Hugging Face's logo

Datasets:
beardymcgee
/
broadbase-v0-r1-64g-tokv2

Tasks:
Text Generation
Languages:
English
Size:
10M<n<100M
Tags:
pretraining
tokenized
sentencepiece
uint16
cascamini
common-pile
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
byte_count: int64
channel_stats: string
content_token_count: int64
document_end_count: int64
document_start_count: int64
domain_stats: string
eod_count: int64
first_document_index: int64
full_target_shard: bool
last_document_index: int64
observed_forbidden_non_eod_control_count: int64
observed_maximum_token_id: int64
observed_minimum_token_id: int64
observed_unknown_token_count: int64
path: string
raw_bytes_credited_at_document_end: int64
schema_version: int64
sha256: string
shard_index: int64
split: string
token_count: int64
verification_failures: list<item: null>
  child 0, item: null
verification_pass: bool
to
{'byte_count': Value('int64'), 'channel_stats': Json(decode=True), 'content_token_count': Value('int64'), 'document_end_count': Value('int64'), 'document_start_count': Value('int64'), 'domain_stats': Json(decode=True), 'eod_count': Value('int64'), 'first_document_index': Value('int64'), 'full_target_shard': Value('bool'), 'last_document_index': Value('int64'), 'path': Value('string'), 'raw_bytes_credited_at_document_end': Value('int64'), 'schema_version': Value('int64'), 'sha256': Value('string'), 'shard_index': Value('int64'), 'split': Value('string'), 'token_count': Value('int64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              byte_count: int64
              channel_stats: string
              content_token_count: int64
              document_end_count: int64
              document_start_count: int64
              domain_stats: string
              eod_count: int64
              first_document_index: int64
              full_target_shard: bool
              last_document_index: int64
              observed_forbidden_non_eod_control_count: int64
              observed_maximum_token_id: int64
              observed_minimum_token_id: int64
              observed_unknown_token_count: int64
              path: string
              raw_bytes_credited_at_document_end: int64
              schema_version: int64
              sha256: string
              shard_index: int64
              split: string
              token_count: int64
              verification_failures: list<item: null>
                child 0, item: null
              verification_pass: bool
              to
              {'byte_count': Value('int64'), 'channel_stats': Json(decode=True), 'content_token_count': Value('int64'), 'document_end_count': Value('int64'), 'document_start_count': Value('int64'), 'domain_stats': Json(decode=True), 'eod_count': Value('int64'), 'first_document_index': Value('int64'), 'full_target_shard': Value('bool'), 'last_document_index': Value('int64'), 'path': Value('string'), 'raw_bytes_credited_at_document_end': Value('int64'), 'schema_version': Value('int64'), 'sha256': Value('string'), 'shard_index': Value('int64'), 'split': Value('string'), 'token_count': Value('int64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

BroadBase v0 R1 64 GiB TokV2 Shards

Release status: public dataset release.

This repository provides the verified TokV2 token shards, tokenizer, manifests, checksums, and provenance metadata for BroadBase v0 R1. The token stream is a lossless mixed-source derivative and is therefore released as license: other. Tokenisation does not remove upstream licence, attribution, privacy, or usage obligations.

BroadBase v0 R1 is a deterministic, broad-domain base-pretraining corpus built for the CascaMini project from 12 pinned Common Pile v0.1 source datasets. This repository distributes the model-ready TokV2 token stream, not raw text.

Core statistics

Split Documents Raw UTF-8 bytes Content tokens EOD tokens Stream tokens Shards
Train 10,434,835 68,720,432,426 18,730,418,617 10,434,835 18,740,853,452 70
Held-out 342,740 2,148,211,724 585,037,863 342,740 585,380,603 3
Total 10,777,575 70,868,644,150 19,315,456,480 10,777,575 19,326,234,055 73

The token shards contain 38,652,468,110 bytes, approximately 36.0 GiB, before the small metadata and tokenizer files.

Domain composition

The mixture was controlled by raw UTF-8 bytes rather than document count. Document lengths differ substantially across domains.

Domain Documents Document share Raw bytes Byte share Content tokens Token share
General web 2,155,115 20.00% 19,842,792,639 28.00% 5,126,346,878 26.54%
Science and technical 279,337 2.59% 10,630,162,514 15.00% 2,746,463,299 14.22%
Code 2,614,975 24.26% 10,630,054,329 15.00% 3,398,955,157 17.60%
Education 325,057 3.02% 8,598,699,776 12.13% 2,081,820,772 10.78%
Reference 2,493,981 23.14% 7,086,707,722 10.00% 1,802,850,761 9.33%
Structured data 1,302,405 12.08% 4,252,025,221 6.00% 1,245,671,345 6.45%
Books and literature 10,289 0.10% 3,913,816,750 5.52% 934,722,491 4.84%
Mathematics 1,489,508 13.82% 3,787,755,476 5.34% 1,199,066,665 6.21%
Dialogue 106,908 0.99% 2,126,629,723 3.00% 779,559,112 4.04%

