Datasets:
Download components/google-math/tools/build_google_math.py from MMLA-ORG/Cayley-Math: direct link, hf CLI and curl.
- Browser
- Download file 93.1 kB
-
https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/resolve/refs%2Fpr%2F9/components/google-math/tools/build_google_math.py
- Command line
-
hf download hf://datasets/MMLA-ORG/Cayley-Math@refs/pr/9/components/google-math/tools/build_google_math.py
-
curl -L -o build_google_math.py https://huggingface.co/datasets/MMLA-ORG/Cayley-Math/resolve/refs%2Fpr%2F9/components/google-math/tools/build_google_math.py
93.1 kB
#!/usr/bin/env python3
"""Build components/google-math: Google math-paper artifacts in Cayley-Math format.
One component, eight upstream subsets. Mirrors components/openai-math exactly:
same record core fields, same Parquet encoding, same SHA256-ordered group split
policy, and the same mmla-type2-packed-v1 packing code.
Nothing here verifies mathematics. No Lean project is compiled, no proof is
checked, and generated QA pairs are generator output rather than research
mathematics.
"""
from __future__ import annotations
import argparse
import gzip
import hashlib
import json
import math
import re
import shutil
import subprocess
from collections import Counter, deque
from pathlib import Path
COMPONENT = 'google-math'
CORE = ('id', 'text', 'group', 'license', 'paper_id', 'split', 'record_type', 'source_url')
SPLITS = ('train', 'dev', 'test')
SEQ_LEN = 2048
SHARD_ROWS = 256
IGNORED_LOCAL = {'.DS_Store', 'manifest.json', 'validation.json'}
# ---------------------------------------------------------------------------
# Source registry. Every included source is a Google or Google DeepMind
# publication that ships math text, and every entry pins a fixed revision.
# ---------------------------------------------------------------------------
SOURCES = {
'alphaproof-nexus': {
'name': 'AlphaProof Nexus results',
'repository': 'https://github.com/google-deepmind/alphaproof-nexus-results',
'commit': '0647711a71183c1ea492ad60860776617ce1ea88',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Lean math proofs generated by AlphaProof Nexus and accompanying natural language prose proofs',
'checkout': 'alphaproof-nexus-results',
'dir': 'alphaproof-nexus-results',
'math_content': 'proofs',
},
'formal-conjectures': {
'name': 'Formal Conjectures',
'repository': 'https://github.com/google-deepmind/formal-conjectures',
'commit': '6751c0f043985ac6647db28a45a63afdf21acffe',
'license': 'Apache-2.0',
'creators': ['The Formal Conjectures Authors', 'Google LLC'],
'paper': 'Formalised statements of mathematical conjectures in Lean, arXiv:2605.13171',
'checkout': 'formal-conjectures',
'dir': 'formal-conjectures',
'math_content': 'formal_statements',
},
'formal-imo': {
'name': 'formal-imo benchmark',
'repository': 'https://github.com/google-deepmind/formal-imo',
'commit': 'a85d84b97c8e88a352c136b1e29c355139609bea',
'license': 'Apache-2.0',
'creators': ['Google LLC'],
'paper': 'Lean 4 formalizations of International Mathematical Olympiad problem statements',
'checkout': 'formal-imo',
'dir': 'formal-imo',
'math_content': 'formal_statements',
},
'formal-putnam-like': {
'name': 'Putnam-like problems in Lean',
'repository': 'https://github.com/google-deepmind/formal-putnam-like',
'commit': '636c2d46b3a7d7e77a9f58209b0c17816f8d172b',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Lean 4 formalizations of Putnam-like problems (eval_hub putnam_like companion)',
'checkout': 'formal-putnam-like',
'dir': 'formal-putnam-like',
'math_content': 'formal_statements',
},
'cryptids': {
'name': 'Cryptids: conjecture-checking and Busy Beaver Turing machines',
'repository': 'https://github.com/google-deepmind/cryptids',
'commit': '3647718b5df22e1a61462dec0376bd204526285a',
'commit_source': 'the checkout as of this build; recorded because the repository moves',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Certified Turing-machine reductions and Busy Beaver proofs, arXiv:2610.22603',
'checkout': 'cryptids',
'dir': 'cryptids',
'math_content': 'proofs',
},
'mathematics-dataset': {
'name': 'Mathematics Dataset generator',
'repository': 'https://github.com/google-deepmind/mathematics_dataset',
'commit': '427f45075f84b8b9774950196ad63867ca20ffb3',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Analysing Mathematical Reasoning Abilities of Neural Models (Saxton, Grefenstette, Hill, Kohli), ICLR 2018',
'checkout': 'mathematics_dataset',
'generated': 'mathematics_dataset_sample.jsonl',
'math_content': 'generated_qa',
},
'alphageometry': {
'name': 'AlphaGeometry geometry problem artifacts',
'repository': 'https://github.com/google-deepmind/alphageometry',
'commit': '6777cb586cbb46beed28db12dc72c69770b68337',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Solving Olympiad Geometry without Human Demonstrations, Nature 2024',
'checkout': 'alphageometry',
'dir': 'alphageometry',
'math_content': 'problem_statements',
},
'mathematics-conjectures': {
'name': 'Mathematics conjectures companion notebooks',
'repository': 'https://github.com/google-deepmind/mathematics_conjectures',
'commit': '8302f29101019434e5a0a25807a4e42cd9d96414',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Advancing mathematics by guiding human intuition with AI, Nature 2021',
'checkout': 'mathematics_conjectures',
'dir': 'mathematics_conjectures',
'math_content': 'natural_language_math',
},
# --- Added after the widened scan's three undecided candidates were measured. ---
# Only the notebook artifacts of these repositories ship mathematical text: the
# problem statements, the discovered constructions and the constructions' code.
# Search harnesses, compressed caches and serialised index dumps are excluded
# below with their measured sizes.
'alphaevolve': {
'name': 'AlphaEvolve mathematical constructions',
'repository': 'https://github.com/google-deepmind/alphaevolve',
'commit': 'c42694909d50f92cb01c2d50916f022b1c7e74d1',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'AlphaEvolve: A coding agent for scientific and algorithmic discovery',
'checkout': 'alphaevolve',
'dir': 'alphaevolve',
'area': 'maths/**/verification.ipynb only',
'math_content': 'constructions',
'notebook_root': 'maths',
},
'alphaevolve_results': {
'name': 'AlphaEvolve results notebook',
'repository': 'https://github.com/google-deepmind/alphaevolve_results',
'commit': '4226acbf237ff9ad10ba7673a2af127a2d8a5971',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'AlphaEvolve results: improved constructions and bounds reported in the AlphaEvolve paper',
'checkout': 'alphaevolve_results',
'dir': 'alphaevolve_results',
'area': 'mathematical_results.ipynb only',
'math_content': 'constructions',
'notebook_root': '',
},
'funsearch': {
'name': 'FunSearch mathematical discoveries',
'repository': 'https://github.com/google-deepmind/funsearch',
'commit': 'cc53f274237d7ab05c19df939edbc1f9616a7c19',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Mathematical discoveries from program search with large language models, Nature 2023',
'checkout': 'funsearch',
'dir': 'funsearch',
'area': 'cap_set, admissible_set, bin_packing, cyclic_graphs notebooks only',
'math_content': 'constructions',
'notebook_root': 'problems',
},
}
# Candidates found by the widened scan and deliberately not included.
EXCLUDED_CANDIDATES = [
{'repository': 'google-deepmind/eval_hub', 'area': 'eval_hub/putnam_like, eval_hub/live_math',
'reason': 'Every problem directory carries a CANARY_ID benchmark-contamination canary '
'(96 putnam_like problems across Sets 1-8, 250 live_math problems across regular/difficult/multimodal). '
'live_math is an actively maintained evaluation benchmark whose multimodal problems depend on images '
'that are not part of this text corpus. Copying them into a pretraining corpus would poison a public '
'benchmark, so the whole repository is excluded despite being Google-published and Apache-2.0.'},
{'repository': 'deepmind/aqua_rat', 'area': 'Hugging Face dataset',
'reason': 'ACC-AQUA-RAT accompanies "Program Synthesis with Large Language Models", a code-generation paper, '
'not a mathematics paper. The 97,467 rows are grade-school arithmetic word problems, so it fails the '
'mathematics-paper test even though the license is Apache-2.0.'},
{'repository': 'deepmind/math_dataset', 'area': 'Hugging Face dataset, 40.0 GB',
'reason': 'A pre-generated mirror of the same mathematics_dataset paper already represented here. Included '
'through the official generator at a pinned commit instead, which keeps provenance exact and avoids '
'pulling 40 GB over the wire.'},
{'repository': 'google-research/math-opt-benchmark',
'area': 'C++ Bazel workspace',
'reason': 'Ships no math text: only C++ sources plus download_data.sh, which fetches external instances at build '
'time. Code-only utility with no shipped statements, proofs or generator.'},
{'repository': 'google/mathsteps', 'area': 'JavaScript library',
'reason': 'A step-by-step equation solver library with 76 lib and 55 test files and no shipped math corpus. '
'It is an equation-solving utility, not an artifact of a mathematics paper.'},
{'repository': 'google-deepmind/cartesian-frames', 'area': 'HOL4 .sml proof scripts',
'reason': 'Formalises the Cartesian Frames sequence from the Alignment Forum, which is a philosophy-of-agency '
'blog series rather than a mathematics paper, and it is HOL4 SML proof script rather than math text.'},
{'repository': 'google-deepmind/miniF2F', 'area': 'Lean 4 fork',
'reason': 'A fork of openai/miniF2F. Only the Lean 4 adaptation is Google work; the corpus itself is not a '
'Google publication, so including it would import a non-Google mirror.'},
{'repository': 'google-deepmind/alphageometry2', 'area': 'DDAR solver, 42 KB',
'reason': 'The 26 olympiad problems live inside Python dict literals in test.py with point coordinates, so they '
'are code text rather than a math dataset. alphageometry (AG1) already ships 261 olympiad geometry '
'problems in the AG language verbatim, so AG2 would add solver code and no new math.'},
{'repository': 'google-deepmind/alphatensor', 'area': 'Python/NumPy',
'reason': '24 entries of matrix-multiplication search code and README files. No math statements, proofs or '
'text-bearing artifacts to project.'},
{'repository': 'google/feabench', 'area': 'Multiphysics benchmark',
'reason': 'NeurIPS 2024 MATH-AI workshop benchmark for multiphysics reasoning, i.e. physics rather than '
'mathematics, and its data ships separately.'},
{'repository': 'google/phasecoder', 'area': 'Audio research code',
'reason': 'Microphone geometry for spatial audio. The word geometry refers to microphone arrays, not mathematics.'},
{'repository': 'google-research/pointdit', 'area': 'Diffusion model code',
'reason': 'Monocular geometry estimation for images. Not a mathematics paper.'},
{'repository': 'google/tree-math, google/mathfu, google/vector_math.dart, google-research/opengrm',
'area': 'Software libraries',
'reason': 'Math-named software libraries (JAX pytree ops, a C++ game math library, a Dart vector package, n-gram '
'libraries). Code-only utilities with no math text and no mathematics paper.'},
{'repository': 'google/research/mathlists, google-deepmind/isabelle_llm',
'area': 'Probed repositories',
'reason': 'Neither resolves: no repository exists at these paths, so there is nothing to include.'},
]
