A Use-Case Specific Dataset for Measuring Dimensions of Responsible Performance in LLM-generated Text

TL;DR

This work addresses the gap in responsible AI evaluation by introducing an application-driven dataset for measuring LLM-generated product descriptions. By aligning quality, veracity, safety, and fairness metrics with a realistic e-commerce use case—constructed from demographic-adjective and category templates—the study demonstrates how to reveal disparities across identity groups and product types. The accompanying analysis with Llama 3.2 illustrates practical metric behavior, cost-performance tradeoffs, and the potential for dataset-guided improvements in end-user experiences. The dataset, open under CC BY 4.0, provides a concrete resource for researchers to study RAI in context and to extend evaluations to additional modalities and languages.

Abstract

Current methods for evaluating large language models (LLMs) typically focus on high-level tasks such as text generation, without targeting a particular AI application. This approach is not sufficient for evaluating LLMs for Responsible AI dimensions like fairness, since protected attributes that are highly relevant in one application may be less relevant in another. In this work, we construct a dataset that is driven by a real-world application (generate a plain-text product description, given a list of product features), parameterized by fairness attributes intersected with gendered adjectives and product categories, yielding a rich set of labeled prompts. We show how to use the data to identify quality, veracity, safety, and fairness gaps in LLMs, contributing a proposal for LLM evaluation paired with a concrete resource for the research community.

Figures (1)

• Figure 1: Bar plots of toxicity from the detoxify unbiased model, showing overall toxicity (left) and sexual_explicit toxicity subtype. "any" is a wildcard value; "all" means all items in the dataset.