minif2f-satp-v4.27

The 488-base-problem view of the miniF2F benchmark used by AlphaProof, packaged for Lean 4.27.0 with initial proof goal_state values generated for SATP v2.

This dataset is intended as the held-out evaluation (test) and hyperparameter-tuning (validation) benchmark for SATP / sketch-and-prove pipelines running in a Lean 4.27 environment.

Source

The single source of the problem statements is Google DeepMind miniF2F@f0a20e1 (MiniF2F/Valid.lean + MiniF2F/Test.lean). Its README identifies this as the benchmark version used to evaluate AlphaProof, and the repository targets Lean 4.27. Theorem declarations are taken byte-verbatim from those files, up to and including their := by proof entry.

• validation = the 244 base theorems of Valid.lean. The file additionally declares 12 *.variants.* restatements of base problems; variants are not separate problems and are not included.
• test = the 244 theorems of Test.lean.
• One statement with a term-mode proof (mathd_numbertheory_66) is closed with := by like every other row; proof bodies are never included.

Self-contained header

Upstream files import the answer( ) term elaborator and Real.nthRoot from google-deepmind/formal-conjectures (@f564f8d7, Apache-2.0). Each row's formal_statement therefore begins with a prelude that vendors those two definitions verbatim and applies the upstream files' open scoped Real Nat Topology Polynomial, followed by the untouched theorem bytes. Every row elaborates against plain import Mathlib — no external packages required.

Lean toolchain	leanprover/lean4:v4.27.0
Mathlib	a3a10db0e9d66acbebf76c5e6a135066525ac900
Goal-state renderer	Lean.Meta.ppGoal at proof entry, row-local environment

Schema

Column	Type	Notes
uuid	string	sha256(canonical(goal_state))[:16]; cross-dataset join key
formal_statement	string	prelude header + verbatim upstream theorem declaration ending in := by
goal_state	string	Pretty-printed Lean goal at proof entry: hypothesis lines + ⊢ + target

Goal-state provenance — read this first

goal_state is the exact (← Lean.Meta.ppGoal (← getMainGoal)).pretty rendering of the statement's initial goal, produced by elaborating the row's own formal_statement under the pinned toolchain. answer( ) terms elaborate to their literal values, so goals read as plain equalities. Consumers that key on bytes (encoders, uuid joins) can rely on this being identical to what a Lean tactic observes at proof entry in the same environment. Do not regenerate goals with a different renderer (e.g. the REPL's sorries[].goal) — notation expansion differs.

Related SATP datasets (same Lean 4.27 environment):

• NuminaMath-LEAN-satp-v4.27 — main training set
• NuminaMath-LEAN-satp-gaps-v4.27 — sketch-hole sub-goal augmentation set
• putnambench-satp-v4.27 — PutnamBench evaluation benchmark