mjbommar/opengloss-v2.3-retrieval-triples

OpenGloss v2.3 — Retrieval Triples

Training triples for an embedding model or reranker. Each row is a query, a positive passage from the sense the query was written for, and one negative. The hard negative is drawn from the graph by a priority-ordered fallback — another sense of the same headword, a confusable_with target, a co-hyponym, or a synonym-of-a-synonym — which is what makes it hard and, unusually, auditable: the contrast paragraph in opengloss-v2.3-contrasts often states in words why the negative is a negative. A direct synonym is never offered as a negative.

Part of the OpenGloss v2.3 release family — 16 datasets built from one store of 160,724 lexemes and 300,787 live senses, all joinable on derived ids. See Related datasets for the rest.

What's new in v2.3 vs v1.3

1. Schema v3. Every lexeme carries a kind discriminator (simplex, compound, phrasal verb, idiom, proper noun, abbreviation, affix, function word); every sense carries a controlled domain leaf from a fixed ~160-leaf taxonomy instead of free text; every example carries the character span of the headword occurrence inside it.
2. Renditions, not one string. A definition is a set: the canonical one plus rewrites at four reading levels and in four registers, each produced in a single call from the canonical text so they say the same thing at different altitudes.
3. A sense graph, not a word graph. Typed relations resolve to sense ids wherever the target's entry exists in the release, so bank --hypernym--> financial institution points at a meaning rather than at a string.
4. Retrieval data is first-class. Synthetic per-sense queries in eight styles, grounded QA pairs, mined word-in-context pairs, MS MARCO-style triples with graph-derived hard negatives, and graded TREC qrels — all derivable from, and consistent with, the same entries.
5. Derivable identifiers everywhere. v1.3 published a positional id for lexemes and senses (3d_model_noun_0) and nothing below that. v2.3 gives every rendition, edge, query, QA pair and provenance record an id computable from the row alone, and never renumbers: a retired sense is tombstoned, so the ids after it keep their meaning.
6. Per-field provenance. Which model wrote a field, how many tokens it took, what it cost — published as its own dataset.

What changed since v2.2

v2.2 (2026-09-07) added tier 5, the WordNet 3.0 gap. v2.3 adds tier 6: named entities. Every tier before it was selected by word frequency or by WordNet membership, and neither signal ranks a name — a name's importance is a fact about the world, not about a corpus — so v2.2 knew Washington and Lincoln but not George Washington, New York City or World War II. Tier 6 is 15,000 candidates ranked by Wikipedia vital-article level, Wikidata sitelink count, WordNet instance membership and US salience, of which 12,078 became entries. Three schema changes come with it:

• Entity types are written rather than defaulted. Every proper noun in v2.2 carried entity_type = other, because the two migrations and the kind classifier all wrote that placeholder and nothing ever replaced it. 28,915 proper nouns now carry a real type — person, place, organization, work, event, product, species — taken from the candidate list where it knew one and bought as a single batched verdict where it did not. lexicon and senses gain an entity_type column, and lexicon gains wikidata_qid, the join key for reconciling an entry against Wikidata.
• Aliases. A name has variants — Lincoln for Abraham Lincoln, the Netherlands for Netherlands, FDR, NASA — and v2.2 had nowhere to put them. A variant that has an entry of its own is now an alias_of edge in opengloss-v2.3-relations (1,267 of them, written by a judged alias pass). The schema also reserves a lexicon.aliases column and alias rows in opengloss-v2.3-inflections for variants with no entry of their own, but no pass populates them yet: aliases is empty on every v2.3 row. An alias_of edge is never demoted, pruned, capped or re-judged by the hygiene passes, unlike every other relation type.
• Two new domain leaves. nature.settlements (cities, towns, villages, neighbourhoods) and law_government.polities (countries, states, provinces, empires, historical polities). A quarter of tier 6 is a settlement or a polity and the taxonomy had no leaf for either; adding a geography root would have been a breaking change to a fixed 15-root vocabulary, so both went under roots that already exist.

