ryanjosephkamp/english-openlist

English OpenList

The largest open-source, validated English word list for NLP and games.

Dataset Description

English OpenList is a comprehensive, continuously updated dictionary of valid English words. It provides:

• ~379,000 validated English words following Scrabble-compatible rules
• Validation provenance for every word: which sources attested it, and when
• Daily updates from authoritative dictionary sources
• Version history with changelogs for every update

What the metadata is. merged_valid_dict.json records how each word was validated — the sources that attested it, the checks it passed, and the dates. It does not contain definitions, parts of speech, pronunciations, or frequency data. If you need those, pair this list with WordNet or Wiktionary.

Why Use English OpenList?

Use Case	Benefit
Spell Checking	High-precision word validation
Word Games	Scrabble/Wordle compatible list
NLP Training	Clean, validated vocabulary
Research	Transparent methodology, full provenance

Dataset Structure

Full Word Lists (data/)

These are the complete, up-to-date word lists that most users will want to download:

data/
├── merged_valid_words.txt      # FULL valid word list (~379,000 words, one per line)
├── merged_valid_dict.json      # FULL dictionary with metadata for all valid words
├── merged_invalid_words.txt    # FULL invalid/rejected entries list
└── merged_invalid_dict.json    # FULL invalid dictionary with rejection reasons

Daily Releases (releases/)

Daily updates with changelog and statistics:

releases/
└── {YYYY-MM-DD}/
    ├── update_stats.json       # Statistics for the update, including the
    │                           # `promoted_words` and `new_words` lists
    └── CHANGELOG.md            # Changelog for the update

A release folder is the record of what changed that day, not a second copy of the dataset — the complete lists are in data/, and the current snapshot is in latest/.

Latest Update Reference (latest/)

The most recent release, plus a copy of the current full lists, for convenience:

latest/
├── merged_valid_words.txt
├── merged_valid_dict.json
├── update_stats.json
└── CHANGELOG.md

Brrrdle Artifacts

Brrrdle-compatible artifacts are generated during daily automation and uploaded to:

latest/brrrdle/
data/brrrdle/

The primary Brrrdle files are words_length_{N}.json for every supported length from 2 through 35. Each file contains metadata.curation, curated answers, and complete validGuesses. The validGuesses array remains the full per-length list, while answers is generated with the deterministic stratified_quality_score_v1 method using seed 42 + length. Both arrays contain plain word strings.

During the transition to length-specific artifacts, the legacy length-5 compatibility files brrrdle_words.txt and brrrdle_words.json are still published. These legacy files should be removed in the next major Brrrdle artifact update, along with any legacy-only manifest or generated README behavior.

Data Fields

Entries are not uniformly shaped. The list was assembled from several intakes, and each kept the fields its own pipeline produced. Read defensively: check for a field before using it.

Tournament word list intake — the largest group:

{
  "word": "broth",
  "source": "twl_scrabble_dictionary",
  "validation_status": "valid",
  "added_date": "2026-01-10",
  "length": 5
}

Verification pipeline intake — carries candidate_source, whose entries are suffixed _valid or _unlikely. Count only the _valid ones; a word can carry eight sources that all say unlikely:

{
  "word": "a",
  "unverified_llm_verdict": "valid",
  "validation_source": "verification_pipeline",
  "candidate_source": ["google_ngrams_valid", "wordnet_valid", "nltk_valid"],
  "advanced_validation": { "passed": true, "confidence": 1.0 },
  "statistical_validation": { "passed": true, "perplexity": 1.0 },
  "proper_noun_check": { "checked": true, "is_proper_noun": false },
  "added_date": "2025-12-17"
}

unverified_llm_verdict was called status until 14 August 2026. It was renamed because the old name made it read like this dataset's own ruling, and it is not one. It records what a single LLM pass (Google Gemini 3 Flash Preview, December 2025) thought of the word, on 137,705 entries. Do not filter on it expecting a validation result. Its accuracy was measured — see Known limitations below — and it was wrong more often than right wherever a real dictionary could check.

