bold (bias in open-ended language generation)

**BOLD (Bias in Open-ended Language Generation Diversity)** is a benchmark designed to evaluate **social biases** in the **open-ended text generation** of language models. Unlike benchmarks that test classification or fill-in-the-blank, BOLD specifically measures biases in **free-form text generation** — the primary use case for modern LLMs. **How BOLD Works** - **Prompts**: The benchmark provides sentence **starters** drawn from Wikipedia articles about people from various demographic groups. For example: - "Marie Curie was a physicist who..." - "Barack Obama served as..." - **Generation**: The model completes each prompt with open-ended text generation. - **Evaluation**: Generated text is analyzed for **sentiment**, **toxicity**, **regard** (positive/negative portrayal), and other bias metrics using automated tools. **Demographic Categories** - **Race**: African American, European American, Hispanic/Latino, Asian American, Native American. - **Gender**: Male, Female. - **Religion**: Christianity, Islam, Judaism, Hinduism, Buddhism. - **Political Ideology**: Left-leaning, Right-leaning. - **Profession**: Various occupations. **Evaluation Metrics** - **Sentiment Analysis**: Is the generated text about certain groups more positive or negative than others? - **Toxicity Scores**: Does the model generate more toxic content when prompted about certain demographics? (Measured using Perspective API.) - **Regard Classifier**: Measures whether generated text portrays the demographic group positively, negatively, or neutrally. **Key Findings** - Models generate **more negative** and **more toxic** text when prompted about certain racial and religious groups. - Gender biases manifest as differences in topics and attributes associated with male vs. female subjects. BOLD is particularly valuable because it evaluates bias in the most natural LLM use case — **open-ended generation** — rather than artificial classification tasks.