cornellbgl/SNAP_scenarios

This is a detailed 500-question benchmarking dataset for measuring LLM or human-in-the-loop accuracy on realistic scenarios related to the U.S. Supplemental Nutrition Assistance Program (SNAP) in California as well as other states (use answer_depends_state column as a starting point for using the data for other states). Questions were generated based on frequencies of types of errors from publicly available national quality control data, which pertains to the accuracy of SNAP benefit amounts. The questions contain their topics to allow for filtering to specific topics of interest. We generated questions using these error frequencies and templated questions written by SNAP quality control auditors. We then had the quality control auditors select these questions from the over 3,400 that we generated for realism and representativeness. We worked with them to edit the questions for clarity and appropriateness. At least two auditors reviewed the answer for every question.

Demographic and circumstantial characteristics allow for generating thousands of related questions and the _fixed columns indicate which variations are possible without changing the correct answer. The data will be made available to government agencies, organizations supporting government agencies, and to researchers.

The benchmark dataset appears in LLMs in social services: How does chatbot accuracy affect human accuracy? Jennah Gosciak (Cornell Tech), Eric Giannella (Georgetown University, Better Government Lab), Zhaowen Guo (Georgetown University, Better Government Lab), Michael Chen (Nava Labs), Allison Koenecke (Cornell Tech)

Data is provided as a parquet, csv, and json file.

Note that there is a held-out 270-question dataset that can be obtained as a secondary request in case there are concerns about benchmark contamination.