MATH_CONTENT = {
'formal_statements': 'formal_statement',
'proofs': 'proof',
'natural_language_math': 'natural_language_math',
'generated_qa': 'generated_qa',
'constructions': 'construction',
'problem_statements': 'problem_statement',
'none': 'none',
}
# Sub-paths of included repositories that were measured and left out. Kept separate from
# `EXCLUDED_CANDIDATES` so that `included` and `excluded_candidates` stay disjoint: every
# repository appears in exactly one of them.
EXCLUDED_SUBPATHS = [
{'repository': 'google-deepmind/alphaevolve',
'included_source_id': 'alphaevolve',
'area': 'omega/',
'math_content': 'none',
'measured_bytes': 45511025,
'reason': 'Excluded on measured evidence: the omega/ directory a first scan flagged is a search harness plus '
'one compressed cache. 44,267,872 of its 45,511,025 bytes are the single .xz archive '
'clean_level4.npz.xz; its 1,126,802-byte verification notebook is 1,098,419 characters of code of '
'which 836,374 characters are decimal digits, i.e. rational vectors written as Python literals for a '
'20 GB verification run. No statements, proofs or prose to project.'},
{'repository': 'google-deepmind/alphaevolve',
'included_source_id': 'alphaevolve',
'area': 'src/, maths/**/*.py, maths/**/config*.yaml, matmul_tensor_decompositions/',
'math_content': 'none',
'measured_bytes': 1165195,
'reason': 'Excluded on measured evidence: src/ is the 11-file, 58,453-byte AlphaEvolve search harness with no '
'math text; maths/**/*.py is 87,725 bytes of initial search programs and maths/**/config*.yaml is '
'60,949 bytes of run configuration; matmul_tensor_decompositions/ is 827,894 bytes of two '
'notebooks whose projections share 4,820 of 4,830 substantial lines with the included '
'alphaevolve_results notebook, so they would be near-duplicates.'},
{'repository': 'google-deepmind/funsearch',
'included_source_id': 'funsearch',
'area': '**/*.txt, implementation/*.py',
'math_content': 'none',
'measured_bytes': 20442546,
'reason': 'Excluded on measured evidence: 13 .txt files hold 20,389,203 bytes of comma-separated integer tuples '
'(admissible_set_n24_w17_size237984.txt alone is 17,372,832 bytes). They serialise exactly the sets the '
'included notebooks build and assert, so they add serialization rather than math text. '
'implementation/ is 53,343 bytes of the FunSearch evolutionary algorithm, a program-search harness.'},
]
PDF_ARTIFACT_RANGES = (
('c0_c1_control', 0x00, 0x1F),
('c1_control', 0x7F, 0x9F),
('replacement_char', 0xFFFD, 0xFFFD),
('private_use', 0xE000, 0xF8FF),
('ligature_block', 0xFB00, 0xFB06),
('combining_mark', 0x0300, 0x036F),
)
def canonical(value):
return json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(',', ':'))
def sha(path):
h = hashlib.sha256()
with Path(path).open('rb') as f:
for chunk in iter(lambda: f.read(1024 * 1024), b''):
h.update(chunk)
return h.hexdigest()
def write_json(path, obj):
Path(path).parent.mkdir(parents=True, exist_ok=True)
Path(path).write_text(json.dumps(obj, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
def write_jsonl(path, rows):
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
if path.suffix == '.gz':
with path.open('wb') as raw, gzip.GzipFile(filename='', mode='wb', fileobj=raw, mtime=0) as gz:
for row in rows:
gz.write((canonical(row) + '\n').encode('utf-8'))
else:
with path.open('w', encoding='utf-8') as f:
for row in rows:
f.write(canonical(row) + '\n')
def write_parquet(path, rows):
import pyarrow as pa
import pyarrow.parquet as pq
columns = {key: [str(row.get(key) or '') for row in rows] for key in CORE}
columns['metadata_json'] = [canonical({k: v for k, v in row.items() if k not in CORE}) for row in rows]
schema = pa.schema([(key, pa.string()) for key in (*CORE, 'metadata_json')])
pq.write_table(pa.Table.from_pydict(columns, schema=schema), path, compression='zstd', row_group_size=1024)
def split_groups(groups):
"""Identical policy to components/openai-math/tools/finalize_openai_math.py."""
ordered = sorted(groups, key=lambda g: (hashlib.sha256(g.encode()).hexdigest(), g))
if len(ordered) < 3:
raise ValueError('Need at least 3 groups for nonempty leakage-safe splits')
holdout = max(1, math.ceil(len(ordered) * .01))
if 2 * holdout >= len(ordered):
holdout = 1
return {g: ('dev' if i < holdout else 'test' if i < 2 * holdout else 'train') for i, g in enumerate(ordered)}
def pack_split(root, split, records, sp, seq_len=SEQ_LEN, shard_rows=SHARD_ROWS):
"""Copied verbatim in behaviour from components/openai-math/tools/finalize_openai_math.py."""
import numpy as np
packed = root / 'packed'
packed.mkdir(exist_ok=True)
shard_entries, pending_rows, pending_labels, pending_attrs = [], [], [], []
buffer, docs, stream_start, total_tokens, label_count, row_number = [], deque(), 0, 0, 0, 0
attr_file = (packed / f'{split}-rows.jsonl').open('w', encoding='utf-8')
token_file = (packed / f'{split}-records.jsonl').open('w', encoding='utf-8')
def flush():
nonlocal pending_rows, pending_labels, pending_attrs
if not pending_rows:
return
shard = len(shard_entries)
fields = {}
arrays = {'input_ids': np.asarray(pending_rows, dtype=np.int32),
'labels': np.asarray(pending_labels, dtype=np.int32),
'role_ids': np.zeros((len(pending_rows), seq_len), dtype=np.int32)}
for name, array in arrays.items():
path = packed / f'{split}-{shard:06d}-{name}.npy'
np.save(path, array, allow_pickle=False)
fields[name] = {'path': path.name, 'sha256': sha(path), 'bytes': path.stat().st_size,
'dtype': 'int32', 'shape': list(array.shape)}
shard_entries.append({'count': len(pending_rows), 'fields': fields})
for index, attr in enumerate(pending_attrs):
attr.update(shard=shard, shard_row=index)
attr_file.write(canonical(attr) + '\n')
pending_rows, pending_labels, pending_attrs = [], [], []
def emit():
nonlocal buffer, stream_start, label_count, row_number
n = min(seq_len, len(buffer))
values = buffer[:n]
targets = buffer[1:min(n + 1, len(buffer))]
if len(targets) < n:
targets.append(-100)
targets = [-100 if t == 1 else t for t in targets]
label_count += sum(t != -100 for t in targets)
pending_rows.append(values + [0] * (seq_len - n))
pending_labels.append(targets + [-100] * (seq_len - n))
stop = stream_start + n
spans = []
for doc in docs:
lo, hi = max(stream_start, doc['start']), min(stop, doc['end'])
if lo < hi:
spans.append({'record_id': doc['id'], 'paper_id': doc['paper_id'],
'row_start': lo - stream_start, 'row_end': hi - stream_start,
'record_start': lo - doc['start'], 'record_end': hi - doc['start']})
pending_attrs.append({'row': row_number, 'stream_start': stream_start, 'length': n, 'segments': spans})
row_number += 1
stream_start = stop
del buffer[:n]
while docs and docs[0]['end'] <= stream_start:
docs.popleft()
if len(pending_rows) == shard_rows:
flush()
for record in records:
tokens = [1] + sp.encode(record['text'], out_type=int) + [2]
array = np.asarray(tokens, dtype='<i4')
token_file.write(canonical({'id': record['id'], 'paper_id': record['paper_id'], 'tokens': len(tokens),
'token_sha256': hashlib.sha256(array.tobytes()).hexdigest()}) + '\n')
docs.append({'id': record['id'], 'paper_id': record['paper_id'], 'start': total_tokens,
'end': total_tokens + len(tokens)})
total_tokens += len(tokens)
buffer.extend(tokens)
while len(buffer) >= seq_len + 1:
emit()
if buffer:
emit()
flush()
attr_file.close()
token_file.close()
return shard_entries, {'rows': row_number, 'tokens': total_tokens, 'loss_bearing_labels': label_count}
def write_manifest(root):
entries = []
for path in sorted(root.rglob('*')):
relative = path.relative_to(root).as_posix()
if path.is_file() and relative not in IGNORED_LOCAL and '.tmp' not in path.name:
entries.append({'path': relative, 'bytes': path.stat().st_size, 'sha256': sha(path)})
write_json(root / 'manifest.json', {'format': 'cayley-contribution-file-manifest-v1',
'excludes': sorted(IGNORED_LOCAL), 'files': entries})
def extract_pdf(path):
import pypdf
reader = pypdf.PdfReader(str(path))
pages = [page.extract_text() or '' for page in reader.pages]
raw = '\n'.join(pages)
control = sum(1 for c in raw if (ord(c) < 32 and c not in '\n\t') or 0x7F <= ord(c) <= 0x9F)
text = ''.join(' ' if (ord(c) < 32 and c not in '\n\t') or 0x7F <= ord(c) <= 0x9F else c for c in raw)
text = '\n'.join(line.rstrip() for line in text.split('\n'))
text = re.sub(r'\n{3,}', '\n\n', text).strip() + '\n'
non_ascii = sum(1 for c in text if ord(c) > 127)
artifacts = Counter()
for name, lo, hi in PDF_ARTIFACT_RANGES:
artifacts[name] = sum(1 for c in text if lo <= ord(c) <= hi)
