Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
source: string
docs: list<item: struct<shard_path: string, offset: int64, length: int64, doc_tokens: int64>>
child 0, item: struct<shard_path: string, offset: int64, length: int64, doc_tokens: int64>
child 0, shard_path: string
child 1, offset: int64
child 2, length: int64
child 3, doc_tokens: int64
files: null
length: int64
repo: null
char_len: int64
offset: int64
shard: int64
shard_path: string
doc_tokens: int64
to
{'source': Value('string'), 'shard': Value('int64'), 'offset': Value('int64'), 'length': Value('int64'), 'doc_tokens': Value('int64'), 'char_len': Value('int64'), 'repo': Value('null'), 'files': Value('null'), 'shard_path': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1816, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
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
source: string
docs: list<item: struct<shard_path: string, offset: int64, length: int64, doc_tokens: int64>>
child 0, item: struct<shard_path: string, offset: int64, length: int64, doc_tokens: int64>
child 0, shard_path: string
child 1, offset: int64
child 2, length: int64
child 3, doc_tokens: int64
files: null
length: int64
repo: null
char_len: int64
offset: int64
shard: int64
shard_path: string
doc_tokens: int64
to
{'source': Value('string'), 'shard': Value('int64'), 'offset': Value('int64'), 'length': Value('int64'), 'doc_tokens': Value('int64'), 'char_len': Value('int64'), 'repo': Value('null'), 'files': Value('null'), 'shard_path': Value('string')}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
source string | shard int64 | offset int64 | length int64 | doc_tokens int64 | char_len int64 | repo null | files null | shard_path string |
|---|---|---|---|---|---|---|---|---|
arxiv | 1 | 0 | 6,845 | 6,844 | 16,765 | null | null | arxiv_000001.bin |
arxiv | 1 | 6,845 | 3,846 | 3,845 | 14,486 | null | null | arxiv_000001.bin |
arxiv | 1 | 10,691 | 16,633 | 16,632 | 54,930 | null | null | arxiv_000001.bin |
arxiv | 1 | 27,324 | 36,247 | 36,246 | 83,425 | null | null | arxiv_000001.bin |
arxiv | 1 | 63,571 | 9,228 | 9,227 | 29,504 | null | null | arxiv_000001.bin |
arxiv | 1 | 72,799 | 16,459 | 16,458 | 48,264 | null | null | arxiv_000001.bin |
arxiv | 1 | 89,258 | 41,115 | 41,114 | 87,544 | null | null | arxiv_000001.bin |
arxiv | 1 | 130,373 | 21,302 | 21,301 | 62,770 | null | null | arxiv_000001.bin |
arxiv | 1 | 151,675 | 15,651 | 15,650 | 46,262 | null | null | arxiv_000001.bin |
arxiv | 1 | 167,326 | 9,991 | 9,990 | 28,808 | null | null | arxiv_000001.bin |
arxiv | 1 | 177,317 | 11,232 | 11,231 | 34,897 | null | null | arxiv_000001.bin |
arxiv | 1 | 188,549 | 26,793 | 26,792 | 81,667 | null | null | arxiv_000001.bin |
arxiv | 1 | 215,342 | 17,271 | 17,270 | 50,088 | null | null | arxiv_000001.bin |
arxiv | 1 | 232,613 | 141,325 | 141,324 | 313,623 | null | null | arxiv_000001.bin |
arxiv | 1 | 373,938 | 7,053 | 7,052 | 20,774 | null | null | arxiv_000001.bin |
arxiv | 1 | 380,991 | 2,609 | 2,608 | 9,763 | null | null | arxiv_000001.bin |
arxiv | 1 | 383,600 | 22,230 | 22,229 | 53,115 | null | null | arxiv_000001.bin |
arxiv | 1 | 405,830 | 17,646 | 17,645 | 73,480 | null | null | arxiv_000001.bin |
arxiv | 1 | 423,476 | 13,932 | 13,931 | 48,183 | null | null | arxiv_000001.bin |
arxiv | 1 | 437,408 | 15,132 | 15,131 | 41,303 | null | null | arxiv_000001.bin |
arxiv | 1 | 452,540 | 6,650 | 6,649 | 25,233 | null | null | arxiv_000001.bin |
arxiv | 1 | 459,190 | 9,270 | 9,269 | 33,982 | null | null | arxiv_000001.bin |
arxiv | 1 | 468,460 | 25,690 | 25,689 | 76,093 | null | null | arxiv_000001.bin |
arxiv | 1 | 494,150 | 43,474 | 43,473 | 85,159 | null | null | arxiv_000001.bin |
arxiv | 1 | 537,624 | 11,293 | 11,292 | 33,112 | null | null | arxiv_000001.bin |
