BLiMP-IT

Dataset Summary

BLiMP-IT is a linguistically motivated benchmark for evaluating Italian language models through minimal pairs. Each example consists of a grammatical sentence paired with a minimally different ungrammatical counterpart that isolates a single morphosyntactic contrast. The benchmark is designed to evaluate whether language models assign higher probability to the grammatical sentence than to the ungrammatical one.

The benchmark is inspired by the English BLiMP while extending and adapting it to Italian morphosyntax. It integrates and adapts several existing resources — the COnVERSA psycholinguistic assessment battery, AcCompl-it, and BLiMP itself — together with newly designed phenomena such as across-the-board (ATB) movement and parasitic gaps.

Rather than measuring downstream task performance, BLiMP-IT measures the acquisition of abstract morphosyntactic generalizations.

Supported Tasks

The dataset is intended for:

• Language model evaluation through sentence probability comparison
• Minimal-pair acceptability evaluation
• Psycholinguistically motivated syntactic evaluation
• Cross-model comparison
• Cross-linguistic comparison with BLiMP and related benchmarks

Typical evaluation consists of assigning higher probability (or lower perplexity) to the grammatical member of each minimal pair.

---

Dataset Structure

The repository exposes one configuration per paradigm, each with a single test split.

Configuration names follow the scheme <source>_<macro-phenomenon>_<paradigm>:

• blimp_it_* — paradigms derived from the original English BLiMP and adapted to Italian
• conversa_a_agreement_*, conversa_b_argument_structure_*, conversa_c_pronouns_*, conversa_d_questions_* — paradigms derived from the COnVERSA battery

Macro-phenomena

Macro-phenomenon	Configurations released
Agreement & Inflection	15
Verbal Class & Argument Structure	6
Pronouns	8
Non-local Dependencies	40
Total	69

The full BLiMP-IT framework targets 78 morphosyntactic phenomena; the 69 configurations released here are the subset for which validated minimal pairs are currently available. Coverage is expanded as generation and expert validation proceed.

---

Phenomena

Agreement & Inflection

Adapted from BLiMP

• DP agreement
  • blimp_it_agreement_and_inflection_dp

COnVERSA

• Determiner–noun agreement

  • conversa_a_agreement_dp

• Subject–verb agreement

  • conversa_a_agreement_subject_verb_attraction
  • conversa_a_agreement_subject_verb_cumulative
  • conversa_a_agreement_subject_verb_trans
  • conversa_a_agreement_subject_verb_unacc
  • conversa_a_agreement_subject_verb_unerg
  • conversa_a_agreement_subject_verb_unerg_trans
  • conversa_a_agreement_verb_subject_unacc

• Predicate agreement

  • conversa_a_agreement_subject_nominal_predicate
  • conversa_a_agreement_subject_nominal_predicate_attraction

• Past participle agreement

  • conversa_a_agreement_past_participle_pre_v_cl
  • conversa_a_agreement_past_participle_unacc

• Psych verbs

  • conversa_a_agreement_psych_verbs_piacere
  • conversa_a_agreement_psych_verbs_preoccupare

---

Verbal Class & Argument Structure

COnVERSA

• Theta-role assignment

  • conversa_b_argument_structure_theta_roles

• Auxiliary selection

  • conversa_b_argument_structure_auxiliary_selection_ditransitives
  • conversa_b_argument_structure_auxiliary_selection_passive_active_diathesis
  • conversa_b_argument_structure_auxiliary_selection_transitives
  • conversa_b_argument_structure_auxiliary_selection_unaccusatives
  • conversa_b_argument_structure_auxiliary_selection_unergatives

---

Pronouns

COnVERSA

• Personal pronouns

  • conversa_c_pronouns_1_2_person_answers
  • conversa_c_pronouns_1_2_person_elicited_declaratives

• Clitics

  • conversa_c_pronouns_clitics_answering_with_acc_cl
  • conversa_c_pronouns_clitics_coordinating_with_acc_cl
  • conversa_c_pronouns_clitics_coordinating_with_dat_cl

