LexVE: lexical minimal pairs for word-level interventions

LexVE is a set of 2,670 English minimal-pair prompts across 38 tasks. In each pair the two prompts differ in exactly one word (the cue), and that word decides the next token. It was built to study how language models use a single word's identity, for example by swapping or steering that word's embedding or value-embedding vectors, and it works as a plain minimal-pair benchmark too.

from datasets import load_dataset

ds = load_dataset("aadityabuilds/lexve")                       # LexVE v2: splits "discovery", "evaluation"
graded = load_dataset("aadityabuilds/lexve", "graded_distance")  # cue-to-prediction distance 0-14 tokens
row = ds["evaluation"][0]

Example row (animacy_relative):

	prompt	cue	expected next token
base	We finally met the doctor,	doctor (human)	" who"
source	We finally met the statue,	statue (inanimate)	" which"

A model that has learned the distinction should prefer " who" after the base prompt and " which" after the source prompt. An intervention on the cue (e.g. replacing the doctor's vector with the statue's) should move the prediction from " who" toward " which".

Configurations

Config	Rows	Discovery / evaluation	What it is
default	2,670	1,376 / 1,294	LexVE v2: 38 tasks, near / far / attractor frames
graded_distance	2,190	1,120 / 1,070	19 tasks; one fixed stem per task with filler phrases so the cue sits 0, 4, 7, 10 or 14 tokens before the prediction

Both configs reuse the same lexical pairs and the same discovery/evaluation assignment.

Tasks

The 38 tasks cover cues of 13 parts of speech and eight linguistic levels.

Task	Level	Cue POS	Tier	Status	Labels (class A / B)	Pairs	Rows	Graded rows	Example (base → label)
adj_prep_of_with	lexicalized_syntax	adjective	open	capable	of / with	17	136	170	Everyone knew she was afraid → of (vs satisfied → with)
adj_prep_to_from	lexicalized_syntax	adjective	open	capable	to / from	17	136	170	Everyone knew she was similar → to (vs different → from)
animacy_relative	lexical_semantics	noun	open	capable	who / which	20	160	–	We finally met the doctor, → who (vs statue → which)
article_phonology	phonology	article	diag	limited	apple / banana	1	8	–	For breakfast, I ate an → apple (vs a → banana)
comparative_than	morphology	adjective	open	capable	than / and	18	144	180	The new house is bigger → than (vs big → and)
complement_clause	lexicalized_syntax	verb	open	capable	whether / that	16	128	160	In the interview, the journalist wondered → whether (vs claimed → that)
container_noun	lexical_semantics	noun	closed	capable	water / paper	5	40	–	She handed me a bottle of → water (vs pile → paper)
correlative	morphosyntax	conjunction	diag	capable	or / nor	1	12	–	You can choose either the red one → or (vs neither → nor)
country_language	world_knowledge	proper_noun	open	capable	pair-specific	11	88	110	In France, most people speak → French (vs Spain → Spanish)
dative_alternation	lexicalized_syntax	verb	open	limited	me / the	16	32	–	Two days ago, my neighbor gave → me (vs donated → the)
determiner_number	morphosyntax	determiner	closed	capable	is / are	8	96	80	The farmer told us that this sheep → is (vs these → are)
entity_type	lexical_semantics	proper_noun	open	capable	who / where	20	120	–	The article mentioned John, → who (vs London → where)
food_selection	lexical_semantics	verb	diag	capable	water / bread	6	36	–	Every morning, my grandmother drank some → water (vs ate → bread)
gender_pronoun	lexical_semantics	noun	open	capable	he / she	11	110	110	The man said that → he (vs woman → she)
gerund_vs_infinitive	lexicalized_syntax	verb	open	capable	going / to	8	64	80	Last year, she enjoyed → going (vs wanted → to)
hypernym_class	lexical_semantics	noun	closed	capable	bird / fish	8	64	80	A eagle is a kind of → bird (vs shark → fish) †
infinitive_vs_object	lexicalized_syntax	verb	open	capable	to / the	16	96	160	After a long day, Maria decided → to (vs enjoyed → the)
irregular_number	morphosyntax	noun	open	capable	was / were	16	160	160	I think the man → was (vs men → were)
light_verb	collocation	verb	diag	limited	decision / walk	1	10	–	Tomorrow I will make a → decision (vs take → walk)
mass_count	morphosyntax	quantifier	closed	capable	money / friends	4	20	–	To be honest, I do not have much → money (vs many → friends)
modal_bare_vs_to	morphosyntax	modal	closed	capable	be / to	9	72	90	Honestly, we should → be (vs ought → to)
noun_prep_for_in	lexicalized_syntax	noun	open	capable	for / in	16	96	160	The report highlighted the need → for (vs increase → in)
noun_prep_on_to	lexicalized_syntax	noun	open	capable	on / to	15	90	150	They discussed the impact → on (vs access → to)
npi_licensing	semantics	quantifier_adverb	closed	capable	any / some	7	40	–	At the party, nobody had → any (vs somebody → some)
numeral_noun_number	morphosyntax	determiner	closed	capable	year / years	7	56	–	The project took one → year (vs three → years)
particle_up_out	collocation	verb	open	capable	up / out	11	88	–	After a while, she gave → up (vs figured → out)
partitive_agreement	morphosyntax	quantifier	closed	capable	is / are	6	48	–	I think one of the boxes → is (vs two → are)
prep_government_mixed	lexicalized_syntax	verb	open	capable	pair-specific	9	72	90	At the time, Alex relied → on (vs looked → at)
pronoun_case	morphosyntax	preposition_vs_subordinator	closed	limited	him / he	8	16	–	They stayed in the room with → him (vs because → he)
range_preposition	lexicalized_syntax	preposition	diag	capable	to / and	1	12	–	Prices range from ten → to (vs between → and)
sensory_modality	lexical_semantics	adjective	closed	capable	sound / light	7	42	–	Suddenly there was a loud → sound (vs bright → light)
temporal_tense	semantics	adverb	closed	capable	will / was	5	30	40	The coach said that tomorrow the team → will (vs yesterday → was)
time_place_relative	lexical_semantics	noun	open	capable	when / where	11	66	–	I still remember the day → when (vs city → where)
transitivity	lexicalized_syntax	verb	open	capable	the / .	15	90	–	A moment later, Maya inspected → the (vs arrived → .)
unit_ago_away	lexical_semantics	noun	closed	capable	ago / away	4	32	–	That was only a few weeks → ago (vs miles → away)
verb_prep_from_of	lexicalized_syntax	verb	closed	capable	from / of	6	48	60	In the end, she suffered → from (vs consisted → of)
verb_prep_in_with	lexicalized_syntax	verb	closed	capable	in / with	6	48	60	In the end, she participated → in (vs dealt → with)
verb_prep_on_to	lexicalized_syntax	verb	open	capable	on / to	8	64	80	In the end, she focused → on (vs referred → to)

