Arabic NLI Binary Classifier — s02-camelbert-nli

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

Base model

CAMeL-Lab/bert-base-arabic-camelbert-mix

Input format

nli — [CLS] gold_answer [SEP] model_answer [SEP] (NLI framing: gold_answer as premise, model_answer as hypothesis — closes the gap left by the official qa baseline format, which ignores gold_answer entirely)

Dev results

• AUC-ROC (official, full dev n=1300): 0.9574
• AUC-ROC (clean dev, n=800, excludes ~100 questions also seen in train): 0.9272
• Macro F1 (official, threshold=0.50): 0.9081

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 (0.9272) is the honest generalization estimate and the number to use for ranking against other runs. It improves on s01-camelbert-qa's clean-dev AUC-ROC of 0.8713 by +5.6pp, confirming that NLI framing (using gold_answer as the premise) substantially closes the gap left by the qa format. Both already 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/s02-camelbert-nli")
model = AutoModelForSequenceClassification.from_pretrained("HassanB4/s02-camelbert-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)