typewriter-ai/TypewriterLM-1913-data

TypewriterLM-1913 data

Cleaned, paragraph-segmented historical English corpora used to train TypewriterLM, a language model trained only on text published up to 1913. Each corpus is stored as Parquet shards in its own folder, one row per document, with the text in a paragraphs column (a list of strings).

Paper: Pretraining Language Models on Historical Text

Date range: some documents fall outside 1700-1913

The cleaned corpora here are not all restricted to 1700-1913. Four of them contain documents from before 1700 or after 1913:

Folder	Documents before 1700	Documents after 1913
ced_clean	87	0
lampeter_clean	24	0
hansard_clean	0	13,659 (up to 1920)
royal_society_clean	0	806 (up to 1920)

In the paper, we filter again to 1700-1913 before training, using the year column that every corpus carries. To reproduce the training data, apply the same filter, for example:

from datasets import load_dataset

ds = load_dataset('typewriter-ai/TypewriterLM-1913-data', data_files='hansard_clean/*.parquet', split='train')
ds = ds.filter(lambda row: 1700 <= row['year'] <= 1913)

Institutional Books is not included

The largest component of the training data is Institutional Books 1.0. We cannot redistribute it: doing so would violate the dataset's terms of service. It is therefore not part of this repository. To reproduce that part of the training data, download Institutional Books yourself and run the pipeline in our code repository (scripts/clean/clean_institutional.py), which keeps books published between 1700 and 1913 whose OCR scores (ocr_score_src and ocr_score_gen) are both at least 90, then strips headers, stamps, call numbers and other non-prose material.

Corpora

Folder	Corpus	Documents	Years in this release
blbooks_clean	British Library Books	19,736	1700-1898
hansard_clean	Historic Hansard (UK Parliament debates)	115,633	1802-1920
royal_society_clean	Royal Society Corpus open v6.0	16,171	1700-1920
old_bailey_clean	Old Bailey Corpus 2.0	637	1720-1913
clmet_clean	Corpus of Late Modern English Texts 3.1	281	1710-1899
ced_clean	Corpus of English Dialogues 1560-1760	167	1600-1759
lampeter_clean	Lampeter Corpus of Early Modern English Tracts	60	1643-1739
clme_prose_clean	Corpus of Late Modern English prose	7	1860-1890
pamphlets_clean	Pamphlets of the American Revolution	5	1750-1765

Schema

Every corpus has year, n_paragraphs, word_count and paragraphs. The remaining metadata columns differ by corpus:

Folder	Metadata columns
blbooks_clean	record_id, title, name, raw_date, place, country, language, page_count, source_page_count, clean_word_count
hansard_clean	date, month, day, house, title, slug
royal_society_clean	doc_id, title, author, type, journal, n_sentences
old_bailey_clean	session_id, date
clmet_clean	filename, text_id, title, author, period, quartcent, decade, genre, subgenre, gender, author_birth, source, language
ced_clean, clme_prose_clean	text_id, title, author, period_code, date_orig
lampeter_clean	text_id, title, author
pamphlets_clean	file_id, title, author, place

Licenses

This repository redistributes processed versions of existing corpora, and each keeps the license of its original source. Check the source before reuse.

• blbooks_clean: CC0 1.0 (public domain).
• clmet_clean: CC BY-NC-SA 4.0 (non-commercial, share-alike).
• All other corpora: see the source linked in the table or the corpus documentation.