Datasets:

xBLiMPs
/

xblimps-pilot-results

model stringclasses 6 values	language stringclasses 6 values	phenomenon stringclasses 138 values	paradigm stringclasses 296 values	n int64 1 10
apertus_70b	en	anaphor_agreement	anaphor_gender_agreement	10
apertus_70b	en	anaphor_agreement	anaphor_number_agreement	10
apertus_70b	en	argument_structure	causative	10
apertus_70b	en	argument_structure	drop_argument	10
apertus_70b	en	argument_structure	inchoative	10
apertus_70b	en	argument_structure	intransitive	10
apertus_70b	en	argument_structure	passive_1	10
apertus_70b	en	argument_structure	passive_2	10
apertus_70b	en	argument_structure	transitive	10
apertus_70b	en	binding	principle_A_c_command	10
apertus_70b	en	binding	principle_A_case_1	10
apertus_70b	en	binding	principle_A_case_2	10
apertus_70b	en	binding	principle_A_domain_1	10
apertus_70b	en	binding	principle_A_domain_2	10
apertus_70b	en	binding	principle_A_domain_3	10
apertus_70b	en	binding	principle_A_reconstruction	10
apertus_70b	en	control_raising	existential_there_object_raising	10
apertus_70b	en	control_raising	existential_there_subject_raising	10
apertus_70b	en	control_raising	expletive_it_object_raising	10
apertus_70b	en	control_raising	tough_vs_raising_1	10
apertus_70b	en	control_raising	tough_vs_raising_2	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_1	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_2	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_irregular_1	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_irregular_2	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_with_adj_2	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_with_adj_irregular_1	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_with_adj_irregular_2	10
apertus_70b	en	determiner_noun_agreement	determiner_noun_agreement_with_adjective_1	10
apertus_70b	en	ellipsis	ellipsis_n_bar_1	10
apertus_70b	en	ellipsis	ellipsis_n_bar_2	10
apertus_70b	en	filler_gap_dependency	wh_questions_object_gap	10
apertus_70b	en	filler_gap_dependency	wh_questions_subject_gap	10
apertus_70b	en	filler_gap_dependency	wh_questions_subject_gap_long_distance	10
apertus_70b	en	filler_gap_dependency	wh_vs_that_no_gap	10
apertus_70b	en	filler_gap_dependency	wh_vs_that_no_gap_long_distance	10
apertus_70b	en	filler_gap_dependency	wh_vs_that_with_gap	10
apertus_70b	en	filler_gap_dependency	wh_vs_that_with_gap_long_distance	10
apertus_70b	en	irregular_forms	irregular_past_participle_adjectives	10
apertus_70b	en	irregular_forms	irregular_past_participle_verbs	10
apertus_70b	en	island_effects	adjunct_island	10
apertus_70b	en	island_effects	complex_NP_island	10
apertus_70b	en	island_effects	coordinate_structure_constraint_complex_left_branch	10
apertus_70b	en	island_effects	coordinate_structure_constraint_object_extraction	10
apertus_70b	en	island_effects	left_branch_island_echo_question	10
apertus_70b	en	island_effects	left_branch_island_simple_question	10
apertus_70b	en	island_effects	sentential_subject_island	10
apertus_70b	en	island_effects	wh_island	10
apertus_70b	en	npi_licensing	matrix_question_npi_licensor_present	10
apertus_70b	en	npi_licensing	npi_present_1	10
apertus_70b	en	npi_licensing	npi_present_2	10
apertus_70b	en	npi_licensing	only_npi_licensor_present	10
apertus_70b	en	npi_licensing	only_npi_scope	10
apertus_70b	en	npi_licensing	sentential_negation_npi_licensor_present	10
apertus_70b	en	npi_licensing	sentential_negation_npi_scope	10
apertus_70b	en	quantifiers	existential_there_quantifiers_1	10
apertus_70b	en	quantifiers	existential_there_quantifiers_2	10
apertus_70b	en	quantifiers	superlative_quantifiers_1	10
apertus_70b	en	quantifiers	superlative_quantifiers_2	10
apertus_70b	en	s-selection	animate_subject_passive	10
apertus_70b	en	s-selection	animate_subject_trans	10
apertus_70b	en	subject_verb_agreement	distractor_agreement_relational_noun	10
apertus_70b	en	subject_verb_agreement	distractor_agreement_relative_clause	10
apertus_70b	en	subject_verb_agreement	irregular_plural_subject_verb_agreement_1	10
apertus_70b	en	subject_verb_agreement	irregular_plural_subject_verb_agreement_2	10
apertus_70b	en	subject_verb_agreement	regular_plural_subject_verb_agreement_1	10
apertus_70b	en	subject_verb_agreement	regular_plural_subject_verb_agreement_2	10
apertus_70b	ja	NPI licensing	NCI	2
apertus_70b	ja	NPI licensing	NPI	2
apertus_70b	ja	argument structure	animacy	1
apertus_70b	ja	argument structure	aspect	3
apertus_70b	ja	argument structure	case	10
apertus_70b	ja	argument structure	internal argument	3
apertus_70b	ja	argument structure	passive	7
apertus_70b	ja	argument structure	scrambling	10
apertus_70b	ja	binding	anaphor	5
apertus_70b	ja	binding	variable binding	1
apertus_70b	ja	binding	weak crossover	7
apertus_70b	ja	control/raising	subject control	2
apertus_70b	ja	ellipsis	nominal ellipsis	10
apertus_70b	ja	ellipsis	parasitic-gap	6
apertus_70b	ja	filler-gap	cleft	3
apertus_70b	ja	filler-gap	intervention effects	3
apertus_70b	ja	filler-gap	relative clause	1
apertus_70b	ja	filler-gap	resumptive pronoun	2
apertus_70b	ja	island effects	adjunct island	5
apertus_70b	ja	island effects	complex NP island	3
apertus_70b	ja	island effects	factive island	1
apertus_70b	ja	island effects	negative island	1
apertus_70b	ja	island effects	specificity island	1
apertus_70b	ja	morphology	honorification	2
apertus_70b	ja	morphology	idiom	8
apertus_70b	ja	morphology	inflection	10
apertus_70b	ja	morphology	nominalization	2
apertus_70b	ja	morphology	part of speech	10
apertus_70b	ja	morphology	reflexive	2
apertus_70b	ja	nominal structure	measure phrase	10
apertus_70b	ja	nominal structure	modifier	9
apertus_70b	ja	quantifiers	classifier	3
apertus_70b	ja	quantifiers	floating quantifiers	5

