PSEB / README.md
stillrolli's picture
Update README.md
6ea178e verified
|
Raw History Blame Contribute Delete
2.02 kB
metadata
license: mit
task_categories:
  - text-retrieval
language:
  - en
tags:
  - lean4
  - mathlib
  - theorem-proving
  - formal-mathematics
pretty_name: PSEB - Proof-State Embedding Benchmark
configs:
  - config_name: t1
    data_files: t1.jsonl
  - config_name: t2
    data_files: t2.jsonl
  - config_name: t3
    data_files: t3.jsonl
  - config_name: t4
    data_files: t4.jsonl
  - config_name: t5
    data_files: t5.jsonl
  - config_name: t6
    data_files: t6.jsonl
  - config_name: t7
    data_files: t7.jsonl
  - config_name: t8
    data_files: t8.jsonl

PSEB — Proof-State Embedding Benchmark

A kernel-free benchmark of eight retrieval tasks over Lean 4 proof states, built from 303,870 steps in 64,031 Mathlib theorems. Labels are mined from proof traces already checked by the Lean kernel, so no task here requires a Lean install or a kernel call to construct or score.

See the paper, "Which Tactics Deserve a Kernel Call? Selection and Benchmarking for Lean 4 Neural Theorem Proving," for task definitions, label-quality measurements against gold annotations, and baseline results.

Tasks

Config Rows Labels
t1 5,695 premises named in tactic [...] brackets; gold-checked
t2 30,856 tactic used at this canonical state (multi-positive)
t3 30,856 tactic for this (before, after) transition
t4 3,516 duplicate-state sanity check (identical canonical hash)
t5 55 duplicate-statement sanity check
t6 154,280 applicability: 1 observed tactic + 4 sampled negatives
t7 30,358 renaming sensitivity: canonical renamed variant vs. hard/random negatives
t8 6,418 subject area, from file path (e.g. Mathlib/Algebra → Algebra)

Split: file-level (every step of a theorem stays on one side), seed 0, 6,422 test theorems / 30,856 test steps out of 64,031 theorems total.

Usage

from datasets import load_dataset

t1 = load_dataset("stillrolli/PSEB", "t1")
print(t1["train"][0])