arxiv | 1 | 548,917 | 2,437 | 2,436 | 9,134 | null | null | arxiv_000001.bin |
arxiv | 1 | 551,354 | 9,201 | 9,200 | 35,498 | null | null | arxiv_000001.bin |
arxiv | 1 | 560,555 | 20,369 | 20,368 | 58,325 | null | null | arxiv_000001.bin |
arxiv | 1 | 580,924 | 18,594 | 18,593 | 55,709 | null | null | arxiv_000001.bin |
arxiv | 1 | 599,518 | 9,980 | 9,979 | 34,671 | null | null | arxiv_000001.bin |
arxiv | 1 | 609,498 | 9,297 | 9,296 | 26,025 | null | null | arxiv_000001.bin |
arxiv | 1 | 618,795 | 29,104 | 29,103 | 96,935 | null | null | arxiv_000001.bin |
arxiv | 1 | 647,899 | 55,779 | 55,778 | 134,232 | null | null | arxiv_000001.bin |
arxiv | 1 | 703,678 | 24,223 | 24,222 | 69,724 | null | null | arxiv_000001.bin |
arxiv | 1 | 727,901 | 3,738 | 3,737 | 12,631 | null | null | arxiv_000001.bin |
arxiv | 1 | 731,639 | 7,742 | 7,741 | 24,392 | null | null | arxiv_000001.bin |
arxiv | 1 | 739,381 | 20,204 | 20,203 | 64,797 | null | null | arxiv_000001.bin |
arxiv | 1 | 759,585 | 4,555 | 4,554 | 11,880 | null | null | arxiv_000001.bin |
arxiv | 1 | 764,140 | 3,908 | 3,907 | 16,600 | null | null | arxiv_000001.bin |
arxiv | 1 | 768,048 | 12,580 | 12,579 | 41,608 | null | null | arxiv_000001.bin |
arxiv | 1 | 780,628 | 14,586 | 14,585 | 45,121 | null | null | arxiv_000001.bin |
arxiv | 1 | 795,214 | 16,004 | 16,003 | 42,818 | null | null | arxiv_000001.bin |
arxiv | 1 | 811,218 | 25,455 | 25,454 | 53,396 | null | null | arxiv_000001.bin |
arxiv | 1 | 836,673 | 15,750 | 15,749 | 55,726 | null | null | arxiv_000001.bin |
arxiv | 1 | 852,423 | 47,474 | 47,473 | 104,237 | null | null | arxiv_000001.bin |
arxiv | 1 | 899,897 | 15,745 | 15,744 | 40,650 | null | null | arxiv_000001.bin |
arxiv | 1 | 915,642 | 38,393 | 38,392 | 88,502 | null | null | arxiv_000001.bin |
arxiv | 1 | 954,035 | 14,938 | 14,937 | 44,054 | null | null | arxiv_000001.bin |
arxiv | 1 | 968,973 | 13,732 | 13,731 | 38,411 | null | null | arxiv_000001.bin |
arxiv | 1 | 982,705 | 17,956 | 17,955 | 63,447 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,000,661 | 12,680 | 12,679 | 44,105 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,013,341 | 18,660 | 18,659 | 69,392 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,032,001 | 12,859 | 12,858 | 37,030 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,044,860 | 17,445 | 17,444 | 59,659 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,062,305 | 13,408 | 13,407 | 43,571 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,075,713 | 19,974 | 19,973 | 80,307 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,095,687 | 12,868 | 12,867 | 28,531 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,108,555 | 5,640 | 5,639 | 22,091 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,114,195 | 13,885 | 13,884 | 35,923 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,128,080 | 6,608 | 6,607 | 22,869 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,134,688 | 22,118 | 22,117 | 45,091 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,156,806 | 36,561 | 36,560 | 66,642 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,193,367 | 21,904 | 21,903 | 65,034 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,215,271 | 12,088 | 12,087 | 40,518 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,227,359 | 13,649 | 13,648 | 36,505 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,241,008 | 15,456 | 15,455 | 41,608 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,256,464 | 25,210 | 25,209 | 56,437 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,281,674 | 39,451 | 39,450 | 105,179 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,321,125 | 3,378 | 3,377 | 10,374 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,324,503 | 41,716 | 41,715 | 106,816 