beyarkay/pre-1950s-text

A dataset of pre-1950 English text

This is a high-quality thoroughly-curated 100+ GB dataset of English-only text written before 1950-01-01. It was collected for the purpose of training LLMs, initially a small 125M model (Archibald-125M) and later a 3B or 7B model (depending on funding).

Why train an LLM on old text?

One unanswered question about LLMs is "can they invent?". Given how much they know about the world, it's somewhat surprising that LLMs seem to have difficulty with making innovative connections about the world (although they get better every day).

Because Archibald-125M has a knowledge cutoff of 1950, we are able to quiz and prompt until Archie figures out something that's a novel invention for 1950 but well-established science in the current day. We can provide more detailed and more helpful hints to Archie until the answer is blindingly obvious, and then examine what is required for an LLM to make a discovery. Importantly, this process can be automated, using modern LLMs to prompt and evaluate the outputs of Archie.

Beyond discovering the nature of invention, text before 1950 has several qualities which make it interesting:

• Significantly different data distribution. Almost all modern LLMs are trained on a corpus of data created mostly after the internet, and a majority of that is after the internet boom in the 2000s. Archibald is possibly the only LLM with a significantly different but still human-generated training distribution.
• Archaic moral views: In the 1950s most women could not open bank accounts, sign leases, or own property in their name without a male guardian. Being homosexual was illegal, racial segregation was commonplace. Probing Archie about the morality of these points might provide insight into how we can probe modern LLMs about our modern moral failures.

The primary goal is to figure out the nature of invention, but it's likely there'll be many interesting side-quests along the way.

