amritanshu05/hallucination-classifier-deberta

Hallucination Classifier — DeBERTa-v3-base

A fine-tuned DeBERTa-v3-base model for detecting hallucinations in LLM outputs. Classifies text as grounded or hallucinated given a claim and optional context.

Model Details

• Base model: microsoft/deberta-v3-base
• Task: Binary sequence classification
• Labels: grounded (0), hallucinated (1)
• Training data: HaluEval + TruthfulQA (~10K balanced samples)
• Author: Amritanshu Yadav (github.com/TechNxt05)

Performance (Test Set)

Metric	Score
Accuracy	0.9371
F1	0.9372
Precision	0.9365
Recall	0.9379

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="amritanshu05/hallucination-classifier-deberta"
)

result = classifier("The Eiffel Tower is located in Berlin. [SEP] Where is the Eiffel Tower?")
print(result)
# [{'label': 'hallucinated', 'score': 0.97}]

Integration

This model powers the hallucination detection layer in AuditAI — a production LLM observability platform.