† This example is one of the flagged ungrammatical rows (see Limitations).

Fields

Field	Description
id	Unique row id: v2__<task>__p<pair>__<template>__<direction> (graded: g__<task>__p<pair>__d<distance>__<direction>).
task	Task name (38 in default, 19 in graded_distance).
base_text	Prompt with the base cue word. The model is asked for the next token.
source_text	The same prompt with the cue word swapped for its partner (the only difference).
base_word, source_word	The cue word in each prompt (no leading space).
base_label, source_label	Expected next token after each prompt, as a string with its leading space (e.g. " who").
base_class, source_class	Class of each cue word (e.g. human / inanimate).
class_labels	True if labels are shared by the whole class (36 of 38 tasks); False if each pair has its own labels (country_language, prep_government_mixed).
direction	a_to_b or b_to_a. Every pair appears in both directions (base and source exchanged).
pair_id, cluster_id	Lexical pair, e.g. animacy_relative::doctor__statue. Use cluster_id to group rows for cluster bootstraps.
pair_index	Index of the pair within its task; even = discovery, odd = evaluation (see Splits).
split	discovery or evaluation (also the HF split name).
template_id, template_text	Sentence template; {w} marks the cue slot.
frame	near (cue within two tokens of the prediction), far (neutral words in between), attractor (agreement attraction). Graded: graded_d0 ... graded_d14.
tier	open (8 or more lexical pairs), closed (function-word cues, few possible pairs), diag (a single lexeme pair; clusters are templates). Tiers should not be pooled.
linguistic_level	One of eight levels: lexical_semantics, world_knowledge, morphology, morphosyntax, lexicalized_syntax, collocation, semantics, phonology.
pos_tag	Part of speech of the cue word.
anchor	(default only) Literature that inspired the task (inspiration only; no text is copied).
calibration_status	(default only) capable or limited: whether the d12/d16/d20 nanochat models passed the discovery-pair baseline rule for this task (see Construction).
filler_tokens	(graded_distance only) Number of filler tokens inserted between cue and prediction: 0, 4, 7, 10 or 14.
schema_version	lexve-v2.0 or lexve-v2.0-graded.
known_issue, known_issue_note	True for rows whose text has "a" before a vowel-initial word (e.g. "A eagle is a kind of"); kept unchanged, see Limitations.
*_token_id, *_label_id, swap_pos, token_count, cue_gap	Tokenizer-specific (nanochat BPE, see below): token ids of the cue and label, cue token position, prompt length in tokens, and tokens between cue and prediction.

