Comprising synthetic logical reasoning traces categorized into correct and flawed classes for AI Mathematical Olympiad (AIMO) problems. It provides a collection of step-by-step mathematical proofs designed to help models identify specific points of failure in complex problem-solving.
Use Cases
- Train a classifier to detect logical inconsistencies in mathematical proofs using the 'Flawed' reasoning traces
- Develop a reward model for Reinforcement Learning from AI Feedback (RLAIF) using the paired correct and flawed traces
- Benchmark the reasoning capabilities of LLMs by requiring them to identify the specific error in 'Flawed' traces
Strengths
- Features synthetic reasoning traces for AI Mathematical Olympiad (AIMO) problems
- Includes binary labels distinguishing between 'Correct' and 'Flawed' logical steps
- Provides step-by-step chain-of-thought sequences for complex mathematical proofs