Arabic NLI Binary Classifier — s05-mbert-nli

Binary Arabic text classifier (0 = faithful, 1 = unfaithful). Fine-tuned for Arabic NLI-based binary text classification.

Base model

google-bert/bert-base-multilingual-cased

Input format

nli — [CLS] gold_answer [SEP] model_answer [SEP] (NLI framing: gold_answer as premise, model_answer as hypothesis)

Dev results

• AUC-ROC (official, full dev n=1300): 0.9594
• AUC-ROC (clean dev, n=800, excludes ~100 questions also seen in train): 0.9310
• Macro F1 (official, threshold=0.50): 0.9171

Note: the official-dev number is inflated by 500 dev rows whose questions also appear in the training set (near-memorization). The clean-dev AUC-ROC (0.9310) is the honest generalization estimate and the number to use for ranking against other runs. It sits above the s01-camelbert-qa baseline (0.8713) and the s02/s03 NLI models (0.927), but below s04-arbert-nli (0.9408, current best single). As a multilingual model (vs the Arabic-specific CAMeLBERT/MARBERT/ARBERT), it adds architectural diversity to the ensemble and achieves the best dev F1-Macro so far (0.9171). All far exceed the published CAMeLBERT baseline (0.7093 dev AUC-ROC).

Training data

Arabic training set — 4,705 Arabic (question, gold_answer, model_answer) triples, 5 source LLMs, 13 knowledge domains.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("HassanB4/s05-mbert-nli")
model = AutoModelForSequenceClassification.from_pretrained("HassanB4/s05-mbert-nli")

inputs = tokenizer(gold_answer, model_answer, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    logits = model(**inputs).logits
score = torch.softmax(logits, dim=-1)[0][1].item()  # unfaithfulness score
predicted_label = int(score > 0.5)