AvrovaDonz's picture
Record publication status, generator shortfall and measured limitations
bb342ec verified
|
Raw History Blame
18.6 kB
# Google math papers contribution to Cayley-Math
An independently versioned component holding source-derived text from eight Google and Google DeepMind
publications about mathematics, formal mathematics and mathematical reasoning. It follows Cayley-Math's
`components/` convention and reuses the record schema, Parquet encoding, split policy and packed-array
format of `components/openai-math/`. The frozen Phase 1 mirror is unchanged.
## What this is, and what it is not
These are **artifacts of specific Google math papers** (Lean libraries, geometry problem files, generator
output, paper notebooks and proof PDFs), not a survey of Google's mathematics output and not a
paper-complete reproduction of any one of them. The corpus is a curated projection: a bounded, seeded
sample for the generator source, a verbatim text projection for the file sources.
Nothing here establishes mathematical truth. **No Lean project was compiled, no proof was checked, no
`lake build` or Comparator run was performed, and no mathematical verification of any kind took place.**
`record_type` labels the artifact kind, not the validity of its content. Records that contain `sorry`
placeholders remain placeholders.
The `mathematics-dataset` subset is output of the Saxton/Grefenstette/Hill/Kohli ICLR 2018 generator at
school-level difficulty. Those pairs are **not research mathematics**, and the subset is a sampled
projection of the paper's curriculum rather than the paper's dataset. Every record carries
`content_origin` so consumers can separate verbatim source text from `generator_output`.
## Included sources
Each source is pinned to a commit or dataset revision, and each ships math text under Apache-2.0
software terms with CC-BY 4.0 for non-software materials where the upstream states so.
| Source | Upstream at pinned revision | Records | Groups | Text characters | Record types |
| --- | --- | ---: | ---: | ---: | --- |
| `alphaproof-nexus` | [alphaproof-nexus-results](https://github.com/google-deepmind/alphaproof-nexus-results) `0647711a7118` | 83 | 51 | 2,229,771 | formal_lean_proof, natural_language_proof |
| `formal-conjectures` | [formal-conjectures](https://github.com/google-deepmind/formal-conjectures) `6751c0f04398` | 2,135 | 2,135 | 8,133,507 | formal_lean_statement |
| `formal-imo` | [formal-imo](https://github.com/google-deepmind/formal-imo) `a85d84b97c8e` | 261 | 261 | 420,948 | formal_lean_statement |
| `formal-putnam-like` | [formal-putnam-like](https://github.com/google-deepmind/formal-putnam-like) `636c2d46b3a7` | 96 | 96 | 117,201 | formal_lean_statement |
| `cryptids` | [cryptids](https://github.com/google-deepmind/cryptids) `3647718b5df2` | 464 | 464 | 31,638,592 | formal_lean_proof |
| `mathematics-dataset` | [mathematics_dataset](https://github.com/google-deepmind/mathematics_dataset) `427f45075f84` | 477,276 | 127 | 35,710,731 | generated_qa_pair |
| `alphageometry` | [alphageometry](https://github.com/google-deepmind/alphageometry) `6777cb586cbb` | 267 | 267 | 70,907 | formal_geometry_problem, geometry_rule_table |
| `mathematics-conjectures` | [mathematics_conjectures](https://github.com/google-deepmind/mathematics_conjectures) `8302f2910101` | 2 | 2 | 43,231 | research_notebook |
`alphaproof-nexus` keeps all 83 records of the earlier standalone build unchanged in content: 71 Lean
proof files plus 12 pypdf-extracted prose proof PDFs over 51 problem groups. Only its split assignment
changed, because splits are recomputed over the merged group set of all eight sources rather than over
its own 51 groups. `formal-conjectures` is the 1,861-file `FormalConjectures/` statement library plus
its 274 helper-library files. `formal-imo` and `formal-putnam-like` are Lean 4 problem statements.
`cryptids` is the whole Lean library, including the Busy Beaver holdout proofs, whose individual files
reach 1.4 MB of mechanically generated state traces. `alphageometry` contributes 265 geometry problems
in the AlphaGeometry language: 231 JGEX problems, 30 olympiad problems and 4 library examples, plus the
`defs.txt` and `rules.txt` rule tables. The two `mathematics-conjectures` notebooks accompany the Nature
2021 knot-theory and representation-theory paper.
## Excluded candidates and why
The widened scan is recorded in full in `sources.json` (`excluded_candidates`) and `audit/source-scan.json`.
