ChristianZ97/NuminaMath-LEAN-satp-gaps

NuminaMath-LEAN-satp-gaps 是一个 Lean 4 子目标数据集，从自然语言草稿到 Lean 草图，再到真实的 Lean 目标状态提取。它包含草图中的开放孔（sorry）以及在该孔处的确切 Lean 目标状态，适用于每个子目标的证明步骤训练。该数据集是主训练集 NuminaMath-LEAN-satp 的增强训练集补充，专注于通过草图分解暴露的内部子目标。数据集包含 uuid、formal_statement 和 goal_state 等特征，并提供了详细的模式、生成流程、来源归属、来源和限制以及引用信息。

NuminaMath-LEAN-satp-gaps is a Lean 4 sub-goal dataset harvested from natural-language drafts to Lean sketches, then to real-Lean goal-state extraction. It contains open holes (sorry) inside a sketch, paired with the exact Lean goal-state at that hole, suitable as a per-sub-goal prove-step training signal. This is the augmented training set complement to the main training set NuminaMath-LEAN-satp, focusing on the internal sub-goals exposed by sketch decomposition. The dataset includes features like uuid, formal_statement, and goal_state, and provides detailed schema, generation pipeline, source attribution, provenance, limitations, and citation information.