Some events after 1950 which Archibald-125M doesn't know about

• 1950: First credit card (Diners Club). Before this: People paid in cash or by cheque. Store credit was local, and debt tracking was manual.
• 1952: Polio vaccine (Salk). Before this: Tens of thousands per year were paralyzed or killed by polio. Children avoided public swimming pools during outbreaks.
• 1953: Death of Stalin. Before this: Stalin's dictatorship controlled the Soviet Union with mass purges and forced labor camps. Eastern Europe remained locked behind the Iron Curtain.
• 1953: Discovery of DNA structure (Watson & Crick). Before this: Heredity was understood abstractly. No genetic engineering, paternity testing, or DNA forensics.
• 1954: Brown v. Board of Education (USA). Before this: Racial segregation was legal in schools. Black children attended underfunded “separate but equal” schools.
• 1955: Rosa Parks arrested / Bus boycott. Before this: Black passengers were legally forced to give up seats for white riders in much of the U.S. South.
• 1957: Launch of Sputnik. Before this: No artificial satellites. Global communications were limited to undersea cables and radio. Weather forecasts were rudimentary.
• 1958: NASA founded. Before this: No civilian space program. Military handled missile research; spaceflight was science fiction.
• 1959: First commercial photocopier (Xerox 914). Before this: Copies were made with carbon paper, mimeographs, or by hand. Reproducing documents was slow and messy.
• 1960: First laser. Before this: No barcode scanning, laser surgery, or optical fiber communication.
• 1961: Yuri Gagarin orbits Earth. Before this: No human had been to space. Space exploration was theoretical; Earth was the only world we’d seen directly.
• 1963: Assassination of JFK. Before this: U.S. politics was in a post-war optimism phase. After: Deepened Cold War tensions and conspiracy culture.
• 1964: U.S. Civil Rights Act. Before this: Legal segregation and open discrimination in housing, employment, and voting. Jim Crow laws were enforced in the South.
• 1965: Moore’s Law proposed. Before this: Computing power was scarce. Computers filled rooms, used punch cards, and served governments or large corporations.
• 1967: First heart transplant (South Africa) Before this: End-stage heart failure meant death. No organ transplants; no immunosuppressive treatment.
• 1969: Apollo 11 Moon Landing. Before this: The Moon was unreachable. Space travel was a Cold War dream and sci-fi trope.
• 1971: Intel 4004 (first commercial microprocessor). Before this: Computers were assembled from separate logic circuits. No personal computing. Embedded electronics were rare.
• 1972: Watergate scandal begins. Before this: Presidential power was largely unchecked in public perception. The scandal triggered a wave of investigative journalism and public distrust.
• 1973: First mobile call (Motorola prototype). Before this: Phones were tethered to landlines. Calling meant finding a telephone booth or home line.
• 1973: Oil Crisis / OPEC embargo. Before this: Western nations assumed oil supply was cheap and endless. After: Gasoline rationing, speed limits, and birth of modern energy policy.
• 1975: Personal computers begin (Altair 8800). Before this: Only corporations or universities used computers. Home computing was unimaginable.
• 1981: IBM PC released. Before this: Hobbyist computers were inconsistent. This standardized architecture for business and home computing.
• 1983: First mobile phone sold commercially (Motorola DynaTAC). Before this: Communication on the move meant CB radio or pagers. Businesspeople were tied to their desks.
• 1984: DNA fingerprinting invented. Before this: Criminal evidence relied on fingerprints, blood types, and eyewitnesses. Paternity was legally disputed without hard evidence.
• 1989: Fall of Berlin Wall. Before this: Germany was split, and Eastern Europe was under Soviet domination. Movement across the Iron Curtain was deadly.
• 1990: World Wide Web invented (Tim Berners-Lee). Before this: The internet existed for scientists and the military, but was text-only, obscure, and difficult to use.
• 1995: GPS becomes publicly available. Before this: Navigation relied on paper maps, compasses, and asking for directions.
• 1996: Dolly the sheep cloned. Before this: Cloning of mammals was thought impossible. Genetics was still largely experimental.
• 1998: Google founded. Before this: Internet search was poor. Engines like AltaVista and Yahoo listed results manually or poorly ranked.
• 1999: Introduction of Bluetooth. Before this: No short-range wireless communication. Devices had to connect physically or over infrared.
• 2001: 9/11 attacks on U.S.. Before this: Air travel was relatively relaxed. Global terrorism was not the focus of national security.
• 2003: Human Genome Project completed. Before this: Human genetics was understood in fragments. Precision medicine was impossible.
• 2004: Facebook launched. Before this: Social life online was fragmented (forums, IRC, email lists). No centralized digital social identity.
• 2007: iPhone released. Before this: Phones were mainly for calling/texting. No universal internet access in your pocket.
• 2008: Global Financial Crisis. Before this: Housing was considered a safe investment. After: Global austerity and mass unemployment.
• 2012: CRISPR used for gene editing. Before this: Gene editing was imprecise, slow, and expensive.
• 2016: Brexit referendum. Before this: EU membership seemed permanent. Britain’s vote marked a turn in global politics toward nationalism.
• 2016: AlphaGo shows that deep learning surpasses human performance in Go. Before this: AI was limited to narrow tasks. After: Widespread fear and hype around general intelligence.

Data sources to investigate

• wikipedia looks like it's got a big list of newspaper archives: https://en.wikipedia.org/wiki/Wikipedia:List_of_online_newspaper_archives
• also see https://github.com/haykgrigo3/TimeCapsuleLLM

Data Sources in use

Project Gutenberg

Download (~8GB), excluding most file types except for .txt (while keeping unknown file types, in case they're useful):

rsync -av --del \
    --include='*/' \
    --include='*.txt' \
    --include='*.TXT' \
    --include='*.text' \
    --exclude='*' \
    --info=progress2 \
    ftp.ibiblio.org::gutenberg \
    data/gutenberg

List all unique file extensions:

find data/gutenberg/ -type f | sed -n 's/.*\.\([^.\/]\+\)$/\1/p' | sort -u

Afterwards (or during) the download, there'll be a lot of non-text files. Remove them using:

find data/gutenberg/ -type f \( -iname '*.m4a' -o -iname '*.m4b' -o -iname '*.gif' -o -iname '*.jpg' -o -iname '*.jpeg' -o -iname '*.html' -o -iname '*.htm' -o -iname '*.png' -o -iname '*.mp3' -o -iname '*.rst' -o -iname '*.rtf' -o -iname '*.doc' -o -iname '*.lit' -o -iname '*.xml' -o -iname '*.iso.*' -o -iname '*.prc' \) -delete

List all txt files and their sizes (in human readable numbers)

find data/gutenberg -type f -iname '*.txt' | xargs du -h -c | sort -h

Get a list of all non-English text files (based on Gutenberg's own Language: $FOOBAR label):

rg -uu '^Language:' data/gutenberg | rg -v 'Language: English' | sed 's/:.*$//'

And to delete them, pipe to xargs rm -v

rg -uu '^Language:' data/gutenberg | rg -v 'Language: English' | sed 's/:.*$//' | sort -u | xargs -r rm -v

Need to remove project Gutenberg header and footer:

*** START OF THE PROJECT GUTENBERG EBOOK 10486 ***

Provided by McGuinn's Folk Den (http://www.ibiblio.org/jimmy/folkden)

[...]

