Datasets:
feat(dataset): add the google-math component
Why
Eight published Google and Google DeepMind mathematics artifacts had no home in Cayley-Math. The component lands them in the same schema, split policy and packed format as components/openai-math, and replaces the superseded google-alphaproof-nexus draft in #2.
What changed
- Adds
components/google-math/: 480,584 records over 3,403 groups from 8 pinned sources, plusdata/{train,dev,test}.jsonl.gzand Parquet, packed arrays, audit files, licenses, tokenizer, tools and README. - Adds a
google-mathconfig and a Collections row to the top-levelREADME.md. - Delivers
audit/publish-check.json, a second publish gate written independently of the build's own verifier.
Scope
Additive only. phase1/, components/openai-math/ and components/README.md are untouched, and there are no deletions.
Tradeoffs
- The corpus is 477,276 of 480,584 records
generated_qa_pair(99.31%), so it is mostly seeded generator output of a school-level curriculum paper. The README says so plainly and flags that the formal subset is 0.63% of records by count. - No Lean compilation or proof verification was performed, and no record carries a checked proof. The README states this as a caveat rather than a validation.
Verification
tools/validate_google_math.pyre-run against the delivered files: passed all 13 checks, packed train/dev/test at 17,691/51/91 rows.audit/publish-check.json: 12 packed rows decoded across all three splits, 263 segments compared, 248 whole-record segments decoded to text identical to source modulo whitespace with token-identical re-encoding, 5 Parquet rows round-tripped exactly, 3,403 groups with 0 spanning two splits.- Every measured record, group, character and packed count matches
statistics.json.
Measured publish evidence
Second gate, written independently of the build's own verifier: components/google-math/audit/publish-check.json.
- Packed decode: 12 rows across dev/test/train, 263 attributed segments. 263 row slices equal their canonical token slice by exact identity. 248 whole-record segments decoded to source text under whitespace normalisation and 248 re-encoded to identical token sequences. 15 row-boundary fragments cannot re-encode per fragment, and all 15 were located verbatim inside their record. Failures: 0.
- Parquet round trip: 5 rows across all three splits, 5 exact dictionary matches against their JSONL records.
- Split separation recomputed from
data/*.jsonl.gz: 3403 groups, 0 spanning two splits, train/dev/test at 3333/35/35 groups. statistics.jsoncomparison: all counts match (True), covering records, groups, characters, per-source and per-split records, and packed rows/tokens/labels for all three splits.
Composition, measured
generated_qa_pair: 477,276 records, 99.3117% of records and 45.5698% of characters.- Lean records: 3,027 (0.6299% of records, 54.2177% of characters).
- Formal, geometry, prose and notebook records: 3,308 (0.6883% of records).
- Lean compilation
not_performed, proof verificationnot_performed, mathematical verificationnot_performed.
The component README states this plainly, and the no-verification caveat is a stated limitation rather than a claim of correctness.
Tool re-run
tools/validate_google_math.py against the delivered files with the pinned source checkouts: passed all 13 checks, packed train/dev/test at 17,691/51/91 rows and 36,230,380/102,872/185,977 tokens.
Blocked step
Merging is refused with 403 Forbidden: You have read access but not the required permissions for this operation. The same token writes fine in AvrovaDonz/CAT-YK, so this is an MMLA-ORG write role, not a token defect. The PR is ready and only needs an org member with write access to merge it.
Final state
- Head commit
d554fa0f47fe7e0d7bb3adb1a06a55ea9e3c7483. All 258 delivered files verified byte-for-byte against the local build: 225 LFS blobs by sha256 and 33 text files by git blob id, 0 mismatches. - Diff against
main(794446092591c428368b2babe0cd8086d3b61b5b): 257 additions, all undercomponents/google-math/, plus the top-levelREADME.mdedit, and 0 deletions.phase1/,components/openai-math/andcomponents/README.mdare untouched. tools/validate_google_math.pyre-run: passed all 13 checks. Packed train/dev/test at 17,691/51/91 rows and 36,230,380/102,872/185,977 tokens.audit/publish-check.json: statuspassed. 12 packed rows, 263 segments, 263 row slices identical, 248 whole-record segments token-identical and whitespace-identical, 0 failures, 5/5 Parquet round-trips exact, 0 groups spanning splits, all 30statistics.jsoncomparisons matching.
Merge blocker
Merging is refused by the Hub itself, not only by policy in this session. The merge endpoint returns HTTP 403 with this body:
{"error":"You have read access but not the required permissions for this operation"}
The same account and token can create branches and commits in AvrovaDonz/CAT-YK, and it could create this pull request and commit to refs/pr/3, so this is an MMLA-ORG write role rather than a bad token. A member of MMLA-ORG with write access can merge this pull request, or re-run hf upload without --create-pr once such access exists.
Pull requests #4 and #5 were duplicates from a concurrent process and are closed as superseded. #2 is closed as superseded.
Final state
- Head commit
d554fa0f47fe7e0d7bb3adb1a06a55ea9e3c7483. All 258 delivered files verified byte-for-byte against the local build: 225 LFS blobs by sha256, 33 text files by git blob id, 0 mismatches. - Diff against
main(794446092591c428368b2babe0cd8086d3b61b5b): 257 additions, all undercomponents/google-math/, plus the top-levelREADME.mdedit, and 0 deletions.phase1/,components/openai-math/andcomponents/README.mdare untouched. tools/validate_google_math.pyre-run: passed all 13 checks. Packed train/dev/test at 17,691/51/91 rows and 36,230,380/102,872/185,977 tokens.audit/publish-check.json: statuspassed, 12 packed rows, 263 segments, 263 row slices identical, 248 whole-record segments token-identical and whitespace-identical, 0 failures, 5/5 Parquet round-trips exact, 0 groups spanning splits, all 30statistics.jsoncomparisons matching.
Merge blocker
Merging is refused by the Hub itself, not by policy in this session:
POST /api/datasets/MMLA-ORG/Cayley-Math/discussions/3/merge -> 403
{"error":"You have read access but not the required permissions for this operation"}
The same account and token can create branches and commits in AvrovaDonz/CAT-YK, and it could create this pull request and commit to refs/pr/3, so this is an MMLA-ORG write role, not a bad token. A member of MMLA-ORG with write access can merge this pull request, or re-run hf upload without --create-pr once such access exists.
Pull requests #4 and #5 were duplicates from a concurrent process and are closed as superseded. #2 is closed as superseded.
Head update
Head is now a456e352029ffc3625348986e8ba705bb15793b5. It restores .gitattributes to the main version, which the first upload had appended two LFS lines to. The change set against main is therefore exactly:
- 257 additions, all under
components/google-math/ - 1 modification: the top-level
README.md(google-mathconfig entry plus Collections row) - 0 deletions
All 257 component files and the top-level README verified byte-for-byte against the built artifacts (225 LFS blobs by sha256, 33 text files by git blob id). phase1/, components/openai-math/ and components/README.md are untouched.
Duplicates #4, #5, #6 and #7 were opened by concurrent processes and are closed as superseded. #3 is the single complete pull request.
#9 is the superseding complete upload
This PR is superseded by #9 (https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/discussions/9). That PR carries the rebuilt google-math component: 11 included sources including the newly decided alphaevolve, alphaevolve_results and funsearch; 480,605 records, 3,424 groups, 87,033,058 characters; a fresh passing validate run and a fresh independent audit. #9's tree matches the local manifest byte for byte across 282 files, with 0 deletions against main. Please merge #9 and close the other google-math PRs.