Strengthening Access to Education Data

Standardized test data across the United States is vast, but its inconsistent formats and scattering across multiple sources have made it very difficult for parents or policymakers to use. Zelma is a research assistant that leverages GPT-4 to help anyone easily query and analyze this education data in everyday language.

It is built on data curated by a Brown University research team and was implemented via the OpenAI API in collaboration with Novy. Technically, it uses function calling to select the appropriate visualization method for the data, provides a question recommendation feature using a fine-tuninged model, and improves answer accuracy by leveraging verified example graphs stored in a vector database.

To optimize the user experience, it offers the following key features:

• Example prompt suggestions: Guide users to understand the scope of Zelma's knowledge and ask appropriate questions.
• Public question system: Helps learning through other users' questions and prevents indiscriminate queries.
• Public SQL code: Allows users to directly verify the logical structure of answers to check reliability.
• Contextual explanations: Provides data definitions and key events (e.g., changes in evaluation methods) together, helping users understand the reasons behind changes in the data.

Through this, school boards can visualize data on the fly during meetings, parents can compare academic achievement by region, and policymakers can quickly grasp education outcomes, enabling data-driven decision-making.