Overview

This test evaluates the model's faithfulness to provided context, specifically focusing on its ability to handle challenging scenarios where the context might be incomplete, contradictory, or counterfactual. Ensuring faithfulness is crucial for the reliability of Retrieval-Augmented Generation (RAG) systems, as retrieved information can vary significantly in quality and may conflict with the model's internal knowledge or other retrieved documents. Unlike factuality tests that assess alignment with established world knowledge, FaithEval specifically measures whether the model's response strictly adheres to the given context, even when that context is flawed or contradicts common sense. Erroneous or unsupported information generated due to a lack of faithfulness can erode user trust and lead to severe consequences, particularly in high-stakes domains.

Dataset

The test utilizes the benchmark from FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows”, a dataset comprising 4.992 question-context pairs designed to probe contextual faithfulness across three distinct task types:

• Unanswerable Context. The provided context contains relevant details but lacks the specific information required to answer the question. A faithful model should recognize this limitation and abstain from answering, typically by responding with unknown or a similar indication.
• Inconsistent Context. The context includes multiple documents or passages that provide conflicting answers to the same question. This simulates scenarios with noisy retrieval from sources with varying credibility. A faithful model should identify the inconsistency and report the conflict, rather than arbitrarily choosing one answer.
• Counterfactual Context. The context contains statements that deliberately contradict common sense or widely accepted facts (e.g., stating that wood is magnetic). The model is expected to answer the question based solely on the provided counterfactual information, ignoring its internal knowledge about the real world. These scenarios test the model's ability to prioritize the immediate context over its parametric knowledge.

The benchmark was constructed using a four-stage framework involving LLM-based context generation and validation, supplemented by human annotation. The underlying data sources include established QA datasets including:

• SQuAD: 100,000+ Questions for Machine Comprehension of Text
• NewsQA: A Machine Comprehension Dataset
• TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension
• Natural Questions: A Benchmark for Question Answering Research
• SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine
• HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
• BioASQ-QA: A manually curated corpus for Biomedical Question Answering
• DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs
• RACE: Large-scale ReAding Comprehension Dataset From Examinations
• Are You Smarter Than a Sixth Grader? Textbook Question Answering for Multimodal Machine Comprehension
• ARC-Challenge

Evaluation