Datasets:

salisai
/

slm-pretrain-corpus

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
text: string
source: string
category: string
id: string
token_est: int64
total_elapsed_seconds: double
peak_ram_mb: double
config: struct<max_ram_gb: double, min_free_disk_gb: double, scratch_cap_gb: double, batch_size: int64, zstd (... 787 chars omitted)
  child 0, max_ram_gb: double
  child 1, min_free_disk_gb: double
  child 2, scratch_cap_gb: double
  child 3, batch_size: int64
  child 4, zstd_level: int64
  child 5, shuffle_buckets: int64
  child 6, dedup_capacity: int64
  child 7, dedup_fp_rate: double
  child 8, min_text_len: int64
  child 9, edu_min_score: double
  child 10, mix: struct<general_web: struct<budget_bytes: int64, fineweb_edu: struct<share: double, dataset: string>, (... 508 chars omitted)
      child 0, general_web: struct<budget_bytes: int64, fineweb_edu: struct<share: double, dataset: string>, dclm_baseline: stru (... 35 chars omitted)
          child 0, budget_bytes: int64
          child 1, fineweb_edu: struct<share: double, dataset: string>
              child 0, share: double
              child 1, dataset: string
          child 2, dclm_baseline: struct<share: double, dataset: string>
              child 0, share: double
              child 1, dataset: string
      child 1, synthetic_textbook: struct<budget_bytes: int64, cosmopedia: struct<share: double, dataset: string>, llm_generated: struc (... 32 chars omitted)
          child 0, budget_bytes: int64
          child 1, cosmopedia: struct<share: double, dataset: string>
              child 0, sh
...
: int64
      child 4, tokens: int64
      child 5, path: string
      child 6, stopped: string
  child 6, instruction_ultrachat: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 34 chars omitted)
      child 0, category: string
      child 1, target_bytes: int64
      child 2, docs: int64
      child 3, bytes: int64
      child 4, tokens: int64
      child 5, path: string
      child 6, config: string
      child 7, stopped: string
  child 7, codeparrot_clean: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 18 chars omitted)
      child 0, category: string
      child 1, target_bytes: int64
      child 2, docs: int64
      child 3, bytes: int64
      child 4, tokens: int64
      child 5, path: string
      child 6, stopped: string
  child 8, synthetic_llm: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 43 chars omitted)
      child 0, category: string
      child 1, target_bytes: int64
      child 2, docs: int64
      child 3, bytes: int64
      child 4, tokens: int64
      child 5, path: string
      child 6, topics_generated: int64
      child 7, stopped: string
  child 9, final: struct<bytes: int64, docs: int64, shards_merged: int64, path: string, done: bool>
      child 0, bytes: int64
      child 1, docs: int64
      child 2, shards_merged: int64
      child 3, path: string
      child 4, done: bool
to
{'pipeline': Value('string'), 'created_at': Value('timestamp[s]'), 'total_elapsed_seconds': Value('float64'), 'peak_ram_mb': Value('float64'), 'final_corpus': {'bytes': Value('int64'), 'docs': Value('int64'), 'shards_merged': Value('int64'), 'path': Value('string'), 'done': Value('bool')}, 'categories': {'general_web': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'synthetic_textbook': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'math': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'instruction': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'code': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}}, 'per_source': {'skipped': {}, 'fineweb_edu': {'category': Value('string'), 'target_bytes': Value('int64'), 'docs': Value('int64'), 'bytes': Value('int64'), 'tokens': Value('int64'), 'path': Value('string'), 'config': Value('string'), 'stopped': Value('string')}, 'dclm_baseline': {'category': Value('string'), 'target_bytes': Value('int64'), 'docs': Value('int64'), 'bytes': Value('int64'), 'tokens': Value('int64'), 'path': Value('string'), 'shards_processed': Value('int64'), 'stopped': Value('string')}, 'cosmopedia': {'category': Value('string'), 'target_bytes': Val
