As the demand for inclusive and responsive AI grows, aligning Large Language Models (LLMs) with global human values becomes an urgent priority. The PRISM Alignment Project addresses this challenge by collecting feedback from 1,500 participants across 75 countries, capturing over 8,000 real-time conversations with 21 Large Language Models (LLMs). This groundbreaking dataset prioritizes inclusivity and personalization, addressing complex, value-laden topics like LGBTQ+ rights, politics, religion, or conflicts between nations to align AI systems with diverse global perspectives.
We Set a New Global Benchmark for AI Inclusivity and Ethical Practices
Factored led the implementation of a real-time conversation interface and assisted the deployment and maintenance of over 20 LLM’s with varying characteristics. The implementation required multiple iterations as different models were tested and the interface got polished. Our engineers provided support throughout the data collection, which took place between November and December 2023.
Shaping the Future of Equitable and Personalized AI Systems
The PRISM dataset has become a cornerstone for researchers worldwide, downloaded over 10,000 times on platforms like Hugging Face. Its findings highlight key variations in how users from different backgrounds interact with AI, influencing the optimization of models to create more equitable and personalized solutions.
Presenting at Neurips 2024
The project achieved significant recognition, including an oral presentation at NeurIPS 2024—one of the most prestigious AI conferences worldwide. Out of 1,820 submissions, only 11 papers (0.60%) were selected for an oral presentation, and this work was further distinguished by winning the Best Paper Award for the Datasets & Benchmarks Track, underscoring its groundbreaking impact and exceptional quality.
To see the full paper click here
