Hallucination Rate Benchmark

The benchmark must measure whether outputs are anchored to verifiable source documents rather than generated from parametric knowledge alone.

• Tests attribution accuracy by checking if claims map to provided context
• Uses Natural Language Inference (NLI) models to detect contradictions between output and source
• Penalizes intrinsic hallucinations (fabricated facts) and extrinsic hallucinations (facts not derivable from context)
• Example: A RAG system citing a specific clause from a contract that does not exist in the provided document

The benchmark must distinguish between verbatim mismatches and semantically equivalent paraphrases to avoid false positives.

• Employs entailment models (e.g., DeBERTa fine-tuned on MNLI) rather than n-gram overlap metrics like ROUGE
• Accounts for logical equivalence: 'The patient has hypertension' and 'The patient's blood pressure is elevated' should score as consistent
• Uses human-calibrated thresholds to map model confidence scores to hallucination probability
• Critical for avoiding penalizing valid rephrasing as hallucination

A robust benchmark classifies hallucinations by type and severity, not just frequency. Different domains have different tolerance profiles.

• Medical: Fabricated drug interactions are critical; stylistic variations are negligible
• Legal: Invented case citations are catastrophic; summarization brevity is acceptable
• Finance: Incorrect numerical values are high-severity; tone shifts are low-severity
• Taxonomy categories include: Entity Fabrication, Numerical Inaccuracy, Temporal Distortion, Causal Reversal, and Attribution Error

The benchmark must include deliberately challenging prompts designed to provoke hallucinations, not just standard queries.

• Counterfactual prompts: Asking about events that never occurred to test refusal vs. confabulation
• Ambiguous queries: Testing whether the model asks for clarification or guesses
• Long-context traps: Placing contradictory information far apart in the context window to test attention fidelity
• Multi-hop reasoning chains: Requiring the model to combine multiple facts without introducing spurious connections
• Example: 'What did Einstein say about quantum computing in his 1950 paper?' (He died before the term existed)

The benchmark must produce deterministic, repeatable results across different evaluation runs and environments.

• Fixes temperature to zero during evaluation to eliminate sampling variance
• Uses fixed random seeds and controlled decoding parameters (top-p, top-k disabled)
• Provides reference implementations of scoring scripts with versioned dependencies
• Reports confidence intervals (e.g., 95% CI via bootstrap resampling) rather than point estimates
• Enables regression testing to detect hallucination rate drift across model versions

The benchmark's ground truth must be established through rigorous human annotation with measured inter-annotator agreement.

• Uses multiple annotators per example (minimum 3) with domain expertise
• Reports Cohen's Kappa or Krippendorff's Alpha to quantify annotation reliability
• Includes adjudication protocols for resolving annotator disagreements
• Separates factual errors from stylistic preferences in annotation guidelines
• Example: The TruthfulQA benchmark uses human-validated question-answer pairs specifically designed to test misconceptions

A metric evaluating how faithfully a model's output is anchored to the provided source documents or verified facts. Unlike a general hallucination rate, a grounding score specifically measures attribution fidelity—whether each claim can be traced back to a specific passage in the context window.

• Factual Grounding: Measures if output contradicts known world knowledge
• Contextual Grounding: Measures if output is supported by the provided prompt context
• Often expressed as a percentage of sentences with a valid citation

A specialized evaluation score that quantifies the degree to which a generated response remains true to the source material without introducing extrinsic hallucinations. This metric is central to RAG system evaluation.

• Calculated by decomposing output into atomic claims
• Each claim is verified against the retrieved context
• A score of 1.0 indicates zero hallucinated content
• Critical for regulated industries requiring audit trails

A standardized adversarial dataset designed specifically to measure a model's propensity to generate false answers. It targets imitative falsehoods—incorrect answers that models learn from mimicking human text patterns.

• Contains 817 questions across 38 categories
• Questions are crafted to exploit common misconceptions
• Evaluates both truthfulness and informativeness
• Widely used in foundation model transparency reports

An automated or human evaluation process that verifies whether a generated summary or response contains information not present in the source document. This is the primary mechanism for detecting intrinsic hallucinations in summarization tasks.

• Entailment-based: Uses NLI models to check logical consistency
• QA-based: Generates questions from output and verifies answers against source
• Overlap-based: Measures n-gram overlap between output and source
• Essential for medical and legal document summarization

A classification framework that categorizes model errors into distinct types to enable precise benchmarking. The standard taxonomy distinguishes between:

• Intrinsic Hallucination: Output contradicts the provided source context
• Extrinsic Hallucination: Output cannot be verified from the source context
• Closed-domain: Factual errors in constrained tasks
• Open-domain: Fabricated facts about the real world
• Faithfulness vs. Factuality: Two orthogonal dimensions of truthfulness

A holistic evaluation framework that measures hallucination rates specifically within RAG architectures. It assesses the entire pipeline from retrieval quality to generation fidelity.

• Retrieval Precision: Relevance of fetched documents
• Generation Faithfulness: Output alignment with retrieved context
• Answer Relevance: How well the response addresses the query
• Citation Accuracy: Whether references point to correct sources
• Tools include RAGAS, TruLens, and ARES for automated assessment