Dataset Viewer
Auto-converted to Parquet Duplicate
name
string
deterministic
bool
sources
dict
total_docs
int64
english_share
float64
doc_counts
dict
vocab_ladder
list
rungs
list
recipe
string
cwt-multilingual-pretrain-mix
true
{ "english": { "repo": "HuggingFaceFW/fineweb", "config": "sample-10BT", "revision": "9bb295ddab0e05d785b879661af7260fed5140fc" }, "non_english": { "repo": "HuggingFaceFW/fineweb-2", "languages": [ "arb_Arab", "ben_Beng", "ces_Latn", "cmn_Hani", "dan_Latn", ...
200,000
0.4
{ "eng(FineWeb)": 80000, "arb_Arab": 4285, "ben_Beng": 4285, "ces_Latn": 4285, "cmn_Hani": 4285, "dan_Latn": 4285, "deu_Latn": 4285, "ell_Grek": 4285, "fas_Arab": 4285, "fin_Latn": 4285, "fra_Latn": 4285, "heb_Hebr": 4285, "hin_Deva": 4285, "hun_Latn": 4285, "ind_Latn": 4285, "ita_Latn":...
[ 32000, 64000, 128000, 256000 ]
[ { "vocab_size": 32000, "occupancy": 0.997125, "dead_tokens": 92, "fertility": 2.536954538289989, "n_words": 7289673, "n_subwords": 18493569, "fertility_per_lang": { "eng": 1.5121450205415778, "arb_Arab": 2.1540646123189187, "ben_Beng": 4.688690933703495, "ces_Latn...
first-N docs of each pinned-revision stream; no sampling/RNG → byte-identical rebuild

CWT Multilingual Pretrain Mix (reproducible recipe)

A deterministic multilingual-including-English pretraining corpus for controlled vocabulary-scaling studies. The raw text is not stored here — it is reproduced byte-identically from manifest.json + the pinned dataset revisions, with no sampling and no RNG (the first-N documents of each stream).

Recipe

  • English: HuggingFaceFW/fineweb config sample-10BT, revision 9bb295ddab0e05d785b879661af7260fed5140fc
  • Non-English: HuggingFaceFW/fineweb-2, revision af9c13333eb981300149d5ca60a8e9d659b276b9
  • 28 non-English languages: arb_Arab, ben_Beng, ces_Latn, cmn_Hani, dan_Latn, deu_Latn, ell_Grek, fas_Arab, fin_Latn, fra_Latn, heb_Hebr, hin_Deva, hun_Latn, ind_Latn, ita_Latn, jpn_Jpan, kor_Hang, nld_Latn, pol_Latn, por_Latn, ron_Latn, rus_Cyrl, spa_Latn, swe_Latn, tha_Thai, tur_Latn, ukr_Cyrl, vie_Latn
  • English share 40%; total 200,000 docs (first-N per language; deterministic).

Reproduce byte-identically

pip install datasets
python rebuild_dataset.py --manifest manifest.json --out corpus.jsonl   # [--total_docs N]

Tokenizers trained on this mix: DaveGabe/cwt-multilingual-tokenizers.

Downloads last month
-