	v2.2 (2026-09-07)	v2.3
Lexemes	148,292	160,724
Live senses	288,304	300,787
Tier 6 lexemes (named entities)	0	12,078
Pretraining documents	1,458,684	1,560,030
Pretraining words	398,029,628	418,161,358
Pretraining tokens (cl100k_base)	565,384,746	594,154,612
Judge score, Opus, 40-entry samples	70.2 (core + tier 2), 66.7 (tier 3), 67.0 (tier 4), 81.3 (tier 5)	70.2 (core + tier 2), 66.7 (tier 3), 67.0 (tier 4), 81.3 (tier 5), 73.0 (tier 6)

Schema. No column was removed or retyped. lexicon gains entity_type, wikidata_qid and aliases; senses gains entity_type; relations gains the alias_of type; inflections gains the alias relation; tier gains the value tier6; and the domain taxonomy gains two leaves (taxonomy version 3).

What changed since v2.1

v2.1 (2026-09-07) added tier 4 and the inflections repo. v2.2 adds tier 5: 43,652 WordNet 3.0 candidate lemmas the earlier tiers lacked — common compounds and technical nouns, adjectives, adverbs and verbs, instances/taxa/organisms excluded — 38,526 of them imported outright, the rest matched against v1.3's own files. The other three changes are about honesty rather than coverage:

• The lemma fold. lexeme-hygiene (D-79) folded 4,377 inflected-form headwords onto the lemma that already carried their meaning ("databases" onto "database", through the store's own recorded morphology) and retired 172 multiword fragments that began or ended on a function word ("is not", "on top of"). Together with D-76's phantom part-of-speech retirements, 4,549 lexemes store-wide now have every sense tombstoned. A lexeme like that is not counted as a lexeme anywhere in this card or in Stats any more — it has no live sense, so it is not a lexeme by this release's own count — but it is not gone: its surface form still resolves through opengloss-v2.2-inflections, and its lexicon row carries retired = true with a retired_reason explaining why.
• Provenance on inherited fields. Every field a migration or import wrote, not only what a model wrote from scratch, now carries a migrate-stage provenance record naming where it came from, so "where did this text come from" is answerable by grep rather than by trusting the pipeline that happened to run.
• A source column on lexicon and senses: opengloss-v1.3 for content this project generated or migrated from its own legacy releases, wordnet-3.0 for the tier-5 entries imported directly from Princeton WordNet 3.0.

	v2.1 (2026-09-07)	v2.2
Lexemes	109,633	160,724
Live senses	250,003	300,787
Tier 5 lexemes (WordNet gap)	0	43,227
Retired lexemes (every sense tombstoned)	0	4,567
Pretraining documents	1,111,044	1,458,684
Pretraining words	331,888,239	398,029,628
Pretraining tokens (cl100k_base)	471,451,693	565,384,746
Judge score, Opus, 40-entry samples	70.2 (core + tier 2), 66.7 (tier 3), 67.0 (tier 4)	70.2 (core + tier 2), 66.7 (tier 3), 67.0 (tier 4), 81.3 (tier 5)

Schema. No column was removed or retyped. lexicon gains source, retired and retired_reason; senses gains source; tier gains the value tier5.

What changed since v2.0

v2.0 (2026-09-05) covered the frequency-ranked single words. v2.3 adds tier 4: the function words the core ranking had excluded on purpose, and every remaining v1.3 entry at Wikipedia frequency ≥ 10 — mostly multiword compounds ("natural selection", "catalog number"), plus names and rarer single words. That doubles the lexeme count and changes the mix: v2.0 was 99.8% single words; a third of v2.3 is multiword.

