Files
lean-bench/leanbench/formalize.py
T
claude-bot 598f7fd662
ci / test (push) Successful in 1m10s
ci / lean-project (push) Successful in 6s
docker / image (push) Successful in 3m6s
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.
                   One Docker layer per Lean step, each long one under `timeout`:
                   the gigabyte cache fetch earns a layer that a later failure never
                   invalidates, and a stalled step fails with a message instead of
                   sitting silent for an hour. There is deliberately no `lake build`
                   of our own library -- verification runs `lake env lean` against
                   Mathlib's oleans and needs nothing else, and building the library
                   was pure cost. 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 build runs with --progress=plain, because BuildKit's
                   default renderer collapses a long step into one rewritten line
                   and a stall then looks exactly like progress.

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 16:22:09 +00:00

285 lines
9.7 KiB
Python

"""Step 3: every detected span becomes Lean 4 code, then Lean gets the last word.
A ``statement`` item becomes whatever declaration fits -- ``def``, ``structure``,
``theorem`` -- with a proof only if the source actually gave one. A ``problem`` item
is an open question, so it becomes a statement with ``sorry`` for the proof and only
the statement's legality is checked.
When Lean rejects the code, the diagnostics go back to the model and it gets another
go: a compiler error is the cheapest, most honest feedback available.
"""
from __future__ import annotations
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass, field
from typing import Any
from .detect import KIND_PROBLEM, Detection
from .llm import extract_code
from .typeindex import index_types
from .verify import MODE_FULL, MODE_STATEMENT, VerifyResult, contains_sorry
SYSTEM_PROMPT = """\
You translate mathematics from books into Lean 4 (Mathlib). You write Lean 4 with
Mathlib naming and syntax -- never Lean 3, never `begin ... end`.
You are given one passage, extracted from a PDF, so expect broken line breaks,
hyphenation and mangled symbols; reconstruct what the mathematics must have been.
Produce one self-contained top-level declaration that captures the passage:
- a definition in the passage -> `def`, `abbrev`, `structure`, `class` or `instance`;
- a theorem / proposition / lemma / corollary -> `theorem`;
- an exercise or open problem -> `theorem` whose statement is the thing to be shown.
Rules:
- Prefer existing Mathlib definitions over re-defining them. Never invent Mathlib
lemma or definition names: if you are unsure a name exists, restate the maths with
primitives you are sure of.
- State the mathematics faithfully. A statement that is trivially true, or that
quantifies over the wrong thing, is a failure even if it compiles.
- Give the declaration a descriptive snake_case name.
- Include a proof ONLY when the passage itself contains the proof and you are
confident in it; otherwise end the declaration with `sorry`.
- If the passage is not formalizable as a single declaration, say so with
"skip": true and a short reason.
Reply with JSON only:
{"skip": false,
"reason": "",
"imports": ["Mathlib"],
"name": "snake_case_name",
"lean_code": "theorem snake_case_name (n : \\u2115) : ... := by\\n sorry",
"has_proof": false,
"informal": "one-line English restatement",
"index_types": ["Int", "Nat"],
"confidence": 0.0}
"index_types" lists the Lean types this item is about: for a theorem, the types of
its inputs and hypotheses; for a definition, the type of the thing being defined.
"""
REPAIR_PROMPT = """\
Lean rejected that code. Fix it and reply with the same JSON shape.
Do not paper over the error with `sorry` in place of a statement, and do not change
what the passage says in order to make it compile. If a Mathlib name does not exist,
restate the mathematics with primitives that do.
Lean diagnostics:
{diagnostics}
"""
@dataclass
class Formalization:
"""The Lean rendering of one detected item, plus what Lean thought of it."""
detection: Detection
lean_code: str = ""
imports: list[str] = field(default_factory=lambda: ["Mathlib"])
name: str = ""
informal: str = ""
has_proof: bool = False
confidence: float = 0.0
skipped: bool = False
skip_reason: str = ""
attempts: int = 0
verify: VerifyResult | None = None
index_types: list[tuple[str, str]] = field(default_factory=list)
@property
def verified(self) -> bool:
return bool(self.verify and self.verify.ok)
@property
def proved(self) -> bool:
return bool(self.verify and self.verify.proved)
def as_dict(self) -> dict[str, Any]:
return {
"detection": self.detection.as_dict(),
"lean_code": self.lean_code,