return {
'text': text,
'pages': len(pages),
'raw_characters': len(raw),
'characters': len(text),
'page_characters': [len(page) for page in pages],
'non_ascii_characters': non_ascii,
'non_ascii_ratio': round(non_ascii / len(text), 6) if text else 0.0,
'control_characters_replaced': control,
'artifact_characters': {k: v for k, v in artifacts.items() if v},
}
# ---------------------------------------------------------------------------
# Per-source record builders.
# ---------------------------------------------------------------------------
def base_record(source_id, group, paper_id, record_type, source_url, path, digest, text, license_id):
return {
'id': None,
'text': text,
'group': group,
'license': license_id,
'paper_id': paper_id,
'split': None,
'record_type': record_type,
'source_url': source_url,
'source_id': source_id,
'source_path': path,
'source_sha256': digest,
'text_sha256': hashlib.sha256(text.encode('utf-8')).hexdigest(),
'creators': SOURCES[source_id]['creators'],
'math_content': SOURCES[source_id].get('math_content', 'none'),
'verification_status': 'unverified_source_claim',
'lean_verified': False,
'mathematical_verification': 'not_performed',
}
def build_alphaproof_nexus(root, source_id, checkout, files, seen):
"""The staged AlphaProof Nexus projection, reused unchanged (83 records)."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records = []
for rel in files:
path = checkout / rel
digest = sha(path)
stem = Path(rel).stem
if rel.endswith('.lean'):
text = path.read_text(encoding='utf-8')
group = lean_group(rel)
record_type, origin = 'formal_lean_proof', 'source_verbatim'
quality, transform = 'verbatim_source_text', 'identity_verbatim_bytes'
metrics = {'source_characters': len(text)}
elif rel.endswith('.pdf'):
extracted = extract_pdf(path)
group = pdf_group(rel)
text = extracted.pop('text')
record_type, origin = 'natural_language_proof', 'source_pdf_extraction'
quality = 'lossy_pdf_text_layer'
transform = 'pypdf_page_text_control_chars_replaced_blank_runs_collapsed'
metrics = extracted
global EXTRACTION_AUDIT
EXTRACTION_AUDIT.append({'source_id': source_id, 'source_path': rel, **metrics,
'extraction_quality': quality,
'source_url': f'{repo}/blob/{commit}/{rel}'})
else:
continue
kind = 'lean' if rel.endswith('.lean') else 'nl'
upstream_group = group
group = f'{source_id}::{upstream_group}'
record = base_record(source_id, group, upstream_group, record_type,
f'{repo}/blob/{commit}/{rel}', rel, digest, text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{upstream_group}:{kind}:{stem}'
record['upstream_group'] = upstream_group
record['content_origin'] = origin
record['transform'] = transform
record['extraction_quality'] = quality
record['extraction_metrics'] = metrics
record['upstream_category'] = rel.split('/')[1]
record['source_commit'] = commit
records.append(record)
return records
def lean_group(rel):
"""Problem-stem group for an APNOutputs Lean file (unchanged from the staged build)."""
parts = Path(rel).parts
category, stem = parts[1], Path(rel).stem
if category == 'ErdosProblems':
return 'erdos_' + stem.split('_')[1].split('.')[0]
if category == 'OEIS':
tokens = stem.split('_')
ident = tokens[1] if tokens[0].lower() == 'oeis' else tokens[0]
return 'oeis_' + ident.lower()
if category == 'StacksProject':
return 'stacks_project'
if category == 'AICollaborator':
subject = parts[2]
if stem.startswith('hilbert_functions'):
return 'ai_hilbert_functions'
if stem == 'bipartite_graph_reconstruction_conjecture_2':
return 'ai_graph_reconstruction'
return 'ai_' + subject.lower()
raise ValueError(f'Unknown Lean category: {rel}')
def pdf_group(rel):
"""Problem-stem group for a NaturalLanguageProofs PDF (unchanged from the staged build)."""
category, stem = Path(rel).parts[1], Path(rel).stem
if category == 'ErdosProblems':
match = re.fullmatch(r'erdos(\d+)(i{1,2})?', stem)
if not match:
raise ValueError(f'Unparsed Erdos PDF name: {rel}')
return 'erdos_' + match.group(1)
if category == 'OEIS':
match = re.fullmatch(r'oeis_(\d+)', stem.lower())
if not match:
raise ValueError(f'Unparsed OEIS PDF name: {rel}')
return 'oeis_' + match.group(1)
if category == 'AICollaborator':
if stem == 'hilbert':
return 'ai_hilbert_functions'
if stem == 'graph_reconstruction':
return 'ai_graph_reconstruction'
raise ValueError(f'Unpaired PDF name: {rel}')
def build_lean_corpus(source_id, checkout, files, record_type):
"""Verbatim projection of Lean source text, one record per file."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records = []
for rel in files:
path = checkout / rel
text = path.read_text(encoding='utf-8')
if not text.strip():
continue
group = f'{source_id}::{rel[:-len(".lean")]}'
record = base_record(source_id, group, group, record_type,
f'{repo}/blob/{commit}/{rel}', rel, sha(path), text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{rel[:-len(".lean")]}'
record['content_origin'] = 'source_verbatim'
record['transform'] = 'identity_verbatim_bytes'
record['extraction_quality'] = 'verbatim_source_text'
record['source_commit'] = commit
record['source_characters'] = len(text)
records.append(record)
return records
def build_alphageometry(source_id, checkout):
"""Verbatim problem statements from the AG language text artifacts."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records = []
pairs = {'examples.txt': 'library_example', 'imo_ag_30.txt': 'olympiad_geometry_problem',
'jgex_ag_231.txt': 'jgex_geometry_problem'}
for name, kind in pairs.items():
path = checkout / name
lines = [line.strip() for line in path.read_text(encoding='utf-8').splitlines()]
body = [line for line in lines if line and not line.startswith('#')]
if len(body) % 2:
raise ValueError(f'{name}: odd number of problem lines ({len(body)})')
for index in range(0, len(body), 2):
problem_id, statement = body[index], body[index + 1]
text = f'{problem_id}\n{statement}\n'
group = f'{source_id}::{name}::{problem_id}'
record = base_record(source_id, group, group, 'formal_geometry_problem',
f'{repo}/blob/{commit}/{name}', name, sha(path), text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{name}:{problem_id}'
record['content_origin'] = 'source_verbatim'
record['transform'] = 'identity_verbatim_lines'
record['extraction_quality'] = 'verbatim_source_text'
record['source_commit'] = commit
record['geometry_problem_id'] = problem_id
record['geometry_problem_kind'] = kind
record['source_line'] = index + 1
records.append(record)
for name in ('defs.txt', 'rules.txt'):
path = checkout / name
text = path.read_text(encoding='utf-8')
group = f'{source_id}::{name}'
record = base_record(source_id, group, group, 'geometry_rule_table',
f'{repo}/blob/{commit}/{name}', name, sha(path), text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{name}'
record['content_origin'] = 'source_verbatim'
record['transform'] = 'identity_verbatim_bytes'
record['extraction_quality'] = 'verbatim_source_text'
record['source_commit'] = commit
records.append(record)
return records
def build_mathematics_dataset(source_id, generated, params):
"""Seeded projection of the official generator's curriculum."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records = []
for row in generated:
group = f'{source_id}::{row["regime"]}::{row["module"]}'
text = row['question'] + '\n' + row['answer'] + '\n'
record = base_record(source_id, group, row['module'], 'generated_qa_pair',
f'{repo}/tree/{commit}', f'generated/{row["regime"]}/{row["module"]}',
None, text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{row["regime"]}:{row["module"]}:{row["index"]:05d}'
record['content_origin'] = 'generator_output'
record['transform'] = 'identity_generator_question_and_answer_lines'
record['extraction_quality'] = 'verbatim_generator_output'
record['source_commit'] = commit
record['generator_regime'] = row['regime']
record['generator_module'] = row['module']
record['generator_index'] = row['index']
record['generator_seed'] = params['seed']
record['question'] = row['question']
record['answer'] = row['answer']
records.append(record)
return records
def build_notebooks(source_id, checkout):
"""Notebook cell sources joined by newlines, verbatim."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records = []
for path in sorted(checkout.glob('*.ipynb')):
notebook = json.loads(path.read_text(encoding='utf-8'))
cells = [(cell['cell_type'], ''.join(cell['source'])) for cell in notebook['cells']]
text = '\n'.join(content for _, content in cells)
group = f'{source_id}::{path.name}'
record = base_record(source_id, group, path.stem, 'research_notebook',
f'{repo}/blob/{commit}/{path.name}', path.name, sha(path), text, source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{path.stem}'
record['content_origin'] = 'source_notebook_projection'
record['transform'] = 'notebook_cell_sources_joined_by_newline'
record['extraction_quality'] = 'verbatim_cell_text'
record['source_commit'] = commit
record['cell_counts'] = dict(Counter(kind for kind, _ in cells))
EXTRACTION_AUDIT.append({'source_id': source_id, 'source_path': path.name,
'cells': len(cells), 'characters': len(text),
'markdown_cells': sum(1 for k, _ in cells if k == 'markdown'),
'code_cells': sum(1 for k, _ in cells if k == 'code'),
'extraction_quality': 'verbatim_cell_text',
'source_url': f'{repo}/blob/{commit}/{path.name}'})
records.append(record)
return records
def strip_notebook_outputs(path, relative):
"""Return the notebook's cell sources with every output and execution counter removed.
