Skip to content

THROWAWAY: smoke-run the 2026-08-25 autoformalization target problems (misformalization check) - #29

Closed
tadamcz wants to merge 1 commit into
bloom-erdos-selectionfrom
erdos-target-smoke
Closed

THROWAWAY: smoke-run the 2026-08-25 autoformalization target problems (misformalization check)#29
tadamcz wants to merge 1 commit into
bloom-erdos-selectionfrom
erdos-target-smoke

Conversation

@tadamcz

@tadamcz tadamcz commented Aug 25, 2026

Copy link
Copy Markdown
Collaborator

Do not merge. Throwaway branch stacked on #28 for a single smoke run.

Vendors the 18 Erdős problems from the 2026-08-25 autoformalization target run (epoch-research/autoformalization results/2026-08-25-erdos-target-run: 17 gate-passing submit/ problems + 1020 from hold/) into apn/data/erdos as subset target18 — 20 statements, since 713 and 1206 each have two parts, one sample per statement per the dataset's convention.

Purpose: these are open problems autoformalized outside FC review. Claude Fable 5 and GPT 5.6 Sol each attempt every statement at a $50/sample cap (configs/erdos-target18-smoke.yaml, one eval set, 40 samples, $2,000 max). Any solve at this budget is a red flag for misformalization (statement too easy / vacuous), not a result.

Pin note: the run's files were produced and gate-checked at FC 9cbe1d3c, an ancestor of this dataset's 488aade2 pin (one day apart; same toolchain and Mathlib, no shared-library declaration touched between the pins is used by these files). Representative isolated specs — including both surgically part-split ones — compile cleanly in the FC image. The existing fc_488aade2 sandbox images in ECR are reused as-is.

Isolation-certificate tests were not updated for the new files (throwaway branch); expect CI to be red.

…s as erdos subset target18

18 Erdős problems (20 statements; 713 and 1206 have two parts each) from
epoch-research/autoformalization results/2026-08-25-erdos-target-run,
autoformalized and gate-passed at FC 9cbe1d3c (an ancestor of this
dataset's 488aade2 pin; same toolchain and Mathlib, no touched shared-lib
declaration is used by these files -- representative specs compile-checked
in the FC image). Sources/ are the run's output verbatim; Isolated/ specs
strip the @[category ...] attributes and sibling part statements, matching
the existing isolation shape.

Smoke-run purpose only: these are open problems formalized outside FC
review, so a solve at a modest budget flags likely misformalization.
Not for merge; stacked on the bloom-erdos-selection branch (PR #28).
@tadamcz tadamcz closed this Aug 25, 2026
@tadamcz

tadamcz commented Aug 25, 2026

Copy link
Copy Markdown
Collaborator Author

Smoke run launched

Eval set: erdos-target18-smoke-3mk4g11s6ghm0o1x
Viewer: https://viewer.hawk.hawkbench.com/eval-set/erdos-target18-smoke-3mk4g11s6ghm0o1x

Task apn_erdos with subset: target18 (20 statements: the 18 target-run problems; 713 and 1206 split into their two parts)
Models epoch/claude-fable-5 and epoch/gpt-5.6-sol, both reasoning_effort: high, 1 epoch
Budget $50/sample cost cap (cost_limit: 50.0), 12h working limit; 40 samples → $2,000 hard ceiling
Config configs/erdos-target18-smoke.yaml at 9ca3aab (snapshot in configs/history/2026-08-25T17-29-15_erdos-target18-smoke-3mk4g11s6ghm0o1x.yaml)
Task package this branch (erdos-target-smoke), commit 9ca3aab
Sandbox images existing ECR LeanOpenProblems_{agent,scorer}_0.1.6_fc_488aade228ec (pulled OK at run start)
Launched 2026-08-25 17:29 UTC; first samples running as of 17:45 UTC

Interpretation: every statement here is an open Erdős problem autoformalized outside FC review. Any CORRECT at this budget is a red flag for misformalization (statement too easy, vacuous, or otherwise not the intended problem) — inspect that sample's transcript and statement, don't celebrate. The expected outcome is 0/40 with most samples burning the cap.

Status: hawk status erdos-target18-smoke-3mk4g11s6ghm0o1x / hawk logs erdos-target18-smoke-3mk4g11s6ghm0o1x.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant