What is the purpose of data collection?

In this lesson, we analyze the importance of data collection in evaluating the behavior of language models. These evaluations are conducted using benchmarks, test sets designed to measure performance and, in particular, detect bias. Bias benchmarks function as controlled experiments containing stereotyped sentences and scenarios to observe how the model responds. However, much of the current research focuses on English, which means metrics, datasets, and mitigation strategies are rooted in specific cultural contexts and do not always represent realities like those in Latin America.

Within this framework, we present the HESEIA 2024 Project, an educational initiative involving teachers and students in Argentina that combined critical AI literacy with collaborative data production. Using the EDIA tool, participants explored biases related to social class, gender, and nationality, building a dataset of nearly 50,000 sentences grounded in local and intersectional contexts. Results showed that many regional stereotypes go undetected by current models. HESEIA thus serves as both a pedagogical and technological intervention: fostering critical thinking while contributing to the creation of fairer, more diverse, and more representative benchmarks.

At the end of the lesson, you are invited to answer a few questions. 

👨‍🏫 Instructor: Guido Ivetta