Sources and pinned revisions

Source dataset Domain/channel Pinned revision
common-pile/cccc_filtered general 03a3de5713a0bb23267d26724346508af0f25327
common-pile/pressbooks_filtered education 1a1d3b50d77f834370f8eb4c0d174668dd1676bb
common-pile/oercommons_filtered education 506b6159dadcbc0dc67611cea024eedb04232fb2
common-pile/wikimedia_filtered reference 0641bb84bd9b7162bcddf8be7836822161a9a342
common-pile/peS2o_filtered science and technical 297747513bfb0ff1fbf61ddad3b03319d0f04597
common-pile/arxiv_papers_filtered science and technical 033cf7f53f9b348deec868c1a5a48484f3ee9e52
common-pile/stackv2_edu_filtered code and structured data c354dbe88469a1153e97c6a63ac50591849654de
common-pile/libretexts_filtered mathematics 70388bca52b4a93515e14b1d56618fd7944988fd
common-pile/stackexchange_filtered mathematics c0ac7373830c688a43fc12d1988c4b19ccd884ab
common-pile/ubuntu_irc_filtered dialogue 84f88c986584f11d672befab542fa4d5123f3e8f
common-pile/doab_filtered education defb24ca72ef6aba6ce0228b669eec06dcfbffbc
common-pile/pre_1929_books_filtered books and literature 23f9d96dbb1db3324bbc9fbfe1f8299cc799c4d1

The Stack v2 source is divided into disjoint code and structured-data channels, which is why the recipe has 13 channels but 12 source repositories.

Construction

  • Frozen recipe: BroadBase-v0-R1-64GiB-TokV2
  • Selection seed: 20260720
  • Global deduplication: exact SHA-256 of UTF-8 document bytes
  • Ordering: fixed channel order, train then held-out, preserving input JSONL order
  • Selection: whole documents with explicit, recorded quota overshoot
  • Tokenizer: TokV2-Broad-01-Lossless-65536
  • SentencePiece version: 0.2.2
  • Vocabulary size: 65,536
  • Stored type: little-endian unsigned 16-bit integer
  • Document boundary: exactly one <|eod|> token, ID 4, after every document
  • Ordinary content: control IDs 0 through 15 are forbidden
  • BOS, EOS, chat, role and tool tokens: not injected
  • Target shard size: 268,435,456 tokens, 512 MiB for a full shard

Documents can cross shard boundaries. The shards form one virtual stream per split and must be read in manifest order.

Loading

After downloading the repository, use the included loader:

from loader.load_uint16 import UInt16ShardStream

train = UInt16ShardStream("/path/to/repository", split="train")
tokens = train.read(start=0, count=2048)

The format is intentionally training-oriented. It is not a standard tabular dataset and is not expected to render raw examples in the Dataset Viewer.

Verification and durable identity

Construction and a separate full-shard verifier both reported Overall: PASS.

Artefact SHA-256
dataset_manifest.json f0d5cdec63a21ca3141be06c4eb64eb6ac1ec29cf6913925e324331f39f54f8d
checksums.sha256 df32b278401fb9770cd8279ab2d81ad5e87262fce48b2f05e8a343cac237a812
verification/verification_report.json 65975c38098cece912738aa06486491ad39138ce822c2469b65db01f8f7e0874
verification/checksums.sha256 c3742b408739d260a991bd7c1690bd11cdb7a7469d9201445a0e018c14503114

The internal manifest candidate repeats TokV2 in its generated name. That is a cosmetic builder artefact; the verified file has not been edited.

Licences and attribution

This is a mixed-licence compilation. license: other is deliberate. Each underlying work retains its original public-domain status or open licence. Tokenisation does not remove those conditions because TokV2 is lossless.

The selected source records retain source identity, URL, authors where available, provenance and licence evidence. The document-level publication export and its audit must pass before this repository is made public. Consult that export and the upstream dataset cards before decoding or redistributing material.

The upstream Common Pile authors explicitly warn that licence laundering and incorrect metadata cannot be eliminated completely. The CCCC records at the pinned revision also omit a per-record licence field despite the source card's claim. BroadBase preserves that limitation and relies on the pinned CCCC licensing documentation together with record URL and provenance evidence.

Risks and limitations

  • The corpus can contain personal information, offensive text, factual errors, harmful instructions, duplicated concepts and biased representation.
  • Code and structured records are untrusted text and must not be executed.
  • Upstream licence evidence can be wrong or incomplete.
  • Exact-hash deduplication does not remove near duplicates.
  • The held-out split is isolated from the selected train documents by exact UTF-8 hash, but no broad benchmark-contamination claim is made.
  • This release is tied to the included TokV2 tokenizer and its exact ID mapping.

Use a repository Discussion to report removal, privacy, provenance or licence issues. Include the source record ID, URL or raw SHA-256 where possible.

Citation

Please cite both this compilation and the Common Pile source work:

@misc{broadbase_v0_r1_2026,
  author       = {Matt K},
  title        = {BroadBase v0 R1 64 GiB TokV2 Shards},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/beardymcgee/broadbase-v0-r1-64g-tokv2}}
}

@article{kandpal2025common,
  title   = {The Common Pile v0.1: An 8T
Downloads last month
140
Total file size:
38.7 GB