	v2.0 (2026-09-05)	v2.1 (2026-09-07)
Lexemes	54,724	109,633
Live senses	137,314	250,003
Multiword entries (compounds, phrasal verbs, idioms)	86	36,366
Proper nouns	10,365	17,073
Function words	114	462
Gloss renditions	1,129,975	1,684,865
Example sentences	1,398,297	2,163,329
Live relations	735,318	1,574,438
Synthetic queries	1,330,311	1,304,650
QA pairs	750,348	736,010
Pretraining documents	617,175	1,111,044
Pretraining words	196,390,946	331,888,239
Pretraining tokens (cl100k_base)	275,659,096	471,451,693
Judge score, Opus, 40-entry samples	70.2 (core + tier 2), 66.7 (tier 3)	70.2 (core + tier 2), 66.7 (tier 3), 67.0 (tier 4)

Schema. No column was added, removed or retyped in any existing dataset. Three things did change:

• tier gains the value tier4 (it was core, tier2 or tier3).
• One new dataset, opengloss-v2.3-inflections: a flat surface-form → lemma lookup (plural, past tense, participles, comparative, superlative, derivations) built from the morphology that the lexicon already carried nested.
• New provenance note prefixes on tombstones and edges, all reversible and all counted in the store audit: phantom_pos: (a v1.3 part-of-speech block whose glosses defined a component word rather than the compound — 11,440 blocks retired), regen: (relations regenerated for senses that had lost every edge to judging), and retyped: contrast (synonym edges the contrast paragraphs showed to be hypernym or hyponym).

Not row-compatible with v2.0. Lexeme, sense, rendition, edge, query and QA ids are stable for every entry v2.0 had. The derived training sets (retrieval-pairs, retrieval-triples, qrels) re-sample negatives over the larger pool, so their rows differ; and the store-wide quality passes run for v2.3 retired ~3,000 senses of the v2.0 entries (phantom part-of-speech blocks and near-duplicate senses), so those senses are now tombstoned rather than live. Treat v2.3 as a new release, not a delta.

Scope: fewer headwords, far more per headword

v2.3 is not a superset of v1.3. It covers 160,724 of v1.3's 205,988 lexemes — every frequency-ranked single word, plus the compounds and names at Wikipedia frequency ≥ 10 — and spends the difference on depth. If you need breadth of vocabulary, use v1.3; if you need graded renditions, resolved relations, spans, or retrieval supervision, use v2.3.

	v1.3	v2.3
Lexemes	205,988	160,724
Senses	565,604	300,787
Definition renditions per sense	1 canonical	1 canonical + up to 8 graded
Relation targets	bare strings	resolved to sense ids
Retrieval training data	companion sets	queries, QA, triples, qrels
Per-field provenance	no	model, tokens and cost per call

Key statistics

Lexemes	160,724
Retired lexemes (every sense tombstoned; not counted above)	4,567
Live senses	300,787
Rows in this dataset	2,870,291
entries_scanned	165,291
queries_considered	1,446,320
senses_considered	300,787
triples_written	2,870,291

By tier

• core — top 10K by composite frequency
• tier2 — ranks to ~42K
• tier3 — the rest of the frequency-ranked single words
• tier4 — stopwords, plus compounds and names at Wikipedia frequency ≥ 10
• tier5 — the WordNet 3.0 lemmas the earlier tiers lacked: common compounds and technical nouns, adjectives, adverbs and verbs (instances, taxa and organisms excluded); 5,126 from v1.3 files, the rest imported from WordNet
• tier6 — named entities — people, places, organizations, works and events ranked by Wikipedia vital-article level, Wikidata sitelinks, WordNet instance membership and US salience, which no frequency list ranks

Tier	Lexemes	Live senses
core	9,427	32,193
tier2	30,346	71,957
tier3	11,452	23,511
tier4	53,841	113,873
tier5	43,227	46,755
tier6	12,078	12,145
unknown	353	353

Coverage by tier

The release was built in 6 frequency-ranked passes (core, tier2, tier3, tier4, tier5 and tier6) and they did not all receive the same stages. This table is per-field and per-tier so the gaps are visible rather than averaged away.

