Corpora

Dated text one could train, extend, or evaluate a point-in-time model on, by domain.

A point-in-time model needs text with trustworthy dates and no later edits. That rules out much of the web as crawled today and favours archives with publication dates: newspapers, filings, court opinions, parliamentary records, preprints. The table gives each source's span, the finest temporal granularity it supports, and its licence. open means public domain or an open licence; academic means access is gated but free for research; commercial means a paid licence; mixed means it varies by sub-source or by API terms.

The sequences on the models page used, respectively: FineWeb (Kelly et al.), chronologically filtered web and news (ChronoBERT, DatedGPT), WMT News Crawl and Wikipedia (Time Machine GPT), Twitter (TimeLMs), American Stories (StoriesLM), and COHA (HistBERT). TypewriterLM's 54B-token pre-1913 corpus (Luo et al. 2026) and the TiC-LM Common Crawl dump series are the newest additions. Everything else below is a candidate.

Domains

Multilingual and Cross-Lingual Corpora

SourceCoverageGranularityLicenceNotes
HPLT (High-Performance Language Technologies)75 languages, 5.6T tokensSnapshot-based (v2.0, 2025-05)open
Open source (OSI compliant)
Multilingual temporal modeling, cross-lingual evolution studies
Common Corpus500B words public-domainSnapshot-basedopen
Public domain
Multilingual pre-training, temporal analysis across languages
FineWeb2013-present (per Common Crawl dump)Crawl dump (roughly monthly)mixed
ODC-By
Chronologically filtered by crawl date for Scaling PiT LMs (Kelly et al. 2026)
Wikipedia dumps2001-presentTwice-monthly dumps; full revision historyopen
CC BY-SA
Yearly snapshots used by Time Machine GPT; consecutive snapshots used by TemporalWiki

News and Media Corpora

SourceCoverageGranularityLicenceNotes
American Stories Dataset1800-1963 (19th-mid-20th century)Yearly clusteringopen
Open source
Historical research, cultural studies, media analysis
New York Times Annotated Corpus1987-2007Daily articlescommercial
Commercial license required
News analysis, temporal event tracking
Reuters News Archive1987-presentReal-time updatescommercial
Commercial subscription required
Financial analysis, event detection
Finance CommonsPublic-domain financial subset of Common CorpusSnapshot-basedopen
Public domain
Financial modeling, open-source financial analysis
Newswire1878-1977Article date (daily)open
Public domain
2.7M unique U.S. newswire articles reconstructed from 138M local-newspaper articles; georeferenced and entity-linked (Silcock et al. 2024)
Chronicling America (Library of Congress)1756-1963Issue date (daily)open
Public domain
Raw scans and OCR behind American Stories; historical research
WMT News Crawl2007-presentYearly releases, article datesopen
Open access for research
Training data for Time Machine GPT

Social Media Corpora

SourceCoverageGranularityLicenceNotes
Twitter TimeLMs Corpus2019-Q4 → 2022-Q4Quarterly snapshotsopen
Open source
Social media analysis, language change studies
Reddit Historical Data2005-presentMonthly dumpsopen
Open source
Community analysis, trend tracking

Historical and Literary Corpora

SourceCoverageGranularityLicenceNotes
Corpus of Historical American English (COHA)1810-2009Decade-by-decade samplingacademic
Academic access required
Historical linguistics, cultural studies
Google Books Ngram Corpus1500-2019Yearly word frequenciesopen
Open access
Lexical change, cultural trends

Scientific and Technical Corpora

SourceCoverageGranularityLicenceNotes
PubMed Central Archive1781-presentPublication dateopen
Open access
Scientific progress tracking, medical history
arXiv Papers1991-presentPublication dateopen
Open access
Scientific progress analysis, knowledge evolution
SourceCoverageGranularityLicenceNotes
U.S. Supreme Court Opinions1791-presentDecision dateopen
Public domain
Legal history, precedent analysis
EU Court of Justice Decisions1952-presentDecision dateopen
Open access
EU legal history, cross-jurisdictional analysis
UN Treaty Collections1945-presentTreaty dateopen
Open access
International law research, treaty analysis
Congressional Record1873-presentDaily proceedingsopen
Public domain
Political history, policy analysis
UK Hansard (Parliamentary Proceedings)1803-presentDaily proceedingsopen
Open access
Political history, policy analysis, parliamentary studies
EU ParlSpeechEuropean Parliament proceedingsSession-basedopen
Open access
EU policy analysis, parliamentary studies
Caselaw Access Project1658-2018Decision dateopen
Public domain
All published U.S. state and federal case law digitised from Harvard Law Library

Financial and Economic Corpora

SourceCoverageGranularityLicenceNotes
SEC EDGAR Database1993-presentFiling dateopen
Public domain
Financial analysis, regulatory compliance
Federal Reserve Economic Data (FRED)1774-presentVaries by seriesopen
Open access
Economic analysis, policy research

Patent and Technical Standards Corpora

SourceCoverageGranularityLicenceNotes
USPTO Bulk Patent Data1790-presentPatent filing dateopen
Public domain
Innovation studies, technical evolution, patent analysis
EPO BigPatentEuropean patentsPatent filing dateopen
Open access
European innovation studies, technical evolution

Climate and Earth Science Corpora

SourceCoverageGranularityLicenceNotes
IPCC Reports1990-presentReport publication dateopen
Open access
Climate science evolution, environmental policy analysis
NASA Technical ReportsNASA research publicationsPublication dateopen
Public domain
Space science evolution, Earth science analysis

How fine can the clock tick?

The cadence of a checkpoint sequence is bounded by the cadence of its data. Most sources support daily dating, but a useful corpus at a given cadence needs enough tokens per step to move the model. The recommendations below, from the granularity guide in the repository, balance temporal precision against tokens per checkpoint.

Data sourceFinest practicalRecommended cadence
Twitter (academic API)hourlyquarterly
Reddit (Pushshift)dailymonthly
Wikipedia dumpsdailymonthly
SEC EDGARdailyquarterly
Financial news (Common Crawl)dailymonthly
FRED economic releasesdailymonthly
arXiv / PubMeddailymonthly
Historical newspapers (Chronicling America)dailyyearly
COHAyearlydecadal
Congressional Record / Hansarddailyyearly
Court opinionsdailyyearly

Source: PIT-CORPUSES.md and CORPUS-GRANULARITY.md. Some entries in the source list are descriptive rather than pointing at a single dataset; where no canonical link exists the name is left unlinked. Corrections welcome.