...
 Value('int64'), 'path': Value('string'), 'done': Value('bool')}}, 'dedup': {'checked': Value('int64'), 'dupes_removed': Value('int64'), 'unique_added': Value('int64'), 'capacity': Value('int64'), 'fp_rate': Value('float64'), 'bloom_bits': Value('int64'), 'bloom_hashes': Value('int64')}, 'skipped_sources': {}, 'config': {'max_ram_gb': Value('float64'), 'min_free_disk_gb': Value('float64'), 'scratch_cap_gb': Value('float64'), 'batch_size': Value('int64'), 'zstd_level': Value('int64'), 'shuffle_buckets': Value('int64'), 'dedup_capacity': Value('int64'), 'dedup_fp_rate': Value('float64'), 'min_text_len': Value('int64'), 'edu_min_score': Value('float64'), 'mix': {'general_web': {'budget_bytes': Value('int64'), 'fineweb_edu': {'share': Value('float64'), 'dataset': Value('string')}, 'dclm_baseline': {'share': Value('float64'), 'dataset': Value('string')}}, 'synthetic_textbook': {'budget_bytes': Value('int64'), 'cosmopedia': {'share': Value('float64'), 'dataset': Value('string')}, 'llm_generated': {'share': Value('float64'), 'model': Value('string')}}, 'code': {'budget_bytes': Value('int64'), 'codeparrot_clean': {'dataset': Value('string')}}, 'math': {'budget_bytes': Value('int64'), 'open_web_math': {'dataset': Value('string')}}, 'instruction': {'budget_bytes': Value('int64'), 'openhermes': {'share': Value('float64'), 'dataset': Value('string')}, 'ultrachat': {'share': Value('float64'), 'dataset': Value('string')}}}, 'llm_budget_bytes': Value('int64'), 'llm_model': Value('string')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              text: string
              source: string
              category: string
              id: string
              token_est: int64
              total_elapsed_seconds: double
              peak_ram_mb: double
              config: struct<max_ram_gb: double, min_free_disk_gb: double, scratch_cap_gb: double, batch_size: int64, zstd (... 787 chars omitted)
                child 0, max_ram_gb: double
                child 1, min_free_disk_gb: double
                child 2, scratch_cap_gb: double
                child 3, batch_size: int64
                child 4, zstd_level: int64
                child 5, shuffle_buckets: int64
                child 6, dedup_capacity: int64
                child 7, dedup_fp_rate: double
                child 8, min_text_len: int64
                child 9, edu_min_score: double
                child 10, mix: struct<general_web: struct<budget_bytes: int64, fineweb_edu: struct<share: double, dataset: string>, (... 508 chars omitted)
                    child 0, general_web: struct<budget_bytes: int64, fineweb_edu: struct<share: double, dataset: string>, dclm_baseline: stru (... 35 chars omitted)
                        child 0, budget_bytes: int64
                        child 1, fineweb_edu: struct<share: double, dataset: string>
                            child 0, share: double
                            child 1, dataset: string
                        child 2, dclm_baseline: struct<share: double, dataset: string>
                            child 0, share: double
                            child 1, dataset: string
                    child 1, synthetic_textbook: struct<budget_bytes: int64, cosmopedia: struct<share: double, dataset: string>, llm_generated: struc (... 32 chars omitted)
                        child 0, budget_bytes: int64
                        child 1, cosmopedia: struct<share: double, dataset: string>
                            child 0, sh
              ...
              : int64
                    child 4, tokens: int64
                    child 5, path: string
                    child 6, stopped: string
                child 6, instruction_ultrachat: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 34 chars omitted)
                    child 0, category: string
                    child 1, target_bytes: int64
                    child 2, docs: int64
                    child 3, bytes: int64
                    child 4, tokens: int64
                    child 5, path: string
                    child 6, config: string
                    child 7, stopped: string
