croqaz/vintage-words

Vintage words

Word-frequency dictionaries built from ~41.4 billion words of English text written before 1900. For every word we counted how many times it appears across a large corpus of historical books, newspapers and pamphlets, and saved the result as simple {word: count} JSON files.

In short: a census of the pre-1900 English vocabulary.

What's in here

Seven per-source frequency dictionaries plus one combined dictionary:

File	Source	Entries	Unique words	Tokens
EEBO	Early English Books Online	34,935	2,415,448	698,964,004
ECCO	Eighteenth Century Collections Online	3,100	548,389	105,508,425
EVANS	Evans Early American Imprints	5,011	329,806	102,063,934
CLMET3.1	Corpus of Late Modern English Texts	278	169,743	30,772,645
gutenberg	Project Gutenberg (pre-1900 books)	240,742	1,961,826	1,107,923,635
long-docs	Long documents (books, BL/LoC/OTA)	198,821	93,305,526	9,093,692,612
short-docs	Short documents (newspapers, notices)	13,077,552	76,154,718	5,025,802,850
english-historical-1800-1875	English historical corpus, 1800–1875	6,753,652	146,007,810	25,277,286,396
ALL	All of the above, merged	20,314,091	320,893,266	41,442,014,501

summary.json lists the token totals, unique-word counts and top-50 words for each source.

How

• Sources. EEBO, ECCO, EVANS and CLMET3.1 are curated historical book collections are from vintage-v2 dataset; Gutenberg, long-docs and short-docs come from vintage-v1 dataset (newspapers, public-domain books and archives).
• Date filtering. Everything is confirmed pre-1900. The only collection containing later material (CLMET3.1, which runs up to ~2000) was filtered by its metadata date field, dropping 55 post-1900 books.
• Tokenisation. Text was lowercased and split into words matching [a-z]+(?:'[a-z]+)* - ASCII letters with internal apostrophes (don't, o'clock). Digits, hyphens, punctuation and non-Latin characters were excluded.
• Counting. Simple occurrence counts per source, summed for the combined file.

OCR and size (words v1)

Much of this material is OCR'd from scans centuries old, so the long tail is full of scanning errors and broken words. Inside the words-v1/ folder, the three largest dictionaries (words-ALL, words-long-docs, words-short-docs) are capped to their 10,000,000 most frequent words to keep the size small; beyond that rank the entries are overwhelmingly OCR noise. The unique_words figures in the table above are the true pre-cap totals.

OCR-noise-filtered dictionaries (words v2)

The words-v2/ folder contains the same dictionaries after a length-aware OCR-noise filter. The v1 long tail is mostly scanning garbage, so those sets were capped at the 10 mil most frequent words. The v2 filter instead trims by word shape, keeping a confident vocabulary at every length:

• drop tokens longer than 34 characters (longer than the longest English word);
• drop tokens with 3+ apostrophes (never valid English);
• keep the rest only if count >= min_count(len) — short words must be frequent (40 occurrences for ≤7-char words), while long words are trusted at low counts (just 5 occurrences for 14+ char words).

File	Words kept	Tokens kept
EEBO	198,280	692,221,365
ECCO	55,380	103,860,030
EVANS	41,326	101,091,974
CLMET3.1	28,401	30,205,008
gutenberg	203,261	1,101,669,797
long-docs	1,347,748	8,932,515,485
short-docs	1,120,123	4,879,798,738
english-historical-1800-1875	3,437,077	25,305,337,457
ALL	Merged from all of the above	3,966,551

Word-pair dictionaries (bigrams)

The bigrams/ folder holds word-pair (bigram) frequency dictionaries in the same JSON format, {"w1_w2": count} (e.g. "of_the": 965416). Bigrams were built with the same English pre-1900 tokeniser (lowercase [a-z]+(?:'[a-z]+)*) and cleaned with the same length-aware OCR-noise filter described above, so each entry is a pair of confident words (in_the, to_the, to_be, …). The 8 per-corpus sets are per-source, exactly mirroring the word dictionaries; most use the soft thresholds, while the large english-historical-1800-1875 set uses stricter thresholds to keep it a manageable size.

File	Bigrams kept	Tokens kept
CLMET3.1	286,460	22,829,774
ECCO	845,508	85,096,706
EVANS	783,940	86,786,528
EEBO	3,369,806	618,589,970
gutenberg	3,656,710	976,317,705
short-docs	11,642,671	4,155,113,825
long-docs	13,459,880	7,631,431,509
english-historical-1800-1875	41,842,311	23,435,881,919

Why bother

A frequency-ranked snapshot of what English actually looked like before 1900 is useful for:

• Anachronism detection -- deciding whether a word plausibly existed in historical text (a modern term like internet or chemotherapy simply never appears here).
• Building or evaluating historical language models and period-accurate text generation.
• Filtering/curating pre-1900 corpora and spotting modern contamination.
• Lexicography and linguistics -- spelling variation, vocabulary shift, and word-frequency studies over time.

Format

Each file is a JSON object mapping a word to its integer count, sorted by descending frequency:

{
  "the": 1086964914,
  "of": 635302823,
  "and": 496624990
}

Citation

@misc{vintage-words,
  title  = {Vintage-words},
  author = {Cristi Constantin},
  year   = {2026},
  url = {https://huggingface.co/datasets/croqaz/vintage-words}
}