A separate dictionary_verdict field appears on 201 entries. That one is a dictionary ruling, from Merriam-Webster or Free Dictionary via the promotion path, and its source field names which.

Synthetic intake — algorithmically constructed candidates, identifiable by source: "synthetic_generation":

{
  "source": "synthetic_generation",
  "category": "Medical",
  "valid": 1,
  "validated": true,
  "notes": "Synthetic candidate awaiting validation",
  "created_date": "2026-01-11T12:19:00"
}

Note the synthetic records have no word field — the word is the object key.

Validation Rules (Scrabble-Compatible)

These are the rules applied to newly discovered words by the daily pipeline:

• ✅ Contain only lowercase letters (a-z)
• ✅ Are recognized by Merriam-Webster Collegiate Dictionary
• ❌ Are NOT proper nouns (unless commonly used as verbs)
• ❌ Are NOT abbreviations or acronyms

Words already in the list arrived through earlier intakes and were not all checked against Merriam-Webster — see Composition below. Lengths run from 1 to 47 characters (a at one end, phosphoribosylaminoimidazolesuccinocarboxamides at the other).

Dataset Statistics

Metric	Value
Total Valid Words	~379,000 (grows daily)
Total Invalid Entries	9,275,000+
Update Frequency	Daily (00:00 UTC)
Primary Source for New Words	Merriam-Webster Collegiate Dictionary

Counts here are approximate by design — the list grows every day. For the exact current figure see latest/update_stats.json.

Composition

The list was assembled from several intakes over time, and they are not interchangeable. Anyone filtering or scoring this data should know the mix:

Intake	Share	How to identify
Tournament word list	~46%	source: "twl_scrabble_dictionary"
Verification pipeline	~35%	validation_source: "verification_pipeline"
Synthetic candidates	~17%	source: "synthetic_generation"
Other / unattested	~2%	none of the above

The synthetic group contains algorithmically constructed forms such as abacteremicer and nonlivabler. They are kept deliberately — this list aims to be broad, and removing them would narrow it — but they are the group most likely to surprise you, and applications wanting only conventional vocabulary should filter on source.

Usage

Python (Hugging Face Datasets)

The default configuration is the valid word list, one word per row in a text column:

from datasets import load_dataset

dataset = load_dataset("ryanjosephkamp/english-openlist", split="train")

for entry in dataset:
    print(entry["text"])

Every other file in the repository — the metadata dictionary, the daily releases, the Brrrdle artifacts — remains browsable and downloadable; they are simply not part of the default load. Point at one explicitly to read it:

dataset = load_dataset(
    "ryanjosephkamp/english-openlist",
    data_files="releases/2026-08-11/merged_valid_words.txt",
    split="train",
)

Or fetch the file directly, which is usually what you want for a word list:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "ryanjosephkamp/english-openlist",
    "data/merged_valid_words.txt",
    repo_type="dataset",
)
words = set(open(path).read().split())
print("hello" in words)   # True

Direct Download

Download the complete word lists:

# Download the FULL valid word list
wget https://huggingface.co/datasets/ryanjosephkamp/english-openlist/resolve/main/data/merged_valid_words.txt

# Download FULL valid dictionary with metadata
wget https://huggingface.co/datasets/ryanjosephkamp/english-openlist/resolve/main/data/merged_valid_dict.json

# Download FULL invalid words list (for reference)
wget https://huggingface.co/datasets/ryanjosephkamp/english-openlist/resolve/main/data/merged_invalid_words.txt

# Download FULL invalid dictionary
wget https://huggingface.co/datasets/ryanjosephkamp/english-openlist/resolve/main/data/merged_invalid_dict.json

Download daily release files:

# Download a specific day's update
wget https://huggingface.co/datasets/ryanjosephkamp/english-openlist/resolve/main/releases/2026-01-19/CHANGELOG.md

