Initial release of ImageBiTe, a framework for evaluating representational harms in text-to-image models. Main features:
- Evaluation capabilities for stereotyping, under-representation and ex-nomination.
- Automatic test case generation, execution and evaluation.
- Prompt template library covering ethical concerns on: ethnicity, age, disability, and gender.
- Dimension dataset covering: occupations, personal and physical traits, socio-economic status, descriptors and activities.
- Connectors to models hosted on OpenAI API and HuggingFace Inference API.