halluscoring-mbert-nli

Multilingual BERT (google-bert/bert-base-multilingual-cased) fine-tuned on HalluScoring 2026 Task 1.1 using the NLI framing ([CLS] gold_answer [SEP] model_answer [SEP]). Internally this is run S05 — our best F1-Macro among single models (0.9171), despite being the only non-Arabic-specific architecture tried. Its predictions add useful architectural diversity and are a component of our best-ever ensemble, S25v2.

Not submitted to the competition — kept as an internal experiment and as a component of the S25v/S25v2 ensembles. See SYSTEM_WRITEUP.md for the officially-submitted models.

How to Use

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "HassanB4/halluscoring-mbert-nli"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

gold_answer = "..."
model_answer = "..."

inputs = tokenizer(gold_answer, model_answer, truncation=True, max_length=512, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
    prob_hallucinated = torch.softmax(logits, dim=-1)[0, 1].item()

print(f"hallucinated={int(prob_hallucinated > 0.5)}, score={prob_hallucinated:.4f}")

Training

Parameter	Value
Base model	google-bert/bert-base-multilingual-cased
Input format	nli (gold_answer + model_answer)
Max sequence length	512
Batch size	16
Epochs	5
Learning rate	2e-5
Warmup ratio	0.1
Weight decay	0.01
Loss	cross-entropy
Seed	42

Evaluation

Split	AUC-ROC	F1-Macro
Dev (official, n=1300)	0.9594	0.9171
Dev (clean, unseen-question subset, n=800)	0.9310	—

Limitations

Not evaluated on the hidden test set — internal dev-only experiment. Lower clean-dev AUC-ROC than halluscoring-arbert-nli despite the higher F1-Macro; used in ensembles primarily for architectural diversity.

Citation

@inproceedings{namaa2026halluscoring,
    title={{NAMAA at HalluScoring 2026: NLI-Framed BERT Classifiers and Ensembling for Model-Agnostic Arabic Hallucination Detection}},
    author={[AUTHOR NAMES TBD]},
    year={2026},
    booktitle={Proceedings of ArabicNLP 2026},
    note={HalluScoring 2026 Shared Task, Track 1}
}