openpd-strict-v1

An ~8.8B-token pretraining mix of public-domain and CC0 text, tokenized with the GPT-2 tokenizer and packed into uint16 NumPy shards. Every input document is public domain (US) or CC0, checked per document where the source records licenses and per source otherwise.

This is the strict variant of openpd-v1, which trained openpd-gpt2-v0: the same sources, sampling, and cleaning, minus Ubuntu IRC, whose public-domain status rests on Ubuntu community policy rather than law or per-author dedication. Without IRC it contains no chat logs and no personal nicknames, and nothing in it depends on the private pseudonymization salt.

This is a derived, tokenized dataset: the underlying texts come from the Common Pile v0.1 filtered collections and the Library of Congress Selected Digitized Books collection (storytracer/LoC-PD-Books). What's added here is the selection, mix weights, light cleaning, and tokenization. Released under CC0-1.0 to keep it as open as its sources.

Files

• openpd-strict-v1_train_000000.npy … 000088.npy: 8,823,603,133 train tokens, 3,781,005 documents (100M tokens per shard; the last is partial)
• openpd-strict-v1_val_000000.npy: 47,295,167 val tokens, 19,305 documents (0.5% of documents, chosen by id hash)
• openpd-strict-v1_meta.json: manifest: mix config, per-source token/document counts, per-shard source counts, licenses seen, cleaning steps, per-input-file sha256, tokens_sha256 fingerprint (e777bfedc277d972…)
• openpd-strict-v1_report.json: content report (table-like / layout-heavy / very long documents)

Each document starts with <|endoftext|> (50256). Documents from all sources are interleaved evenly through the stream (by document count), and each shard's manifest entry lists how many tokens of each source it holds. Load a shard with np.load(path); decode with tiktoken.get_encoding("gpt2").

Sources

Source	License basis	Train tokens	Share	Cleaning
Project Gutenberg	PD (US)	2.176 B	24.7%	hard-wrapped prose unwrapped, verse kept
Pre-1929 books (Internet Archive)	PD (US)	2.113 B	24.0%	de-hyphenated
LoC Selected Digitized Books	CC0 / PD (US)	1.156 B	13.1%	de-hyphenated, prose unwrapped
Caselaw Access Project	PD (US)	0.724 B	8.2%	none
arXiv abstracts	CC0	0.550 B	6.2%	none
US Government Publishing Office	PD (US gov works)	0.469 B	5.3%	de-hyphenated
Biodiversity Heritage Library	PD	0.435 B	4.9%	de-hyphenated
PubMed Central, CC0 subset	CC0	0.410 B	4.6%	none
peS2o, CC0/PD subset	CC0 / PD	0.301 B	3.4%	none
arXiv papers, PD/CC0 subset	PD / CC0	0.284 B	3.2%	de-hyphenated
Regulations.gov	PD (US gov works)	0.202 B	2.3%	de-hyphenated
Python Enhancement Proposals	PD	0.003 B	0.0%	none
Total		8.824 B

Each source's share is its sampled fraction of the upstream collection (repeat in the mix config; e.g. 44% of Project Gutenberg, 4% of case law), chosen to weight books and prose over legal and government text. License filters use each document's own license field where the source has one (metadata.license; peS2o's oa_license).

Personal information

No chat or social-media sources. Case law and government documents name people as public records do. No additional PII scrubbing was applied.

Known issues

From openpd-strict-v1_report.json:

• Very long documents. 739 documents (7.5% of tokens) are longer than one 524,288-token training batch, up to 7.2M tokens; each fills several consecutive batches by itself. Training on openpd-v1 (same documents) showed occasional very-low-loss steps and gradient-norm spikes from these (handled by gradient clipping).
• Layout-heavy government text. 4.2% of tokens (30,845 documents) are in documents where ≥60% of tokens contain no letters: indentation runs, dot leaders, rule lines, column layouts.
• Numeric data tables. 0.3% of tokens (242 documents) are in documents where ≥30% of tokens contain digits (tariff codes, emissions tables).
• OCR noise in the book, LoC, BHL, and case-law text; BHL is the noisiest. Greek and other non-Latin passages in older books are often garbled.
• No cross-source deduplication. The same work can appear in Project Gutenberg, the pre-1929 books, and LoC.
• Period bias. Most text predates 1930, with its language, knowledge, and prejudices.
• US-centric public domain. Some works may still be under copyright outside the United States.
• Tokenizer. The GPT-2 tokenizer was trained by OpenAI on WebText, which is not open data.
• Build provenance. Built from commit 1466a87 with uncommitted local changes (-dirty in the manifest); the per-input-file hashes in the manifest identify the exact inputs.

Citation

Common Pile v0.1: Kandpal et al. 2025, The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text. Library of Congress Selected Digitized Books compiled by Sebastian Majstorovic (storytracer).