Flaglab/ESNLIR-dataset

ESNLIR

A Spanish multi-genre NLI dataset with explicit causal annotation. 4.4M training pairs drawn from 34 corpora across 8 genres, with a fourth label — reasoning — for cause–effect relationships that the usual entailment / contradiction / neutral scheme cannot express.

ESNLIR: Expanding Spanish NLI Benchmarks with Multi-Genre and Causal Annotation Johan R. Portela, Nicolás Pérez-Terán, Rubén Manrique — Universidad de los Andes, Bogotá Applied Informatics, Springer, 2026, pp. 345–361 — doi:10.1007/978-3-032-07175-0_23

Construction pipeline: jd-rodriguezp1234/esnlir · Archive: Zenodo 15002575

Labels

label	how it is assigned
contrasting	discourse connector signalling contradiction, comparison or opposition
entailment	connector signalling generalization, specification or equivalence
reasoning	connector signalling cause, reason or condition — the class this dataset adds
neutral	not connector-derived: sentences paired across different paragraphs

neutral being built differently is a known bias in its favour — every baseline finds it the easiest class. Keep that in mind when reading per-class numbers.

Files

All files are JSON Lines — one object per line — despite the .json extension.

file	rows	what it is
train.json	4,387,968	training split
val.json	67,400	validation split
test.json	80,216	test split
test_length_mismatch.json	80,216	stress test: premise padded
test_negation.json	80,216	stress test: negation added to hypothesis
test_overlap.json	80,216	stress test: forced lexical overlap
test_spelling.json	80,216	stress test: misspelling in premise
full_labeled_final_dataset.json	2,136	human-annotated sample, before filtering
labeled_final_dataset.json	924	pairs where the human majority label matched the connector label
dataset_metadata.xlsx	—	per-corpus metadata

Every split is class-balanced by construction: train, val and test each hold exactly equal counts of the four labels (test: 20,054 each).

Fields

{"sentence_1": "...", "sentence_2": "...", "connector": "por eso",
 "connector_type": "reasoning", "extraction_strategy": "linking_phrase", "distance": 1.0,
 "sentence_1_paragraph": 46, "sentence_1_position": 25,
 "sentence_2_paragraph": 46, "sentence_2_position": 26,
 "id": "esnews__spanish_pd_news__365259", "dataset": "esnews__spanish_pd_news",
 "genre": "news", "domain": "spanish_public_domain_news"}

The gold label is connector_type. The human-validated files add connection_type, the human majority label — where the two disagree, the connector-derived label did not survive review.

Note that id is not unique (it identifies the source article, not the pair). Join on (sentence_1, sentence_2) or on row position.

Genres and splits

Genres: articles, books, comments, legal, clinical, news, talks, theses — 34 corpora in total, spanning encyclopedia and web text, public-domain literature, social media, Colombian legal documents, clinical case reports, news, TED talks and university theses.

Splits are doubly stratified: source articles are partitioned first, so no article — and therefore no sentence pair — leaks between splits; each split is then downsampled to class balance.

Ten corpora are held out to the test split only, so out-of-domain generalization can be measured directly: esbooks__eswikibooks, esbooks__libreriaunal, escomments__reddit, escomments__suicide, eslegal__entrevistas_comision_verdad, esmedical__BVS, esmedical__SPACC, esmedical__TEI_ES, esnews__eswikinews, estalks__TED.

Baselines

model	accuracy	macro F1
Majority class	0.250	0.250
XGBoost	0.350	0.348
BERTIN	0.663	0.664
XLM-RoBERTa	0.676	0.676

For reference, no prompted 7–8B LLM evaluated on this test set comes within 25 points of the XLM-RoBERTa baseline — see Flaglab/esnlir-llm-predictions.

Stress tests (XLM-RoBERTa, per-class accuracy)

test	contrasting	entailment	neutral	reasoning
test	0.664	0.676	0.695	0.667
test_length_mismatch	0.513	0.524	0.755	0.654
test_negation	0.580	0.575	0.744	0.656
test_overlap	0.624	0.637	0.719	0.665
test_spelling	0.687	0.646	0.677	0.659

Spelling noise is harmless, but padding the premise or negating the hypothesis costs contrasting and entailment about 15 points each while leaving neutral and reasoning alone — evidence that models lean on length and negation cues precisely where they should be reasoning semantically.

Known limitations

Connector labels are not human labels. Of 2,136 pairs annotated by 27 students, fewer than half kept a human majority label matching the connector-derived one. Treat connector_type as a distant-supervision signal, not ground truth. Any absolute number on the full test set carries that label noise.

Validation quality is genre-dependent. Scoring XLM-RoBERTa on validated data lifts comments from 0.647 to 0.719 but drops legal from 0.727 to 0.667 — lay annotators handle informal text well and formal legal text badly, so validation there removed correct labels. Domain-expert annotators are needed for reliable validation of formal genres.

neutral is constructed differently from the other three classes, which biases per-class comparisons in its favour.

Usage

from datasets import load_dataset

ds = load_dataset("Flaglab/ESNLIR-dataset")                          # train / validation / test
stress = load_dataset("Flaglab/ESNLIR-dataset", "stress_negation")   # one stress set

Or file by file, which is lighter if you only want the test split:

import json
from huggingface_hub import hf_hub_download

p = hf_hub_download("Flaglab/ESNLIR-dataset", "test.json", repo_type="dataset")
rows = [json.loads(l) for l in open(p, encoding="utf-8") if l.strip()]

Citation

@InProceedings{portela2025esnlirspanishmultigenredataset,
  author    = {Portela, Johan R. and P{\'e}rez-Ter{\'a}n, Nicol{\'a}s and Manrique, Rub{\'e}n},
  editor    = {Florez, Hector and Peluffo-Ordo{\~{n}}ez, Diego},
  title     = {{ESNLIR}: Expanding Spanish {NLI} Benchmarks with Multi-genre and Causal Annotation},
  booktitle = {Applied Informatics},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {345--361},
  isbn      = {978-3-032-07175-0},
  doi       = {10.1007/978-3-032-07175-0_23},
}