• Reflexives

  • conversa_c_pronouns_reflexives_other
  • conversa_c_pronouns_reflexives_psych_verbs
  • conversa_c_pronouns_reflexives_unacc

---

Non-local Dependencies

COnVERSA — question formation and comprehension

• conversa_d_questions_answering_questions_with_rc_obj
• conversa_d_questions_answering_questions_with_rc_subj
• conversa_d_questions_answering_wh_adjunct_questions
• conversa_d_questions_answering_wh_argument_questions
• conversa_d_questions_answering_wh_argument_questions_number_disambiguation
• conversa_d_questions_answering_why_questions
• conversa_d_questions_answering_yes_no_questions
• conversa_d_questions_question_formation_subject_position

Adapted from BLiMP and newly designed

• Relative clauses

  • blimp_it_non_local_dependencies_atb_rc_object_gap_nogap
  • blimp_it_non_local_dependencies_atb_rc_object_nogap_gap
  • blimp_it_non_local_dependencies_atb_rc_object_nogap_nogap
  • blimp_it_non_local_dependencies_rc_object
  • blimp_it_non_local_dependencies_rc_object_nogap
  • blimp_it_non_local_dependencies_rc_subject
  • blimp_it_non_local_dependencies_rc_subject_nogap

• WH extraction

  • blimp_it_non_local_dependencies_wh_extraction_embedd2_clitic_inanimate
  • blimp_it_non_local_dependencies_wh_extraction_embedd2_np_inanimate
  • blimp_it_non_local_dependencies_wh_extraction_embedd2_np_inanimate_dem
  • blimp_it_non_local_dependencies_wh_extraction_embedd_clitic_inanimate
  • blimp_it_non_local_dependencies_wh_extraction_embedd_np_inanimate
  • blimp_it_non_local_dependencies_wh_extraction_embedd_np_inanimate_dem

• WH islands

  • blimp_it_non_local_dependencies_wh_island_embedd
  • blimp_it_non_local_dependencies_wh_island_root
  • blimp_it_non_local_dependencies_wh_island_root_affirmative_dove

• Across-the-board (ATB) movement

  • blimp_it_non_local_dependencies_atb_affirmative2
  • blimp_it_non_local_dependencies_atb_affirmative_mi
  • blimp_it_non_local_dependencies_atb_affirmative_ti
  • blimp_it_non_local_dependencies_atb_interrogative_gap_nogap_clitic
  • blimp_it_non_local_dependencies_atb_interrogative_gap_nogap_np_aux
  • blimp_it_non_local_dependencies_atb_interrogative_nogap_gap_clitic
  • blimp_it_non_local_dependencies_atb_interrogative_nogap_gap_np_aux
  • blimp_it_non_local_dependencies_atb_interrogative_nogap_nogap_clitic
  • blimp_it_non_local_dependencies_atb_interrogative_nogap_nogap_np_aux
  • blimp_it_non_local_dependencies_atb_rc_object_gap_nogap
  • blimp_it_non_local_dependencies_atb_rc_object_nogap_gap
  • blimp_it_non_local_dependencies_atb_rc_object_nogap_nogap

• Other islands

  • blimp_it_non_local_dependencies_adjunctisland
  • blimp_it_non_local_dependencies_complexnpisland
  • blimp_it_non_local_dependencies_coordinatestructureconstraint_complex_object_extraction
  • blimp_it_non_local_dependencies_left_branch_island_echo_question
  • blimp_it_non_local_dependencies_sententialsubjectisland

• Parasitic gaps

  • blimp_it_non_local_dependencies_parasiticgap_adjunctisland1
  • blimp_it_non_local_dependencies_parasiticgap_adjunctisland2

---

Data Fields

Each configuration contains sentence pairs with the following fields:

Field	Type	Description
pairID	string	Identifier of the minimal pair within the paradigm
sentence_good	string	The grammatical member of the pair
sentence_bad	string	The minimally different ungrammatical member
UID	string	Paradigm identifier shared by all items in the configuration
simple_LM_method	bool	Whether the pair can be scored by direct full-sentence probability comparison

