This research fine-tunes smaller, locally-deployable language models with synthetic data to improve physical common sense reasoning during disaster scenarios.

• Created FRIDA (Field Ready Instruction Dataset), a synthetic dataset for earthquake response scenarios
• Demonstrated that synthetic data can effectively improve smaller models' performance in specialized domains
• Achieved significant improvements in physical reasoning capabilities with limited training resources
• Established a repeatable pipeline for creating domain-specific instruction datasets

This work addresses critical security challenges by enabling more capable AI assistants that can operate locally during disasters when connectivity and power may be limited, potentially saving lives through better emergency response.

FRIDA to the Rescue! Analyzing Synthetic Data Effectiveness in Object-Based Common Sense Reasoning for Disaster Response