5,000 headwords · A1–C1 · 1,000 per level · ranked on 13,632,611,970 words from five corpora
Docs · Methodology · Provenance · Data dictionary · Validation · Known issues · Artefacts · العربية
Important
All 5,000 LexiCore words plus ~700 high-utility additions (≈5,700 words, A1–C1), with 5 different example sentences per word (28,000+ sentences) that rotate during review, 38+ hours of human-recorded sentence audio, headword audio, IPA, an illustration for every word, and translations in 12 languages.
Built on this list, documented with measured evidence: description · card model & sample audit · the ~700 additions
LexiCore is a core-vocabulary wordlist for English, built the way a frequency list is built — by counting — but then selected the way a teaching list should be selected: by balancing raw frequency against how widely a word is dispersed across sources, how useful it is for learners (NGSL / NAWL / BSL membership) and — in the design, though not yet in the data of this release — how modern its evidence is (KI-2).
Every word carries a CEFR level (A1–C1) taken from published, citable profiles wherever a profile exists, and from a calibrated classifier only for the remainder.
| Headwords | 5,000, unique, single-token, lower-case |
| Levels | A1 1,000 · A2 1,000 · B1 1,000 · B2 1,000 · C1 1,000 |
| Corroborated in all five corpora | 4,953 words (99.06%); in ≥4 corpora: 99.98% |
| Occurrences of these 5,000 lemmas | 9,705,062,231 (71.2% of the 13.63 B words counted) |
| Level provenance | CEFR-J v1.5 → 3,898 · Octanove C1/C2 → 665 · multinomial-logit model → 437 |
| vs Oxford 5000 (comparison only) | 3,526 shared (70.5% of LexiCore), Jaccard 0.5508, quadratic κ 0.7124 |
| Licence | CC BY 4.0 — see NOTICE.md for third-party terms |
git clone https://github.com/X-Trivle/LexiCore.git && cd LexiCore
pip install -e . # optional: only needed for the CLI / helpersfrom lexicore import load_entries, entry_for, words_by_level
a1 = words_by_level("A1") # 1,000 words, release order
e = entry_for("agreement") # -> Entry(...) or None
print(e.cefr_level, e.total_count, f"{e.per_million:.1f} ppm") # B1 1100444 80.7 ppmpython -m lexicore words --level B2 --sort alpha --limit 10 # a printable handout
python -m lexicore show --word agreement # one row, as JSON
python -m lexicore validate # the CI integrity gate
python -m lexicore stats # headline numbersFor research you may also want the release checkpoint (743 MB of per-source counters, the full
feature table, labels and the selection detail) — it is an asset of
release v1.0.0, not tree content, and
data/derived/ in this repository is a deterministic rebuild of the parts you are most likely to
need.
No dependency, no download step: the list is a plain CSV, so curl works too.
curl -LO https://raw.githubusercontent.com/X-Trivle/LexiCore/main/data/LexiCore_5000.csvdata/
├── LexiCore_5000.csv canonical 5,000-row list (schema: docs/DATA_DICTIONARY.md)
├── LexiCore_5000.json same rows + per-corpus ppm; machine-friendly
├── LexiCore_5000.txt words only, one per line, in rank order
├── oxford_comparison_sets.csv 1,474 LexiCore-only + 1,402 Oxford-only words
├── MANIFEST.json upstream release manifest: method, sources, hashes
├── SHA256SUMS.txt integrity ledger for every file below data/, docs/, …
├── lists/ per-level handouts (rank order + alphabetical)
└── derived/
├── candidate_features.csv.gz 215,416 candidates × the full feature vector
├── candidate_labels.csv.gz CEFR level, label source, confidence, class probabilities
└── surface_to_lemma.csv.gz 283,502 inflected surface forms → lemma
docs/ methodology, provenance, validation, known issues, how to rebuild
reports/upstream/ the release's own reports and stage logs, unmodified
ecosystem/lexideck-multilingual/ the Anki deck built on this list (+ ~700 extra words)
src/lexicore/ stdlib-only loader, validator, re-scoring experiment harness
tests/ · scripts/ · Makefile the checks behind the badge
Why data/derived/ matters: it contains the whole candidate pool — all 215,416 lemmas
with the exact features the release scored them on. You can re-run the selection with your own
weights without re-processing a single corpus file:
python -m lexicore reweight --freq 0.5 --disp 0.25 --util 0.25 # diff against the shipped mixflowchart LR
A[5 corpora<br/>192 files] --> B[tokenise + en-filter<br/>128 parallel shards]
B --> C["surface counts<br/>freq / df / per-shard"]
C --> D[lemma merge<br/>283,502 surface→lemma]
D --> E["feature table<br/>215,416 candidates"]
E --> F["CEFR labels<br/>CEFR-J + Octanove, model for the rest"]
F --> G["gate + score + quota fill<br/>1,000 per level"]
G --> H["5,000 headwords<br/>+ validation & Oxford comparison"]
Score = 0.34·freq + 0.18·disp + 0.18·util + 0.10·ent + 0.08·rng + 0.06·df + 0.06·rec
(z-scored over the eligible pool). Full description, stage-by-stage timings and the funnel
(215,416 → 119,155 eligible → 5,000 selected) live in docs/METHODOLOGY.md.
Each source entry in MANIFEST.json records the exact file URL, byte size and SHA-256. We
re-checked some of them against the live servers, and the numbers matched exactly — e.g.
sample/10BT really contains 15 parquet shards (the release used all 15) at
2,147,292,183 bytes for shard 000_00000.parquet, whose git-LFS oid equals the recorded
SHA-256. Details, live-check log and the licence table: docs/PROVENANCE.md.
