XLM-RoBERTa fine-tuned on ESNLIR through pool-based active learning with the LER acquisition strategy. This is the best-performing round: 800,000 labels (80% of the 1M pool), selected over 8 of 8 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	0.795	0.789	0.787	0.777	0.846	0.746
full supervision (`Flaglab/ESNLIR-XLM-RoBERTa`)	—	0.822

96.0% of full-supervision macro F1 using 80% 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

Low-Energy Removal (LER) filters rather than ranks: each round it permanently discards the 25,000 lowest-energy instances — the ones the model already handles — then draws the acquisition batch uniformly at random from the survivors. Energy decides only what is removed, never what is selected, which makes LER a near-random acquisition control.

In the code this strategy is named Rem.

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-XLM-RoBERTa-LER"
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 FacebookAI/xlm-roberta-base; its vocabulary matches this model's embedding table exactly.

Training


base model	`FacebookAI/xlm-roberta-base`
pool	`Flaglab/ESNLIR-dataset` train split
seed set	10,000 (balanced)
acquisition size	100,000 per round
rounds	8 total; this checkpoint is round 7
labels used	800,000
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.