Flaglab/esnlir-fewshot-pool

ESNLIR — few-shot example pool (es_wikibooks)

144 human-validated sentence pairs from the es_wikibooks source of ESNLIR, used only as the pool to draw few-shot examples from in An Analysis of the Performance of Large Language Models in Spanish NLI Datasets with Causal Relationships (IBERAMIA 2026, to appear). Part of the ESNLIR-LLM collection, used in Pacolas/NLI-via-LLM.

Validated by two annotators; only pairs where both assigned the same label as the dataset were kept.

Why it is separate

This pool is kept apart from the evaluation splits so the few-shot examples cannot leak labels into the evaluation. It is disjoint from both: 0 pairs overlap with Flaglab/esnlir-human-validated, and the 2 pairs that overlap with Flaglab/esnlir-test are discarded at sampling time.

Format

JSON Lines — one object per line, three fields only. The gold label is connector_type.

{"sentence_1": "algunos dicen una joya preciosa, y otros dicen una ventana",
 "sentence_2": "como dijimos, hay dos niveles",
 "connector_type": "neutral"}

Label distribution

label	n
neutral	43
reasoning	42
entailment	30
contrasting	29

Note: the file has 144 rows but 133 unique (sentence_1, sentence_2) pairs — 11 are repeats. Sampling is by row, so a duplicated pair is slightly likelier to be drawn. With the default FEW_SHOT_K this is immaterial, but deduplicate first if you need exact per-pair sampling weights.

Usage

from huggingface_hub import hf_hub_download
import json

p = hf_hub_download("Flaglab/esnlir-fewshot-pool",
                    "es_wikibooks_validated.jsonl", repo_type="dataset")
rows = [json.loads(l) for l in open(p) if l.strip()]

Citation

These splits were packaged for the following paper:

@InProceedings{perez2026llmspanishnlicausal,
  author    = {P{\'e}rez, Nicol{\'a}s and Portela, Johan R. and Manrique, Ruben},
  title     = {An Analysis of the Performance of Large Language Models in Spanish
               NLI Datasets with Causal Relationships},
  booktitle = {Advances in Artificial Intelligence -- IBERAMIA 2026},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  note      = {To appear},
}

The data itself is ESNLIR, released here under CC BY 4.0 — attribution to the source corpus is a condition of that licence:

@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},
}