A record must carry the notebook's source text, never executed output or embedded
base64. When a notebook has output blobs, the projection drops only those, keeps
every cell's code and markdown verbatim, and records the byte cost of what it
stripped so the drop is measurable rather than silent.
"""
notebook = json.loads(Path(path).read_text(encoding='utf-8'))
cells = [(cell['cell_type'], ''.join(cell['source'])) for cell in notebook['cells']]
output_cells = sum(1 for cell in notebook['cells'] if cell.get('outputs'))
output_count = sum(len(cell.get('outputs') or []) for cell in notebook['cells'])
output_bytes = sum(len(json.dumps(cell.get('outputs') or [], ensure_ascii=False).encode('utf-8'))
for cell in notebook['cells'])
image_bytes = 0
for cell in notebook['cells']:
for output in cell.get('outputs') or []:
for mime, value in (output.get('data') or {}).items():
if mime.startswith('image/'):
image_bytes += len(value if isinstance(value, str) else ''.join(value))
text = '\n'.join(content for _, content in cells)
stripped = {
'cells_with_outputs': output_cells,
'output_objects_removed': output_count,
'output_bytes_removed': output_bytes,
'embedded_image_bytes_removed': image_bytes,
'projection_rule': 'cell sources joined by a newline; outputs and execution counters dropped',
}
return text, cells, stripped
def build_notebook_files(source_id, checkout, files):
"""Verbatim cell-source projection of the notebooks a source actually ships."""
source = SOURCES[source_id]
commit, repo = source['commit'], source['repository']
records, audit = [], []
for relative in files:
path = checkout / relative
text, cells, stripped = strip_notebook_outputs(path, relative)
group = f'{source_id}::{relative}'
record = base_record(source_id, group, relative, 'research_notebook',
f'{repo}/blob/{commit}/{relative}', relative, sha(path), text,
source['license'])
record['id'] = f'{COMPONENT}:{source_id}:{relative[:-len(".ipynb")]}'
record['content_origin'] = 'source_notebook_projection'
record['transform'] = 'notebook_cell_sources_joined_by_newline_outputs_stripped'
record['extraction_quality'] = 'verbatim_cell_text'
record['source_commit'] = commit
record['math_content'] = source['math_content']
record['cell_counts'] = dict(Counter(kind for kind, _ in cells))
record['stripped_outputs'] = stripped
records.append(record)
audit.append({'source_id': source_id, 'source_path': relative, 'cells': len(cells),
'characters': len(text),
'markdown_cells': sum(1 for k, _ in cells if k == 'markdown'),
'code_cells': sum(1 for k, _ in cells if k == 'code'),
'extraction_quality': 'verbatim_cell_text', **stripped,
'source_url': f'{repo}/blob/{commit}/{relative}'})
return records, audit
EXTRACTION_AUDIT = []
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--component', type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument('--source-root', type=Path, required=False,
help='Directory holding the pinned checkouts of every included repository')
parser.add_argument('--generated-root', type=Path, required=False,
help='Directory holding the pinned mathematics_dataset generator sample')
parser.add_argument('--tokenizer', type=Path,
default=Path('/Users/zxy/Cayley-Math/tokenizer/tokenizer.model'))
parser.add_argument('--pack', action='store_true')
parser.add_argument('--manifest-only', action='store_true',
help='Rewrite manifest.json from the current directory contents and exit')
args = parser.parse_args()
root = args.component.resolve()
if args.manifest_only:
stats = json.loads((root / 'statistics.json').read_text(encoding='utf-8'))
write_readme(root, stats, stats['per_source'])
write_manifest(root)
print(json.dumps({'manifest': 'rewritten',
'files': len(json.loads((root / 'manifest.json').read_text())['files'])}, indent=2))
return
src = args.source_root.resolve()
# --- pin every checkout and refuse to build against a drifting revision ---
for source_id, source in SOURCES.items():
checkout = src / source['checkout']
if not checkout.is_dir():
raise ValueError(f'Missing pinned checkout for {source_id}: {checkout}')
if source.get('generated') or not source.get('commit'):
continue
live = subprocess.check_output(['git', '-C', str(checkout), 'rev-parse', 'HEAD'], text=True).strip()
if live != source['commit']:
raise ValueError(f'{source_id} checkout {live} differs from pinned {source["commit"]}')
generated_path = args.generated_root.resolve() / SOURCES['mathematics-dataset']['generated']
generated = [json.loads(line) for line in generated_path.read_text(encoding='utf-8').splitlines() if line]
gen_params = json.loads(Path(str(generated_path) + '.params.json').read_text(encoding='utf-8'))
all_files = []
records = []
# 1. AlphaProof Nexus (verbatim staged projection, 83 records)
apn = src / SOURCES['alphaproof-nexus']['checkout']
apn_files = sorted(p.relative_to(apn).as_posix() for p in apn.rglob('*')
if p.is_file() and '.git' not in p.parts and p.suffix in ('.lean', '.pdf'))
records += build_alphaproof_nexus(root, 'alphaproof-nexus', apn, apn_files, None)
all_files += [{'source_id': 'alphaproof-nexus', 'path': rel} for rel in apn_files]
# 2. Formal Conjectures library (statements only; tests and build scripts excluded)
fc = src / SOURCES['formal-conjectures']['checkout']
fc_files = sorted(p.relative_to(fc).as_posix() for p in fc.rglob('*.lean')
if '.git' not in p.parts
and p.relative_to(fc).parts[0] in ('FormalConjectures', 'FormalConjecturesForMathlib',
'FormalConjecturesUtil'))
records += build_lean_corpus('formal-conjectures', fc, fc_files, 'formal_lean_statement')
all_files += [{'source_id': 'formal-conjectures', 'path': rel} for rel in fc_files]
# 3. formal-imo
imo = src / SOURCES['formal-imo']['checkout']
imo_files = sorted(p.relative_to(imo).as_posix() for p in imo.rglob('*.lean')
if '.git' not in p.parts and p.relative_to(imo).parts[0] == 'Imo')
records += build_lean_corpus('formal-imo', imo, imo_files, 'formal_lean_statement')
all_files += [{'source_id': 'formal-imo', 'path': rel} for rel in imo_files]
# 4. formal-putnam-like
pl = src / SOURCES['formal-putnam-like']['checkout']
pl_files = sorted(p.relative_to(pl).as_posix() for p in pl.rglob('*.lean')
if '.git' not in p.parts and p.relative_to(pl).parts[0] == 'PutnamLike')
records += build_lean_corpus('formal-putnam-like', pl, pl_files, 'formal_lean_statement')
all_files += [{'source_id': 'formal-putnam-like', 'path': rel} for rel in pl_files]
# 5. Cryptids (whole Lean library, including the Busy Beaver holdout proofs)
cr = src / SOURCES['cryptids']['checkout']
cr_files = sorted(p.relative_to(cr).as_posix() for p in cr.rglob('*.lean') if '.git' not in p.parts)
records += build_lean_corpus('cryptids', cr, cr_files, 'formal_lean_proof')
all_files += [{'source_id': 'cryptids', 'path': rel} for rel in cr_files]
# 6. Mathematics Dataset generator sample
records += build_mathematics_dataset('mathematics-dataset', generated, gen_params)
all_files.append({'source_id': 'mathematics-dataset', 'path': f'generated/{generated_path.name}',
'generated_bytes': generated_path.stat().st_size,
'generated_sha256': sha(generated_path)})
# 7. AlphaGeometry text artifacts
ag = src / SOURCES['alphageometry']['checkout']
ag_files = [name for name in ('examples.txt', 'imo_ag_30.txt', 'jgex_ag_231.txt', 'defs.txt', 'rules.txt')]
records += build_alphageometry('alphageometry', ag)
all_files += [{'source_id': 'alphageometry', 'path': rel} for rel in ag_files]
# 8. Mathematics conjectures notebooks
mc = src / SOURCES['mathematics-conjectures']['checkout']
mc_files = sorted(p.name for p in mc.glob('*.ipynb'))
records += build_notebooks('mathematics-conjectures', mc)
all_files += [{'source_id': 'mathematics-conjectures', 'path': rel} for rel in mc_files]
# 9. AlphaEvolve maths/ verification notebooks (16 notebooks over 15 problem directories)
ae = src / SOURCES['alphaevolve']['checkout']
ae_files = sorted(p.relative_to(ae).as_posix() for p in (ae / SOURCES['alphaevolve']['notebook_root']).rglob('verification*.ipynb'))
ae_records, ae_audit = build_notebook_files('alphaevolve', ae, ae_files)
records += ae_records
all_files += [{'source_id': 'alphaevolve', 'path': rel} for rel in ae_files]
EXTRACTION_AUDIT.extend(ae_audit)
# 10. AlphaEvolve results notebook (one notebook, 36 markdown cells and 92 code cells)
aer = src / SOURCES['alphaevolve_results']['checkout']
aer_files = ['mathematical_results.ipynb']
aer_records, aer_audit = build_notebook_files('alphaevolve_results', aer, aer_files)
records += aer_records
all_files += [{'source_id': 'alphaevolve_results', 'path': rel} for rel in aer_files]
EXTRACTION_AUDIT.extend(aer_audit)
# 11. FunSearch problem notebooks (four: cap_set, admissible_set, bin_packing, cyclic_graphs)
fs = src / SOURCES['funsearch']['checkout']
fs_dirs = ('cap_set', 'admissible_set', 'bin_packing', 'cyclic_graphs')
fs_files = sorted(p.relative_to(fs).as_posix() for p in fs.rglob('*.ipynb')
if p.relative_to(fs).parts[0] in fs_dirs)
fs_records, fs_audit = build_notebook_files('funsearch', fs, fs_files)
records += fs_records
all_files += [{'source_id': 'funsearch', 'path': rel} for rel in fs_files]
EXTRACTION_AUDIT.extend(fs_audit)
# --- exact text dedup ---
duplicates, seen = [], {}
kept = []
for record in records:
digest = record['text_sha256']
if digest in seen:
duplicates.append({'excluded_id': record['id'], 'kept_id': seen[digest],
'source_id': record['source_id'], 'group': record['group'],
'text_sha256': digest})
continue
seen[digest] = record['id']
kept.append(record)
records = kept
ids = [r['id'] for r in records]
if len(set(ids)) != len(ids):
raise ValueError('Duplicate record IDs')
# --- splits, assigned per group ---
groups = sorted({r['group'] for r in records})
split_of = split_groups(groups)
empty = [s for s in SPLITS if s not in split_of.values()]
if empty:
raise ValueError(f'Empty splits: {empty}')
for record in records:
record['split'] = split_of[record['group']]
# a group key is namespaced by source, so paper_id stays the upstream work unit
records.sort(key=lambda r: (r['source_id'], r['group'], r['record_type'], r['source_path'], r['id']))
split_records = {}
for split in SPLITS:
selected = [r for r in records if r['split'] == split]
if not selected:
raise ValueError(f'Empty {split} split')
write_jsonl(root / f'data/{split}.jsonl.gz', selected)
write_parquet(root / f'data/{split}.parquet', selected)
split_records[split] = selected
# --- licenses, one per source subdirectory plus an index ---
for source_id, source in SOURCES.items():
checkout = src / source['checkout']
destination = root / 'licenses' / source_id
destination.mkdir(parents=True, exist_ok=True)
license_file = checkout / 'LICENSE'
if license_file.is_file():
shutil.copy2(license_file, destination / 'LICENSE')
else:
raise ValueError(f'{source_id}: upstream LICENSE missing')
root_license = (root / 'licenses' / 'alphaproof-nexus' / 'LICENSE').read_text(encoding='utf-8')
(root / 'LICENSE').write_text(root_license, encoding='utf-8')
# --- audit ---
audit = root / 'audit'
audit.mkdir(parents=True, exist_ok=True)
source_file_rows = []
for item in all_files:
source = SOURCES[item['source_id']]
if 'generated_sha256' in item:
source_file_rows.append({
'source_id': item['source_id'], 'path': item['path'],
'sha256': item['generated_sha256'], 'bytes': item['generated_bytes'],
'source_url': f'{source["repository"]}/tree/{source["commit"]}',
'kind': 'generated_by_pinned_generator',
})
continue
path = (src / source['checkout'] / item['path'])
source_file_rows.append({
'source_id': item['source_id'], 'path': item['path'], 'sha256': sha(path),
'bytes': path.stat().st_size,
'source_url': f'{source["repository"]}/blob/{source["commit"]}/{item["path"]}',
'kind': 'upstream_blob',
})
write_jsonl(audit / 'source-files.jsonl', source_file_rows)
write_jsonl(audit / 'extraction.jsonl', EXTRACTION_AUDIT)
write_jsonl(audit / 'excluded-duplicates.jsonl', duplicates)
write_json(audit / 'group-splits.json', {
'format': 'cayley-group-splits-v1',
'policy': 'sorted SHA256(group), first ceil(1%) groups to dev, next ceil(1%) to test, rest train',
'group_namespacing': 'every group key starts with its source_id so two upstreams never share a group',
'splits': split_of,
'groups_by_source': dict(Counter(g.split('::')[0] for g in groups)),
'records_by_source': dict(Counter(r['source_id'] for r in records)),
})
write_json(audit / 'mathematics-dataset-generation.json', {
'source_id': 'mathematics-dataset',
'generator_repository': SOURCES['mathematics-dataset']['repository'],
'generator_commit': SOURCES['mathematics-dataset']['commit'],
'driver_script': 'tools/generate_mathematics_dataset_sample.py',
'projection': 'sampled projection of the paper curriculum, not the paper-complete dataset',
'measured': {
'records': len(generated),
'records_by_regime': dict(Counter(r['regime'] for r in generated)),
'modules': len({r['module'] for r in generated}),
'characters': sum(len(r['question']) + len(r['answer']) + 2 for r in generated),
'records_retained_after_exact_dedup': sum(1 for r in records if r['source_id'] == 'mathematics-dataset'),
'records_removed_as_exact_duplicates': sum(1 for d in duplicates
if d['source_id'] == 'mathematics-dataset'),
'modules_stopped_early': [k for k, v in gen_params.get('module_yield', {}).items() if v['stopped_early']],
},
'parameters': gen_params,
})
write_json(audit / 'source-scan.json', {
'format': 'cayley-source-scan-v1',
'queries': [
'GitHub search repositories: org:google-deepmind + math|proof|theorem|formal|lean|geometry|olympiad|conjecture',
'GitHub search repositories: org:google + math|theorem|lean|proof',
'GitHub search repositories: org:google-research + math|proof|theorem|lean|geometry|olympiad|formal',
'Direct probes of every candidate named in the brief, plus every google-deepmind repository page 1-2 '
'filtered on math/theorem/proof/formal/lean/geometry/olympiad/conjecture/arithmetic/algebra keywords',
'Hugging Face dataset listing: authors google (76 datasets), google-research-datasets (37), deepmind (6)',
],
'included': [{'source_id': sid, **{k: s[k] for k in ('name', 'repository', 'commit', 'license',
'paper', 'creators', 'math_content')}}
for sid, s in SOURCES.items()],
'excluded_candidates': EXCLUDED_CANDIDATES,
'excluded_subpaths': EXCLUDED_SUBPATHS,
})
# --- packed arrays ---
packed_stats = {}
if args.pack:
import sentencepiece as spm
sp = spm.SentencePieceProcessor(model_file=str(args.tokenizer))
if (sp.pad_id(), sp.bos_id(), sp.eos_id(), sp.unk_id(), sp.vocab_size()) != (0, 1, 2, 3, 32768):
raise ValueError('Unexpected tokenizer special IDs or vocabulary size')
tokenizer_dir = root / 'tokenizer'
tokenizer_dir.mkdir(exist_ok=True)
shutil.copy2(args.tokenizer, tokenizer_dir / 'tokenizer.model')
packed = root / 'packed'
packed.mkdir(exist_ok=True)
for stale in packed.glob('*'):
if stale.is_file():
stale.unlink()
shards = {}
for split in SPLITS:
shards[split], packed_stats[split] = pack_split(root, split, split_records[split], sp)
write_json(packed / 'manifest.json', {
'format': 'mmla-type2-packed-v1', 'phase': 'pretrain', 'layout': 'lm', 'seq_len': SEQ_LEN,
'tokenizer': {'kind': 'sentencepiece', 'vocab_size': 32768, 'sha256': sha(args.tokenizer),
'path': '../tokenizer/tokenizer.model',
'ids': {'pad': 0, 'bos': 1, 'eos': 2, 'unk': 3, 'boundary': 4}},
'quota_units': 'tokens', 'source': COMPONENT, 'splits': shards, 'statistics': packed_stats,
'packing': {'stride': SEQ_LEN, 'bos_per_record': True, 'eos_per_record': True,
'mask_target_bos': True, 'mask_final_target': True, 'loss_mask': -100,
'row_attribution': '{split}-rows.jsonl', 'record_tokens': '{split}-records.jsonl'},
})
write_json(audit / 'packing.json', {
'packer': 'tools/build_google_math.py:pack_split (copied from components/openai-math)',
'tokenizer_sha256': sha(tokenizer_dir / 'tokenizer.model'),
'tokenizer_note': 'SentencePiece decode is whitespace-normalising for runs of spaces; every packed row is '
'verified by re-encoding.',
'sequence_length': SEQ_LEN, 'shard_rows': SHARD_ROWS,
})
# --- measured upstream bytes per included source, from the pinned checkouts ---
measured_source_bytes = {}
for source_id, item in SOURCES.items():
checkout = src / item['checkout']
if item.get('notebook_root') == 'problems':
paths = sorted(p for d in ('cap_set', 'admissible_set', 'bin_packing', 'cyclic_graphs')
for p in (checkout / d).glob('*.ipynb'))
elif 'notebook_root' in item and item['notebook_root']:
paths = sorted((checkout / item['notebook_root']).rglob('verification*.ipynb'))
elif 'notebook_root' in item:
paths = [checkout / 'mathematical_results.ipynb']
elif item.get('generated'):
paths = [args.generated_root.resolve() / item['generated']]
else:
paths = [p for p in checkout.rglob('*') if p.is_file() and '.git' not in p.parts]
measured_source_bytes[source_id] = {
'included_files': len(paths),
'included_bytes': sum(p.stat().st_size for p in paths),
'whole_repository_bytes': sum(p.stat().st_size for p in checkout.rglob('*')
if p.is_file() and '.git' not in p.parts),
}
# --- statistics ---
source_stats = {}
for source_id in SOURCES:
rows = [r for r in records if r['source_id'] == source_id]
packed_rows = [r for r in rows if r['split'] == 'train']
source_stats[source_id] = {
'math_content': SOURCES[source_id]['math_content'],
'records': len(rows),
'groups': len({r['group'] for r in rows}),
'record_types': dict(Counter(r['record_type'] for r in rows)),
'text_characters': sum(len(r['text']) for r in rows),
'upstream_files': sum(1 for x in source_file_rows if x['source_id'] == source_id),
'split_records': {s: sum(1 for r in rows if r['split'] == s) for s in SPLITS},
}
stats = {
'component': COMPONENT,
'sources': len(SOURCES),
'source_commits': {sid: s['commit'] for sid, s in SOURCES.items()},
'records_before_deduplication': len(records) + len(duplicates),
'records': len(records),
'excluded_exact_duplicates': len(duplicates),
'groups': len(groups),
'record_types': dict(Counter(r['record_type'] for r in records)),
'record_type_characters': {t: sum(len(r['text']) for r in records if r['record_type'] == t)
for t in sorted({r['record_type'] for r in records})},
'math_content_records': dict(Counter(r['math_content'] for r in records)),
'text_characters': sum(len(r['text']) for r in records),
'per_source': source_stats,
'splits': {s: {'records': len(v), 'groups': len({r['group'] for r in v}),
'sources': len({r['source_id'] for r in v}),
'text_characters': sum(len(r['text']) for r in v)} for s, v in split_records.items()},
'packed': packed_stats,
'notebook_projection': {
'records': sum(1 for r in records if r['record_type'] == 'research_notebook'),
'cells_with_outputs': sum(a.get('cells_with_outputs', 0) for a in EXTRACTION_AUDIT
if a.get('cells_with_outputs') is not None),
