NiviG/hallucination-detector

LLM Hallucination Detector Model

This is a fine-tuned version of Microsoft's DeBERTa-v3-base trained on 25,000+ examples of the FEVER NLI dataset.

It is designed to evaluate whether an AI-generated statement (Claim) is supported, contradicted, or unresolvable based on a provided context paragraph (Evidence).

Output Labels

• —LABEL_0 (FACTUAL): The claim is supported by the evidence.
• —LABEL_1 (UNCERTAIN): The evidence is neutral or does not contain enough information to verify the claim.
• —LABEL_2 (HALLUCINATION): The claim is refuted/contradicted by the evidence.

Hyperparameters

• —Epochs: 3
• —Batch Size: 8
• —Learning Rate: 2e-5
• —Optimizer: AdamW