awhecmu/canonical-drafter-extract-mathlib

Canonical Drafter — extract data (raw)

Ground-truth drafter/premise data lifted from existing Mathlib proofs by Training/ExtractData.lean. This is the raw pool (429303 draft rows, 664445 premise rows from: Mathlib). Every row is a real have / closed subgoal taken from a checked proof, so there are no success/used flags to filter on. Columns follow the uniform schema shared with the rollout dataset, so the two pools concatenate cleanly.

config drafts

One row per have recovered from a Mathlib proof.

column	type	meaning
goal	str	pretty-printed goal the have addresses
type	str	the have type
name	str	the have hypothesis name
removals	list[str]	hypotheses the have cleared
tactic	str	the ground-truth tactic that produced this step
kind	str	what the step is (e.g. goal, hypothesis)
messages	list	single user->assistant turn; the assistant turn is the have tool call
tools	list	tool schema offered to the model
const, roundtrip, success, used, usedTactic, typeFromPp	—	null for extract rows (rollout-only columns)

Recommended drafter-SFT use: every row is a used, ground-truth draft — keep all rows and project to {messages, tools} (loss on the final assistant turn only). Optionally filter by kind.

config premises

One row per subgoal closed within an extracted proof, with the closing tactic and the library constants it used.

column	type	meaning
goal	str	pretty-printed goal that was closed
tactic	str	the closing tactic / term
premises	list[str]	library constant names used by the closing proof
const, roundtrip	—	null for extract rows (rollout-only columns)

Recommended premise-selection use: many rows have an empty premises list (goal closed by rfl/simp with no library lemma); drop those if you only want goals that needed premises.

Splitting

Both configs ship train and validation splits. The split is module-level (all rows from one Mathlib file land in the same split, so no train/val leakage) and is the canonical project split shared with the rollout pipeline — an explicit list of modules bundled in this repo at mathlib_v1_splits/{train,val}_modules.txt, not a hash. Every row also carries:

column	type	meaning
module	str	source Mathlib module (file) the row came from
split	str	train or val, from data/splits/mathlib_v1/