'output_objects_removed': sum(a.get('output_objects_removed', 0) for a in EXTRACTION_AUDIT),
'output_bytes_removed': sum(a.get('output_bytes_removed', 0) for a in EXTRACTION_AUDIT),
'embedded_image_bytes_removed': sum(a.get('embedded_image_bytes_removed', 0) for a in EXTRACTION_AUDIT),
'bytes_kept': sum(len(r['text'].encode('utf-8')) for r in records
if r['record_type'] == 'research_notebook'),
},
'math_content': {sid: s['math_content'] for sid, s in SOURCES.items()},
'mathematical_verification': 'not_performed',
'lean_compilation': 'not_performed',
}
write_json(root / 'statistics.json', stats)
sources_index = {
'component': COMPONENT,
'format': 'cayley-sources-v1',
'all_records_license': 'per-record',
'license_index': {sid: {'license': s['license'],
'license_file': f'licenses/{sid}/LICENSE',
'creators': s['creators']} for sid, s in SOURCES.items()},
'included': {sid: {
'name': s['name'],
'repository': s['repository'],
'commit': s['commit'],
'fixed_revision_url': f'{s["repository"]}/tree/{s["commit"]}',
'license': s['license'],
'license_file': f'licenses/{sid}/LICENSE',
'creators': s['creators'],
'paper': s['paper'],
'math_content': s['math_content'],
'math_content_scope': s.get('area', 'whole repository'),
'measured_bytes': measured_source_bytes.get(sid, {}),
'records': source_stats[sid]['records'],
'text_characters': source_stats[sid]['text_characters'],
} for sid, s in SOURCES.items()},
'excluded_candidates': EXCLUDED_CANDIDATES,
'excluded_subpaths': EXCLUDED_SUBPATHS,
'math_content_classes': list(MATH_CONTENT),
'math_content_note': 'constructions is the honest class for AlphaEvolve and FunSearch findings: the '
'decompositions, coordinate sets and step functions are the results, not proofs of them',
'excluded_subpaths': EXCLUDED_SUBPATHS,
'changes': [
'Verbatim text projection of Lean sources, geometry problem language, notebooks and PDFs; no rewriting',
'Deterministically seeded sample from the official mathematics_dataset generator at a pinned commit',
'Group keys namespaced per source so a proof and its prose companion and a generated QA family stay together',
'Exact text deduplication and deterministic SHA256-ordered group split assignment',
'Lossless JSONL/Parquet export; SentencePiece causal-window packing in the existing openai-math format',
],
'record_changes': [
'No text was generated, summarised, translated or repaired beyond PDF control-character handling',
'Notebook execution outputs and embedded images were dropped; code and markdown cells stay verbatim',
'PDF text control characters were replaced and blank-line runs collapsed; original bytes stay identified by sha256',
'No Lean compilation, proof checking or mathematical verification was performed',
],
'mathematical_verification': 'not_performed',
}
write_json(root / 'sources.json', sources_index)
# --- README ---
write_readme(root, stats, source_stats)
write_manifest(root)
print(json.dumps(stats, ensure_ascii=False, indent=2))
def _sum_by_split(mapping, key, splits):
"""Sum one per-split counter out of a gate section keyed by split name."""
if not isinstance(mapping, dict):
return 0
return sum(int((mapping.get(split) or {}).get(key, 0) or 0) for split in splits)
def read_gate_evidence(root, records):
"""Normalise the publish gate into one evidence record, whichever schema it uses.
The gate has been produced by two different generators. The earlier one reports
whole-run totals (`rows_decoded`, `segments_compared`, `parquet_roundtrip.rows_checked`,
`split_separation`). The later one reports the same measurements per split
(`packed_decode.train/dev/test`, `parquet_roundtrip.rows/exact`, `group_separation`).
They measure identical things, so nothing here is reinterpreted: per-split counts are
summed and legacy totals are read directly, which keeps every field meaning exactly what
the gate that wrote it meant. Fields the later schema adds (how many boundary fragments
carried text, carried none, or needed a multi-byte edge trim) stay None under the legacy
schema; the README omits those clauses rather than inventing numbers for them.
Returns None when the gate is absent or describes a different corpus than `statistics.json`.
"""
splits = ('train', 'dev', 'test')
path = root / 'audit/publish-check.json'
if not path.is_file():
return None
gate = json.loads(path.read_text(encoding='utf-8'))
totals = gate.get('measured_totals') or {}
if totals.get('records') != records:
return None
packed = gate.get('packed_decode') or {}
rounds = gate.get('parquet_roundtrip') or {}
separation = gate.get('group_separation') or gate.get('split_separation') or {}
per_split = 'train' in {key for key in packed if isinstance(key, str)}
if per_split:
rows_per_split = {split: int((packed.get(split) or {}).get('rows', 0) or 0) for split in splits}
evidence = {
'schema': 'per-split',
'rows_decoded': sum(rows_per_split.values()),
'rows_per_split': rows_per_split,
'splits_covered': sorted(split for split, n in rows_per_split.items() if n),
'segments_compared': _sum_by_split(packed, 'segments', splits),
'token_slice_exact': _sum_by_split(packed, 'token_slice_exact', splits),
'whole_record_segments': _sum_by_split(packed, 'whole_record_segments', splits),
'whole_record_text_identical': _sum_by_split(packed, 'whole_record_text_exact', splits),
'whole_record_reencoded_identical': _sum_by_split(packed, 'whole_record_reencode_exact', splits),
'fragment_segments': _sum_by_split(packed, 'fragment_segments', splits),
'fragment_verbatim_contained': _sum_by_split(packed, 'fragment_verbatim_contained', splits),
'fragment_no_text': _sum_by_split(packed, 'fragment_no_text', splits),
'fragment_boundary_trimmed': _sum_by_split(packed, 'fragment_boundary_trimmed', splits),
'parquet_rows': int(rounds.get('rows', 0) or 0),
'parquet_exact': int(rounds.get('exact', 0) or 0),
}
else:
evidence = {
'schema': 'legacy',
'rows_decoded': int(packed.get('rows_decoded', 0) or 0),
'rows_per_split': None,
'splits_covered': sorted(packed.get('splits_covered') or []),
'segments_compared': int(packed.get('segments_compared', 0) or 0),
'token_slice_exact': int(packed.get('segments_whose_row_slice_equals_canonical_token_slice', 0) or 0),
'whole_record_segments': int(packed.get('whole_record_segments_decoded', 0) or 0),
'whole_record_text_identical': int(
packed.get('whole_record_segments_with_identical_whitespace_normalised_text', 0) or 0),
'whole_record_reencoded_identical': int(
packed.get('whole_record_segments_reencoded_token_identically', 0) or 0),
'fragment_segments': int(packed.get('fragment_segments_not_reencoded_per_segment', 0) or 0),
'fragment_verbatim_contained': None,
'fragment_no_text': None,
'fragment_boundary_trimmed': None,
'parquet_rows': int(rounds.get('rows_checked', 0) or 0),
'parquet_exact': int(rounds.get('exact_matches', 0) or 0),
}
evidence['records'] = totals.get('records')
evidence['groups_total'] = int(separation.get('groups_total', 0) or 0)
evidence['groups_spanning'] = len(separation.get('groups_spanning_two_or_more_splits') or [])
evidence['status'] = gate.get('status')
return evidence
def write_readme(root, stats, source_stats):
records = stats['records']
chars = stats['text_characters']
types = stats['record_types']
type_chars = stats['record_type_characters']
gen_records = types.get('generated_qa_pair', 0)
gen_chars = type_chars.get('generated_qa_pair', 0)
lean_records = types.get('formal_lean_proof', 0) + types.get('formal_lean_statement', 0)
lean_chars = type_chars.get('formal_lean_proof', 0) + type_chars.get('formal_lean_statement', 0)
other_records = records - gen_records
other_chars = chars - gen_chars
sorted_types = sorted(type_chars.items(), key=lambda kv: (-kv[1], kv[0]))
def pct(numerator, denominator):
return f'{100 * numerator / denominator:.2f}%'
def ratio(numerator, denominator):
return f'{numerator / denominator:.6f}'
def count(number):
return f'{number:,}'
composition = '\n'.join(
f'{pct(chars_of, chars)} of text characters (ratio {ratio(chars_of, chars)}) in `{name}`'
for name, chars_of in sorted_types)
per_source_table = '\n'.join(
f'| `{sid}` | [{SOURCES[sid]["repository"].rsplit("/", 1)[-1]}]({SOURCES[sid]["repository"]}) '
f'`{SOURCES[sid]["commit"][:12]}` | {src["records"]:,} | {src["groups"]:,} | '
f'{src["text_characters"]:,} | {", ".join(sorted(src["record_types"]))} | {src["math_content"]} |'
for sid, src in source_stats.items())
new_sources = ('alphaevolve', 'alphaevolve_results', 'funsearch')
new_records = sum(source_stats[s]['records'] for s in new_sources)
new_chars = sum(source_stats[s]['text_characters'] for s in new_sources)
new_groups = sum(source_stats[s]['groups'] for s in new_sources)
notebook = stats.get('notebook_projection', {})
stripped_outputs = notebook.get('output_objects_removed', 0)
stripped_bytes = notebook.get('output_bytes_removed', 0)
stripped_images = notebook.get('embedded_image_bytes_removed', 0)
stripped_cells = notebook.get('cells_with_outputs', 0)
construction_chars = type_chars.get('research_notebook', 0)
stats_json = json.dumps(stats, indent=2)
publication = {}
publication_note = (
'The pull request that carries this build is recorded in `audit/publication.json` with its number,\n'
'URL and the parent commit it was opened against. It adds the whole component and the top-level\n'
'`README.md` config row and has not been merged into `main`.')