We publish the flaws as loudly as the features, because a wordlist that hides them is useless to teachers. Summary — full detail in docs/KNOWN_ISSUES.md:
| # | Issue | Effect |
|---|---|---|
| KI-2 | The recency axis was never computed: recent_share_2022plus and newest_share_2025_2026 are empty in 5,000/5,000 rows and in 0/215,416 candidates |
The 0.06 rec weight is inert; do not describe this release as "recency-ranked" |
| KI-1 | The upstream report's "notable modern additions, ranked by 2025-26 evidence" list is in fact alphabetical | Treat that section as illustrative, not measured |
| KI-3 | pos comes from a dictionary lookup, not from tagging: at/be/a are labelled noun, 72% of the list is noun |
Don't ship POS to learners without review |
| KI-4 | The classifier that levels 437 words scores CV accuracy 0.3814 / macro-F1 0.3829 (majority class ≈ 0.32) | ~8.7% of levels are weak evidence |
| KI-5 | confidence is a constant 0.97 (CEFR-J) or 0.95 (Octanove) for profile labels |
The mean 0.9205 is a mix of caps and real posteriors; not calibrated |
| KI-6 | C1 is a catch-all: 97,963 C1 candidates vs 1,380 B2 | C1 membership ≠ rarity ordering; the level pool is heavily imbalanced |
| KI-7 | subs dispersion is measured over 198.8 M subtitle lines, not documents |
Cross-source disp/dfr are not on one scale |
| KI-8 | Reported "Files" counts mix downloaded files with compute chunks (115 = 15 + 100) | Use the per-source file table in PROVENANCE.md |
- Oxford 5000 is used externally for comparison only — it never entered the construction
of LexiCore (see
MANIFEST.json → oxford_comparison.note). Agreement on shared words is 0.466 exact / 0.8659 within one band. - NGSL / NAWL / BSL / BNC-COCA / CEFR-J / Octanove are used as anchors and validity gates. Their own licences apply to those files, which are not redistributed here — see NOTICE.md.
Product page: https://whop.com/lexideck/lexideck-multilingual/
Free 50-word sample: application.apkg (5,877,461 bytes, verified live)
Sample mirror (this repo's release asset, so the link cannot rot): LexiDeck-Multilingual-free-sample-50-words.apkg (5,877,461 bytes, SHA-256 8338bafb1118d9fb…, byte-identical to the file served by the product page)
In this repository: README of the deck ·
measured card structure ·
the ~700 additions ·
screenshots
LexiDeck Multilingual is an Anki-based English vocabulary system designed around context, repetition, audio, images, pronunciation, and varied examples rather than isolated word memorization.
It contains more than 5,700 English words from A1 to C1, with 5 different example sentences for every word, giving you more than 28,000 example sentences in total. The examples rotate during reviews, so you repeatedly encounter the same vocabulary in different contexts instead of seeing one sentence every time.
The deck also contains more than 38 hours of human-recorded sentence audio, plus audio for the headwords, IPA pronunciation, and a visual illustration for every word. The sentence collection contains roughly 1.8 million characters of written content, providing a substantial amount of contextual English practice.
- 5,700+ English words, covering A1–C1
- 28,000+ example sentences
- 5 different examples for each word
- Rotating examples during reviews
- 38+ hours of human-recorded sentence audio
- Human-recorded audio for the example sentences
- Headword pronunciation audio
- IPA pronunciation
- An illustration for every word
- Translations in 12 languages
- Anki spaced-repetition system
- Context-based learning rather than isolated vocabulary
- Examples covering different ways words are naturally used
中文 · हिन्दी · Español · العربية · Português · 한국어 · 日本語 · Bahasa Indonesia · Русский · اردو · Français · Deutsch
The amount of written example material is roughly comparable to several full-length books, while the 38+ hours of recorded sentence audio represents dozens of hours of listening practice. The exact book or video equivalent depends on the average length of the books or videos being compared, so these figures are best understood as a rough indication of the amount of material rather than a fixed conversion.
The goal is straightforward: give you enough vocabulary, repeated exposure, context, listening, pronunciation, and visual information to make Anki reviews more useful than simply memorizing a list of translations.
How it relates to the list: every deck note is tagged LexiCore and carries this release's
CEFR band, so the lineage is checkable rather than claimed. In the free sample, 44 of 50 notes
match a LexiCore headword and 6 come from the additions; the deck's order field is its own
sequence over ≈5,700 items and is not this list's rank_global
(measured in DECK_STRUCTURE.md).
Start with CONTRIBUTING.md. Everything a change must satisfy is a command, not a convention:
make install && make validate && make test && make lint
make checksums # verify the ledger after re-pointing data/
make ledger # regenerate data/SHA256SUMS.txt after an approved data changeData changes go through an issue first (see the data-request template); the ledger and CI
make silent, unreviewed edits to data/ impossible.
The list and the code in this repository are released under CC BY 4.0. Third-party corpora and anchor lists keep their own terms (NOTICE.md — not legal advice).
@misc{lexicore2026,
title = {LexiCore-5000: an open, evidence-ranked English core vocabulary list (A1--C1)},
year = {2026},
author = {{LexiCore maintainers}},
url = {https://github.com/X-Trivle/LexiCore},
note = {release 1.0.0, generated 2026-09-14; CC BY 4.0}
}