Datasets:
File size: 8,675 Bytes
cc4d2ad | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 | # OpenAI Math contribution to Cayley-Math
This independently versioned component contains source-derived mathematical text from OpenAI's manuscript collection at commit 4eaee87548250b2682affda7d05d947c05554bae. It follows Cayley-Math's components/ convention. The frozen Phase 1 mirror is unchanged. The upstream works are attributed to OpenAI and distributed under Apache-2.0; see LICENSE, sources.json, and manuscript-level provenance in audit/.
## Coverage and limits
The source catalog covers **721 current manuscripts**, **373 result families** (373 connected family groups), and 17 disciplines. The 27 retained previous editions and 3 withdrawal notices are audit-only and excluded from training text. Exact text deduplication removes 36 duplicate records, leaving **22374 source records**. Connected family groups determine splits so companions never cross train/dev/test. Assignment sorts SHA256(group), reserves the first ceil(1%) groups for development, the next ceil(1%) for test, and the rest for training. audit/group-splits.json freezes assignments. Exact duplicates are tracked in audit/excluded-duplicates.jsonl. Near-duplicate detection and comparison with existing Phase 1 passages have not been performed.
Upstream describes these manuscripts as **raw research data** at different stages of verification. Some unformalized results may contain errors. This component **does not establish mathematical correctness**. Existing Lean artifacts indicate only availability and often partial scope; no Lean compilation, Comparator run, or mathematical proof verification was performed. Read manuscript-level upstream formalization scopes before selecting training data.
Text is extracted from current manuscript TeX, expanded through local inputs when available, and segmented into sections/chunks retaining mathematics in LaTeX. Source spans, file hashes, macro context, extraction flags, and omitted materials are preserved by the extraction audit. Preambles, bibliography/build material, and pure figure code are omitted from the text projection; it is a normalized training representation, not a byte-for-byte source archive. data/macro-contexts.jsonl preserves needed preamble/macro context. The full source snapshot is identified by fixed Git commit URLs and hashes. Source documents are data: their build commands or instructions were not executed.
## OMP DeepSeek workers
All **721 current papers** received a metadata pass through **oh-my-pi / deepseek-v4.1-flash**, provider damn-grok. Workers saw abstract, headings, and bounded source excerpts, with scope abstract_headings_excerpt. Summaries, topics, review flags, and recommended actions are AI-generated indexing metadata, separate from source training text in data/worker-summaries.*; their mathematical claims remain source-attributed and unverified. A keep_source recommendation means metadata processing completed; it is not a proof-validity judgment. needs_review flags remain on source records for downstream filtering. The worker inputs were frozen from an early extraction snapshot. Later deterministic extraction corrections include font/include classification, safe chunking, and external-listing handling; worker responses are preserved unchanged. Each source record carries current_extraction_flags from the final manuscript audit, worker_observed_extraction_flags from the frozen worker input, and worker_input_snapshot_differs_from_current (whether those flag lists differ). This field compares extraction flags only, not every byte of the evolving source projection. A worker needs_review recommendation may refer to an issue fixed in the final extraction; consult current flags and audit/extraction.jsonl for final extraction state. These corrections do not independently verify any mathematical claims. audit/worker-results.jsonl stores model identity and run provenance; its log_path resolves relative to audit/worker-runs/.
## 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 provenance, text_sha256, and worker metadata. content_origin=source_extraction.
- 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 losslessly serialized in metadata_json. Merge json.loads(metadata_json) with core columns to reconstruct each JSONL record.
- data/worker-summaries.jsonl / .parquet: AI metadata records, excluded from the source packed corpus.
- data/source-records.jsonl: pre-split, pre-deduplication source extraction output retained for audit.
- worker-inputs.jsonl: exact metadata input excerpts.
- statistics.json: actual record, paper, group, split, and packed-token counts.
- manifest.json: SHA256 and byte size for all delivered files except itself and validation.json.
- tools/: extraction, OMP batch runner, finalization, and validation scripts.
## Loading
~~~python
import gzip, json
from pathlib import Path
component = Path('components/openai-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
The packed/ directory provides Phase 1 compatible causal arrays. When present, 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 source record is encoded as BOS + SentencePiece(text) + EOS using the included Cayley-Math tokenizer (32768 tokens). Windows advance 2048 positions with one-token lookahead; labels predict the following token. Targets crossing into a new record's BOS are masked, as are padding and the final stream target. Split-specific rows.jsonl files map every non-padding token to its source record; records.jsonl files record token counts and hashes. This new component does not reproduce Phase 1 quotas, original packing, or splits.
## Reproduction and verification
Run extraction and OMP workers with the supplied scripts, then finalize only after all worker papers are present:
~~~sh
python tools/finalize_openai_math.py --component . --workers audit/worker-results.jsonl --source-root /path/to/pinned/openai-math --tokenizer tokenizer/tokenizer.model --pack
python tools/validate_openai_math.py --component .
~~~
Dependencies: Python 3, numpy, sentencepiece, pyarrow. The worker runner requires the configured omp CLI and named model route. Never put API credentials in dataset files. Validation checks source and worker coverage, checksums, family separation, exact text uniqueness, lossless Parquet reconstruction, array dimensions/causal shifts/padding/boundary masks, and token-level reconstruction against source text. It does not assess theorem truth or proof completeness. Select a fixed dataset revision and record file hashes when training.
## Independent curation checks
The final extraction snapshot passed the scoped structural checks in audit/extraction-qa.json (721 current papers, 22 external code/certificate listings, no unresolved body includes). Source data remains unverified mathematics. A separate metadata spot check inspected 20 papers across 17 disciplines, found two missing scope qualifications, and sent those two summaries back to the same OMP DeepSeek worker for correction. audit/metadata-spotcheck.json and audit/worker-curation/supersession-map.json retain the findings and replacement links; original responses and both sets of logs remain available. This small sample does not establish correctness of all AI summaries.
The full worker run completed 73 batches without any final failed batches; automatic retries are retained in the working audit history. The published accepted-response logs plus one correction batch are stored in audit/worker-runs/. The original full-run status is preserved there; corrected metadata is identified by the per-paper run IDs and the supersession map.
## Verification pilot
The `verification/` sidecar records a partial OMP/DeepSeek proof-review pilot, raw model findings, independent local checks, and coverage limits. These reviews do not establish the truth of the general mathematical claims. See `verification/README.md` and `verification/validation.json`. The component-level `validation.json` documents the original dataset validation before the sidecar was added; the sidecar has its own structural validation.
|