if (root / 'audit/publication.json').is_file():
publication = json.loads((root / 'audit/publication.json').read_text(encoding='utf-8'))
publication_note = (
f'The component is delivered as Hugging Face dataset pull request\n'
f'[MMLA-ORG/Cayley-Math discussion #{publication["pr_num"]}]({publication["pr_url"]})\n'
f'on ref `{publication["pr_revision"]}`, opened against main at `{publication["parent"][:12]}`.\n'
f'It adds the whole component and the top-level `README.md` config row, and it has not been merged\n'
f'into `main`. Consult that discussion for the current merge state.')
import textwrap
def para(*sentences):
return textwrap.fill(' '.join(sentences), width=100, break_long_words=False,
break_on_hyphens=False)
notebook_sources = sum(1 for v in source_stats.values() if v['record_types'].get('research_notebook'))
composition_paragraph = para(
'**This corpus is mostly seeded generator output, not research mathematics.**',
f'Measured over the delivered data, `generated_qa_pair` records are {count(gen_records)} of',
f'{count(records)} records, which is {pct(gen_records, records)} of records',
f'(ratio {ratio(gen_records, records)}) and {count(gen_chars)} of {count(chars)} text characters',
f'({pct(gen_chars, chars)}, ratio {ratio(gen_chars, chars)}).',
f'Lean records total {count(lean_records)}, namely {count(types.get("formal_lean_proof", 0))}',
f'`formal_lean_proof` plus {count(types.get("formal_lean_statement", 0))} `formal_lean_statement`,',
f'which is {pct(lean_records, records)} of records (ratio {ratio(lean_records, records)}), and lean',
f'text carries {pct(lean_chars, chars)} of the characters (ratio {ratio(lean_chars, chars)}).',
f'Geometry problems, prose proofs and notebook records add {count(other_records - lean_records)} more',
f'records, so every non-generator record together is {count(other_records)} records,',
f'{pct(other_records, records)} of the corpus (ratio {ratio(other_records, records)}), carrying',
f'{count(other_chars)} characters ({pct(other_chars, chars)}, ratio {ratio(other_chars, chars)}).',
'Lean files are long, so a character-weighted view inverts the record-weighted balance, and the',
'construction notebooks are longer still because their coordinates and step-function coefficients',
'are decimal literals. Read the composition from records first, then from characters.',
'**No Lean compilation and no proof verification was performed on any subset**, so no record in',
'this corpus has a checked proof or a checked statement, and no construction in it was re-verified',
'here.')
math_content_paragraph = para(
f'`math_content` classifies each record\'s mathematical payload as one of formal_statements, proofs,',
'natural_language_math, generated_qa, constructions, problem_statements or none. Measured by records:',
f'{json.dumps(stats.get("math_content_records", {}), sort_keys=True)}.')
notebook_paragraph = para(
'Notebook records are the cell sources of each file joined by a newline and nothing else.',
f'Across the {count(notebook_sources)} sources that ship notebooks, {count(stripped_cells)} cells',
f'carried {count(stripped_outputs)} output objects totalling {count(stripped_bytes)} bytes, of which',
f'{count(stripped_images)} bytes were embedded images; every one of those outputs was removed and',
'every code and markdown cell stayed verbatim.',
'`audit/extraction.jsonl` records the same counts per notebook.')
gate_paragraph = para('[' + '`audit/publish-check.json`' + '](audit/publish-check.json) is a second publish gate '
'with its own code that re-derives every total in this file from the delivered '
'`data/*.jsonl.gz` and `packed/` files.')
gate_note = ('decodes packed rows from each split with token identity and whitespace-normalised text '
'identity required for whole-record segments, matches Parquet rows to their JSONL records, and '
'recomputes group separation from the data with zero groups spanning two splits. '
'`audit/publish-check-previous.json` keeps the gate as it stood before this build so the growth '
'of the corpus can be diffed rather than asserted.')
evidence = read_gate_evidence(root, records)
if evidence is not None:
rows_desc = (f'{count(evidence["rows_decoded"])} packed rows, '
f'{count(evidence["rows_per_split"]["train"])} per split across train, dev and test,'
if evidence['rows_per_split'] else
f'{count(evidence["rows_decoded"])} packed rows across '
f'{", ".join(evidence["splits_covered"])}')
fragments = evidence['fragment_segments']
if evidence['fragment_no_text'] is None:
fragment_clause = (f'and the remaining {count(fragments)} row-boundary fragments are covered by '
f'that same token-slice identity')
else:
fragment_clause = (
f'and of the {count(fragments)} row-boundary fragments, '
f'{count(evidence["fragment_verbatim_contained"])} decoded to text appearing verbatim inside '
f'their source record and {count(evidence["fragment_no_text"])} carried no text at all '
f'because they contributed only BOS, EOS or bare whitespace')
if evidence['fragment_boundary_trimmed']:
fragment_clause += (
f'. Boundary fragments can start or end inside a multi-byte glyph, so their edges are '
f'trimmed of replacement characters before that containment test')
note = (f'Its latest measured run decoded {rows_desc} and compared '
f'{count(evidence["segments_compared"])} attributed segments: every row slice equalled its '
f'canonical token slice ({count(evidence["token_slice_exact"])} of '
f'{count(evidence["segments_compared"])} by exact identity), the '
f'{count(evidence["whole_record_segments"])} segments that cover a whole record all decoded to '
f'text identical to the source record under whitespace normalisation and all re-encoded to the '
f'identical token sequence, {fragment_clause}. It matched '
f'{count(evidence["parquet_exact"])} of {count(evidence["parquet_rows"])} Parquet round-trips '
f'exactly and found {count(evidence["groups_spanning"])} groups spanning two splits out of '
f'{count(evidence["groups_total"])}.')
gate_note = para(note)
gate_paragraph = para('[' + '`audit/publish-check.json`' + '](audit/publish-check.json) is a second '
'publish gate with its own code that re-derives every total in this file from '
'the delivered `data/*.jsonl.gz` and `packed/` files.', note)
record_mix_paragraph = para(
'The record mix is deliberately stated because it is skewed: `generated_qa_pair` records are',
f'{pct(gen_records, records)} of records and {pct(gen_chars, chars)} of characters, while Lean records',
f'are {pct(lean_records, records)} of records and {pct(lean_chars, chars)} of characters, and all',
f'non-generator records together are {pct(other_records, records)} of records.',
'A consumer wanting formal, prose or construction mathematics should filter on `record_type` or',
'`math_content` rather than read the corpus whole.')
packed_paragraph = para(
f"Measured packed counts: train {stats['packed'].get('train', {}).get('rows', 0)} rows /",
f"{stats['packed'].get('train', {}).get('tokens', 0)} tokens /",
f"{stats['packed'].get('train', {}).get('loss_bearing_labels', 0)} loss-bearing labels; dev",
f"{stats['packed'].get('dev', {}).get('rows', 0)} /",
f"{stats['packed'].get('dev', {}).get('tokens', 0)} /",
f"{stats['packed'].get('dev', {}).get('loss_bearing_labels', 0)}; test",
f"{stats['packed'].get('test', {}).get('rows', 0)} /",
f"{stats['packed'].get('test', {}).get('tokens', 0)} /",
f"{stats['packed'].get('test', {}).get('loss_bearing_labels', 0)}.")
split_paragraph = para(
'`audit/group-splits.json` freezes every assignment: train',
f"{stats['splits']['train']['records']:,} records over {stats['splits']['train']['groups']:,} groups,",
f"dev {stats['splits']['dev']['records']:,} over {stats['splits']['dev']['groups']:,},",
f"test {stats['splits']['test']['records']:,} over {stats['splits']['test']['groups']:,}.")
new_sources_paragraph = para(
'`alphaevolve`, `alphaevolve_results` and `funsearch` are the constructions behind their papers,',
'labelled `math_content = constructions`. They add',
f'{count(new_records)} records over {count(new_groups)} groups and {count(new_chars)} characters,',
f'which is {pct(new_chars, chars)} of the corpus characters.')
readme = f"""# Google math papers contribution to Cayley-Math
An independently versioned component holding source-derived text from {stats['sources']} 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, proof PDFs and the constructions their search programs found), 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` and `math_content` so consumers can separate verbatim source text from
`generator_output`, and formal statements from constructions.
{composition_paragraph}
{math_content_paragraph}
## 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 | math_content |
| --- | --- | ---: | ---: | ---: | --- | --- |
{per_source_table}
`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 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.
{new_sources_paragraph}
- `alphaevolve` ships 16 `verification.ipynb` notebooks over the 15 problem directories under `maths/`
(cap_sets, circle_packing, erdos_minimum_overlap, extremal_girth_5, first_autocorrelation_inequality,
heilbronn_convex, heilbronn_triangle, hexagon_packing, kissing, max_min_ratio,
second_autocorrelation_inequality, shannon_cycle, sums_and_differences,
third_autocorrelation_inequality, uncertainty_principles, the last with separate Hermite and Laguerre
notebooks). Each notebook states the problem, records the coordinates the search found and carries the
code that checks them. One record per notebook.
