athanor-ai/formal-anytime-valid-stats

Formal-AVS是一个用于随时有效置信序列（CS）定理证明的Lean-4基准测试数据集。它包含60个定理，覆盖了四个CS家族（Howard-Ramdas、betting、Whitehouse vector、asymptotic CLT），并分为六个能力层级T0到T5。该数据集主要用于测试大型语言模型（LLMs）是否能生成或完成此类证明，特别是在数学库（Mathlib）和统计学习（kairos-stats-lean）的上下文中。数据集还包含详细的评估协议、层级结构、主要结果和引用信息。

Formal-AVS is a Lean-4 benchmark for anytime-valid confidence-sequence (CS) theorem proving. It includes 60 theorems across four CS families (Howard-Ramdas, betting, Whitehouse vector, asymptotic CLT), grouped into six capability tiers T0 to T5. The dataset is designed to test whether large language models (LLMs) can produce or complete such proofs, particularly in the context of Mathlib and kairos-stats-lean. It also provides detailed evaluation protocols, tier structures, headline results, and citation information.