Shout-out to my wonderful co-authors: Alina Leidinger, Yacine Jernite, Atoosa Kasirzadeh, Sasha Luccioni, @mmitchell.bsky.social
You can also read more about the project on TechCrunch (techcrunch.com/2024/06/06/s...huggingface.co/blog/giadap/...)
According to a new study, AI models hold opposing views on topics like LGBTQ+ rights depending on how they're trained -- and who's training them.
Good news: CIVICS is now available to the public! Access it here: huggingface.co/datasets/CIV...huggingface.co/spaces/CIVIC...
Perfect de-biasing is unattainable, but our research stresses the need for broader social impact evaluations beyond traditional metrics. We're eager to see what future research will do with datasets like this one!
The CIVICS dataset aims to foster AI development that respects global cultural diversities and value pluralism. We encourage further research in this crucial area by making the dataset and tools available under open licenses.
We also encountered significant variation in cultural bias among different open-weight models. Refusal to respond to prompts on LGBTQI rights and immigration varied widely, suggesting that models from diverse cultural contexts show varying sensitivity and ethical considerations.
Some key findings: beyond refusal rates, our experiments using CIVICS show diverse responses across LLMs on sensitive topics -- e.g., immigration, LGBTQI rights, and social welfare triggered varied reactions.
The dataset has undergone a dynamic annotation process from native speakers: annotators, co-authors of the research, applied multiple labels to each prompt, reflecting the diverse values inherent in the topics.
Spanning five languages (Turkish, German, Italian, French, English) and nine national contexts (Singapore, Canada, and Australia for English; France and Canada for French), CIVICS captures different cultural perspectives and reveals the diverse ethical views embedded in LLMs.
We designed a dataset to evaluate social and cultural variations in LLM responses. Hand-curated with value-laden prompts in multiple languages, it covers sensitive topics like LGBTQI rights, social welfare, immigration, disability rights, and surrogacy.