---

Evaluation

BLiMP-IT is available as a task suite in lm-evaluation-harness:

lm_eval --model hf --model_args pretrained=<model> --tasks blimp_it

Task names match the configuration names in this repository exactly, so any single paradigm can be run on its own:

lm_eval --model hf --model_args pretrained=<model> \
        --tasks blimp_it_non_local_dependencies_wh_island_embedd

Four sub-groups aggregate the paradigms by macro-phenomenon: blimp_it_agreement_and_inflection, blimp_it_argument_structure, blimp_it_pronouns, blimp_it_non_local_dependencies. Aggregation is an unweighted mean over paradigms at both levels, following the convention of the original blimp task — paradigm sizes range from 4 to 344 pairs, so size-weighted means would be dominated by a handful of the larger non-local-dependency conditions.

---

Data Creation

Source

BLiMP-IT combines:

• manually designed minimal pairs
• automatic lexical generation using grammar templates
• a manually annotated Italian lexicon
• human validation

The lexicon was derived from a 2.4M-token corpus of Italian child-directed and naturalistic language (CHILDES, film subtitles, children's songs, telephone conversations and fairy tales), annotated with POS and Universal POS tags, morphological features, and manual animacy labels. Automatic generation instantiates abstract tag templates against this lexicon, producing lexical variants while preserving the targeted morphosyntactic contrast. Generated pairs are reviewed independently by at least two linguists, with disagreements resolved by discussion.

---

Dataset Characteristics

The benchmark follows two design principles:

1. Minimality of contrasts

   Each grammatical and ungrammatical sentence differs only in the manipulation responsible for (un)grammaticality.

2. Linguistically meaningful errors

   Ungrammatical sentences contain theoretically motivated morphosyntactic violations rather than random perturbations or word permutations.

---

Intended Uses

BLiMP-IT is intended for

• evaluating Italian language models
• probing syntactic generalization
• comparing language models
• comparing humans and language models
• psycholinguistic research
• computational linguistics research

---

Limitations

• The benchmark focuses exclusively on morphosyntax rather than semantics.
• Some generated sentences may be semantically unusual while remaining syntactically well formed.
• Human acceptability judgments currently exist only for subsets overlapping with COnVERSA.
• Paradigm sizes are uneven (4 to 344 pairs); the smallest configurations give noisy per-paradigm estimates and are best read at the macro-phenomenon level.
• The source corpus is largely child-directed speech, which limits lexical register.
• Automatic lexical generation is still expanding coverage across all 78 target phenomena.

---

Citation

Bressan et al. (2026) Language models assessment through linguistically motivated contrasts: A benchmark for Italian (BLiMP-IT). Proceedings of GLOW 47.

Barbini et al. (2025) BLiMP-IT: Harnessing Automatic Minimal Pair Generation for Italian Language Model Evaluation. Proceedings of CLiC-it 2025.

@inproceedings{bressan2026blimpit,
  title={Language models assessment through linguistically motivated contrasts: A benchmark for Italian (BLiMP-IT)},
  author={Bressan, Veronica and Barbini, Matilde and Fusco, Achille and Neri, Sofia and Piccini Bianchessi, Maria Letizia and Rossi, Sarah and Chesi, Cristiano},
  booktitle={Proceedings of GLOW 47},
  year={2026}
}

@inproceedings{barbini2025blimpit,
  title={BLiMP-IT: Harnessing Automatic Minimal Pair Generation for Italian Language Model Evaluation},
  author={Barbini, Matilde and Piccini Bianchessi, Maria Letizia and Bressan, Veronica and Fusco, Achille and Neri, Sofia and Rossi, Sarah and Sgrizzi, Tommaso and Chesi, Cristiano},
  booktitle={Proceedings of the 11th Italian Conference on Computational Linguistics (CLiC-it 2025)},
  year={2025}
}

---

License

CC BY-SA 4.0