Benchmarking LLM math skills without leaking the test

Discussion: - Datasets: How do you prevent contamination? Create fresh test sets and keep private. - Methods: Chain-of-thought vs program-aided reasoning; how do you evaluate fairly? - Metrics: Exact answers, reasoning validity, and runtime. - Tools: Sympy, unit checkers, and hidden variants. Share your practices to get honest signals and avoid overfitting to well-known benchmarks.