Generating Zero-Shot Hard-Case Hallucinations: A Synthetic and Open Data Approach
Overview
Experience | In Person |
---|---|
Type | Lightning Talk |
Track | Artificial Intelligence |
Industry | Enterprise Technology |
Technologies | AI/BI, Llama, PyTorch |
Skill Level | Intermediate |
Duration | 20 min |
We present a novel framework for designing and inducing controlled hallucinations in long-form content generation by LLMs across diverse domains. The purpose is to create fully-synthetic benchmarks and mine hard cases for iterative refinement of zero-shot hallucination detectors. We will first demonstrate how Gretel Navigator can be used to design realistic, high-quality long-context datasets across various domains. Second, we will describe our reasoning-based approach to hard-case mining. Specifically, our methodology relies on chain-of-thought-based generation of both faithful and deceptive question-answer pairs based upon long-context samples. Subsequently, a consensus labeling and detector framework is employed to filter synthetic examples to zero-shot hard cases. The result of this process is a fully-automated system, operating under open data licenses such as Apache-2.0, for the generation of hallucinations at the edge-of-capabilities for a target LLM to detect.
Session Speakers
IMAGE COMING SOON
Eric Tramel
/Principal Research Scientist
Nvidia