Files
lean-bench/leanbench/verify.py
T
claude-bot b9e37952e6
ci / test (push) Successful in 1m22s
ci / lean-project (push) Successful in 6s
docker / image (push) Failing after 3h0m9s
Rebuild lean-bench: pages in, verified indexed Lean out
Replaces the previous tree entirely. The pipeline is four separable pieces, each
testable on its own:

1. leanbench/llm.py      provider-neutral OpenAI-compatible connector, stdlib only,
                        with retries, JSON coercion and a scripted client so the
                        rest of the code is testable without a network.
2. leanbench/detect.py   per-page detection of formalizable mathematics. Case 1 is
                        mathematics asserted as fact (definition/theorem/lemma/...,
                        deliberately not sub-classified); case 2 is a problem left
                        open to the reader. Hits come back as spans -- start page
                        and index, end page and index -- so several per page is
                        normal. A hit cut off by the page end makes the detector
                        re-read that page together with the next, repeatedly, so an
                        item spread over three pages is found whole.
                        Offsets come from verbatim quotes resolved by
                        leanbench/anchors.py (exact -> whitespace-insensitive ->
                        fuzzy), because models quote well and count badly.
3. leanbench/formalize.py + leanbench/verify.py
                        each span becomes one Lean 4 declaration, judged by the real
                        compiler: full elaboration for case 1, statement legality
                        only for case 2, which has no proof to check. Lean's own
                        diagnostics feed back for a repair attempt. "verified" (Lean
                        accepted it) and "proved" (accepted with no sorry) are kept
                        strictly apart.
4. leanbench/db.py       indexed SQLite plus JSONL export. Every item is indexed by
                        the types of its inputs, or by its own type when it is not a
                        proof, and records the source book (sha256) and the page it
                        starts on -- so `drop --source H --pages A-B` followed by a
                        re-ingest redoes exactly those pages and nothing else.

Infrastructure:

- lean/            Lake project pinned to Mathlib.
- docker/          image carrying Lean 4 and a prebuilt Mathlib. The Mathlib
                   revision and the toolchain are resolved against each other at
                   build time rather than both hardcoded, so the build does not rot.
                   docker/smoke.py runs the whole pipeline inside the image against
                   the real compiler, with the LLM scripted.
- .gitea/workflows ci (lint + tests in a venv, since the runner image's Python is
                   Debian-managed and refuses installs into itself), docker (build,
                   verify, optionally push), ingest (submit a book by hand).

The docker workflow is not a throwaway build: it leaves the image, the layer cache,
an idle `lean-bench` container and the `lean-bench-data` / `lean-bench-books`
volumes in place on the runner, and never prunes, so the environment is there to be
used afterwards. Pushing to the registry is optional and cannot fail the build --
Gitea's container registry does not accept the built-in Actions token, so that needs
a REGISTRY_TOKEN secret with write:package.

58 tests; they need neither network nor Lean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ci5T3gWDhhmqDec3e1XtYZ
2026-09-19 07:16:37 +00:00

149 lines
5.0 KiB
Python

"""Step 3, second half: hand the generated Lean to the real compiler and believe it.
Two verification modes, matching the two kinds of item:
``full`` elaborate the declaration as written. If the model supplied a proof,
this is a genuine proof check.
``statement`` the item is an open problem, so there is no proof to check: we only
ask whether the *statement* is legal Lean -- it elaborates with
``sorry`` standing in for the proof.
``verified`` never means "proved". ``proved`` does, and it is only true when Lean
accepted the declaration with no ``sorry`` anywhere in it.
"""
from __future__ import annotations
import os
import re
import shutil
import subprocess
import tempfile
from dataclasses import dataclass
from pathlib import Path
MODE_FULL = "full"
MODE_STATEMENT = "statement"
_ERROR = re.compile(r"^.*?:\d+:\d+: error:", re.MULTILINE)
_SORRY_WARNING = re.compile(r"declaration uses 'sorry'")
_SORRY_TOKEN = re.compile(r"(?<![A-Za-z0-9_'])sorry(?![A-Za-z0-9_'])")
@dataclass
class VerifyResult:
ok: bool
mode: str
used_sorry: bool
diagnostics: str
skipped: bool = False
@property
def proved(self) -> bool:
return self.ok and not self.used_sorry and self.mode == MODE_FULL
def as_dict(self) -> dict:
return {
"ok": self.ok,
"mode": self.mode,
"used_sorry": self.used_sorry,
"proved": self.proved,
"skipped": self.skipped,
"diagnostics": self.diagnostics[-4000:],
}
def contains_sorry(code: str) -> bool:
return bool(_SORRY_TOKEN.search(code))
@dataclass
class LeanChecker:
"""Runs ``lake env lean`` (or whatever you configure) on a scratch file."""
command: list[str]
project: str | None = None
timeout: float = 120.0
header: str = "import Mathlib\n"
@classmethod
def from_settings(cls, settings) -> LeanChecker: # noqa: ANN001 - LeanSettings
return cls(
command=list(settings.command),
project=settings.project,
timeout=settings.timeout,
header=settings.header,
)
def available(self) -> bool:
if not self.command:
return False
if shutil.which(self.command[0]) is None:
return False
if self.project and not Path(self.project).is_dir():
return False
return True
def source_for(self, code: str, imports: list[str] | None = None) -> str:
header = self.header
for module in imports or []:
module = module.strip()
if not module:
continue
line = module if module.startswith("import ") else f"import {module}"
if line not in header:
header += line + "\n"
return header.rstrip("\n") + "\n\n" + code.strip() + "\n"
def check(
self,
code: str,
*,
mode: str = MODE_FULL,
imports: list[str] | None = None,
) -> VerifyResult:
if not self.available():
return VerifyResult(
ok=False,
mode=mode,
used_sorry=contains_sorry(code),
diagnostics=(
f"Lean not available: command {self.command!r}"
+ (f", project {self.project!r}" if self.project else "")
),
skipped=True,
)
source = self.source_for(code, imports)
workdir = self.project or os.getcwd()
with tempfile.TemporaryDirectory(prefix="leanbench-", dir=workdir) as directory:
path = Path(directory) / "LeanBenchCheck.lean"
path.write_text(source, encoding="utf-8")
try:
completed = subprocess.run( # noqa: S603 - command is operator-supplied
[*self.command, str(path)],
cwd=workdir,
capture_output=True,
text=True,
timeout=self.timeout,
)
except subprocess.TimeoutExpired:
return VerifyResult(False, mode, contains_sorry(code),
f"Lean timed out after {self.timeout}s")
except FileNotFoundError as error:
return VerifyResult(False, mode, contains_sorry(code),
f"Lean not runnable: {error}", skipped=True)
diagnostics = (completed.stdout or "") + (completed.stderr or "")
ok = completed.returncode == 0 and not _ERROR.search(diagnostics)
used_sorry = bool(_SORRY_WARNING.search(diagnostics)) or contains_sorry(code)
return VerifyResult(ok, mode, used_sorry, diagnostics[-8000:])
class NullChecker:
"""Stands in when Lean is unavailable (unit tests, a machine without elan)."""
def available(self) -> bool:
return False
def check(self, code: str, *, mode: str = MODE_FULL, imports=None) -> VerifyResult:
return VerifyResult(False, mode, contains_sorry(code), "Lean checking disabled",
skipped=True)