The short version:
- `google-deepmind/eval_hub` (putnam_like and live_math) is excluded because every problem carries a
`CANARY_ID` benchmark-contamination canary and live_math is an actively maintained benchmark whose
multimodal problems depend on images absent from this text corpus.
- `deepmind/aqua_rat` accompanies a program-synthesis paper, not a mathematics paper.
- `deepmind/math_dataset` is a 40 GB pre-generated mirror of the generator already used here.
- `google-research/math-opt-benchmark`, `google/mathsteps`, `google-deepmind/alphatensor` and every
math-named software library (`tree-math`, `mathfu`, `vector_math.dart`, `opengrm`) ship code and no
math text.
- `google-deepmind/cartesian-frames` formalises an alignment blog sequence in HOL4 SML, not a
mathematics paper.
- `google-deepmind/miniF2F` is a fork of a non-Google corpus.
- `google-deepmind/alphageometry2` embeds 26 problems in Python dict literals, and AG1 already ships
the same olympiad problems verbatim in the AlphaGeometry language.
- `google/feabench`, `google/phasecoder` and `google-research/pointdit` are physics, audio and vision
papers whose titles happen to contain mathematical words.
## Files and schema
- `data/{train,dev,test}.jsonl.gz`: source records with required fields text, group, license; unique id,
paper_id, split, record_type, source_url, and `source_id` naming the upstream.
- `data/{train,dev,test}.parquet`: identical records. Core fields
(id,text,group,license,paper_id,split,record_type,source_url) are string columns; all remaining JSON
fields are serialised in `metadata_json`. Merge `json.loads(metadata_json)` with the core columns to
reconstruct each JSONL record.
- `sources.json`: every included source with its pinned revision, license and measured counts, plus every
excluded candidate with its reason.
- `licenses/<source_id>/LICENSE`: one upstream license per source; `LICENSE` at the component root is the
Apache-2.0 text shared by all eight.
- `audit/source-files.jsonl`: path, sha256, bytes, source_id, kind and pinned source_url for every blob and
for the generated corpus file.
- `audit/group-splits.json`: the group assignment, the policy, and per-source group counts.
- `audit/extraction.jsonl`: per-PDF and per-notebook extraction metrics and quality flags.
- `audit/mathematics-dataset-generation.json`: generator commit, seed, per-regime quotas, module list, and
measured yield.
- `audit/excluded-duplicates.jsonl`, `audit/packing.json`, `audit/source-scan.json`.
- `audit/validation-evidence.json`: the independent decode and round-trip evidence described below.
- `statistics.json`: measured records, groups, splits, characters and packed counts, per source.
- `manifest.json`: SHA256 and byte size for every delivered file except itself and validation.json.
- `packed/`: mmla-type2-packed-v1 arrays, manifest, row attribution and per-record token counts.
- `tools/`: the generator driver, the build script and the validation scripts actually run.
## Loading
~~~python
import gzip, json
from pathlib import Path
component = Path('components/google-math')
with gzip.open(component / 'data/train.jsonl.gz', 'rt', encoding='utf-8') as f:
first_record = json.loads(next(f))
import pyarrow.parquet as pq
row = pq.read_table(component / 'data/train.parquet').slice(0, 1).to_pylist()[0]
metadata = json.loads(row.pop('metadata_json'))
first_parquet_record = {**row, **metadata}
~~~
## Packed training arrays
`packed/manifest.json` uses `mmla-type2-packed-v1`: input_ids, labels and role_ids are int32 arrays of
shape [rows, 2048], pad=0, role=0, ignored labels=-100. Each record is encoded as
BOS + SentencePiece(text) + EOS with the included Cayley-Math tokenizer (32768 pieces); window stride is
2048 with one-token lookahead, labels predict the following token, and targets that cross into a new
record's BOS or into padding are masked. Split-specific `rows.jsonl` maps every non-padding token back to
its record; `records.jsonl` stores token counts and hashes. This component does not reproduce Phase 1
quotas, packing or splits.
Measured packed counts: train 17691 rows /
36230380 tokens /
35751064 loss-bearing labels; dev
51 / 102872 /
101639; test
91 / 185977 /
185942.
## Splits
Splits are assigned per group, never per record, so companions and families stay together: a Lean proof
and its prose proof, a conjecture file and its helper library, or one generated question family. Every
group key is namespaced by its `source_id`. Assignment sorts SHA256(group) as hex, reserves the first
ceil(1%) of groups for dev, the next ceil(1%) for test, and the rest for train.
`audit/group-splits.json` freezes every assignment: train
479316 records over 3333 groups,
dev 1233 over 35,
test 35 over 35.