- `alphaevolve_results` ships `mathematical_results.ipynb`, 128 cells (36 markdown, 92 code), holding the
paper's improved tensor decompositions, autocorrelation and uncertainty-principle bounds, Erdős minimum
overlap, sums and differences, hexagon packing, Heilbronn, kissing-number, circle-packing and
max-to-min-ratio constructions, each with its verification code. One record.
- `funsearch` ships four problem notebooks (cap_set, admissible_set, bin_packing, cyclic_graphs) with the
skeleton the search evolved, the discovered `priority` functions and the explicit constructions they
justify. Four records.
{notebook_paragraph}
## Excluded candidates and why
The widened scan is recorded in full in `sources.json` (`excluded_candidates` for whole repositories and
`excluded_subpaths` for parts of included ones) and `audit/source-scan.json`. A repository appears in
`included` or in `excluded_candidates`, never both. 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. `alphatensor` remains excluded on its original measurement of 24 search-code entries: it is
a search harness, the same class of artifact as the `alphaevolve` sub-paths excluded below.
- `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.
Two included repositories also ship sub-paths that were measured and left out in part, so those exclusions
are recorded separately in `excluded_subpaths` and each reason names the measured bytes:
- `alphaevolve` `omega/` (45,511,025 bytes measured) is a search harness plus one compressed cache:
44,267,872 bytes are the single `.xz` archive `clean_level4.npz.xz`, and its 1,126,802-byte
`verification.ipynb` is 1,098,419 characters of code, 836,374 of them decimal digits holding rational
vectors for a 20 GB verification run. Code-only, so nothing from it is projected.
- `alphaevolve` `src/`, `maths/**/*.py`, `maths/**/config*.yaml` and `matmul_tensor_decompositions/`
(1,165,195 bytes measured) are the 58,453-byte search harness, 87,725 bytes of initial programs,
60,949 bytes of run configuration, and 827,894 bytes of two tensor-rank notebooks whose projections
share 4,820 of 4,830 substantial lines with the included `alphaevolve_results` notebook. Code-only or
near-duplicate, so none of it is projected.
- `funsearch` `**/*.txt` and `implementation/*.py` (20,442,546 bytes measured): 13 `.txt` files hold
20,389,203 bytes of comma-separated integer tuples, one alone 17,372,832 bytes, serialising exactly the
sets the included notebooks build and assert; `implementation/` is 53,343 bytes of the search algorithm.
Serialization and harness code, so none of it is projected.
## 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, `math_content` class and
measured counts, plus every excluded candidate with its measured reason and every excluded sub-path of
an included repository.
- `licenses/<source_id>/LICENSE`: one upstream license per source; `LICENSE` at the component root is the
Apache-2.0 text shared by all {stats['sources']}.
- `audit/source-files.jsonl`: path, sha256, bytes, source_id, kind and pinned source_url for every blob and
for the generated corpus file.
- `audit/source-scan.json`: the scan queries, the included set and the whole-repository exclusions, plus
the measured sub-path exclusions of included repositories.
- `audit/group-splits.json`: the group assignment, the policy, and per-source group counts.
- `audit/extraction.jsonl`: per-PDF and per-notebook extraction metrics, quality flags, and for notebooks
the exact number of output cells and bytes removed.
- `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.
- `audit/publish-check.json`: the second publish gate, measured from the delivered files after the build;
`audit/publish-check-previous.json` preserves the gate as it stood before this build so the delta is
auditable.
- `statistics.json`: measured records, groups, splits, characters and packed counts, per source, plus
per-record-type and per-`math_content` counts.
- `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, {SEQ_LEN}], 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
{SEQ_LEN} 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.
{packed_paragraph}
## 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.
{split_paragraph}
## 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, and only
execution output is dropped where it existed. `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. The construction notebooks
carry `math_content = constructions` because their decimals are the construction itself, not a proof of
it; the code in them checks the construction, and this component never ran it.
## 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 an independent check against the delivered files, and its output is
[`audit/validation-evidence.json`](audit/validation-evidence.json). It re-encodes and decodes sampled
packed rows across all three splits, compared against the records they are attributed to rather than
against the build's own manifest, and it round-trips sampled Parquet rows back to their JSONL records. It
also recomputes split separation from the delivered data instead of trusting `audit/group-splits.json`.
{gate_paragraph}
{record_mix_paragraph}
## Publication
{publication_note}
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 open pull requests, but a direct commit to `main` is rejected with
`403 Forbidden: You have read access but not the required permissions for this operation`. Merging the
pull request, or re-running an 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
construction notebooks' verifier code was read as text and never executed, so no construction here was
re-checked against its claim.
- The build ran in a sandbox-free subprocess only for the generator; every other source was 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`.
- The construction notebooks record decimal literals for coordinates and coefficients at the precision
upstream wrote them. They are verbatim, so rounding past 16 digits is upstream's, not this component's;
no construction was recomputed in exact arithmetic.
- `funsearch` sets whose only form upstream ships is the integer tuple list in `.txt` are represented here
through the notebook that constructs them, not through the serialized list.
- `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 is limited to exact text duplicates and the recorded line overlap between the
`matmul_tensor_decompositions` notebooks and `alphaevolve_results`. Phase 1 overlap analysis and
editorial review were not performed. {stats['excluded_exact_duplicates']:,} 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 characters because a handful of Busy Beaver traces are enormous, and the `kissing`
notebook alone contributes 5.8M of the {count(construction_chars)} notebook characters because its
593-sphere configuration is 5.8M characters of coordinates. Both are machine-generated artifacts rather
than mathematics prose.
## Statistics
~~~json
{stats_json}
~~~
"""
(root / 'README.md').write_text(readme, encoding='utf-8')
if __name__ == '__main__':
main()
#!/usr/bin/env python3
"""Build components/google-math: Google math-paper artifacts in Cayley-Math format.
One component, eight upstream subsets. Mirrors components/openai-math exactly:
same record core fields, same Parquet encoding, same SHA256-ordered group split
policy, and the same mmla-type2-packed-v1 packing code.
Nothing here verifies mathematics. No Lean project is compiled, no proof is
checked, and generated QA pairs are generator output rather than research
mathematics.
"""
from __future__ import annotations
import argparse
import gzip
import hashlib
import json
import math
import re
import shutil
import subprocess
from collections import Counter, deque
from pathlib import Path
COMPONENT = 'google-math'
CORE = ('id', 'text', 'group', 'license', 'paper_id', 'split', 'record_type', 'source_url')
SPLITS = ('train', 'dev', 'test')
SEQ_LEN = 2048
SHARD_ROWS = 256
IGNORED_LOCAL = {'.DS_Store', 'manifest.json', 'validation.json'}
# ---------------------------------------------------------------------------
# Source registry. Every included source is a Google or Google DeepMind
# publication that ships math text, and every entry pins a fixed revision.
# ---------------------------------------------------------------------------
SOURCES = {
'alphaproof-nexus': {
'name': 'AlphaProof Nexus results',
'repository': 'https://github.com/google-deepmind/alphaproof-nexus-results',
'commit': '0647711a71183c1ea492ad60860776617ce1ea88',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Lean math proofs generated by AlphaProof Nexus and accompanying natural language prose proofs',
'checkout': 'alphaproof-nexus-results',
'dir': 'alphaproof-nexus-results',
'math_content': 'proofs',
},
'formal-conjectures': {
'name': 'Formal Conjectures',
'repository': 'https://github.com/google-deepmind/formal-conjectures',
'commit': '6751c0f043985ac6647db28a45a63afdf21acffe',
'license': 'Apache-2.0',
'creators': ['The Formal Conjectures Authors', 'Google LLC'],
'paper': 'Formalised statements of mathematical conjectures in Lean, arXiv:2605.13171',
'checkout': 'formal-conjectures',
'dir': 'formal-conjectures',
'math_content': 'formal_statements',
},
'formal-imo': {
'name': 'formal-imo benchmark',
'repository': 'https://github.com/google-deepmind/formal-imo',
'commit': 'a85d84b97c8e88a352c136b1e29c355139609bea',
'license': 'Apache-2.0',
'creators': ['Google LLC'],
'paper': 'Lean 4 formalizations of International Mathematical Olympiad problem statements',
'checkout': 'formal-imo',
'dir': 'formal-imo',
'math_content': 'formal_statements',
},
'formal-putnam-like': {
'name': 'Putnam-like problems in Lean',
'repository': 'https://github.com/google-deepmind/formal-putnam-like',
'commit': '636c2d46b3a7d7e77a9f58209b0c17816f8d172b',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Lean 4 formalizations of Putnam-like problems (eval_hub putnam_like companion)',
'checkout': 'formal-putnam-like',
'dir': 'formal-putnam-like',
'math_content': 'formal_statements',
},
'cryptids': {
'name': 'Cryptids: conjecture-checking and Busy Beaver Turing machines',
'repository': 'https://github.com/google-deepmind/cryptids',
'commit': '3647718b5df22e1a61462dec0376bd204526285a',
'commit_source': 'the checkout as of this build; recorded because the repository moves',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Certified Turing-machine reductions and Busy Beaver proofs, arXiv:2610.22603',
'checkout': 'cryptids',
'dir': 'cryptids',
'math_content': 'proofs',
},
'mathematics-dataset': {
'name': 'Mathematics Dataset generator',
'repository': 'https://github.com/google-deepmind/mathematics_dataset',
'commit': '427f45075f84b8b9774950196ad63867ca20ffb3',
'license': 'Apache-2.0',
'creators': ['Google DeepMind'],
'paper': 'Analysing Mathematical Reasoning Abilities of Neural Models (Saxton, Grefenstette, Hill, Kohli), ICLR 2018',
'checkout': 'mathematics_dataset',