End of preview. Expand in Data Studio

xBLiMPs — cross-lingual BLiMP pilot (results)

Scoring outputs for a small-scale, reviewer-defensible pilot that probes where multilingual LMs systematically fail on minimal pairs, and whether those failures reflect grammatical competence or a surface confound (sentence length / subword tokenisation). Harness: github.com/suchirsalhan/xblimps.

What was scored

Published BLiMP-family benchmarks, one common scoring/confound/sampling pipeline, 10 pairs per paradigm (the BLiMP-NL "small" set):

language	benchmark	paradigms
en	BLiMP (`nyu-mll/blimp`)	67
nl	BLiMP-NL (`juletxara/blimp-nl`)	81
ru	RuBLiMP (`RussianNLP/rublimp`)	45
ja	JBLiMP (`polm-stability/jblimp`)	37
zh	ZhoBLiMP (`chinese-babylm/zhoblimp`) + SLING	11 + 43
tr	TurBLiMP (`ezgibasar/turblimp`)	16

Models (this batch): aya_expanse_8b, gemma2_9b, qwen3_8b (multilingual, all languages) and goldfish_nl (monolingual Dutch anchor). apertus_70b and gemma2_27b are deferred.

Metrics

Primary — whole-sequence log-probability accuracy (correct iff log P(grammatical) > log P(ungrammatical)), via minicons.
SLOR reported alongside (length/frequency control; TurBLiMP).
Mandatory confound logging — per pair we log the subword length delta and the cue-span subword delta, then regress each model's log-prob difference on them per (model, paradigm). R² ≥ 0.5 ⇒ the paradigm is length/tokenisation confounded and is flagged for redesign (confounds.csv, warning).

Configs (dataset viewer)

config	file	one row per
`accuracy` (default)	`results/pilot_small/accuracy.csv`	(model, paradigm): log-prob & SLOR accuracy
`pairs`	`results/pilot_small/pairs.csv`	(model, minimal pair): log-prob diff, deltas, correctness
`confounds`	`results/pilot_small/confounds.csv`	(model, paradigm): regression R², betas, warning
`regions`	`results/pilot_small/regions.csv`	pair: sampled region + cross-model agreement

Per-model raw outputs are also under results/small/<model>/, and the glossed hard-case hunt (reliably-wrong / high-disagreement pairs) is in results/pilot_small/hard_cases.md.

Headline

gemma2_9b strongest (nl .88, ru .95, tr .92, ja .85, zh .81, en .76); goldfish_nl reaches .86 on Dutch despite being tiny and monolingual. 431 reliably-wrong + 84 high-disagreement pairs feed the hard-case hunt; 114 paradigms flagged length/tokenisation-confounded (R² ≥ 0.5). Failures cluster on the literature's known weak spots — island effects, reflexives/external possessor, Japanese honorification, Dutch te/om-te.

Citation of the source benchmarks

Please cite the underlying benchmarks (BLiMP, BLiMP-NL, RuBLiMP, JBLiMP, ZhoBLiMP, SLING, TurBLiMP) — this dataset only redistributes derived scoring results, not their minimal pairs.

Downloads last month: 116