## Text projection
Lean, geometry-problem and rule files are byte-identical to the pinned blob, license headers included.
Notebook records join the cell sources of each notebook with a newline; no cell text is altered.
`natural_language_proof` records come from pypdf page-text extraction and are **lossy**: math symbols
arrive as Unicode substitutions, ligature artifacts appear, and inline formulas lose their two-dimensional
layout. C0/C1 control characters were replaced and blank-line runs collapsed. Each PDF record carries
`extraction_quality = lossy_pdf_text_layer` with page and character counts, and `audit/extraction.jsonl`
holds the same per-file audit.
## Verification evidence
`tools/validate_google_math.py` checks file hashes, source coverage, unique record IDs and unique text,
group split separation, lossless Parquet reconstruction, packed array shapes, causal shift, padding and
boundary masks, and re-encodes every record's token stream against `data/*.jsonl.gz`. It writes
`validation.json`; the last run passed all 13 checks
(`all_delivered_file_hashes_and_sizes`, `pinned_sources_metadata_and_excluded_candidates`,
`source_file_inventory_local_bytes`, `group_split_policy_recomputed_from_scratch`,
`lossless_jsonl_parquet_exact_text_uniqueness_and_record_provenance`, `group_split_separation`,
`verbatim_lean_source_equality_local_bytes`, `lean_and_generated_records_survive`,
`per_file_extraction_audit`, `dedup_audit_present`, `generator_audit_seed_and_quota`,
`all_packed_tokens_reencoded_source_shift_padding_masks_and_attribution`, `statistics_match_artifacts`).
`tools/verify_google_math.py` is the independent check against the delivered files, and its output is
[`audit/validation-evidence.json`](audit/validation-evidence.json). It decoded 96 packed rows spread over
every shard of all three splits, and for each row (a) confirmed every row slice equals the re-encoding of
the corresponding span of `[BOS] + SentencePiece(text) + [EOS]` for its attributed source record and
(b) confirmed the decoded text equals the decoded source text modulo whitespace. It recorded 8 literal
unified diffs of a decoded row against its source record; every diff had 0 lines, whitespace identical.
It round-tripped 3 Parquet rows (indices 0, 239,658 and 479,315) back to their JSONL records with exact
match, and confirmed a `formal_lean_proof`, a `formal_lean_statement`, a `generated_qa_pair`, a
`formal_geometry_problem`, a `natural_language_proof` and a `research_notebook` record all survive with
matching `text_sha256`. Split separation was recomputed from the data rather than trusted: 3,403 groups
in the audit and 3,403 in the data, no group spanning two splits, train/dev/test at 3,333/35/35 groups.
## Publication
The component is published as Hugging Face dataset pull request
[MMLA-ORG/Cayley-Math discussion #3](https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/discussions/3),
on ref `refs/pr/3`, commit `99f09edf1f380be5dd8074c621b551aea67ad5a0`. It has not been merged into `main`.
A token with `write` scope or org-admin rights is required to merge it. The token used for this build can
read the repository and create the pull request, but direct commits to `main` are rejected with
`403 Forbidden: You have read access but not the required permissions for this operation`. Merging, or re-running
`hf upload` without `--create-pr` once such a token is available, completes the submission.
## Limitations
- No Lean compilation, proof checking, Comparator run or mathematical verification of any kind.
- The build ran in a sandbox-free subprocess only for the generator; `cryptids`, `formal-conjectures`,
`formal-imo`, `formal-putnam-like`, `alphageometry`, `alphaproof-nexus` and the notebooks were never
built or executed, only read as text.
- PDF prose proofs are noisy pypdf text; symbol and ligature fidelity is not guaranteed.
- The generator subset is a seeded sample of one curriculum, not the paper's full dataset, and its
questions are school-level rather than research mathematics. One module, `algebra__linear_1d`, stopped
at 3,055 of its 8,000 quota when upstream's own `assert entropy == 0` fired; nothing was repaired, and
the shortfall is recorded in `audit/mathematics-dataset-generation.json`.
- `formal-conjectures` upstream states that some content derives from third-party sources (Wikipedia,
MathOverflow and OEIS under CC-BY-SA 4.0). Per-file reference URLs are retained inside the Lean text; no
separate per-file license audit was performed.
- Near-duplicate detection, Phase 1 overlap analysis and editorial review were not performed. 1,629
exact duplicate texts were dropped and are listed in `audit/excluded-duplicates.jsonl`.
