feat(dataset): rebuild google-math with the AlphaEvolve and FunSearch constructions

#9

Why

Three AlphaEvolve and FunSearch repositories were measured by an earlier scan and never decided, so the component had a coverage gap. They are now judged on measured evidence: included where they ship math text, excluded with a named size where they ship code only.

What changed

  • Adds alphaevolve (16 verification notebooks over the 15 maths/ problem directories), alphaevolve_results (mathematical_results.ipynb) and funsearch (four problem notebooks): 21 records, 21 groups, 8,668,170 characters.
  • Adds math_content to every source and record, from formal_statements, proofs, natural_language_math, generated_qa, constructions, problem_statements, none. The new sources are constructions.
  • Records excluded_subpaths separately from excluded_candidates, so a repository is in included or excluded_candidates, never both. omega/ (45,511,025 bytes, 44,267,872 of them one .xz archive) and the search harnesses are excluded with their measured sizes.
  • Rebuilds data, packed arrays, statistics, sources, manifest, README and validation. 21 new records; all 480,584 pre-existing records keep byte-identical text and only 2 of them changed split.
  • Replaces the stale audit/publish-check.json with a fresh independent gate, keeps the old one as audit/publish-check-previous.json.

Scope

Additive only. phase1/, components/openai-math/, components/README.md and every other path on main are untouched. The top-level README.md change is the google-math config row that the component's Parquet files need.

Tradeoffs

  • The new records are research_notebook because the results ship as notebooks. 5,826,569 of the 7,311,423 alphaevolve characters come from one 593-sphere kissing configuration, so the addition is coordinate-heavy rather than prose-heavy. The README states this.
  • 12 execution outputs totalling 3,087 bytes were dropped from funsearch notebooks; no embedded images existed. Code and markdown cells are verbatim.
  • The corpus is still 99.31% seeded school-level generator output by record count, and the README says so.

Blast Radius

components/google-math/ only, plus one config block and one table row in the top-level README.md. No deletions, no moves, no rewrites of any other component.

Verification

  • tools/validate_google_math.py against the delivered files with the pinned checkouts: 13 of 13 checks passed.
  • audit/publish-check.json, independent code: 42 packed rows decoded across train/dev/test, 613 segments compared with 613 exact token-slice matches, 558 whole-record segments matching text under whitespace normalisation and re-encoding identically, 55 fragments covered by the same token-slice identity, 9 of 9 Parquet round-trips exact, 0 of 3,424 groups spanning two splits.
  • Rebuild delta against the earlier totals: 480,584 to 480,605 records, 3,403 to 3,424 groups, 78,364,888 to 87,033,058 characters.
  • Remote tree diff after this PR: additions under components/google-math/ plus the README.md config row, nothing else.

Independent publish gate, re-run against the delivered files

Status passed. independent, written from scratch, no component tool code reused.

Totals recomputed from data/*.jsonl.gz: 480,605 records, 3,424 groups, 87,033,058 characters, 11 sources, 480,605 unique ids, 480,605 unique text digests.

split records groups characters packed rows tokens loss-bearing labels
train 479,337 3,354 86,369,833 19,676 40,295,101 39,815,764
dev 1,233 35 234,753 51 103,210 101,977
test 35 35 428,472 92 186,986 186,951

Packed decode per split

split rows segments whole-record text exact re-encode exact fragments (verbatim) fragments (no text) edge-trimmed token-slice exact mismatches
train 14 302 282 282 282 20 0 2 302 0
dev 14 293 273 273 273 18 2 0 293 0
test 14 18 3 3 3 15 0 2 18 0

Parquet round-trips exact: 9 of 9. Groups spanning two splits: 0 of 3,424.

Declared-number comparisons

13 checks against statistics.json, validation.json and packed/manifest.json; disagreements: none.

Record mix, measured

  • generated_qa_pair: 477,276 records = 99.31% of records, 35,710,731 characters = 41.03% of characters.
  • Lean (formal_lean_proof + formal_lean_statement): 3,027 records = 0.63%.

