disco-jack-basement/byob-pd-book-corpus

Dataset Card for byob_llm Public-Domain Book Corpus

A multilingual, public-domain corpus of literary, philosophical, and scientific works, published for character-level language-model pre-training. It ships in two forms: raw per-author .txt files (one file per author) and a pre-tokenized, memmap-ready cache (train.bin / val.bin / meta.json) for fast training.

Dataset Sources

• Project Gutenberg (https://www.gutenberg.org)
• Wikisource (https://wikisource.org)

All texts are public domain in their source jurisdiction. Project Gutenberg boilerplate headers/footers and trademark notices are stripped, and text is NFKD-normalized. Verify public-domain status in your own jurisdiction.

Dataset Structure

Three nested tiers - medium is a subset of large, which is a subset of xlarge:

tier	characters	authors	vocab
medium	523,561,434	109	97
large	1,079,432,702	248	198
xlarge	2,050,985,583	431	199

(Size is measured in characters; size_categories counts characters as units.)

• Raw text: <tier>/<author>-complete.txt.
• Per-tier manifests: indices/<tier>.md (authors, works, char counts).
• Prepared caches: prepared/<tier>_bin/{train.bin,val.bin,meta.json}, where meta.json = {vocab_size, stoi, itos, train_len, val_len, dtype:"uint8"}. These are a byob-internal convenience cache, fully reproducible from the raw .txt; pull them with hf_hub_download, not load_dataset.

Uses

Language-model pre-training, especially character-level GPTs. The prepared/<tier>_bin/ caches are flat uint8 streams (one char = one byte, vocab < 256) meant to be memmapped (see byob_llm CharDataset.from_bin). Not annotated; not intended for supervised tasks needing labels.

from datasets import load_dataset
ds = load_dataset("<repo-id>", data_dir="medium")   # a raw tier

Dataset Creation

Curation Rationale

PUBLIC DOMAIN ONLY. The corpus enforces, in code, a strict rule: NO Russian content anywhere - no Russian authors, no Russian-language works, no Russia-themed material. This is a deliberate, code-enforced curation policy.

Source Data

Harvested from Project Gutenberg (via Gutendex) and Wikisource; Gutenberg boilerplate stripped; cleaned, de-duplicated, and NFKD-normalized. The tiers are frozen. See indices/<tier>.md for the full per-author manifest. The corpus spans English, American, French, German, Italian, Latin, Ancient Greek, and Ukrainian literature and philosophy, among others.

Bias, Risks, and Limitations

Historical texts reflect the dated and biased language of their eras. The corpus is multilingual but English-dominant. No content filtering beyond the curation rules.

License and Attribution

Released under CC0-1.0 (public-domain dedication). Note: Wikisource editorial apparatus may be CC-BY-SA, and the NFKD normalization is a derivative. Please credit Project Gutenberg and Wikisource.