- The SentencePiece model normalises runs of spaces, so packed text decodes whitespace-normalised rather
than byte-identically to the sources.
- Only problems that upstream published are present, so several subsets are selection-biased. `cryptids`
contributes 31.6M of the 78.4M characters because a handful of Busy Beaver traces are enormous, so the
corpus is dominated by machine-generated proof script rather than mathematics prose.
## Statistics
~~~json
{
"component": "google-math",
"sources": 8,
"source_commits": {
"alphaproof-nexus": "0647711a71183c1ea492ad60860776617ce1ea88",
"formal-conjectures": "6751c0f043985ac6647db28a45a63afdf21acffe",
"formal-imo": "a85d84b97c8e88a352c136b1e29c355139609bea",
"formal-putnam-like": "636c2d46b3a7d7e77a9f58209b0c17816f8d172b",
"cryptids": "3647718b5df22e1a61462dec0376bd204526285a",
"mathematics-dataset": "427f45075f84b8b9774950196ad63867ca20ffb3",
"alphageometry": "6777cb586cbb46beed28db12dc72c69770b68337",
"mathematics-conjectures": "8302f29101019434e5a0a25807a4e42cd9d96414"
},
"records_before_deduplication": 482213,
"records": 480584,
"excluded_exact_duplicates": 1629,
"groups": 3403,
"record_types": {
"geometry_rule_table": 2,
"formal_geometry_problem": 265,
"formal_lean_proof": 535,
"natural_language_proof": 12,
"formal_lean_statement": 2492,
"research_notebook": 2,
"generated_qa_pair": 477276
},
"text_characters": 78364888,
"per_source": {
"alphaproof-nexus": {
"records": 83,
"groups": 51,
"record_types": {
"formal_lean_proof": 71,
"natural_language_proof": 12
},
"text_characters": 2229771,
"upstream_files": 83,
"split_records": {
"train": 83,
"dev": 0,
"test": 0
}
},
"formal-conjectures": {
"records": 2135,
"groups": 2135,
"record_types": {
"formal_lean_statement": 2135
},
"text_characters": 8133507,
"upstream_files": 2135,
"split_records": {
"train": 2088,
"dev": 23,
"test": 24
}
},
"formal-imo": {
"records": 261,
"groups": 261,
"record_types": {
"formal_lean_statement": 261
},
"text_characters": 420948,
"upstream_files": 261,
"split_records": {
"train": 252,
"dev": 6,
"test": 3
}
},
"formal-putnam-like": {
"records": 96,
"groups": 96,
"record_types": {
"formal_lean_statement": 96
},
"text_characters": 117201,
"upstream_files": 96,
"split_records": {
"train": 93,
"dev": 1,
"test": 2
}
},
"cryptids": {
"records": 464,
"groups": 464,
"record_types": {
"formal_lean_proof": 464
},
"text_characters": 31638592,
"upstream_files": 464,
"split_records": {
"train": 457,
"dev": 2,
"test": 5
}
},
"mathematics-dataset": {
"records": 477276,
"groups": 127,
"record_types": {
"generated_qa_pair": 477276
},
"text_characters": 35710731,
"upstream_files": 1,
"split_records": {
"train": 476076,
"dev": 1200,
"test": 0
}
},
"alphageometry": {
"records": 267,
"groups": 267,
"record_types": {
"geometry_rule_table": 2,
"formal_geometry_problem": 265
},
"text_characters": 70907,
"upstream_files": 5,
"split_records": {
"train": 265,
"dev": 1,
"test": 1
}
},
"mathematics-conjectures": {
"records": 2,
"groups": 2,
"record_types": {
"research_notebook": 2
},
"text_characters": 43231,
"upstream_files": 2,
"split_records": {
"train": 2,
"dev": 0,
"test": 0
}
}
},
"splits": {
"train": {
"records": 479316,
"groups": 3333,
"sources": 8,
"text_characters": 77705566
},
"dev": {
"records": 1233,
"groups": 35,
"sources": 6,
"text_characters": 233131
},
"test": {
"records": 35,
"groups": 35,
"sources": 5,
"text_characters": 426191
}
},
"packed": {
"train": {
"rows": 17691,
"tokens": 36230380,
"loss_bearing_labels": 35751064
},
"dev": {
"rows": 51,
"tokens": 102872,
"loss_bearing_labels": 101639
},
"test": {
"rows": 91,
"tokens": 185977,
"loss_bearing_labels": 185942
}
},
"mathematical_verification": "not_performed",
"lean_compilation": "not_performed"
}
~~~