Field	Of	core	tier2	tier3	tier4	tier5	tier6	unknown
Canonical gloss	sense	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%
Controlled domain tag	sense	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%
Gloss at 4 reading levels	sense	100.0%	99.9%	99.9%	100.0%	99.9%	100.0%	0.0%
Gloss in 4 registers	sense	100.0%	100.0%	0.0%	0.0%	0.0%	0.0%	0.0%
At least one example	sense	100.0%	99.9%	99.8%	98.4%	100.0%	100.0%	100.0%
Examples at 4 reading levels	sense	99.0%	99.6%	99.7%	97.6%	99.9%	100.0%	0.0%
At least one relation	sense	99.3%	99.2%	99.2%	99.2%	94.4%	98.5%	100.0%
Synthetic retrieval queries	sense	100.0%	100.0%	0.0%	0.0%	0.0%	0.0%	0.0%
Grounded QA pairs	sense	99.8%	99.6%	0.0%	0.0%	0.0%	0.0%	0.0%
Etymology	lexeme	100.0%	100.0%	99.8%	100.0%	100.0%	100.0%	100.0%
Lexical explanation	lexeme	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%
Encyclopedia (neutral)	lexeme	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%
Encyclopedia at grade 5 + college (core entries also carry grade 1 and grade 10)	lexeme	100.0%	100.0%	100.0%	100.0%	100.0%	100.0%	0.0%
Contrast paragraphs	lexeme	74.9%	56.2%	0.0%	0.0%	0.0%	0.0%	0.0%

Files

Files	Config	Rows	Shards	Size
data/train-*.parquet	default	2,870,291	6	139.6 MB

Fields

2,870,291 rows, one row per (query, positive, negative) triple.

Field	Type	Description
query	string	The query text.
positive	string	A passage the query should retrieve.
negative	string	A passage it should not.
negative_kind	string	other_sense, confusable, co_hyponym, synonym_of_synonym or easy.
query_id	string	{sense_id}#q{n} for a written query, or the id of the gloss rendition standing in for one.
positive_id	string	A sense id, a {sense_id}#example id, or a {lexeme_id}:encyclopedia id.
negative_id	string	Same id space as positive_id.
query_source	string	generated for a written query, gloss_pseudo when a gloss rendition stood in — so pseudo-queries can be filtered or reweighted.
sense_id	string	The sense the query belongs to.
lexeme_id	string	That sense's entry id. Join key.
live_senses	int32	Live sense count of the query's own entry.
tier	string	Tier of the query's entry.

One real row:

{
  "query": "Showing that none of the counted items are present.",
  "positive": "a zero score",
  "negative": "a mathematical element that when added to another number yields the same number",
  "negative_kind": "other_sense",
  "query_id": "0:adjective:0#grade_5/plain",
  "positive_id": "0:adjective:0#example",
  "negative_id": "0:noun:0",
  "query_source": "gloss_pseudo",
  "sense_id": "0:adjective:0",
  "lexeme_id": "0",
  "live_senses": 2,
  "tier": "tier5"
}

Loading it

from datasets import load_dataset

ds = load_dataset("mjbommar/opengloss-v2.3-retrieval-triples", split="train")
print(ds)
print(ds[0])

The shards are plain parquet, so nothing forces you through datasets — read them straight, locally or over hf://:

import polars as pl

df = pl.read_parquet("hf://datasets/mjbommar/opengloss-v2.3-retrieval-triples/data/train-*.parquet")
print(df.head())

import duckdb

duckdb.sql("SELECT count(*) FROM 'hf://datasets/mjbommar/opengloss-v2.3-retrieval-triples/data/train-*.parquet'").show()

Train a bi-encoder with MultipleNegativesRankingLoss

from datasets import load_dataset
from sentence_transformers import SentenceTransformer, SentenceTransformerTrainer, losses

triples = load_dataset("mjbommar/opengloss-v2.3-retrieval-triples", split="train")
# MultipleNegativesRankingLoss expects exactly (anchor, positive, negative) columns.
triples = triples.select_columns(["query", "positive", "negative"]).rename_column(
    "query", "anchor"
)

model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
trainer = SentenceTransformerTrainer(
    model=model,
    train_dataset=triples,
    loss=losses.MultipleNegativesRankingLoss(model),
)
trainer.train()

Identifiers, and how they compose

Every id is derived from structure, never randomly minted, so a consumer can recompute one from a row and join across the whole family without a lookup table. Sense positions are stable across regenerations: a retired sense is tombstoned, not removed, so the indices after it never shift.

