Datasets:

birgermoell
/

oellm-longctx-tokenized-streamed-all-v2

Tasks:

Text Generation

Languages:

Bulgarian

Bosnian

Catalan

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

train

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
alpha: null
floor: null
suffix: string
bin_dir: string
total_bin_bytes: int64
tiers: struct<16k_plus: struct<min: int64, weight: double>, 4_16k: struct<min: int64, max: int64, weight: d (... 52 chars omitted)
  child 0, 16k_plus: struct<min: int64, weight: double>
      child 0, min: int64
      child 1, weight: double
  child 1, 4_16k: struct<min: int64, max: int64, weight: double>
      child 0, min: int64
      child 1, max: int64
      child 2, weight: double
  child 2, under4k: struct<max: int64, weight: double>
      child 0, max: int64
      child 1, weight: double
languages: struct<lc16k_smoke_20260507_mt_s0000_c00000_16k_plus: struct<weight: double, bin_bytes: int64, natur (... 538 chars omitted)
  child 0, lc16k_smoke_20260507_mt_s0000_c00000_16k_plus: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
      child 0, weight: double
      child 1, bin_bytes: int64
      child 2, natural_p: double
      child 3, prefix: string
      child 4, tier: string
      child 5, language: string
      child 6, source: string
      child 7, docs: int64
      child 8, source_token_count: int64
  child 1, lc16k_smoke_20260507_mt_s0000_c00000_4_16k: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
      child 0, weight: double
      child 1, bin_bytes: int64
      child 2, natural_p: double
      child 3, prefix: string
      child 4, tier: string
      child 5, language: string
      child 6, source: string
      child 7, docs: int64
      child 8, source_token_count: int64
  child 2, lc16k_smoke_20260507_mt_s0000_c00000_under4k: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
      child 0, weight: double
      child 1, bin_bytes: int64
      child 2, natural_p: double
      child 3, prefix: string
      child 4, tier: string
      child 5, language: string
      child 6, source: string
      child 7, docs: int64
      child 8, source_token_count: int64
data_path_args: string
artifact_format: string
tokenizer_type: string
source_dataset: string
entries: list<item: struct<name: string, language: string, tier: string, source: string, docs: int64, source_ (... 128 chars omitted)
  child 0, item: struct<name: string, language: string, tier: string, source: string, docs: int64, source_token_count (... 116 chars omitted)
      child 0, name: string
      child 1, language: string
      child 2, tier: string
      child 3, source: string
      child 4, docs: int64
      child 5, source_token_count: int64
      child 6, bin_bytes: int64
      child 7, idx_bytes: int64
      child 8, bin_sha256: string
      child 9, idx_sha256: string
      child 10, weight: double
      child 11, prefix: string
tokenizer_model: string
created_at_unix: int64
vocab_size: int64
to
{'created_at_unix': Value('int64'), 'artifact_format': Value('string'), 'source_dataset': Value('string'), 'tokenizer_type': Value('string'), 'tokenizer_model': Value('string'), 'vocab_size': Value('int64'), 'entries': List({'name': Value('string'), 'language': Value('string'), 'tier': Value('string'), 'source': Value('string'), 'docs': Value('int64'), 'source_token_count': Value('int64'), 'bin_bytes': Value('int64'), 'idx_bytes': Value('int64'), 'bin_sha256': Value('string'), 'idx_sha256': Value('string'), 'weight': Value('float64'), 'prefix': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
                  return get_rows(
                         ^^^^^^^^^
                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 77, in get_rows
                  rows_plus_one = list(itertools.islice(ds, rows_max_number + 1))
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2690, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2227, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2251, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 494, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 384, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 299, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 128, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2321, in table_cast
                  return cast_table_to_schema(table, schema)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2249, in cast_table_to_schema
                  raise CastError(
              datasets.table.CastError: Couldn't cast
              alpha: null
              floor: null
              suffix: string
              bin_dir: string
              total_bin_bytes: int64
              tiers: struct<16k_plus: struct<min: int64, weight: double>, 4_16k: struct<min: int64, max: int64, weight: d (... 52 chars omitted)
                child 0, 16k_plus: struct<min: int64, weight: double>
                    child 0, min: int64
                    child 1, weight: double
                child 1, 4_16k: struct<min: int64, max: int64, weight: double>
                    child 0, min: int64
                    child 1, max: int64
                    child 2, weight: double
                child 2, under4k: struct<max: int64, weight: double>
                    child 0, max: int64
                    child 1, weight: double
              languages: struct<lc16k_smoke_20260507_mt_s0000_c00000_16k_plus: struct<weight: double, bin_bytes: int64, natur (... 538 chars omitted)
                child 0, lc16k_smoke_20260507_mt_s0000_c00000_16k_plus: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
                    child 0, weight: double
                    child 1, bin_bytes: int64
                    child 2, natural_p: double
                    child 3, prefix: string
                    child 4, tier: string
                    child 5, language: string
                    child 6, source: string
                    child 7, docs: int64
                    child 8, source_token_count: int64
                child 1, lc16k_smoke_20260507_mt_s0000_c00000_4_16k: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
                    child 0, weight: double
                    child 1, bin_bytes: int64
                    child 2, natural_p: double
                    child 3, prefix: string
                    child 4, tier: string
                    child 5, language: string
                    child 6, source: string
                    child 7, docs: int64
                    child 8, source_token_count: int64
                child 2, lc16k_smoke_20260507_mt_s0000_c00000_under4k: struct<weight: double, bin_bytes: int64, natural_p: double, prefix: string, tier: string, language:  (... 63 chars omitted)
                    child 0, weight: double
                    child 1, bin_bytes: int64
                    child 2, natural_p: double
                    child 3, prefix: string
                    child 4, tier: string
                    child 5, language: string
                    child 6, source: string
                    child 7, docs: int64
                    child 8, source_token_count: int64
              data_path_args: string
              artifact_format: string
              tokenizer_type: string
              source_dataset: string
              entries: list<item: struct<name: string, language: string, tier: string, source: string, docs: int64, source_ (... 128 chars omitted)
                child 0, item: struct<name: string, language: string, tier: string, source: string, docs: int64, source_token_count (... 116 chars omitted)
                    child 0, name: string
                    child 1, language: string
                    child 2, tier: string
                    child 3, source: string
                    child 4, docs: int64
                    child 5, source_token_count: int64
                    child 6, bin_bytes: int64
                    child 7, idx_bytes: int64
                    child 8, bin_sha256: string
                    child 9, idx_sha256: string
                    child 10, weight: double
                    child 11, prefix: string
              tokenizer_model: string
              created_at_unix: int64
              vocab_size: int64
              to
              {'created_at_unix': Value('int64'), 'artifact_format': Value('string'), 'source_dataset': Value('string'), 'tokenizer_type': Value('string'), 'tokenizer_model': Value('string'), 'vocab_size': Value('int64'), 'entries': List({'name': Value('string'), 'language': Value('string'), 'tier': Value('string'), 'source': Value('string'), 'docs': Value('int64'), 'source_token_count': Value('int64'), 'bin_bytes': Value('int64'), 'idx_bytes': Value('int64'), 'bin_sha256': Value('string'), 'idx_sha256': Value('string'), 'weight': Value('float64'), 'prefix': Value('string')})}
              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.