Boundary-fragment rule change

Four fragments needed their multi-byte edges trimmed and two carried no text (BOS/EOS or bare whitespace only). The earlier gate counted all 55 fragments as covered by token-slice identity alone; this run additionally requires verbatim containment of the decoded fragment text, which the 53 text-bearing fragments satisfy.

Schema adapter fixed: the full rebuild path no longer aborts

tools/build_google_math.py --manifest-only used to crash with KeyError: 'rows_decoded' at write_readme line 1242, because the README generator read the gate's legacy keys while the gate now emits per-split keys.

Reproduced first (same traceback), then fixed at the root: one boundary adapter, read_gate_evidence(root, records), replaces every direct gate-key access.

legacy schema per-split schema
rows rows_decoded sum of packed_decode.*.rows
segments segments_compared sum of packed_decode.*.segments
token slices segments_whose_row_slice_equals_canonical_token_slice sum of *.token_slice_exact
whole-record segments whole_record_segments_decoded sum of *.whole_record_segments
whole-record text whole_record_segments_with_identical_whitespace_normalised_text sum of *.whole_record_text_exact
whole-record re-encode whole_record_segments_reencoded_token_identically sum of *.whole_record_reencode_exact
fragments fragment_segments_not_reencoded_per_segment sum of *.fragment_segments
fragments with text not measured — stays None, clause omitted sum of *.fragment_verbatim_contained
fragments without text not measured — stays None, clause omitted sum of *.fragment_no_text
edge-trimmed fragments not measured — stays None, clause omitted sum of *.fragment_boundary_trimmed
Parquet round-trips rows_checked / exact_matches rows / exact
groups split_separation group_separation

Nothing is reinterpreted: per-split counters are summed, legacy totals are read as totals, so each field keeps the meaning its writer gave it. Fields the legacy gate never measured stay None and the README drops those clauses rather than inventing numbers. The gate is also rejected when measured_totals.records disagrees with statistics.json, so stale evidence cannot be quoted.

Both schemas exercised through the adapter:

  • per-split: 42 rows (14/14/14), 613 segments, 613 token slices, 558 whole-record, 55 fragments (53 with text, 2 without, 4 edge-trimmed), 9 of 9 round-trips, 3,424 groups, 0 spanning
  • legacy: 12 rows, 263 segments, 263 token slices, 248 whole-record, 248 text-exact, 248 re-encode-exact, 15 fragments, 5 of 5 round-trips, 3,403 groups, 0 spanning

End-to-end proof

  • The tool now exits 0 and reports 280 files; regenerate-then-validate completes.
  • Independently, the tool downloaded fresh from this branch ran in a clean directory and completed without aborting.
  • tools/validate_google_math.py: passed, 13 checks, 480,605 records, 3,424 groups.

Corpus unchanged. Regenerated manifest and README describe the identical dataset: 480,605 records, 3,424 groups, 87,033,058 characters, identical record-type counts (generated_qa_pair 477,276 / formal_lean_statement 2,492 / formal_lean_proof 535 / formal_geometry_problem 265 / research_notebook 23 / natural_language_proof 12 / geometry_rule_table 2), identical packed totals (19,676 / 40,295,101 / 39,815,764 train, 51 / 103,210 / 101,977 dev, 92 / 186,986 / 186,951 test) and identical per-split numbers. Manifest diff is two entries only: README.md 29,545 to 29,535 bytes and tools/build_google_math.py 87,577 to 93,052 bytes. The README shrinks 10 bytes because the regenerated paragraph ends one line-break earlier; all prose and every number in it are unchanged under whitespace normalisation.

Commit 425a80034fc1773254e3d5285d33ef132b09087b on refs/pr/9. Branch tree still matches the local manifest with 280 files and zero byte mismatches. main is untouched at 794446092591c428368b2babe0cd8086d3b61b5b with zero google-math files.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment