Search papers, labs, and topics across Lattice.
This paper investigates the limitations of existing AI alignment metrics when evaluating fine-tuned LLMs in industry-specific contexts, particularly within the Banking BIAN domain. It evaluates metrics like Correctness, Faithfulness, and Harmfulness on QnA data tuples and LLM responses after fine-tuning, highlighting their shortcomings in providing actionable insights. The study proposes a dual approach integrating general and domain-specific evaluation methodologies to address these gaps and improve the quality and fairness of QA pair generation.
Off-the-shelf AI alignment metrics can fail spectacularly when evaluating fine-tuned LLMs in real-world industry applications, demanding a more nuanced, domain-aware approach.
Evaluating Large Language Models (LLMs) for AI alignment necessitates methodologies that go beyond general-purpose benchmarks to address domain-specific challenges and ethical complexities. This study investigates alignment metrics tailored to industry-specific contexts, utilizing large datasets from subject matter experts and synthetic data to fine-tune LLMs. Existing metrics, when used out-of-the-box, often fail to offer actionable insights or maintain relevance in real-world applications. To mitigate this, we evaluate a range of AI alignment metrics, including Correctness, Faithfulness, Completeness, Conciseness, Harmfulness, and Maliciousness for fine-tuning QnA data tuples and fine-tuned LLM's responses after training. Furthermore, we address issues of disparate impacts and historical biases to improve the quality and fairness of QA pair generation. Grounded in a theoretical framework, we propose that AI alignment requires a dual approach, integrating general and domain-specific evaluation methodologies. Experimental results from fine-tuning LLMs in the Banking BIAN domain reveal significant shortcomings in existing frameworks. In this paper we propose an approach to emphasize and mitigate these gaps, ensuring nuanced and ethically grounded evaluation practices. This work advances alignment methodologies aimed at fostering transparency, trustworthiness, and responsible LLM deployment in high-stakes, regulated domains.