OpenEuroLLM long-context Megatron streamed tokenized artifacts

This dataset contains Megatron-LM indexed data (.bin/.idx) uploaded shard by shard from longctx stream-upload. It is intended as a transport format for continued pretraining on LUMI or another training machine; it is not raw text.

Source dataset: HuggingFaceFW/finepdfs-edu
Tooling: openeuro-longctx-datamix
Run namespace: runs/lc16k_full_20260507/
Tokenizer type: HuggingFaceTokenizer
Tokenizer: OpenEuroLLM 256k tokenizer path used during preprocessing
Format: Megatron-LM *_text_document.bin and *_text_document.idx files plus per-chunk JSON manifests
Uploaded artifacts: 3,305 chunk manifests, 3,305 .bin files, 3,305 .idx files
Tokenized artifact size: about 1,115.0 GiB (.bin + .idx, uncompressed file sizes)

Actual Token Length Distribution

The figure below bins documents by actual tokenized Megatron document length, computed from the uploaded .idx document spans rather than the FinePDFs-Edu source token_count field. Cell labels are summed token counts per language and bucket. The exact values, including document counts per bucket, are also available as CSV.

Exact CSV: runs/lc16k_full_20260507/reports/language_length_bins_by_language.csv
Total actual Megatron tokens from .idx: 299,090,791,852
Total documents: 43,510,918
Buckets: <=16k, 16-32k, 32-64k, 64-128k, >128k tokens per document

Included Languages And Length Buckets

This snapshot includes the 35 OpenEuroLLM target languages available in FinePDFs-Edu. The three target languages not present in FinePDFs-Edu (ga, sq, lb) are not included in this artifact snapshot and need separate sourcing before tokenization.

Length buckets are based on the FinePDFs-Edu source token_count field:

Bucket	Source document length
`<4k`	fewer than 4,000 source tokens
`4-16k`	4,000 to fewer than 16,000 source tokens
`>=16k`	16,000 or more source tokens

Sizes below are uploaded Megatron artifact sizes in GiB, not compressed source parquet size. Each bucket cell is chunks / GiB.

