ORBIT’s 20,000 answer-first riddles for training search agents
Weaviate Podcast describes a dataset that starts with short, verifiable answers and wraps each in four or five clues. Solving a question properly means checking every clue, one search at a time.
TLDR
Weaviate Podcast says Nandan Thakur and colleagues at the University of Waterloo built ORBIT, a 20,000-question dataset for training search agents. Each riddle works backward from a short, verifiable answer, adding four or five clues that narrow the field. The podcast says external search agents then re-verified the answers to avoid training on incorrect ground truth.
Combined views
—
1 Source, first seen ago
