cortex-20b

A 20 B-token (100 B-character), cap-balanced English/Japanese pretraining mixture assembled for general-purpose language model training. Release tag v0.9 (mixture build balanced-pretraining-v8, finalized 2026-09-15).

The dataset is partitioned into three splits (train, validation, test) and stores 22.4 M text records on disk in zstd-compressed Parquet. The source stream spans web text, encyclopedias, news, public-domain books, academic articles, patents, court documents, code, Reddit submissions, Japanese web fiction, dialogue/game transcripts, and OpenStreetMap/Overture spatial data.

At a glance

Split	Records	Characters (text)	Bytes (raw JSONL)	Parquet shards
train	22,426,697	96,963,026,897	127,541,814,421 (127 GB)	113 × ~400 MB
validation	17,087	124,708,316	145,751,270 (140 MB)	1
test	16,773	103,121,121	123,351,487 (118 MB)	1
Total	22,460,557	97,190,856,334	127,810,917,178	115

Approximate tokens (4 chars/token heuristic): 24.3 B — round-figure label cortex-20b is what the card sticks to; the heuristic under-counts multilingual content.

Load

from datasets import load_dataset

# Full mix (streaming-friendly, 22.4 M rows).
ds = load_dataset("m8than/cortex-20b", split="train")

# Just validation / test.
val = load_dataset("m8than/cortex-20b", split="validation")
test = load_dataset("m8than/cortex-20b", split="test")

# Or by category / source dataset:
dclm = load_dataset("m8than/cortex-20b", split="train").filter(
    lambda r: "HuggingFaceTB/dclm-edu" in r["source"]
)

# Streaming (recommended for the full train split):
ds = load_dataset("m8than/cortex-20b", split="train", streaming=True)
for row in ds:
    ...

Each row has the schema below:

Column	Type	Description
id	string	Stable per-record id (preserved from upstream datasets where possible).
kind	string	Always "text" for this release.
source	string (JSON)	Upstream provenance block as JSON (dataset, license, URL, IDs, etc.).
quality	string (JSON)	Quality/filtering metadata as JSON (version, tier, score, flags, metrics).
mixture	string (JSON)	Per-record mixture-bucket info as JSON (input source, cap, license gate).
extra	string (JSON)	Any record-level fields that don't fit the columns above (e.g. legal_moves, action, messages, annotations). Empty for most rows.
text	string	The training text itself.
split_group	string	Stable cross-split identity used for split construction (provenance key, family id, etc.). Rows lacking one are train-only.

Build summary

• Mixture version: 8 (mixture_config: data/config/mixture-v8.json, sha256 8554981b…51de96).
• Release tag: v0.9 (mixture build is balanced-pretraining-v8).
• Seed: balanced-v8 (sampling), balanced-v3-split (splitting).
• Split ratios: train 98 % / validation 1 % / test 1 %, split at the source.split_group level where one exists; otherwise train-only.
• Determinism: exact text deduplication (142,339 dupes dropped) and deterministic per-bucket rank-based subsampling so the release is reproducible bit-for-bit given the same seed.
• Verification: python scripts/build_release.py --audit release-v8 confirms malformed_records=0, cross_split_group_records=0, expected_split_mismatches=0.

Category breakdown (final, post-dedup)

Category	Records	Text chars	Share
general (web, encyclopedias, news, speech, social)	15,686,992	41,030,478,036	42.3 %
code (Stack v3 + Nemotron-CC-Code)	5,910,855	43,385,459,535	44.7 %
academic (arXiv, PubMed, USPTO, free law, MedQA)	155,000	3,834,484,147	3.9 %
books (Gutenberg, US-PD, LoC-PD)	6,629	2,301,415,159	2.4 %
fiction (Japanese web novels)	215,854	4,999,969,532	5.1 %
dialogue (CRD3, FIREBALL, Hanabi, blackjack, …)	196,802	424,223,265	0.4 %
spatial (OpenStreetMap, Overture)	254,565	986,997,223	1.0 %

Source attribution

All inputs are listed below with their upstream HF identifier, license, upstream URL, and the post-dedup record count that ended up in this release.

