Dataset Viewer
The dataset viewer is not available for this dataset.
Cannot get the config names for the dataset.
Error code:   ConfigNamesError
Exception:    ValueError
Message:      
Expected data_files in YAML to be either a string or a list of strings
or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'heldout', 'path': 'data/heldout-[0-9]*'}, {'split': 'heldout-cc-by-sa-4.0', 'path': 'data/heldout-cc-by-sa-4.0-*'}, {'split': 'train', 'path': 'data/train-[0-9]*'}, {'split': 'train-cc-by-sa-4.0', 'path': 'data/train-cc-by-sa-4.0-*'}]
Examples of data_files in YAML:

   data_files: data.csv

   data_files: data/*.png

   data_files:
    - part0/*
    - part1/*

   data_files:
    - split: train
      path: train/*
    - split: test
      path: test/*

   data_files:
    - split: train
      path:
      - train/part1/*
      - train/part2/*
    - split: test
      path: test/*

PS: some symbols like dashes '-' are not allowed in split names

Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1182, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1157, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 613, in get_module
                  metadata_configs = MetadataConfigs.from_dataset_card_data(dataset_card_data)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 153, in from_dataset_card_data
                  cls._raise_if_data_files_field_not_valid(metadata_config)
                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 100, in _raise_if_data_files_field_not_valid
                  raise ValueError(yaml_error_message)
              ValueError: 
              Expected data_files in YAML to be either a string or a list of strings
              or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'heldout', 'path': 'data/heldout-[0-9]*'}, {'split': 'heldout-cc-by-sa-4.0', 'path': 'data/heldout-cc-by-sa-4.0-*'}, {'split': 'train', 'path': 'data/train-[0-9]*'}, {'split': 'train-cc-by-sa-4.0', 'path': 'data/train-cc-by-sa-4.0-*'}]
              Examples of data_files in YAML:
              
                 data_files: data.csv
              
                 data_files: data/*.png
              
                 data_files:
                  - part0/*
                  - part1/*
              
                 data_files:
                  - split: train
                    path: train/*
                  - split: test
                    path: test/*
              
                 data_files:
                  - split: train
                    path:
                    - train/part1/*
                    - train/part2/*
                  - split: test
                    path: test/*
              
              PS: some symbols like dashes '-' are not allowed in split names

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.

KIP pretrain w48

Status: numbered release - all eight publish gates pass. Pure-knowledge pretraining corpus built by the KIP pipeline (std-lib only, $0 lanes): 14 domains x the language registry, quality-gated (C4/Gopher battery + script-aware LID + near-dup), deduplicated, license-routed against the release target, and exported with a per-record provenance spine (source, source_id, license, quality, formatting provenance).

Composition

domain records tokens_est
facts 4,006 80,971
encyclopedic 2,978 16,859,401
reference 114 37,520
civic 10 306,295
textbooks 4 142,857

Languages (certification)

Diacritic/NFC/numeral audit from formatting.certify_multilingual - the corpus keeps its linguistic detail (no ASCII flattening).

lang docs diacritic share NFC violations quotes
ar 626 0.008 0 cjk_brackets, curly, guillemet, straight
cs 120 0.123 0 cjk_brackets, curly, guillemet, straight
de 643 0.015 0 cjk_brackets, curly, guillemet, straight
el 134 0.000 0 cjk_brackets, curly, guillemet, straight
en 851 0.001 0 cjk_brackets, curly, guillemet, straight
es 731 0.023 0 cjk_brackets, curly, guillemet, straight
fa 15 0.003 0 cjk_brackets, curly, guillemet, straight
fr 806 0.037 0 cjk_brackets, curly, guillemet, straight
he 145 0.002 0 cjk_brackets, curly, guillemet, straight
hi 157 0.559 0 straight
id 147 0.000 0 cjk_brackets, curly, guillemet, straight
it 141 0.006 0 cjk_brackets, curly, guillemet, straight
ja 107 0.000 0 cjk_brackets, curly, guillemet, straight
ko 95 0.000 0 cjk_brackets, curly, straight
my 1 1.151 0 -
ne 8 0.412 0 cjk_brackets, curly, straight
nl 1 0.001 0 straight
pl 140 0.059 0 cjk_brackets, curly, guillemet, straight
pt 143 0.036 0 cjk_brackets, curly, guillemet, straight
ro 140 0.057 0 cjk_brackets, curly, guillemet, straight
ru 698 0.000 0 cjk_brackets, curly, guillemet, straight
si 13 0.451 0 cjk_brackets, curly, straight
sv 339 0.043 0 cjk_brackets, curly, straight
th 67 0.241 0 cjk_brackets, curly, straight
tr 141 0.107 0 cjk_brackets, curly, guillemet, straight
uk 133 0.000 0 cjk_brackets, curly, guillemet, straight
vi 138 0.132 0 cjk_brackets, curly, guillemet, straight
zh 432 0.000 0 cjk_brackets, curly, straight

Licenses (routing)

Release target apache-2.0 at policy attribution-ok: the primary splits (train/heldout) hold 4,027 records (56.6%) whose licenses are covered by that declaration; everything else is routed losslessly into per-license splits. Covering solver: cc0-1.0 + NOTICE.md ledger (NOTICE.md ships in this repo)

