EPFL-LARA/lean-evaluation-benchmark

Evaluation benchmark

The paper's evaluation benchmark for learned Lean 4 evaluators contains 68,932 deduplicated candidate programs from three public provers. Each row contains a complete source program and the acceptance label obtained with Lean v4.15.0.

Subset	Programs	Valid
miniF2F	41,638	25,702 (61.7%)
ProofNet	27,294	2,443 (9.0%)
Combined	68,932	28,145 (40.8%)

The remaining 40,787 programs are invalid. Candidates come from Pythagoras-Prover-4B, Goedel-Prover-V2-8B, and DeepSeek-Prover-V2-7B, each run on miniF2F and ProofNet. This benchmark is separate from the training corpus and the 27,210-program in-house validation set used to select thresholds and the self-consistency voting rule.

Labels and collection

passed maps to label 1; failed and wrong_declaration map to 0. Timeouts, infrastructure exceptions, not_run, and precheck failures are excluded. Labels reflect the collection verifier's acceptance policy, including expected-theorem matching and forbidden constructs, rather than successful elaboration of an arbitrary statement alone.

Programs are deduplicated by (source, lean_version) and conflicting verdicts are removed. The collection uses a 180-second per-command timeout, 64 concurrent verification jobs, and reuse of loaded Lean environments. miniF2F and ProofNet validity rates differ substantially; report metrics per benchmark as well as pooled results. The release does not assert that theorem statements are disjoint from all evaluator training material.

Columns

Column	Meaning
run_id	Source prover evaluation run
benchmark	minif2f or proofnet
prover_model	Prover that generated the candidate
lean_version	Toolchain used for the reference verdict
problem	Benchmark problem identifier
sample_id	Generator sample identifier, when available
candidate_index	Candidate position within its source run
source	Complete unannotated Lean source
label	1 for valid, 0 for invalid
status	passed, failed, or wrong_declaration
error_kind	Error category for invalid candidates; null for valid candidates

Load

from datasets import load_dataset

benchmark = load_dataset("EPFL-LARA/lean-evaluation-benchmark", split="test")
source = benchmark[0]["source"]
label = benchmark[0]["label"]

Use source as the model input and reserve label, status, and error_kind for evaluation.

Results reported in the paper

Evaluator	Accuracy	Macro F1	MCC	AUROC
Encoder: ModernBERT-large	0.586	0.563	0.372	0.862
One-token 8B	0.752	0.752	0.592	0.905
One-token 8B, calibrated threshold 0.86	0.831	0.831	0.683	0.905
Feedback 4B, greedy	0.851	0.851	0.733	—
Cascade: one-token → greedy feedback	0.878	0.877	0.772	—
Cascade with 5-sample feedback, 4 valid votes required	0.888	0.887	0.783	—

Default classifier and gate thresholds are 0.5. Calibrated thresholds and the four-of-five self-consistency rule are selected on the separate in-house validation set. Self-consistency uses five traces at temperature 0.8. See the model cards for exact prompts, parsing rules, and serving settings.

Integrity

SHA-256 of data/test-00000-of-00001.parquet:

5a59f709b1466d53c72412cb8c7b7bb53cadd131b3a5b4ce8596677f70e2b3aa

Paper and related artifacts

This is the evaluation benchmark described in Learning Lean Feedback for Efficient Proof-Candidate Evaluation on GPUs.

• Encoder (ModernBERT-large)
• One-token (Qwen3-8B)
• Feedback (Qwen3-4B)