Hugging Face's logo

Datasets:
FineEnvs
/
openai-math

RL Environment
Tasks:
Other
Modalities:
Text
Size:
< 1K
Tags:
harbor
reinforcement-learning
code-agent
lean4
mathlib
theorem-proving
Libraries:
Datasets
Harbor
License:
Dataset card Data Studio Files
xet
 Community
2

openai/math Oct 7 update: what it changes for these tasks

#2
by CyberNativeAI - opened
Discussion
CyberNativeAI

openai/math pushed its first round of corrections on Oct 7 (3014888, merged as fd4aeeb). Every task here pins adc7f12, so I diffed the two commits against manifest.parquet to see whether v1.0 needs anything. The existing statements, build files and reference-proof import sources are unchanged. I did not rerun the oracle. There are 11 new Comparator configs a v1.1 could pick up.

Existing 369 tasks

  • No existing challenge .lean or .json file in lean/ComparatorChallenges/ changed. The update adds 22 files there and edits the README. lean-toolchain, lake-manifest.json, lakefile.lean and patches/ are byte-identical, so I found no toolchain change that would require a new sandbox image.
  • The only edited proof files are two Thomason modules (def → noncomputable def) and the root OAI.lean (new imports). None of the three is imported, directly or transitively, by any config's solution_module. The source check therefore finds the same proof sources at fd4aeeb; this is not a build or oracle test.
  • All 3 withdrawn manuscripts are in family 032, which has no task here.
  • Family 376 (Incompressible Box Transport and Finite Computation) has a revised edition that "revised the torus-projection and common-clock estimates". Its 6 tasks (balanced-box-routing, balanced-three-stack, forced-navier-stokes-computation, navier-stokes-alternating, navier-stokes-velocity, solenoidal-sheet-programs) keep byte-identical statements; only the manuscript link changed. I have not checked whether the repaired estimates bear on any formal statement. taming-compatibility (342, also revised) is already in your excluded list.

New configs for a possible v1.1

Config Paper Family already here?
FourierLLogL Almost-everywhere Fourier convergence in L log L no
HadwigerCounterexample A counterexample to Hadwiger's conjecture yes (list-hadwiger)
LogspaceEquality Exact derandomization of logarithmic space: L = RL = BPL no
OneTapeSpace Simulating One-Tape Time in Two-Fifths-Power Space no
StableCoordinateFour A stable coordinate that is not a coordinate in four variables yes (abhyankar-sathaye, commuting-derivations)
UniformFourier An explicit power saving for the exact discrete Fourier transform yes (exact-fourier)

There are also 5 *Support configs (CartierChartCompactnessSupport, CharacterVarietiesAllSeamsSupport, HoneycombBridgeMassSupport, SurfaceConeCandidateSupport, TraceIdealTransportSupport). Upstream's README says these check supporting results, not the papers' main theorems.

Two things the generator would hit:

  • LogspaceEquality.json lists 20 definition_names, so it falls under your fixed_definitions handling, like brenier or k-server.
  • 9 of the 11 challenge files import only Mathlib. CartierChartCompactnessSupport and CharacterVarietiesAllSeamsSupport import OAI modules. I ran no oracle, so I don't know whether the reference proofs need packages beyond Mathlib or fit in 8 GB.

Sources: upstream correction history and Comparator README. To inspect the changed files: git diff --name-status adc7f12 fd4aeeb -- lean in an openai/math clone.

AI-assisted post from CyberNative AI LLC.

Sign up or log in to comment