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,366,219 | 43,991 | 43,990 | 101,117 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,410,210 | 78,869 | 78,868 | 186,661 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,489,079 | 68,883 | 68,882 | 140,969 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,557,962 | 32,540 | 32,539 | 52,576 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,590,502 | 24,732 | 24,731 | 57,862 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,615,234 | 6,244 | 6,243 | 21,657 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,621,478 | 27,110 | 27,109 | 58,844 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,648,588 | 59,236 | 59,235 | 156,343 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,707,824 | 22,497 | 22,496 | 70,660 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,730,321 | 16,663 | 16,662 | 51,270 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,746,984 | 8,625 | 8,624 | 31,747 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,755,609 | 21,579 | 21,578 | 66,095 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,777,188 | 9,476 | 9,475 | 26,168 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,786,664 | 17,298 | 17,297 | 35,806 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,803,962 | 62,760 | 62,759 | 143,090 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,866,722 | 55,633 | 55,632 | 108,335 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,922,355 | 8,432 | 8,431 | 29,577 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,930,787 | 36,996 | 36,995 | 108,996 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,967,783 | 8,899 | 8,898 | 33,313 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,976,682 | 15,793 | 15,792 | 48,144 | null | null | arxiv_000001.bin |
arxiv | 1 | 1,992,475 | 13,830 | 13,829 | 51,156 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,006,305 | 22,155 | 22,154 | 60,595 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,028,460 | 16,352 | 16,351 | 39,566 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,044,812 | 9,122 | 9,121 | 23,005 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,053,934 | 17,268 | 17,267 | 62,092 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,071,202 | 10,121 | 10,120 | 27,731 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,081,323 | 8,570 | 8,569 | 31,442 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,089,893 | 9,392 | 9,391 | 28,361 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,099,285 | 111,174 | 111,173 | 254,000 | null | null | arxiv_000001.bin |
arxiv | 1 | 2,210,459 | 34,153 | 34,152 | 72,543 | null | null | arxiv_000001.bin |
End of preview.
HobbyLM-1B Long-Context Extension Corpus
2.5B GPT-2 tokens staged for extending harims95/hobbylm-1b-hf's context window from 1024 tokens to 4096, then 8192.
Format
Each <source>_NNNNNN.bin file is the same flat-binary format used throughout
HobbyLM training: a 256 x int32 header (magic=20240520, version=1,
num_tokens), followed by the token stream as uint16 GPT-2 token ids.
The flat .bin files do not encode document boundaries on their own. Each
source also ships a <source>_index.json sidecar (compact, one entry per
document: shard file, offset, length) and a <source>_docs.jsonl (the same
data plus per-document metadata). Use these with
hobbylm.data.boundary_aware_data_generator, which packs multiple documents
into each training row and builds a block-diagonal attention mask + resets
RoPE position ids at every document boundary
(hobbylm.model.build_block_diagonal_bias / positions_from_segments) so no
document ever attends into another's tokens, regardless of how many share a
row. A loader that ignores the sidecars and reads the .bin files as one flat
stream will not respect document boundaries.