"imports": self.imports,
"name": self.name,
"informal": self.informal,
"has_proof": self.has_proof,
"confidence": self.confidence,
"skipped": self.skipped,
"skip_reason": self.skip_reason,
"attempts": self.attempts,
"verify": self.verify.as_dict() if self.verify else None,
"index_types": [{"name": n, "role": r} for n, r in self.index_types],
}
def verification_mode(detection: Detection) -> str:
"""Open problems carry no proof, so only their statement can be checked."""
return MODE_STATEMENT if detection.kind == KIND_PROBLEM else MODE_FULL
def _user_prompt(detection: Detection, document_title: str = "") -> str:
where = (
f"pages {detection.start_page}-{detection.end_page}"
if detection.end_page != detection.start_page
else f"page {detection.start_page}"
)
kind_note = (
"This is an exercise / open problem: state it, do not solve it. End with `sorry`."
if detection.kind == KIND_PROBLEM
else "This is asserted as mathematics in the book."
)
header = f"Source: {document_title or 'book'}, {where}"
if detection.label:
header += f", labelled {detection.label!r}"
return (
f"{header}\n{kind_note}\n\nPassage:\n<<<BEGIN>>>\n{detection.text}\n<<<END>>>"
)
def _clean_code(payload: dict) -> str:
code = payload.get("lean_code") or payload.get("code") or ""
if not isinstance(code, str):
return ""
code = extract_code(code, "lean")
# models occasionally keep the import lines inside the declaration body
lines = [line for line in code.splitlines() if not re.match(r"^\s*import\s", line)]
return "\n".join(lines).strip()
def _imports(payload: dict) -> list[str]:
raw = payload.get("imports")
if isinstance(raw, str):
raw = [raw]
if not isinstance(raw, list):
raw = []
modules = [str(m).strip() for m in raw if str(m).strip()]
return modules or ["Mathlib"]
def formalize(
client: Any,
detection: Detection,
*,
checker: Any = None,
document_title: str = "",
repair_attempts: int = 1,
) -> Formalization:
"""One detected span -> one Lean declaration, verified as far as it can be."""
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": _user_prompt(detection, document_title)},
]
result = Formalization(detection=detection)
mode = verification_mode(detection)
for attempt in range(repair_attempts + 1):
payload = client.complete_json(messages)
if not isinstance(payload, dict):
payload = {}
result.attempts = attempt + 1
if payload.get("skip"):
result.skipped = True
result.skip_reason = str(payload.get("reason") or "model declined")
return result
code = _clean_code(payload)
if not code:
result.skipped = True
result.skip_reason = "model returned no Lean code"
return result
result.lean_code = code
result.imports = _imports(payload)
result.name = str(payload.get("name") or "").strip()
result.informal = str(payload.get("informal") or "").strip()
result.has_proof = bool(payload.get("has_proof")) and not contains_sorry(code)
try:
result.confidence = float(payload.get("confidence") or 0.0)
except (TypeError, ValueError):
result.confidence = 0.0
extra = payload.get("index_types")
result.index_types = index_types(code, extra if isinstance(extra, list) else [])
if checker is None:
return result
result.verify = checker.check(code, mode=mode, imports=result.imports)
if result.verify.ok or result.verify.skipped or attempt >= repair_attempts:
return result
messages = messages + [
{"role": "assistant", "content": _echo(payload, code)},
{
"role": "user",
"content": REPAIR_PROMPT.format(
diagnostics=(result.verify.diagnostics or "")[-3000:]
),
},
]
return result
def _echo(payload: dict, code: str) -> str:
import json
return json.dumps({**payload, "lean_code": code}, ensure_ascii=False)[:6000]
def formalize_all(
client: Any,
detections: list[Detection],
*,
checker: Any = None,
document_title: str = "",
repair_attempts: int = 1,
progress: Callable[[str], None] | None = None,
on_error: str = "skip",
) -> list[Formalization]:
results: list[Formalization] = []
for position, detection in enumerate(detections, 1):
try:
formalization = formalize(
client,
detection,
checker=checker,
document_title=document_title,
repair_attempts=repair_attempts,
)
except Exception as error: # noqa: BLE001 - one bad item must not kill a book
if on_error == "raise":
raise
formalization = Formalization(
detection=detection, skipped=True, skip_reason=f"formalizer failed: {error}"
)
results.append(formalization)
if progress:
status = (
"skipped"
if formalization.skipped
else "proved"
if formalization.proved
else "verified"
if formalization.verified
else "unverified"
)
progress(
f" [{position}/{len(detections)}] {detection.label or detection.kind} "
f"@ {detection.locator}: {status}"
)
return results
def as_dicts(formalizations: list[Formalization]) -> list[dict]:
return [f.as_dict() for f in formalizations]
__all__ = [
"Formalization",
"formalize",
"formalize_all",
"verification_mode",
"as_dicts",
"asdict",
]