Model-agnostic vs tokenizer-specific. Every text, word, label-string, class, task and metadata field is model-agnostic. The token-id and position fields refer to the nanochat BPE tokenizer (Karpathy's nanochat, MIT license; tokenizer tokenizer.pkl sha256 387cfc082b0bee45467774fd6f1310a922ad170886a58ccddcb468f275e06a6c). With another tokenizer, re-tokenize base_text / source_text and the labels. Cue words and labels were chosen to be single tokens with a leading space under nanochat, so some may split into several tokens elsewhere.

Splits

• discovery and evaluation are assigned by lexical pair: even pair_index → discovery, odd → evaluation. Exceptions: article_phonology splits by template and food_selection by lexeme.
• Within a task (excluding diag tier), no cue word appears in both splits, so held-out results test generalization to new words.
• Use discovery to choose anything (heads, directions, hyperparameters) and report on evaluation.
• Each pair appears in both directions (a_to_b, b_to_a). For confidence intervals, resample by cluster_id, not by row.

Construction

• Hand-written. Class word lists, pairings and sentence templates were written by hand (specs_v2.py). Cross-class pairs were matched roughly by corpus frequency. No row was generated by a language model. Pairs were never selected or dropped based on any model's intervention results.
• Tokenizer screen. Every cue and label had to be a single nanochat token with a leading space and to align in every template. Pairs that failed BPE alignment were rejected (listed in provenance/manifest_*.json).
• Template calibration (baseline only, discovery pairs only). Templates were kept if, averaged over three nanochat models (d12, d16, d20 with value embeddings), at least 50% of discovery pairs were predicted correctly in both directions and the two labels held at least 2% of probability mass. Evaluation pairs never influenced selection. Tasks with fewer than two surviving near templates are kept and marked calibration_status: limited. The per-template numbers are in provenance/template_selection.json, and baseline accuracy and label mass per task and model are in provenance/baseline_competence.json.
• Graded distance. Uses the same pairs with one stem per task plus fixed filler phrases (", it seems,", ", as far as we know,", ...). The stems and fillers are in provenance/manifest_graded.json.
• Inspirations, not sources. Task designs draw on CausalGym, BLiMP, RAVEL, COMPS, Ettinger (2020), Warstadt et al. (2019) and Chierchia (1998) (field anchor). No text was copied from these resources.

Limitations

• Ungrammatical article rows. 22 rows in default (11 distinct texts) and 30 rows in graded_distance (15 texts) put "a" before a vowel-initial word: "A eagle / eel / owl is a kind of" (hypernym_class), "a aircraft" (determiner_number), "a audible" (sensory_modality). They are kept unchanged so the data match published results, and flagged with known_issue. Filter them out with ds.filter(lambda r: not r["known_issue"]).
• Calibration circularity. Templates were chosen using baselines of the same nanochat models the dataset was first used to analyse (discovery pairs only). Results on other models do not inherit this bias, but results on those three models should be read with it in mind.
• Small tasks. Four diag tasks have a single lexical pair (article_phonology, correlative, light_verb, range_preposition); their clusters are templates. Several closed tasks have only 4-9 pairs.
• Weak baselines. Four tasks are limited (article_phonology, dative_alternation, light_verb, pronoun_case). In the calibration models, label mass is low for some tasks (pronoun_case about 2%, article_phonology 1-2%, light_verb 3-5%).
• Cue reuse. Some cue words appear in more than one task (e.g. "actor").
• Scope. English only. Gender pairs are binary (masculine/feminine nouns). Labels are a single next token, so this is a two-way forced choice, not a test of fluent generation.

Provenance and integrity

The Parquet files are a lossless conversion of the frozen JSONL files plus the two known_issue columns, verified value by value. The source hashes are:

Config	Source file	sha256
default	lexve_v2.jsonl	3b8a30a8f989a900026bfaf0c480a7e00a516a400270db6b2c79db94b7d957e2
graded_distance	lexve_v2_graded.jsonl	72028c0b5fdc7067f568722b25c0d39c549d598be3af9011891d9a592fcff6b9

SHA256SUMS covers every file in this repository.

License

CC BY 4.0.

Citation

@misc{lexve2026,
  title  = {LexVE: Lexical Minimal Pairs for Word-Level Interventions},
  author = {AUTHORS_TBD},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/aadityabuilds/lexve}}
}