Python (Raw Files)

import json

# Load word list
with open("merged_valid_words.txt", "r") as f:
    words = set(line.strip() for line in f)

# Check if a word is valid
print("hello" in words)  # True
print("asdf" in words)   # False

# Load the dictionary for validation provenance.
# Note this file is ~290 MB; stream it if memory is tight.
with open("merged_valid_dict.json", "r") as f:
    dictionary = json.load(f)

entry = dictionary["broth"]
print(entry["source"])             # twl_scrabble_dictionary
print(entry.get("candidate_source"))  # None on this intake -- check before use

Methodology

Phase 1: Corpus Acquisition (December 2025)

Aggregated 9.8 million candidate words from 15+ open sources:

• Wiktionary (6.5M words)
• WordNet 3.1 (150K words)
• SCOWL 2020 (500K words)
• Google Books Ngrams (1M+ words)
• Collins Complete Dictionary (800K words)

Phase 2: Validation Pipeline (December 2025 - January 2026)

Multi-stage AI validation using Gemini 2.0/2.5 Flash:

• Pattern-based screening
• LLM classification with iterative convergence
• Statistical sampling for quality assurance
• Synthetic word generation and validation

Phase 3: Continuous Updates (January 2026 - Ongoing)

Daily automated pipeline:

1. Discover new words from Merriam-Webster RSS feed and manual additions
2. Validate ~1,000 words from invalid list against dictionary APIs
3. Promote validated words to the valid list
4. Update full word lists and dictionaries on Hugging Face
5. Generate changelog and statistics

Phase 4: Correcting the list itself (August 2026 - Ongoing)

Until August 2026 this dataset was only ever added to. Words are now also corrected, with every decision recorded in corrections/ in the source repository before any list is rewritten.

13 August 2026. 150 words were in merged_valid_words.txt and in merged_invalid_words.txt at the same time — words the daily pipeline had promoted onto one side without removing them from the other. Every one already carried a dictionary API ruling in its own record (Merriam-Webster for 63, Free Dictionary for 86, MW Medical for 1), so all 150 were cleared as valid and removed from the invalid list. None was demoted, and the valid word list did not change. 201 entries whose status field (now dictionary_verdict) read "WordStatus.VALID" — a stringified Python enum — were normalised to "valid" in the same change.

14 August 2026 — the status field was renamed, and measured.

status read like this dataset's verdict on a word. It was not: on 137,705 entries it recorded what one LLM pass (Google Gemini 3 Flash Preview, December 2025) thought. It is now unverified_llm_verdict, so nobody has to read the documentation to find that out.

Before deciding what to do about the 20,052 words it calls invalid, its accuracy was measured. 400 were sampled, stratified by how many corpora attested them, and looked up in Merriam-Webster Collegiate and Medical:

Sampled	400
Absent from Merriam-Webster entirely	383 (95.8%)
Could be checked at all	18
Of those, the LLM was wrong	13
Of those, the LLM was right	5 (all proper nouns)

Words it wrongly rejected include clorazepate, antinociceptive, hemoconcentrations, esophagogastroplasty, rotifera and palliasse.

No word was moved, and none can be. Merriam-Webster answered every word we asked and had no entry for 95.8% of them — this vocabulary is chemical, medical and taxonomic, outside what these dictionaries cover. Adjudicating all 20,052 would return "no entry" for roughly 19,200 of them at any budget. So the verdict is kept with its provenance and clearly labelled, rather than acted on or quietly deleted.

Two known issues therefore remain, deliberately not acted on:

• 20,052 entries carry unverified_llm_verdict: "invalid" while listed as valid. Treat it as one machine's opinion from December 2025, not a validation result. Where it could be checked it was wrong more often than right.
• 64,837 entries from the synthetic intake carry no attestation at all — no corpus source and no validation record — while marked validated: true. If you need only words a human source attested, filter on source != "synthetic_generation".

Citation

@dataset{english_openlist_2026,
  title = {English OpenList: A Comprehensive Validated English Word List},
  author = {English OpenList Project Team},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/english-openlist/english-openlist}
}

License

This dataset is released under the MIT License.

The underlying word data is derived from open sources with compatible licenses.

Contact

• Issues: GitHub Issues
• Updates: Check the releases/ folder for version history

---

Last Updated: January 2026