calibre 9.17.0 (ebook-convert), chapters split on <h2>, title, author, language and year set as metadata. The exact extra options for each file are in the options column.

• 8-bit formats. PalmDOC and zTXT are written in cp1252 for the English and Spanish books and in cp1251 for the Russian one (see text_encoding). Read them with that code page.
• Gaps. Calibre's PML writer, which is also used for eReader PDB, keeps only cp1252 and writes ? for every Cyrillic letter, so those two files are not included for the Russian book. Calibre also writes RocketBook files for the Spanish and Russian texts that it cannot read back itself, so those were left out too.
• Round-trip check. Every file was converted back to text with the same Calibre and compared with the ground truth. roundtrip_word_recall is the share of ground-truth words found in that text (lowest value in the set: 0.994). It is a sanity check that the file really contains the book, not a benchmark of other tools.

Keep in mind that all samples come from one converter. Files made by the original vendor tools (Microsoft Reader, Sony, Mobipocket Creator, Kindle) can use features and quirks that Calibre does not produce.

Ground truth

UTF-8 plain text: title, author, then each chapter title and its paragraphs, separated by blank lines. The footnote reference is written as [1] and the note itself follows its chapter. Illustrations and scene breaks carry no text. Non-breaking spaces are replaced with ordinary spaces.

Usage

from datasets import load_dataset
from huggingface_hub import hf_hub_download

meta = load_dataset("Toolster/ebook-format-samples", split="train")
row = meta.filter(lambda r: r["variant"] == "lit")[0]
path = hf_hub_download("Toolster/ebook-format-samples", row["file_name"], repo_type="dataset")
truth = hf_hub_download("Toolster/ebook-format-samples", row["ground_truth"], repo_type="dataset")

Sources and license

• Texts: Project Gutenberg eBooks #11, #132 and #2000 (only the chapters themselves were taken, without the Gutenberg header and license), and Russian Wikisource (Chekhov, Complete Works edition texts). All are in the public domain.
• Illustrations: John Tenniel (1865), from Wikimedia Commons, public domain.
• The converted files, ground truth and metadata are released under CC0 1.0.