                child 7, codeparrot_clean: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 18 chars omitted)
                    child 0, category: string
                    child 1, target_bytes: int64
                    child 2, docs: int64
                    child 3, bytes: int64
                    child 4, tokens: int64
                    child 5, path: string
                    child 6, stopped: string
                child 8, synthetic_llm: struct<category: string, target_bytes: int64, docs: int64, bytes: int64, tokens: int64, path: string (... 43 chars omitted)
                    child 0, category: string
                    child 1, target_bytes: int64
                    child 2, docs: int64
                    child 3, bytes: int64
                    child 4, tokens: int64
                    child 5, path: string
                    child 6, topics_generated: int64
                    child 7, stopped: string
                child 9, final: struct<bytes: int64, docs: int64, shards_merged: int64, path: string, done: bool>
                    child 0, bytes: int64
                    child 1, docs: int64
                    child 2, shards_merged: int64
                    child 3, path: string
                    child 4, done: bool
              to
              {'pipeline': Value('string'), 'created_at': Value('timestamp[s]'), 'total_elapsed_seconds': Value('float64'), 'peak_ram_mb': Value('float64'), 'final_corpus': {'bytes': Value('int64'), 'docs': Value('int64'), 'shards_merged': Value('int64'), 'path': Value('string'), 'done': Value('bool')}, 'categories': {'general_web': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'synthetic_textbook': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'math': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'instruction': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}, 'code': {'bytes': Value('int64'), 'docs': Value('int64'), 'tokens': Value('int64'), 'sources': List(Value('string'))}}, 'per_source': {'skipped': {}, 'fineweb_edu': {'category': Value('string'), 'target_bytes': Value('int64'), 'docs': Value('int64'), 'bytes': Value('int64'), 'tokens': Value('int64'), 'path': Value('string'), 'config': Value('string'), 'stopped': Value('string')}, 'dclm_baseline': {'category': Value('string'), 'target_bytes': Value('int64'), 'docs': Value('int64'), 'bytes': Value('int64'), 'tokens': Value('int64'), 'path': Value('string'), 'shards_processed': Value('int64'), 'stopped': Value('string')}, 'cosmopedia': {'category': Value('string'), 'target_bytes': Val
              ...
               Value('int64'), 'path': Value('string'), 'done': Value('bool')}}, 'dedup': {'checked': Value('int64'), 'dupes_removed': Value('int64'), 'unique_added': Value('int64'), 'capacity': Value('int64'), 'fp_rate': Value('float64'), 'bloom_bits': Value('int64'), 'bloom_hashes': Value('int64')}, 'skipped_sources': {}, 'config': {'max_ram_gb': Value('float64'), 'min_free_disk_gb': Value('float64'), 'scratch_cap_gb': Value('float64'), 'batch_size': Value('int64'), 'zstd_level': Value('int64'), 'shuffle_buckets': Value('int64'), 'dedup_capacity': Value('int64'), 'dedup_fp_rate': Value('float64'), 'min_text_len': Value('int64'), 'edu_min_score': Value('float64'), 'mix': {'general_web': {'budget_bytes': Value('int64'), 'fineweb_edu': {'share': Value('float64'), 'dataset': Value('string')}, 'dclm_baseline': {'share': Value('float64'), 'dataset': Value('string')}}, 'synthetic_textbook': {'budget_bytes': Value('int64'), 'cosmopedia': {'share': Value('float64'), 'dataset': Value('string')}, 'llm_generated': {'share': Value('float64'), 'model': Value('string')}}, 'code': {'budget_bytes': Value('int64'), 'codeparrot_clean': {'dataset': Value('string')}}, 'math': {'budget_bytes': Value('int64'), 'open_web_math': {'dataset': Value('string')}}, 'instruction': {'budget_bytes': Value('int64'), 'openhermes': {'share': Value('float64'), 'dataset': Value('string')}, 'ultrachat': {'share': Value('float64'), 'dataset': Value('string')}}}, 'llm_budget_bytes': Value('int64'), 'llm_model': 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.

