To overcome these challenges, we prioritize the representativeness of our training data, aligning it with the global population it serves. Our approach involved creating a diverse dataset with a minimum of 50 distinct demographic groups, ensuring comprehensive coverage. For each of these groups, we meticulously collected and profiled hundreds of individuals from publicly available sources. Drawing insights from social media content, biographies, and personality forums, we established rigorous and validated personality labels.
In parallel, for our text model, we followed a similar strategy. We amassed thousands of publicly accessible social media posts authored by individuals spanning various demographics, specifically focusing on English language content to train our text model. This model, based on natural language processing (NLP), uses linguistic patterns, analyzing word usage across different individuals and geographies to infer the most probable personality traits.