README.md

metadata

license: mit
task_categories:
  - text-generation
  - text-classification
  - token-classification
  - question-answering
  - zero-shot-classification
  - summarization
  - feature-extraction
  - fill-mask
  - sentence-similarity
language:
  - en
tags:
  - NLP
pretty_name: bookcorpus
size_categories:
  - 10M<n<100M

BookCorpus — Cleaned for Pre-training LLMs

A cleaned, deduplicated, document-segmented version of SamuelYang/bookcorpus

TL;DR

Property	Value
Rows (sentences)	33,649,142
Documents (books)	4,086
Format	CSV, 3 columns: doc_id, sent_id, text
Language	English (lowercased)
Source	SamuelYang/bookcorpus (74,004,228 raw rows)

Schema

Column	Type	Description
doc_id	int	Inferred document/book identifier. Sentences sharing the same doc_id come from the same book.
sent_id	int	Sentence position within its document (0-indexed). Preserves original order.
text	string	Cleaned sentence text (lowercased, normalized).

How to use it

Quick load

from datasets import load_dataset

ds = load_dataset("kd13/bookcorpus-clean", split="train")
print(ds[0])
# {'doc_id': 0, 'sent_id': 0, 'text': 'i wish i had a better answer ...'}

Cleaning pipeline

Applied in this order to the source dataset:

1. Unicode + whitespace normalization — NFKC normalization, collapse consecutive whitespace, strip.
2. Document segmentation — since the source is a flat stream of sentences without book IDs, document boundaries are inferred from telltale markers at the start of books:
   • ISBN lines (e.g. isbn : 1492913731)
   • Copyright declarations (copyright 2013 ...)
   • all rights reserved
   • chapter 1
3. Line-level filters — sentences are dropped if they:
   • have fewer than 20 or more than 1000 characters
   • match boilerplate patterns (copyright, ISBN, "all rights reserved")
   • have an alphabetic-character ratio below 0.6
   • have a digit ratio above 0.3
   • contain no alphabetic characters
4. Language filter — cheap English stop-word ratio check (≥ 5% of tokens must be in a small English stop-word set; short lines pass through).
5. Within-document exact dedup — SHA-1 hashing drops repeated sentences inside the same book (e.g. recurring chapter headers, section dividers). Note: dedup is not applied globally — sentences like "he nodded." occur legitimately across many books.
6. Document filter — books with fewer than 8 surviving sentences are dropped (not enough context for NSP).
7. Cross-document near-duplicate removal — a SHA-1 fingerprint of each document's first 5 sentences identifies same-book re-uploads; duplicates are dropped.

Cleaning statistics

Metric	Value
Raw rows (sentences) in source	74,004,228
Documents detected	6,779
Documents kept	4,086
Documents dropped (< 8 sentences)	973
Documents dropped (near-duplicate)	1,720
Sentences kept	33,649,142

Drop rate: ~40% of detected documents removed (mostly same-book re-uploads and too-short documents).

Source & licensing

• Source dataset: SamuelYang/bookcorpus
• Original corpus: BookCorpus (Zhu et al., 2015), originally scraped from Smashwords. The original BookCorpus has well-documented provenance and consent concerns; downstream users should review them before commercial use.
• This cleaned derivative is released under the MIT License for the cleaning code and structuring effort. The underlying text retains whatever rights apply to the upstream source.