ISO	Language	Chunks	Tokenized GiB	`>=16k` chunks/GiB	`4-16k` chunks/GiB	`<4k` chunks/GiB
`bg`	Bulgarian	15	4.6	4 / 3.4	4 / 0.9	7 / 0.3
`bs`	Bosnian	9	3.7	3 / 3.0	3 / 0.5	3 / 0.2
`ca`	Catalan	20	4.3	4 / 2.6	4 / 1.1	12 / 0.6
`cs`	Czech	53	15.2	8 / 10.5	14 / 3.3	31 / 1.5
`cy`	Welsh	3	0.3	1 / 0.1	1 / 0.1	1 / 0.0
`da`	Danish	22	6.3	4 / 4.2	4 / 1.4	14 / 0.7
`de`	German	313	62.7	45 / 34.4	45 / 16.8	223 / 11.6
`el`	Greek	20	8.7	5 / 6.7	5 / 1.3	10 / 0.7
`en`	English	302	109.9	20 / 63.0	60 / 29.5	222 / 17.3
`es`	Spanish	254	106.1	54 / 76.6	54 / 20.7	146 / 8.8
`et`	Estonian	7	1.5	2 / 1.0	2 / 0.3	3 / 0.2
`eu`	Basque	6	1.1	2 / 0.8	2 / 0.2	2 / 0.1
`fi`	Finnish	20	7.3	5 / 5.7	5 / 1.1	10 / 0.5
`fr`	French	129	33.6	21 / 19.4	21 / 8.6	87 / 5.6
`gl`	Galician	4	0.9	1 / 0.6	1 / 0.2	2 / 0.1
`hr`	Croatian	16	6.1	4 / 4.8	4 / 0.9	8 / 0.4
`hu`	Hungarian	38	18.2	10 / 15.0	10 / 2.3	18 / 0.8
`is`	Icelandic	7	0.9	2 / 0.7	2 / 0.2	3 / 0.1
`it`	Italian	146	38.1	25 / 23.7	25 / 8.7	96 / 5.8
`lt`	Lithuanian	11	3.3	3 / 2.5	3 / 0.6	5 / 0.2
`lv`	Latvian	7	2.1	2 / 1.5	2 / 0.4	3 / 0.2
`mk`	Macedonian	6	1.0	2 / 0.8	2 / 0.1	2 / 0.0
`mt`	Maltese	3	0.4	1 / 0.2	1 / 0.1	1 / 0.0
`nl`	Dutch	74	18.3	13 / 10.5	13 / 4.9	48 / 2.9
`no`	Norwegian Bokmal	16	4.9	4 / 3.4	4 / 1.0	8 / 0.4
`pl`	Polish	85	28.1	15 / 20.1	25 / 5.7	45 / 2.3
`pt`	Portuguese	118	47.9	25 / 32.5	43 / 11.9	50 / 3.5
`ro`	Romanian	28	16.5	7 / 13.2	7 / 2.4	14 / 0.9
`ru`	Russian	148	63.2	37 / 47.9	37 / 11.0	74 / 4.3
`sk`	Slovak	20	6.9	4 / 4.8	6 / 1.5	10 / 0.6
`sl`	Slovenian	12	2.3	3 / 1.6	3 / 0.4	6 / 0.3
`sr`	Serbian	12	7.2	4 / 6.3	4 / 0.6	4 / 0.2
`sv`	Swedish	33	12.2	7 / 8.5	7 / 2.5	19 / 1.2
`tr`	Turkish	20	9.1	5 / 6.8	5 / 1.8	10 / 0.6
`uk`	Ukrainian	28	17.8	7 / 14.2	7 / 2.8	14 / 0.8
Total	35 languages	2,005	670.4	359 / 450.8	435 / 145.8	1,211 / 73.8

Repository Layout

The run is stored under runs/lc16k_full_20260507/:

runs/lc16k_full_20260507/
├── bin/                 # Megatron *_text_document.bin/.idx files
└── manifests/chunks/    # one JSON manifest per tokenized chunk

Chunk names follow:

lc16k_full_20260507_<lang>_s<source-shard>_c<chunk>_<tier>

For example:

lc16k_full_20260507_sv_s0000_c00000_under4k_text_document.bin
lc16k_full_20260507_sv_s0000_c00000_under4k_text_document.idx
manifests/chunks/lc16k_full_20260507_sv_s0000_c00000_under4k.json

Use On LUMI Or Another Training Machine

Download the artifact snapshot with the helper in openeuro-longctx-datamix:

python -m longctx.cli artifacts download \
  --repo-id birgermoell/oellm-longctx-tokenized-streamed-all-v2 \
  --output-dir /scratch/project_462000353/$USER/oellm-longctx-tokenized

This repository contains chunk artifacts only for this run. The final global mix metadata was intentionally skipped during the low-disk continuation upload, so rebuild the Megatron data mix from the downloaded .bin/.idx files before training.

Caveats

This is not raw text; it is Megatron indexed binary data.
FinePDFs-Edu license and filtering constraints apply to the source data.
The three target languages missing from FinePDFs-Edu (ga, sq, lb) are not included here.
The uploaded sizes are uncompressed .bin + .idx file sizes as reported by Hugging Face file metadata.

Downloads last month: 7,774

Total file size:

1.2 TB