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:    TypeError
Message:      Couldn't cast array of type
struct<fineweb_edu_train_00001_00000_tokens.bin: struct<indexed_documents: int64, indexed_tokens: int64, sha256: string, size_bytes: int64>, fineweb_edu_train_00001_00000_tokens.idx: struct<sha256: string, size_bytes: int64>>
to
{'fineweb_edu_train_00000_00000_tokens.bin': {'indexed_documents': Value('int64'), 'indexed_tokens': Value('int64'), 'sha256': Value('string'), 'size_bytes': Value('int64')}, 'fineweb_edu_train_00000_00000_tokens.idx': {'sha256': Value('string'), 'size_bytes': Value('int64')}}
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 2303, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1852, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2149, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<fineweb_edu_train_00001_00000_tokens.bin: struct<indexed_documents: int64, indexed_tokens: int64, sha256: string, size_bytes: int64>, fineweb_edu_train_00001_00000_tokens.idx: struct<sha256: string, size_bytes: int64>>
              to
              {'fineweb_edu_train_00000_00000_tokens.bin': {'indexed_documents': Value('int64'), 'indexed_tokens': Value('int64'), 'sha256': Value('string'), 'size_bytes': Value('int64')}, 'fineweb_edu_train_00000_00000_tokens.idx': {'sha256': Value('string'), 'size_bytes': Value('int64')}}

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.

FineWeb-Edu GPT-2 Megatron

This repository contains a deterministic, pretokenized Megatron indexed dataset built from a pinned revision of the globally shuffled FineWeb-Edu 100BT subset.

Important split-overlap disclosure

The split source files are disjoint, but their text is not content-deduplicated across splits. The artifact is published unchanged with the measured overlap preserved:

Held-out split Exact matches with training MinHash candidate overlap with training
development 375 documents (4.9349%); 362,367 IDs (4.5294%) 598 documents (7.8695%); 639,308 IDs (7.9911%)
final 4,909 documents (5.0697%); 4,743,920 IDs (4.7439%) 7,873 documents (8.1307%); 8,491,487 IDs (8.4914%)

Exact matches were confirmed with SHA-256 and raw UTF-8 equality. MinHash results are candidate clusters, not confirmed near duplicates. Do not describe the development or final split as contamination-free.

This behavior is consistent with the upstream dataset: FineWeb-Edu inherits FineWeb's independent per-CommonCrawl-dump deduplication, while the source used here was subsequently shuffled globally.

Contents

Split Prefix Documents Available indexed IDs UTF-8 source bytes
training data/train 7,290,286 7,556,553,510 35,173,581,642
development data/development 7,599 8,000,266 37,310,247
final data/final 96,830 100,001,217 465,855,889

The training artifact contains complete source shards 0 through 7. The available indexed-ID count is the physical capacity of those selected shards; downstream users choose their own consumption budget.

Each nonempty source document was encoded independently with the pinned GPT-2 tokenizer and followed by exactly one EOD ID 50,256. No external normalization, implicit BOS, or other implicit special token was applied. Token IDs are stored as uint16.

Pinned inputs

Item Revision
Source HuggingFaceFW/fineweb_edu_100BT-shuffled@be6b2a50d3a9c60d330c45384e80c7863cd3a25d
Tokenizer openai-community/gpt2@607a30d783dfa663caf39e06633721c8d4cfcd7e
DataTrove 87f7bad5c4a56ec648265fbf0b91d7d226bad428
Megatron-LM / MCore 571370c829ca768fe37244f4e2e7f28d8accc4ab

The tokenizer vocabulary size is 50,257. The pinned tokenizer.json SHA-256 is 8414cab924d8b9b33013f0d221c5862f365ee9be39c5c2bfae8a5a9e970478a6. Environment and source-file identities are recorded under provenance/ and in artifact-manifest.json.

Loading with pinned MCore

Use independent prefixes:

--train-data-path /path/to/fineweb-edu-gpt2-megatron/data/train
--valid-data-path /path/to/fineweb-edu-gpt2-megatron/data/development
--test-data-path /path/to/fineweb-edu-gpt2-megatron/data/final
--tokenizer-type NullTokenizer
--vocab-size 50257
--null-tokenizer-eod-id 50256
--null-tokenizer-pad-id -1

Do not combine these independent split prefixes with --split or --data-path.

The artifact was verified with pinned MCore as a memory-mapped uint16 IndexedDataset. Document boundaries are contiguous and the first and last document of every split end in EOD 50,256.

Integrity

File SHA-256
data/train.bin 94c1cd2266a162a615d27b6b11ac1c1cc5887d629742eefad8efff1976b46222
data/train.idx 564815271149b0322bd262b9f681e3feb30a96d7c677162ad5ef9f2ee3458614
data/development.bin b806f2f1ba34e8d86c3f6517f9aa305fa554d850c44362790bf6b84b88f47d55
data/development.idx aac74ffcf4d090a3bc96c320227a25805f3b0666f4595e36ac6f98a0c828d713
data/final.bin 9fb1f3d21deabe8b9cc85675b21cebb7434f1221ea8cc8ebb825069729e97a49
data/final.idx efca13527d24b44423fc521947ce8c630b6739e15f5d8139ad84d9efc56be85d

artifact-manifest.json records every uploaded payload file's size and SHA-256, plus a deterministic tree hash.

Audits

The exhaustive exact audit scanned all selected training documents and the complete held-out prefixes. It found zero SHA-1-64 collision candidates after SHA-256 and raw-byte confirmation.

The full MinHash audit covered 7,394,715 documents, 7,664,554,993 indexed IDs, and 35,676,747,778 UTF-8 bytes. Every document received a 5-gram signature. The configuration was English spaCy tokenization, 5-grams, 14 buckets, 8 hashes per bucket, SHA-1 at 64-bit precision, seed 1, and four CPU workers. It reported 8,439 training-to-held-out candidate clusters containing 19,640 documents. The retained audit tree's SHA-256 is d550804eb3869d5c99cf3466383264088cf0a6b9ac92e7377012f1056a49d92b.

Compact reports and candidate metadata are under audits/. No filtering, removal, or backfill step was applied.

License

The source dataset is released under ODC-By. This derived tokenized artifact retains that license.

Downloads last month
31