How to measure model hallucination rate using NLI (Natural Language Inference) entailment models?

Short answer: Measuring hallucination using Natural Language Inference (NLI) fine-tunes small classification models (e.g. DeBERTa-v3) to score whether generated output statements are logically entailed by (Entailment), neutral to (Neutral), or contradicted by (Contradiction) retrieved source context passages.

Detail

LLM-as-a-Judge evaluations using GPT-4o are accurate but expensive for high-volume production monitoring.

NLI cross-encoder models evaluate hallucination deterministically at $100\times$ lower latency and cost.

Premise (Retrieved Context): "Company revenue grew 15% in Q3 to $5M."
Hypothesis (LLM Output):    "Company revenue was $5M in Q3."
                                        │
                                        ▼
                  [DeBERTa NLI Cross-Encoder Classifier]
                                        │
                                        ▼
         Probabilities: [Entailment: 0.98, Neutral: 0.01, Contradiction: 0.01]
         ──► Status: GROUNDED (Zero Hallucination)

NLI Classification Categories

• Entailment ($P_{entail}$): The premise logically proves the hypothesis is true.
• Neutral ($P_{neutral}$): The hypothesis introduces extra un-verifiable details not in the premise.
• Contradiction ($P_{contra}$): The hypothesis directly contradicts the premise (Hallucination!).

Example

PyTorch Transformers NLI entailment evaluation snippet:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("cross-encoder/nli-deberta-v3-base")
model = AutoModelForSequenceClassification.from_pretrained("cross-encoder/nli-deberta-v3-base")

premise = "The refund policy permits returns within 14 days."
hypothesis = "Customers have 30 days to return items." # Contradiction!

inputs = tokenizer(premise, hypothesis, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)

# Labels: [0: contradiction, 1: entailment, 2: neutral]
print(f"Contradiction Probability: {probs[0][0].item():.4f}") # High contradiction -> Hallucination!

Interview tips

• Highlight that NLI models provide fast, objective, reproducible hallucination scores without prompt engineering variability.
• Explain claim-level decomposition: splitting long LLM responses into atomic sentence claims before running NLI checks.

Related Concepts

• [[How does Self-RAG train models to dynamically decide when to retrieve, evaluate, and critique documents?]] (#130): How does Self-RAG train models to dynamically decide when to retrieve, evaluate, and critique documents?
• [[How to build an automated RAG evaluation harness measuring Context Precision and Recall?]] (#178): How to build an automated RAG evaluation harness measuring Context Precision and Recall?

⬅ Back to Evaluation and Testing · All topics