*** END OF THE PROJECT GUTENBERG EBOOK 10486 ***

And also some transcriber's notes:

[Transcriber's Note: Printers' errors have been marked with the notation
** . There are a few special characters in the section on Erasmus Darwin;
macrons (a straight line over a letter) are denoted [=x] and breves
(the bottom half of a circle over a letter) are denoted [)x].]

And also anything published after 1950

And also anything not in English

Hmm. A bit problematic, Project Gutenberg explicitly does not include the original publication date of the items in their catalogue link:

Project Gutenberg metadata does not include the original print source publication date(s). Because Project Gutenberg eBooks are substantially different from the source book(s), we track the Project Gutenberg publication date (“release date”), but do not include print source information in the metadata.

So we'll need to date all the items manually. Hrmm

Chronicling America

Information: https://chroniclingamerica.loc.gov/ocr/

JSON listing of files: https://chroniclingamerica.loc.gov/ocr.json

Download the full dataset, one archive at a time (total size is 2 115 GB):

uv run src/download_chronicling_america.py

Conveniently, they're all organised by date, so we can find all directories indicating content after 1950 and delete them. Use -depth to not traverse too deep into all the subdirectories:

(this is after the shuffle around to reduce inode usage)

# preview
$ find . -regextype posix-extended -mindepth 4 -maxdepth 4 -type f \
  -regex '.*/(1950|19[5-9][0-9]|20[0-9]{2})-(0[1-9]|1[0-2])-(0[1-9]|[12][0-9]|3[01])/.+' -print

# DELETE
$ find . -regextype posix-extended -mindepth 4 -maxdepth 4 -type f \
  -regex '.*/(1950|19[5-9][0-9]|20[0-9]{2})-(0[1-9]|1[0-2])-(0[1-9]|[12][0-9]|3[01])/.+' -print -delete

We'll also want to delete all the XML files:

find -type f -iname '*.xml' -delete

(this will clear a few hundred GBs, current versions of the download python script will auto-delete after extraction)

TODO the resulting files are pretty bad. The OCR has many many artefacts, and not all of them are obvious how to fix, since the source scans/images aren't available apparently. Not sure how to fix these without using modern LLMs and potentially infecting the dataset.

At some point, I ran out of inodes. Removing XML files helped, but didn't solve the issue. This shell command was able to remove leaf directories:

find data -mindepth 1 -depth -type d -empty -print -delete

This command lists inode usage:

$ df -i /
Filesystem                        Inodes IUsed IFree IUse% Mounted on
/dev/mapper/ubuntu--vg-ubuntu--lv    60M   60M  313K  100% /

See where most of the inodes are going

$ du --inodes -d1 data/gutenberg | sort -nr | head
154K    total
154K    data/gutenberg
44K     data/gutenberg/1
43K     data/gutenberg/4
42      data/gutenberg/0
35K     data/gutenberg/3
33K     data/gutenberg/2

Go through az_bentonite_ver02, and flatten the directory structure from az_bentonite_ver02/thing/year/month/day/edition/sequence/orc.txt down to az_bentonite_ver02/thing/year-month-day/edition/sequence/orc.txt

find az_bentonite_ver02 -regextype posix-extended -type d -depth \
  -regex '.*/[0-9]{4}/[0-9]{2}/[0-9]{2}$' \
| while IFS= read -r d; do
    y=$(basename "$(dirname "$(dirname "$d")")")
    m=$(basename "$(dirname "$d")")
    dd=$(basename "$d")
    root=$(dirname "$(dirname "$(dirname "$d")")")
    tgt="$root/$y-$m-$dd"
    [ -e "$tgt" ] && { echo "skip (exists): $tgt"; continue; }
    mv "$d" "$tgt"
    rmdir -p "$(dirname "$d")" 2>/dev/null || true   # remove empty month/year
  done

Biodiversity Heritage Library

60+ million pages of OCR content (~41 GB compressed, 138GB uncompressed)

Download:

BHL_URL="https://smithsonian.figshare.com/ndownloader/files/52893371"
mkdir -p data/bhl && curl -L "$BHL_URL" | tar -xj -C data/bhl

NOTE: the download was weird for me, and I struggled. Eventually I got it to work, but there's a bunch of redirects and really short lifetimes on the links. Good luck.