Source dataset	License	Records	Upstream
nvidia/Nemotron-CC-Code-v1	CC-BY-4.0	5,677,001	https://huggingface.co/datasets/nvidia/Nemotron-CC-Code-v1
Zyphra/Zyda-2	CC-BY-4.0 / ODbL	3,414,053	https://huggingface.co/datasets/Zyphra/Zyda-2
HuggingFaceTB/dclm-edu	CC-BY-4.0	2,408,652	https://huggingface.co/datasets/HuggingFaceTB/dclm-edu
common-pile/wikiteam_filtered	CC-BY-SA 3.0/4.0	594,267	https://huggingface.co/datasets/common-pile/wikiteam_filtered
HuggingFaceFW/finewiki	CC-BY-SA 4.0	396,723	https://huggingface.co/datasets/HuggingFaceFW/finewiki
HuggingFaceCode/stack-v3-train	per-file	233,854	https://huggingface.co/datasets/HuggingFaceCode/stack-v3-train
tvtropes	CC-BY-NC-SA-3.0	233,704	https://tvtropes.org (WikiTeam extract, see manifest for provenance)
syosetu	LicenseRef-HF-p1atdev-syosetu711k	215,854	https://huggingface.co/datasets/p1atdev/syosetu711k
open2roam/openstreetmaps	ODbL-1.0	202,488	https://huggingface.co/datasets/open2roam/openstreetmaps
lara-martin/FIREBALL	MIT (derived)	121,042	https://huggingface.co/datasets/lara-martin/FIREBALL
common-pile/news_filtered	per-source	105,818	https://huggingface.co/datasets/common-pile/news_filtered
common-pile/uspto_filtered	CC-BY-4.0	62,377	https://huggingface.co/datasets/common-pile/uspto_filtered
common-pile/youtube_filtered	CC-BY-4.0	62,147	https://huggingface.co/datasets/common-pile/youtube_filtered
overturemaps/places	ODbL-1.0	52,077	https://huggingface.co/datasets/overturemaps/places
common-pile/pubmed_filtered	CC0 / per-article	44,564	https://huggingface.co/datasets/common-pile/pubmed_filtered
common-pile/caselaw_access_project_filtered	public-domain (US)	34,007	https://huggingface.co/datasets/common-pile/caselaw_access_project_filtered
Mahesh111000/Hanabi_dataset	MIT	29,965	https://huggingface.co/datasets/Mahesh111000/Hanabi_dataset
microsoft/crd3	CC-BY-4.0	27,744	https://huggingface.co/datasets/microsoft/crd3
steam/appreviews-wide	Steam-Valve-Terms	12,880	https://huggingface.co/datasets/steam/appreviews
GBaker/MedQA-USMLE-4-options	MIT	10,176	https://huggingface.co/datasets/GBaker/MedQA-USMLE-4-options
common-pile/project_gutenberg_filtered	public-domain	5,941	https://huggingface.co/datasets/common-pile/project_gutenberg_filtered
scholarweave/arxiv-latex	per-paper (allowlisted: CC0/CC-BY/CC-BY-SA)	3,876	https://huggingface.co/datasets/scholarweave/arxiv-latex
steam/appreviews	Steam-Valve-Terms	2,704	https://huggingface.co/datasets/steam/appreviews
blackjack-open-rules	public-domain	2,443	upstream rules document
storytracer/US-PD-Books	public-domain	527	https://huggingface.co/datasets/storytracer/US-PD-Books
storytracer/LoC-PD-Books	public-domain	161	https://huggingface.co/datasets/storytracer/LoC-PD-Books
Idrinth/gamemasterai	MIT (upstream roleplay-ai)	24	https://huggingface.co/datasets/Idrinth/gamemasterai
reddit/pushshift-submissions	LicenseRef-Pushshift-Research (research only)	8,471,628	Pushshift mirror

The Pushshift/Reddit slice is included for research use under the Pushshift research license. Downstream commercial model weights derived from this dataset may need to remove or replace the Reddit slice; consult your legal counsel.

License

This dataset is released as a mixture under the Open Data Commons Attribution License 1.0 (ODC-BY 1.0). Each individual record retains its upstream license — see the source.license field on every row and the rights review embedded in auxiliary_metadata/manifest.json. Applying this mixture to train a model requires that downstream use respects the most restrictive per-input license (in particular: NC clauses from TVTropes, research-use terms from Pushshift, per-file Stack v3 licenses).

Rights review

The full per-source rights review (date 2026-09-05) is included in auxiliary_metadata/manifest.json under rights_review. Highlights:

• Excluded: mastermind, europarl (pending provenance), wikimedical (derivative of FineWiki), terraria-wiki, minecraft-wiki (NC on primary sources), dnd-dm-v3 (no verifiable upstream).
• Per-paper allowlist: arXiv only includes CC0/CC-BY/CC-BY-SA submissions; NC/ND/unknown were rejected before sampling (83,711 records / 31.6 B characters dropped).
• Pushshift/Reddit: included under research license only.
• Steam reviews: included under Steam/Valve terms.

Repository layout

.
├── README.md
├── LICENSE                                # ODC-BY 1.0
├── dataset_infos.json                     # features / split sizes for load_dataset
├── auxiliary_metadata/
│   ├── manifest.json                      # Full v8 build manifest
│   ├── finalize.manifest.json             # Finalize-stage manifest
│   ├── verification.json                  # Audit script output
│   └── SHA256SUMS                         # Per-file SHA256s of the original JSONL
└── data/
    ├── train-00000.parquet
    ├── train-00001.parquet
    ├── …
    ├── train-00112.parquet
    ├── validation-00000.parquet
    └── test-00000.parquet

Parquet files are zstd-compressed, ~400 MB each (200,000 records per train shard). Load with datasets.load_dataset(..., split="train") or stream shard-by-shard via streaming=True.

Provenance and reproducibility

• Built with python scripts/build_release.py --mixture data/config/mixture-v8.json.
• Sampling seed: balanced-v8. Splitting seed: balanced-v3-split.
• Each shard's SHA256 is recorded in auxiliary_metadata/shard_stats.jsonl (generated alongside the upload).
• The original JSONL split files are hashed in auxiliary_metadata/SHA256SUMS.

Cite

@misc{cortex-20b,
  title  = {cortex-20b: a 100B-character / 20B-token cap-balanced pretraining mixture (v0.9)},
  author = {Wilce, Nathan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/m8than/cortex-20b},
}