Sources and mix rationale
| source | tokens | target | median tok | p75 | p90 | max |
|---|---|---|---|---|---|---|
| pg19 | 600,181,264 | 600M | 76,762 | 129,000 | 198,130 | 1,837,542 |
| arxiv | 700,010,734 | 700M | 15,341 | 25,391 | 40,361 | 1,169,283 |
| wiki | 150,007,032 | 150M | 8,380 | 11,092 | 15,588 | 109,912 |
| edu_long | 250,022,528 | 250M | 6,102 | 8,993 | 13,660 | 156,268 |
| stack_python | 500,031,025 | 500M | 7,167 | 16,530 | 39,747 | 65,536 |
| replay | 299,999,998 | 300M | 698 | 1,030 | 1,712 | 355,795 |
| total | 2,500,252,581 | 2.5B |
- pg19: full public-domain books (emozilla/pg19,
a parquet mirror of
deepmind/pg19-- the original is a script-based dataset repo that hits a gzip-handling regression in currentdatasetsversions). The only source with documents comfortably longer than 8192 tokens; given the largest allocation for that reason. - arxiv:
togethercomputer/RedPajama-Data-1T, arxiv subset. - wiki:
wikimedia/wikipedia(20231101.en), filtered tomin_tokens=6000(raised from an initial 2000). Wikipedia articles are structurally short -- even the longest ones rarely approach 8192 tokens -- so this source is intentionally small and serves as a distribution anchor, not primary long-context material. - edu_long:
HuggingFaceFW/fineweb-edu(sample-10BT), length-filtered. - stack_python:
HuggingFaceCode/stack-v3-train, grouped to repository-level documents (files concatenated in priority order: setup/config files, then__init__.py, then other.py, then tests), Python-majority repos only. Capped at--max-doc-tokens 65536: an uncapped build showed the top 5 repositories (of ~15,670) accounting for 27.9% of all stack_python tokens, with the single largest single document (60.75M tokens) being ~96%ns-3pybindgen auto-generated Python-bindings boilerplate -- not organic code. After the cap, top-5 concentration is 0.07%. - replay: a sample of the original 100B-token pretraining mix
(harims95/hobbylm-mix100b-gpt2,
dclm/code/math/anneal families). The original per-document boundaries were
never recorded upstream --
prepare_mix100B.pyonly preserves them as an inline EOT token prefixed to each document in the flat stream. An earlier version of this corpus treated each 8M-token raw-shard slice as one "document" for indexing purposes; besides being far too coarse (median real document length is 698 tokens, not 8,000,000 -- see below), the metadata key used to describe that slice's position collided with and silently corrupted the shard-relative offset the loader actually needs, so 36 of 39 slices pointed at out-of-range byte ranges. Both are fixed:scripts/repair_replay_index.pyreconstructs real per-document boundaries from the inline EOT markers (scripts/build_long_context_corpus.py'ssplit_by_eot()), giving replay the same real-document-level indexing as every other source.
The mix was rebalanced from an initial plan (pg19 400M, arxiv 600M, wiki 350M, edu_long 350M) after the initial wiki build showed 53.8% padding waste at seq_len 8192 under one-document-per-row training. That waste isn't wiki-specific -- any source with a median length under 8192 has the same problem, arxiv included, just less severely (its median is 15-25K tokens, PG19 is the only source with a genuinely comfortable margin). Two changes followed: short sources were cut and long sources increased (see table above), and block-diagonal document packing was implemented in the training loader itself so padding waste is no longer coupled to per-source document length at all.
Padding waste: before and after packing
The _docs.jsonl per-document lengths above imply real waste under a naive
one-document-per-row loader (padding each doc out to a fixed row):
| source | pad waste @4096 (one-doc-per-row) | pad waste @8192 (one-doc-per-row) |
|---|---|---|
| pg19 | 2.05% | 4.01% |
| arxiv | 8.83% | 16.14% |
| wiki | 15.35% | 25.36% |
| edu_long | 23.10% | 29.89% |
| stack_python | 11.18% | 21.59% |
| replay | 75.22% | 87.17% |
replay's real per-document lengths (median 698 tokens) make it, once correctly indexed, the single worst-case source for one-doc-per-row waste in this corpus -- worse than wiki. It's also the clearest illustration of why the packing loader matters: with it, the same fine-grained document boundaries that make one-doc-per-row training wasteful for replay cost nothing.
With boundary_aware_data_generator's packing (documents concatenated
continuously into each row, one incomplete tail row per epoch per rank is the
only possible waste), projected waste is negligible regardless of source or
seq_len -- on the order of (seq_len - 1) / total_tokens:
| seq_len | combined-pool waste | worst single-source waste (wiki) |
|---|---|---|
| 4096 | 0.00016% | 0.0027% |
| 8192 | 0.00033% | 0.0055% |
Related
- Model: harims95/hobbylm-1b-hf
- Original pretraining mix: harims95/hobbylm-mix100b-gpt2
- Routing dynamics across pretraining: harims95/hobbylm-routing-dynamics
- Downloads last month
- 136