From

https://smithsonian.figshare.com/articles/dataset/BHL_Optical_Character_Recognition_OCR_-_Full_Text_Export_new_/21422193?file=52893371

Remove non-text files:

find data/bhl/ -type f \( -iname '*.jpg' -o -iname '*.jpeg' -o -iname '*.html' -o -iname '*.htm' -o -iname '*.png' -o -iname '*.mp3' -o -iname '*.rst' -o -iname '*.rtf' -o -iname '*.doc' -o -iname '*.lit' -o -iname '*.xml'  -o -iname '*.prc' \) -delete

Archive.org

The script src/download_archive_dot_org.py will download an index of all the archive files matching the above query. These indices take up about 259MB, and will be stored to data/archive-dot-org/indices/, containing the date, the id, and the size of the item in bytes. The ID can then be used to download the actual files. To download all the text files associated with the IDs listed in a file ./itemlist.txt, you can use this command:

wget \
    --recursive \
    --span-hosts \
    --no-clobber \
    --no-parent \
    --no-host-directories \
    --cut-dirs=1 \
    --accept=txt \
    --execute robots=off \
    --level=1 \
    --input-file=./itemlist.txt \
    --base='http://archive.org/download/'

• Dataset query (1800-1950): https://archive.org/search?query=date%3A%5B1800-01-01%20TO%201949-12-31%5D
• Advanced search: https://archive.org/advancedsearch.php
• Query: mediatype:(texts) AND language:(English) AND date:[1800-01-01 TO 1949-12-31]

Better query:

• mediatype:(texts) AND (language:eng OR language:"English") AND date:[1800-01-01 TO 1949-12-31]
• fields of interest:
  • creator
  • date
  • downloads
  • identifier
  • item_size
  • subject
  • title

US Post Office

(requires an API key)

https://data.uspto.gov/apis/getting-started

Hathi Trust

HathiTrust was founded in 2008 as a not-for-profit collaborative of academic and research libraries now preserving 18+ million digitized items in the HathiTrust Digital Library. We offer reading access to the fullest extent allowable by U.S. and international copyright law, text and data mining tools for the entire corpus, and other emerging services based on the combined collection.

Looks like it has a lot of information, although this might all just be duplicated from the data available in the Internet Archive. Also it's less easy to download, Google Books has some pretty restrictive licensing

https://babel.hathitrust.org/cgi/pt?id=mdp.39015082239875&seq=26&format=plaintext

Advanced search URL

Cleaning up

Will need to remove all references to google, the internet, any URLs, OCR, etc

List all file containing at least 1% lines of non-English characters (there's a lot of Hebrew newspapers)

PAT='[\p{Hebrew}\p{Cyrillic}\p{Greek}\p{Arabic}\p{Hangul}\p{Hiragana}\p{Katakana}]'
join -t: -j1 \
  <(rg -u -g '*.txt' -P -c "$PAT" | sort -t: -k1,1) \
  <(rg -u -g '*.txt'    -c '^'    | sort -t: -k1,1) \
    | awk -F: '{ if ($2 > 0.01*$3) print $1 }' > non-english.txt

These files you'll probably want to manually check, because OCR does funky stuff imagining non-English characters, and old English texts often had Latin/Greek/Hebrew/French/etc.

