Flaglab/esnlir-annotation-candidates

ESNLIR — confidence-stratified annotation candidates

The 2,664 candidate pairs sent to human annotators, each carrying the model confidence score and the stratum it was drawn from. This is the pool before annotation; the labelled result is Flaglab/esnlir-al-annotated-test.

Active Learning for Spanish Natural Language Inference on a Heterogeneous Multi-Domain Corpus Diego Ortiz, Johan R. Portela, Ruben Manrique — Universidad de los Andes, Bogotá Advances in Artificial Intelligence — IBERAMIA 2026 (to appear)

Selection code: jd-rodriguezp1234/esnlir-active-learning (notebooks/annotation/)

Why this exists

Annotating uniformly at random wastes effort on instances every model already gets right. Candidates were instead stratified by model confidence, so annotation concentrated where the labels were most informative. Releasing the pool lets you see exactly what was offered for labelling, and compare it against what survived — useful for anyone studying selection effects in evaluation sets.

The strata

Confidence is the mean maximum softmax probability across both ESNLIR baselines — Flaglab/ESNLIR-RoBERTa and Flaglab/ESNLIR-XLM-RoBERTa — not a single model's score.

confidence_config	criterion	n	confidence range	mean
high	both models correct	1,856	0.965 – 0.999	0.991
medium	both models correct	284	0.600 – 0.749	0.688
low	both models correct	248	0.298 – 0.494	0.398
no	not both correct	276	0.337 – 0.983	0.626

no is not a confidence band. It is defined by the models being wrong, so its confidence spans almost the entire interval — a pair can sit here at 0.98, meaning the models were confidently mistaken. Those are the most diagnostic instances in the pool, and they are invisible to any scheme that stratifies on confidence alone.

Labels are balanced at 666 per class across the four relations.

genre	n		genre	n
theses	852		articles	256
books	620		legal	208
comments	380		talks	8
news	340

Fields

Standard ESNLIR fields plus two:

field	meaning
confidence	mean max-softmax across the two baselines, in [0, 1]
confidence_config	the stratum: high, medium, low or no

connector_type is the connector-derived label used as ground truth when deciding whether the models were correct. There is no human label in this file — that is the point; these are pre-annotation candidates.

What happened next

2,664 candidates → annotated by 12 annotators, 3 labels per instance via overlapping assignment windows → majority vote → retained only where the majority matched connector_type → 151 disagreements discarded → 1,695 released as Flaglab/esnlir-al-annotated-test.

Not every candidate was annotated; the labelled subset is smaller than this pool. Comparing the two files shows both which candidates reached annotators and which survived the agreement filter.

Usage

from datasets import load_dataset

pool = load_dataset("Flaglab/esnlir-annotation-candidates", split="train")

hard = pool.filter(lambda r: r["confidence_config"] == "no")
print(len(hard), "instances both baselines got wrong")

confident_and_wrong = pool.filter(
    lambda r: r["confidence_config"] == "no" and r["confidence"] > 0.9)
print(len(confident_and_wrong), "confidently wrong")

Citation

@InProceedings{ortiz2026activelearningspanishnli,
  author    = {Ortiz, Diego and Portela, Johan R. and Manrique, Ruben},
  title     = {Active Learning for Spanish Natural Language Inference
               on a Heterogeneous Multi-Domain Corpus},
  booktitle = {Advances in Artificial Intelligence -- IBERAMIA 2026},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  note      = {To appear},
}

The corpus is ESNLIR (Portela, Pérez-Terán & Manrique, 2026), a separate project with its own code and paper.