Bertin-RoBERTa fine-tuned on ESNLIR through pool-based active learning with the NegE acquisition strategy. This is the best-performing round: 710,000 labels (71% of the 1M pool), selected over 7 of 10 acquisition rounds.

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)

Code: jd-rodriguezp1234/esnlir-active-learning

Results

On the 1,695-pair human-annotated test set:

	accuracy	macro F1	contrasting	entailment	neutral	reasoning
this model	0.771	0.764	0.749	0.767	0.819	0.722
full supervision (`Flaglab/ESNLIR-RoBERTa`)	—	0.811

94.2% of full-supervision macro F1 using 71% of the labels. The gap to full supervision is never closed by any AL configuration in this study.

Classes are uneven in the test set (neutral 37.4%, contrasting 18.4%), so macro F1 is the primary metric.

Strategy

Negative Energy (NegE) ranks the unlabeled pool by energy $-\log\sum_k \exp(z_k)$ and acquires the highest-scoring instances — those to which the model assigns no class a large logit. Unlike softmax entropy, energy is not distorted by probability normalization.

Both strategies perform indistinguishably (within 0.7 points of each other for a given encoder). Since LER acquires uniformly at random from a filtered pool, energy-based acquisition provides no measurable advantage over near-random selection — the paper's central negative result.

Labels

id	label
0	`contrasting`
1	`entailment`
2	`neutral`
3	`reasoning`

reasoning (cause–effect) is ESNLIR's departure from the usual three-class schema and is the hardest class for every model here.

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

name = "Flaglab/ESNLIR-AL-BERTIN-NegE"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForSequenceClassification.from_pretrained(name)

inputs = tok("se pueden retener 10 pasaportes alemanes al mismo tiempo",
             "estos pasaportes son válidos por solo 6 años",
             return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits
print(model.config.id2label[int(logits.argmax(-1))])

The bundled tokenizer is an unmodified copy of the one from bertin-project/bertin-roberta-base-spanish; its vocabulary matches this model's embedding table exactly.

Training


base model	`bertin-project/bertin-roberta-base-spanish`
pool	`Flaglab/ESNLIR-dataset` train split
seed set	10,000 (balanced)
acquisition size	100,000 per round
rounds	10 total; this checkpoint is round 6
labels used	710,000
learning rate	2e-5
batch size	32
epochs per round	3
max sequence length	256
warm start	yes (model carried across rounds)
class weighting	balanced

Full round-by-round metrics and the exact pool indices acquired at each step are published at Flaglab/esnlir-al-trajectories.

Citation

bibtex

@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.