Hallucination Detection

# pip install transformers torch spacy
# python -m spacy download en_core_web_sm
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch
import spacy
from typing import TypedDict

class ClaimResult(TypedDict):
    claim: str
    label: str          # ENTAILMENT | NEUTRAL | CONTRADICTION
    confidence: float
    is_hallucination: bool

# Use DeBERTa-v3 (not BART-MNLI) — purpose-built for NLI,
# consistently outperforms BART on MNLI and ANLI benchmarks
MODEL = "cross-encoder/nli-deberta-v3-base"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForSequenceClassification.from_pretrained(MODEL)
nlp = spacy.load("en_core_web_sm")

LABELS = ["CONTRADICTION", "ENTAILMENT", "NEUTRAL"]
HALLUCINATION_THRESHOLD = 0.75  # flag if contradiction confidence > 75%

def decompose_claims(text: str) -> list[str]:
    """Split text into individual claims using spaCy sentence segmentation.
    Each sentence is treated as a separate claim to verify."""
    doc = nlp(text)
    return [sent.text.strip() for sent in doc.sents if len(sent.text.strip()) > 10]

def check_claims_batch(
    source: str,
    claims: list[str],
    threshold: float = HALLUCINATION_THRESHOLD
) -> list[ClaimResult]:
    """Batch-verify multiple claims against a source document.
    Returns per-claim verdicts with confidence scores."""
    results: list[ClaimResult] = []

    # Tokenize all pairs at once for efficient GPU batching
    pairs = [(source, claim) for claim in claims]
    encoded = tokenizer.batch_encode_plus(
        pairs,
        padding=True,
        truncation=True,
        max_length=512,
        return_tensors="pt"
    )

    with torch.no_grad():
        logits = model(**encoded).logits
        probs = torch.softmax(logits, dim=-1)

    for i, claim in enumerate(claims):
        label_idx = probs[i].argmax().item()
        label = LABELS[label_idx]
        confidence = probs[i][label_idx].item()

        results.append({
            "claim": claim,
            "label": label,
            "confidence": round(confidence, 3),
            "is_hallucination": (
                label == "CONTRADICTION" and confidence > threshold
            ),
        })

    return results

# --- Example usage ---
source = """TechCorp reported Q3 2024 revenue of $45.2 million,
a 12% increase year-over-year. Founded in 2018, headquarters
in Austin, Texas. Employees: 342."""

generated = """TechCorp had revenue of $45.2 million in Q3 2024,
growing 15% from last year. The Austin-based company now employs
over 500 people. CEO John Martinez praised the results."""

claims = decompose_claims(generated)
results = check_claims_batch(source, claims)

for r in results:
    flag = " ** HALLUCINATION **" if r["is_hallucination"] else ""
    print(f"[{r['label']} {r['confidence']:.0%}] {r['claim']}{flag}")

# Output:
# [ENTAILMENT 0.94] TechCorp had revenue of $45.2 million in Q3 2024,
#   growing 15% from last year.
# [CONTRADICTION 0.89] The Austin-based company now employs over 500
#   people. ** HALLUCINATION **
# [NEUTRAL 0.82] CEO John Martinez praised the results.