SLM Pretraining Corpus

A curated English pretraining corpus for small language models: 1,131,717 documents, ~1.28 billion tokens (estimated), 5.36 GB uncompressed text delivered as a 1.87 GB zstd-compressed JSONL file. Every record carries a text, source, category, id, and a token estimate, so you can filter, sample, or shard without extra tooling.

The corpus mixes general web, synthetic textbooks, code, math, and instruction data in a deliberate ratio, built streaming-first from public sources on an 8 GB RAM machine.

Quickstart

pip install datasets

from datasets import load_dataset

# Full corpus, streamed (no full download to disk)
ds = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/salisai/slm-pretrain-corpus/resolve/main/final_corpus.jsonl.zst",
    split="train",
    streaming=True,
)

for doc in ds:
    print(doc["category"], doc["source"], doc["token_est"])
    break

Small and fast for prototyping — a 100-document preview file:

sample = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/salisai/slm-pretrain-corpus/resolve/main/sample_100.jsonl.zst",
    split="train",
)

Filter by category before use:

code_docs = (doc for doc in ds if doc["category"] == "code")

Record format

One JSON object per line (zstd compressed), no nested structures:

Field	Type	Description
`text`	string	The document content
`source`	string	Upstream source name, e.g. `fineweb_edu`
`category`	string	One of `general_web`, `synthetic_textbook`, `code`, `math`, `instruction`
`id`	string	Stable document id (`<source>-<index>`)
`token_est`	int	Token estimate, `len(text) // 4` (no tokenizer needed)

Category	Documents	Tokens	Raw bytes	Share of docs	Share of tokens
general_web	560,238	687,862,243	2.85 GB	49.5%	53.5%
instruction	260,000	176,918,211	0.75 GB	23.0%	13.8%
synthetic_textbook	197,079	180,359,022	0.75 GB	17.4%	14.0%
math	64,000	120,535,235	0.50 GB	5.7%	9.4%
code	50,400	119,133,903	0.50 GB	4.5%	9.3%
Total	1,131,717	1,284,808,614	5.36 GB	100%	100%

Per-source breakdown

Source	Category	Documents	Tokens	Raw bytes
FineWeb-Edu (sample-10BT)	general_web	409,600	484,127,502	2.00 GB
DCLM-baseline-1.0	general_web	150,638	203,734,741	0.84 GB
Cosmopedia	synthetic_textbook	196,800	180,236,466	0.75 GB
DeepSeek-generated passages	synthetic_textbook	279	122,556	0.5 MB
OpenHermes-2.5	instruction	196,800	85,980,200	0.37 GB
UltraChat-200k	instruction	63,200	90,938,011	0.37 GB
OpenWebMath	math	64,000	120,535,235	0.50 GB
CodeParrot-clean	code	50,400	119,133,903	0.50 GB

Upstream sources

Source	Upstream dataset	Used for
FineWeb-Edu	HuggingFaceFW/fineweb-edu	general web (edu score >= 3.0)
DCLM-baseline	mlfoundations/dclm-baseline-1.0	general web
Cosmopedia	HuggingFaceTB/cosmopedia	synthetic textbooks (8 configs)
DeepSeek-generated	deepseek-chat API	synthetic passages
OpenHermes	teknium/OpenHermes-2.5	instruction / dialogue
UltraChat	HuggingFaceH4/ultrachat_200k	instruction / dialogue
OpenWebMath	open-web-math/open-web-math	math / reasoning
CodeParrot	codeparrot/codeparrot-clean	code

Curation details

The corpus was produced by a streaming pipeline (load_dataset(..., streaming=True)), so no source dataset was ever fully downloaded or held in memory:

FineWeb-Edu filtered with an edu score cutoff of 3.0 (scale 0-5).
Minimum document length of 200 characters (300 for instruction records).
Dedup: on-disk Bloom filter (8,000,000 capacity, 0.0005 false-positive rate) over SHA1 of normalized text prefixes, removing near-duplicate web boilerplate between FineWeb and DCLM.
Shuffling: bucket-based (256 zstd buckets, shuffled in RAM, concatenated).
Build statistics: 1,999 seconds elapsed, 979.6 MB peak RAM, zstd level 3, no source skipped.

The full machine-readable build report is in dataset_manifest.json.

Files

File	Size	Description
`final_corpus.jsonl.zst`	1.87 GB	The full corpus, shuffled, 1,131,717 records
`sample_100.jsonl.zst`	~140 KB	First 100 records for fast prototyping
`dataset_manifest.json`	6 KB	Sizes, per-category breakdown, dedup stats, config snapshot

Licensing

This corpus is a mix of independently licensed public datasets. Each document retains the terms of its upstream source — see the links in the table above for the license of each component (e.g. FineWeb-Edu and DCLM are ODC-By, Cosmopedia, CodeParrot and UltraChat are Apache-2.0/MIT-family, OpenHermes-2.5 is MIT). This repository does not impose an additional license on the mixed corpus; verify the source licenses against your intended use before redistributing derivative works.

Downloads last month: 26

Total file size:

1.87 GB