Id	Shape	Example
Lexeme	slugify(headword)	abseil
Sense	{lexeme_id}:{pos}:{index} (zero-based)	abseil:verb:0
Rendition	{owner_id}#{reading_level}/{register}	abseil:verb:0#grade_5/plain
Entry-level owner	{lexeme_id}:encyclopedia / :explanation	abseil:encyclopedia
Edge	{source_sense_id}-{type}->{target_lexeme_id}	abseil:verb:0-synonym->rappel
Query	{sense_id}#q{n} (zero-based)	abseil:verb:0#q3
QA pair	{sense_id}#qa{n} (zero-based)	abseil:verb:0#qa3
Provenance record	p{n} within its entry (one-based)	p12

An edge id keys on the target's slug, not on the target's sense, so resolving a target never changes the id of the edge that found it.

Reading levels and registers

A rendition is keyed on a (reading_level, register) pair. The canonical rendition of every field is (neutral, plain); everything else is a rewrite of it.

reading_level	Who it is written for	Rough CCSS band
neutral	The canonical text: an adult general reader, no level targeted	—
grade_1	Beginning readers; short sentences, common words	K–1
grade_5	Upper elementary	4–5
grade_10	Secondary	9–10
college	Undergraduate and above; technical vocabulary allowed	11–CCR

register	What changes	Reading it
plain	Nothing — the neutral register	The default
informal	Conversational, contractions, everyday words	How you'd say it to a friend
formal	Full forms, precise hedging, no contractions	How you'd write it in a report
technical	Domain vocabulary, exact conditions	How a specialist would state it
marketing	Benefit-first, persuasive framing	A genre, not a formality level

marketing sits on the register axis for convenience but is a genre value rather than a point on the formality scale — worth remembering if you train a formality classifier on this column.

Related datasets

Everything below is built from the same store and joins on lexeme_id / sense_id.

Dataset	Grain	What it holds
opengloss-v2.3-lexicon	one row per lexeme	One row per lexeme: kind, morphology, etymology, encyclopedia, contrasts, sense ids, provenance summary.
opengloss-v2.3-senses	one row per live sense	One row per live sense: canonical gloss, 8 gloss renditions, examples, resolved relations, synthetic queries, grounded QA pairs.
opengloss-v2.3-definitions	one row per gloss rendition	One row per gloss rendition (canonical included): reading level, register, text, readability grade.
opengloss-v2.3-examples	one row per example rendition	One row per example sentence with the headword's character span, its reading level and register.
opengloss-v2.3-encyclopedia	one row per encyclopedia rendition · one row per lexical-explanation rendition	One row per encyclopedia article rendition, plus an explanation config for the "why this word" prose.
opengloss-v2.3-etymology	one row per entry with an etymology	One row per entry with an etymology: prose summary, ordered language trail, cognates, references.
opengloss-v2.3-inflections	one row per inflected, derived or lemma form	One row per inflected or derived form, plus the lemma itself: a flat form→lemma lookup.
opengloss-v2.3-relations	one row per live relation edge · one row per removed relation edge	One row per semantic edge, resolved to target sense ids; a tombstoned config recovers the edges the reconcile pass removed.
opengloss-v2.3-queries	one row per synthetic query	One row per synthetic retrieval query, across eight query styles, tagged to the sense it should retrieve.
opengloss-v2.3-qa-pairs	one row per question/answer pair	One row per grounded question/answer pair, with the rendition ids the answer cites.
opengloss-v2.3-contrasts	one row per contrast paragraph	One row per "X vs Y" paragraph on a synonym/antonym/confusable edge, with a verdict on the edge.
opengloss-v2.3-provenance	one row per provenance record	One row per recorded generation call: stage, model, tokens, cost, run id — the audit trail.
opengloss-v2.3-retrieval-pairs	one row per mined pair	Word-in-context and doc2query-shaped (text_a, text_b, label) pairs mined from the store for free.
opengloss-v2.3-retrieval-triples (this one)	one row per (query, positive, negative) triple	MS MARCO-style (query, positive, negative) triples whose hard negatives come from the graph.
opengloss-v2.3-qrels	one row per query, with its whole graded candidate list · one row per document in the retrieval corpus	Graded TREC relevance judgements (0–3) plus the document corpus and listwise candidate lists.
opengloss-v2.3-pretrain	one row per rendered document	Entries serialised into plain-prose dictionary, thesaurus, encyclopedia and usage-note documents.

