This dataset provides 100,000,000,000 training input positions and 10,000,000 held-out validation positions. It is a custom memory-mapped token corpus, not a Parquet dataset for datasets.load_dataset. The exact build is identified by corpus-v2/manifest.json, corpus-v2/verification.json, and the UPLOAD_COMPLETE.json receipt. That receipt's data_revision pins the immutable commit whose new corpus files were verified.
The frozen tokenizer has SHA-256 fc240b741112488b639a711f502305601e3b8a1b01c5ae93a48b9ae55692dd33, vocabulary size 2,048, and structural document-boundary ID 2,047. The tokenizer JSON and its review/provenance artifacts are under corpus-v2/tokenizer/.
Download and load
Download the corpus-v2/ prefix at the data_revision in UPLOAD_COMPLETE.json; preserve the prefix in the local directory. See corpus-v2/LOADER.md for a pinned download example, file formats, and the project's reproducible TokenLoader entry point. Training should use the manifest and source indexes; the physical token file contains stored positions that are not all selected into a split.
The selectors and fetched-file receipts are recorded in corpus-v2/plan.json and corpus-v2/source-manifest.json. Four UltraData-Math L3 synthetic styles divide the 15B math allocation equally. The direct Nemotron component uses the original Common Crawl-hosted Nemotron-CC high-quality actual and diverse-QA variants; it is not Nemotron-CC-Math. UltraData-Math acknowledges NVIDIA, MegaMath, and FineMath among its seed resources, so removing a direct FineMath stream does not remove all FineMath ancestry. See corpus-v2/ATTRIBUTION.md and corpus-v2/provenance/source-research.md.
Format
corpus-v2/tokens.bin: little-endian unsigned 16-bit token IDs (<u2). A stored document begins with ID 2,047.
corpus-v2/-.idx: little-endian unsigned 64-bit triples (offset, selected_length, full_length), 24 bytes per selected document. Split is train or validation. Offsets and lengths count token positions.
The manifest binds each index, digest and origin sidecar. Use only the source files named by the manifest; do not treat every physical token position as training data.
Verification checks and results
Publication was independently checked on 30 September 2026 (UTC). The local finalizer and public-repository verification passed the checks below.
Check
Recorded result
Selected positions and source quotas
PASS — exactly 100,000,000,000 training and 10,000,000 validation positions; all 11 component quotas matched.
Full local output hashes
PASS — SHA-256 recomputed for all 69 builder output files.
Token range and document boundaries
PASS — stored token IDs checked against the 2,048-ID vocabulary; boundary ID 2,047 checked for all 96,948,408 selected documents.
Physical token span coverage
PASS — full document spans cover the token file without overlaps, gaps or an unreferenced tail.
Digest, split and source provenance
PASS — all recorded digest-to-split assignments checked; source download receipts bound to the pinned plan.
Fast tokenizer parity
PASS — 9,479 sampled Splintr/reference comparisons matched. This was sampled parity, not dual-tokenization of every record.
Production loader
PASS — 311,296-position batches and exact state save/resume.
Held-out loader coverage
PASS — one complete 10,000,000-position validation traversal.
Published corpus integrity
PASS — all 131 corpus files, totaling 207,186,190,336 bytes, matched remote size and content-hash evidence at immutable revisions.
Completion receipt and legacy retirement
PASS — public receipt matched the local receipt; all 180 authorized legacy paths were absent at the retirement revision.
Verified data revision: 76c88a54094708e776e688d6f745c92b540167e7. Retirement revision: 41d30eaf4fe4d2b26650dfcee5fced52d6a2fb41. Large LFS files were checked against remote SHA-256 and size metadata. For small Git files, their local SHA-256 was checked against the plan and their Git blob SHA-1 against the remote object. The independent publication check fetched the public receipt and metadata anonymously; it did not redownload and rehash the entire 207 GB payload.
Evidence: finalizer results, loader readiness, completion receipt, and independent publication-check results. The receipt covers 132 files / 207,186,195,962 bytes when the original dataset card is included. It pins that original card at the data revision; this later documentation update does not change the corpus payload or the receipt. Legacy retirement removes files from the latest tree, not from historical commits.
Data screening results
Examined 98,192,743 source records and retained 96,948,408 documents (96,938,586 training and 9,822 validation). Hygiene, exact deduplication and split/evaluation safeguards excluded 1,224,235 records (1.2468%). Another 13,673 records were skipped because their assigned split quota was full, and 6,427 buffered records were not selected after quotas filled; these are not quality failures.
Recorded exclusion reason
Records
Exact content duplicates
416,177
Environment, VCS, standard-library or binary copies
252,315
English-language checks
275,360
Too short
125,602
Python 3.11 syntax check failures
81,431
Excessive repetition
39,566
Excess markup, URLs or UI/commercial boilerplate
15,804
Tokenizer fit/heldout conflicts
5,971
Oversized records
5,250
Broken encoding or control characters
3,968
Conspicuously unfinished code answers
1,809
Embedded or chunked encoded payloads
607
Evaluation prompt matches
375
Counts are source records, not tokens, and each excluded record is counted under its first failing check. Exact rejected-token volume was not measured. The 375 evaluation-prompt matches comprised 258 ARC-Easy, 94 ARC-Challenge and 23 PIQA records; zero recorded matches for the other tasks does not establish absence of contamination. Python syntax screening is not execution or correctness validation.
Download the complete cleaning counts and per-source breakdown. Exclusion rates apply only to consumed records and use source-specific rules; they are not a quality ranking of upstream datasets. Mathematical/factual correctness, proof validity, code execution, semantic deduplication and exhaustive benchmark decontamination were not established by these checks.
Quality and limitations
This corpus was structurally screened, exact-content deduplicated, assigned to train/validation using content and available source identity, and checked against tokenizer-fit and exact evaluation-prompt exclusions. The builder and finalizer receipts describe these checks. They do not certify semantic correctness, mathematical proof validity, QA entailment, semantic near-duplicate removal, complete identity isolation, or complete benchmark decontamination. The accepted UltraData-Math L3 sample audit found mathematical and answer errors; see corpus-v2/provenance/AUDIT.md. One-shard samples are bounded evidence and are not a corpus-wide error-rate estimate.
No raw source text or sample previews are included as separate files. Upstream terms remain applicable to source content; this combined tokenized derivative does not impose a new blanket license. Consult each upstream card/license and the detailed source-terms discussion in corpus-v2/provenance/source-research.md.
This repository contains no model weights, training code package, or model card. This is a dataset card.