Declaration candidates - which licenses CAN be applied to this build:

  • apache-2.0 (+NOTICE.md ledger): viable - covers every primary family; attribution rows ride via the NOTICE.md ledger
  • mit: NOT viable - attribution-family rows are present and MIT has no NOTICE mechanism - declare apache-2.0 (+NOTICE.md) or route them out The final MIT-vs-Apache choice is a one-flag decision (--license-target mit|apache-2.0); the recommended pick is the most-permissive covering license (a MIT+Apache mix declares apache-2.0).
raw license normalized family records goes to
CC0-1.0 (4,006) cc0-1.0 public-domain 4,006 primary (train/heldout)
CC BY-SA 4.0 (2,978); CC BY-SA 4.0 / GFDL (dual) (107) cc-by-sa-4.0 share-alike 3,085 route split 3085
EU reuse with attribution (10) eu-reuse attribution 10 primary (train/heldout)
IETF Trust Legal Provisions (attribution) (7) ietf-trust attribution 7 primary (train/heldout)
CC BY 4.0 (OpenStax origin, per-title verify) (4) cc-by-4.0 attribution 4 primary (train/heldout)

Mixture fit (observed vs target)

fit score 0.357 (1 = perfect; total-variation distance per axis against the curriculum targets). The gap table below is the fetcher's shopping list for the next build.

  • domain tvd 0.812
  • language tvd 0.390
  • programming_language tvd 0.727
  • record_type (no target): raw_filtered: 7,112
axis under-supplied bucket deficit
domain web 30.0%
domain code 15.0%
programming_language java 10.0%
language nl 9.6%
programming_language cpp 9.0%
domain books 8.0%
programming_language javascript 7.7%
language it 7.3%

Release gates

  • attribution - 7112/7112 records carry source+source_id+license
  • min_corpus - 7,112 exported (gate 2,000 for a numbered release)
  • contamination - 0 eval-bank overlaps (12-gram)
  • quality - 0 records with gate violations
  • license_allowlist - licenses: ['cc-by-4.0', 'cc-by-sa-4.0', 'cc0-1.0', 'eu-reuse', 'ietf-trust']
  • license_homogeneity - target apache-2.0 / policy attribution-ok: primary 4,027 (56.6%), routed 3,085 [cc-by-sa-4.0:3085], covering=cc0-1.0 +NOTICE
  • multilingual_certification - 28 languages certified; 0 NFC violations
  • dataset_card - README.md in this repo (frontmatter + reviewer body)

Splits

  • format: parquet (nested spine fields ride along as lossless JSON columns)
  • heldout: 118 records (1 shards) - primary (release target)
  • heldout-cc-by-sa-4.0: 93 records (1 shards) - license route bucket
  • train: 3,909 records (1 shards) - primary (release target)
  • train-cc-by-sa-4.0: 2,992 records (1 shards) - license route bucket
  • heldout assignment: 3% deterministic sha1(id) split, applied before license routing
  • eval bank: excluded at export by a 12-gram contamination gate

Method

  1. fetch per-domain free sources (catalog: $0, license story per source)
  2. extract/format per domain rulebook (formatting.py: keep structure, strip boilerplate, raw code, verbatim math)
  3. script-aware normalization (NFC, ZWJ/ZWNJ policies)
  4. quality gates per domain profile + 13-gram near-dup (exact substring level)
  5. MinHash-LSH fuzzy dedup (dedup.py: 128 perms, 32x4 banding, exact-Jaccard 0.8 verification, union-find clusters, verifier-aware survivor policy - the FineWeb/Nemotron-CC highest-value stage, script-aware for CJK)
  6. linear verifier pass (verifier.py: DCLM-style hashed n-gram logistic ensemble; per-record clf_prob/tier in the spine; calibrated retention gate; 20-bucket quality tiers drive the mixture boosts; the SDG lane reads tier>=18 only)
  7. mixture allocation (composition.py water-filling over the language x domain grid; mixtures.py IPF + multi-axis fit against the curriculum targets)
  8. license routing (licensing.py: normalize -> family -> route bucket; NOTICE.md attribution ledger)
  9. export with provenance spine + contamination gate + deterministic split

Limitations

  • the primary splits are license-homogeneous under the release target; route splits carry their own licenses and users must respect each split's terms (per-record license column + NOTICE.md)
  • tokens_est is a character-based estimate, not a tokenizer count
  • capsule/doc_class=records derived from teacher paraphrase carry license 'derived (teacher-paraphrased)' and route to their own split (midtrain lane by default)

Reproduce

PYTHONPATH=src python3 demo/run_pretrain_pack.py --preset pretrain_micro --raw-dir data/pretrain/raw \
    --capsules data/fullscale/out/<WEEK>/views_train.jsonl \
    --out data/pretrain/out/shard.jsonl
PYTHONPATH=src python3 -m kip.fullscale.export_hf --in data/pretrain/out/shard.jsonl --out data/hf/kip-beta --repo IO-AIR/kip-beta --version beta --repo-kind beta

built 2026-11-19T23:04:19Z - KIP v3.0 (repo: IO-AIR/kip-dataset)

Downloads last month
28