Known limitations

• It is synthetic. Every string here was written by a language model against a schema, not transcribed from a corpus or checked by a lexicographer. It is well-formed and internally consistent; it is not attested usage, and it will contain confident errors. Do not use it as ground truth about what a word means.
• Judge scores 70.2/100 (core + tier 2) and 66.7/100 (tier 3). A different model family (Claude Opus) scored fixed 40-entry stratified samples at the close of each build. Sample statistics, not per-entry guarantees, and the judge is itself a model.
• Relation precision is the weakest axis. Relations were judged for validity and the ones that failed were demoted rather than asserted; symmetric reciprocity finished at 94.2% for synonyms and 94.3% for antonyms, and 4,722 senses were left with no relation at all. Treat a single edge as a hypothesis, not a fact; treat the aggregate graph as usable.
• core, tier2, tier3, tier4, tier5 and tier6 are deliberately partial. 160,371 lexemes across core, tier2, tier3, tier4, tier5 and tier6 received the text stages (glosses, examples, encyclopedia) but not the queries, QA pairs, contrasts or register renditions. The coverage table above gives the exact per-field share; nothing is hidden behind an average.
• The encyclopedia is entry-level. One article per headword, about the headword as a whole. On a polysemous entry it is not a description of any one sense, and it is never used as a positive for one (D-71). It is entry-level reference prose, not a specialist article.
• Pseudo-queries. A sense with no written query falls back to its grade_5/plain gloss standing in as one; query_source says which happened, and a pseudo-query is a paraphrase of the document it is supposed to retrieve, which makes it easier than a real query. Filter on query_source == 'generated' for the harder set.

Sources and licences

This release is Creative Commons Attribution 4.0 International (CC-BY 4.0). Of 160,724 lexemes in this release, 40,643 (the tier-5 entries whose source column reads wordnet-3.0) are derived from Princeton WordNet 3.0: their glosses, examples, relations and derivationally related forms, plus WordNet's own capitalisation of the headword (D-78).

The WordNet License permits use, copying, modification and distribution without fee, provided its notice is preserved:

The WordNet License notice is quoted in full on the opengloss-v2.3-lexicon and opengloss-v2.3-senses cards; this repo's WordNet-derived rows are governed by the same terms.

Citation

@misc{bommarito2025opengloss,
  title  = {OpenGloss: A Synthetic Encyclopedic Dictionary and Semantic Knowledge Graph},
  author = {Bommarito, Michael J., II},
  year   = {2025},
  eprint = {2511.18622},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2511.18622}
}

Tier-5 entries additionally derive from Princeton WordNet 3.0 (D-78):

@article{miller1995wordnet,
  title   = {WordNet: A Lexical Database for English},
  author  = {Miller, George A.},
  journal = {Communications of the ACM},
  volume  = {38},
  number  = {11},
  pages   = {39--41},
  year    = {1995}
}

@book{fellbaum1998wordnet,
  title     = {WordNet: An Electronic Lexical Database},
  editor    = {Fellbaum, Christiane},
  publisher = {MIT Press},
  year      = {1998}
}

License

Released under Creative Commons Attribution 4.0 International (CC-BY 4.0). Attribution to the OpenGloss project is required; commercial use is permitted. See Sources and licences above for the Princeton WordNet License that additionally covers this release's tier-5 entries.