s1: Simple Test-Time Scaling

A 1,000-example SFT set (s1K) plus "budget forcing" — appending "Wait" to extend thinking — lifts Qwen2.5-32B to o1-preview-level math (6.6K★). With Muennighoff, Hashimoto, Liang, and Fei-Fei Li; launched the open test-time-scaling wave of 2025 alongside the lab's earlier "Large Language Monkeys" repeated- sampling result.