Introducing VERITAS: A Unified Approach to Reliability Evaluation
Veritas is a suite of Reliability Judges for Batch and Real-time use cases
Large language models (LLMs) have revolutionized fields such as search, question answering, and natural language understanding. Trained on vast amounts of data, these models can generate impressively coherent and relevant responses. However, there is a major drawback: they sometimes produce factually incorrect information—typically known as hallucinations.
What Are Hallucinations in LLMs?
In the context of LLMs, hallucinations refer to instances where a model generates false or misleading content that seems plausible but is factually inaccurate. This is particularly concerning when using LLMs for knowledge-heavy tasks like fact-checking, summarization, or customer support. However, in creative contexts like storytelling or brainstorming, hallucinations can add value by introducing novel and imaginative ideas.
Hallucinations are especially common in closed-book settings, where the model has to rely solely on the knowledge it’s stored during training, without access to external sources. In contrast, in open-book settings, the model can pull information from external sources via approaches like Retrieval-Augmented Generation (RAG), which helps reduce hallucinations by grounding responses in real data. Yet, even in open-book settings, LLMs can still contradict the provided sources.
Tackling the Hallucination Problem
Researchers are actively working on understanding and fixing the hallucination issue. The causes range from flawed training data to biases in the information fed to the model. Several benchmarks have been designed to evaluate how well LLMs can distinguish fact from fiction, with human evaluators or LLMs acting as judges to assess content accuracy.
Current hallucination detection models specialize in identifying hallucinations in specific tasks such as question answering (QA) and natural language inference (NLI). However, these models don’t generalize well across various tasks or domains, highlighting the need for a more comprehensive solution.
VERITAS: A Unified Solution
To address this, we propose VERITAS that aims to unify hallucination detection across different formats, including QA, NLI, and dialogue. We present a family of hallucination detection models in different sizes (440M, 3B and 8B).
By trained on a high-quality diverse training dataset, VERITAS can detect hallucinations reliability across 19 different datasets and multiple task formats. This makes it a more versatile and robust solution for detecting hallucinations in LLM-generated content.
Key Results:
VERITAS demonstrates exceptional generalization across multiple task formats, outperforming single-task models like Minicheck DeBERTa [1], Bespoke MiniCheck 7B [2], and Lynx 8B [3]. While these models excel in either question answering (QA) or natural language inference (NLI), they struggle to generalize to unseen data formats, such as dialogue.
VERITAS Nano, our smallest model, performs competitively on LLMAggreFact [4] and effectively generalizes across all formats, unlike specialized models.
Our largest model, VERITAS 8B, competes closely with GPT-4, underscoring its effectiveness in detecting hallucinations across various formats.
Ensuring that LLMs are reliable is crucial, especially as these models are increasingly used in critical applications like legal advice, medical information, and customer service. Advances in fine-tuning techniques like RLAIF (Reinforcement Learning with AI Feedback) help models learn to prioritize factual accuracy, but this process depends heavily on the hallucination detection system’s ability to distinguish truth from fiction. VERITAS, by providing a cross-task framework for hallucination detection, is a major step toward making LLMs more dependable.
Try it out for yourself here !
Check out the full paper here!
References
At Collinear AI, we’re not just participants in the Enterprise LLM market—we’re setting new standards, ensuring our customers have access to an effective, scalable and automated feedback for their AI Applications.
Try it out for yourself -
https://app.collinear.ai/