Flaglab/esnlir-al-trajectories

ESNLIR — active learning trajectories

The complete round-by-round record of four active-learning runs on ESNLIR: which pool instances were acquired, which were discarded, and how validation performance moved after every acquisition.

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

With these you can replay any run's label budget exactly, audit what each strategy selected, or re-derive the learning curves without a GPU.

Start here: learning_curves.csv

One row per run per round, 36 rows total:

column	meaning
encoder	xlm-roberta-base or bertin-roberta-base-spanish
strategy	NegE (Negative Energy) or LER (Low-Energy Removal)
iteration	acquisition round, 0-indexed
acquired	instances labeled this round
removed_from_pool	instances discarded this round (LER only; 0 for NegE)
n_labeled	cumulative labeled set size, seed set included
val_accuracy, val_macro_f1, val_loss	evaluation after retraining this round
is_best_iteration	the checkpoint released on the Hub

Validation here is the ESNLIR validation split (connector labels), which is why these numbers sit near 0.62–0.66 while the paper's headline figures — 0.75–0.80 — are macro F1 on the human-annotated test set. Different evaluation sets; do not compare them directly.

What the curves show

Both encoders peak mid-run and then decline:

run	best round	labels at best	val macro F1 at best	val macro F1 at final round
XLM-R + NegE	5	610,000	0.6590	0.6520 (round 9)
XLM-R + LER	7	800,000	0.6534	0.6534 (round 7, is final)
Bertin + NegE	6	710,000	0.6248	0.6152 (round 9)
Bertin + LER	7	800,000	0.6201	0.6201 (round 7, is final)

XLM-R with NegE gains through round 5 and then loses ground for four consecutive rounds while continuing to acquire 100k labels each time — 390,000 additional labels that make the model slightly worse. That is the paper's label-efficiency result visible directly in the raw metrics.

NegE runs 10 rounds to exhaust the 1M pool; LER runs 8, because it also discards 25,000 instances per round. The final round of every run acquires 90,000 rather than 100,000 — the pool runs out.

Files

xlm-roberta-base_NegE/                    bertin-roberta-base-spanish_NegE/
xlm-roberta-base_LER/                     bertin-roberta-base-spanish_LER/
learning_curves.csv

Each run directory holds, per round N:

file	contents
metrics_iter_N.json	validation loss / accuracy / macro F1 after retraining
selected_indices_iter_N.csv	pool indices acquired this round
removed_indices_iter_N.csv	pool indices discarded this round — LER only
val_report_iter_N.csv	per-class precision / recall / F1 on a 2,000-instance validation sample (500 per class)
best_iteration.json	which round produced the released checkpoint

Note the two metric files are computed on different evaluation sets: metrics_iter_N.json uses the full 67,400-row validation split, while val_report_iter_N.csv uses a balanced 2,000-instance sample. Their aggregate numbers therefore will not match exactly.

Indices refer to row positions in the ESNLIR train split, Flaglab/ESNLIR-dataset — needed to reconstruct any labeled set.

Usage

import pandas as pd
from huggingface_hub import hf_hub_download

curves = pd.read_csv(hf_hub_download("Flaglab/esnlir-al-trajectories",
                                     "learning_curves.csv", repo_type="dataset"))

# the peak-then-decline, per run
for (enc, strat), g in curves.groupby(["encoder", "strategy"]):
    peak = g.loc[g.val_macro_f1.idxmax()]
    print(f"{enc:30s} {strat:5s} peaks at round {peak.iteration} "
          f"({peak.n_labeled:,} labels), ends at {g.val_macro_f1.iloc[-1]:.4f}")

Rebuilding the labeled set as it stood after round t:

import pandas as pd
from huggingface_hub import hf_hub_download

idx = pd.concat([
    pd.read_csv(hf_hub_download("Flaglab/esnlir-al-trajectories",
                                f"xlm-roberta-base_NegE/selected_indices_iter_{i}.csv",
                                repo_type="dataset"))
    for i in range(6)                      # rounds 0..5 = the best checkpoint's budget
])
print(len(idx), "acquired + 10,000 seed = 610,000")

Related

Flaglab/esnlir-al-annotated-test	the 1,695-pair evaluation set
Flaglab/ESNLIR-dataset	the pool these indices refer to
Flaglab/ESNLIR-AL-XLM-RoBERTa-NegE	the checkpoint at the best round

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},
}