Also remove files with >50% lines containing German diatribes:

PAT='[äöüÄÖÜß]'
join -t: -j1 \
  <(rg -u -g '*.txt' -P -c "$PAT"   data/ | sort -t: -k1,1) \
  <(rg -u -g '*.txt'    -c '^'      data/ | sort -t: -k1,1) \
| awk -F: '{ if ($2 > 0.1*$3) print $1 }' > german.txt

And again with Danish/Swedish

PAT='[æøåØ]'
join -t: -j1 \
  <(rg -u -g '*.txt' -P -c "$PAT"   data/ | sort -t: -k1,1) \
  <(rg -u -g '*.txt'    -c '^'      data/ | sort -t: -k1,1) \
| awk -F: '{ if ($2 > 0.5*$3) print $1 }' > neurope.txt

Mega regex to find all dates after 1950:

rg -n -i -P '\b((jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s+\d{1,2}(st|nd|rd|th)?\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|(jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|\d{1,2}(st|nd|rd|th)?\s+(jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))[-/](0?[1-9]|1[0-2])[-/](0?[1-9]|[12]\d|3[01])|(0?[1-9]|[12]\d|3[01])[-/](0?[1-9]|1[0-2])[-/](1950|19[5-9]\d|20(0\d|1\d|2[0-5])))\b' data/

and to delete the files:

rg --null -n -i -P '\b((jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s+\d{1,2}(st|nd|rd|th)?\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|(jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|\d{1,2}(st|nd|rd|th)?\s+(jan(uary)?|feb(ruary)?|mar(ch)?|apr(il)?|may|jun(e)?|jul(y)?|aug(ust)?|sep(t(ember)?)?|oct(ober)?|nov(ember)?|dec(ember)?)\s*,?\s*(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))|(1950|19[5-9]\d|20(0\d|1\d|2[0-5]))[-/](0?[1-9]|1[0-2])[-/](0?[1-9]|[12]\d|3[01])|(0?[1-9]|[12]\d|3[01])[-/](0?[1-9]|1[0-2])[-/](1950|19[5-9]\d|20(0\d|1\d|2[0-5])))\b' data/ -l | xargs -0 rm -v

Removing empty directories:

find data -mindepth 1 -depth -type d -empty -print -delete

Find some non-english language suspect characters (lots of false-positives due to OCR though):

rg -m 10 -P '[àáâãäåçèéêëìíîïñòóôõöùúûüÿÀÁÂÃÄÅÇÈÉÊËÌÍÎÏÑÒÓÔÕÖÙÚÛÜŸßæøåÆØÅčďěňřšťžČĎĚŇŘŠŤŽąćęłńśźżĄĆĘŁŃŚŹŻășțâîĂȘȚÂÎğşİıĞŞ]' data/

Will also want to remove the euro symbol, since that didn't exist before 1950.

Maybe just run a spell-checker over the whole thing and give every file a quality score? or just looking at words based on some dictionary of known English words?

Estimated training costs

NanoGPT:

• 10B FineWeb tokens -> GPT2 ~accuracy
• 40B FineWeb tokens -> GPT3 ~accuracy

TinyLlama 1.1B:

• 3T tokens (data set size: 950B)
• just 90 days using 16 A100-40G GPUs
• Note that they're training for longer than Meta's Llama trained for

SmolLM3:

• Hugging Face
• 3B
• 11T tokens pretraining
• 220k GPU hours: 48 nodes, 8xH100 GPUs, 24 days
• https://huggingface.co/blog/smollm3

Lambda AI pricing:

• $2.69/H100/hour

Better cleaning

Okay it looks like I'll have to do some proper cleaning myself.

Ideas:

• Use regex to un-hard-wrap the words
• use basic Regex to find non-English languages and characters
• use a dictionary to help correct words

Cleaning test cases

Some test cases for cleaning data:

==> ./ct_jackson_ver01/sn92051126/1911/10/05/ed-1/seq-3/ocr.txt <==
ipv\ ■ ....
:5 Until a short time
ago, scarcely one
person in a thousand
had ever tasted a
§;■■ really good soda
cracker—as it came
'• fresh and crisp from
the oven.
Now every man,

==> ./ct_jackson_ver01/sn92051126/1911/10/05/ed-1/seq-4/ocr.txt <==
New HavenUnion
i; >T/ ' - .. '
Iqdnusu bv aluxandeb troop.
• '4*-----—
^.THURSDAY, OCTOBER 5, 1911.
r V Notice to Advertisers.
vChunge of advertisements must be
Ityby 10 o’clock In tbe morning, to ln
iim0 the change being nruuie tbe same
TTt cannot be guaranteed that