priyammaz/WrenPretraining100B

Wren Pretraining Mix (~116.14B tokens)

Pre-tokenized, ready-to-train pretraining data for the Wren model — so you can skip the prep pipeline entirely and start training. Tokens are packed as Megatron-style indexed datasets (.bin = concatenated uint16 token ids, .idx = per-document offset table), each document prefixed with the <|begin_of_text|> BOS token. Tokenized with the wren_tokenizer (retrained from Llama-3, vocab ≤ 65536 → uint16).

Layout

WrenPretraining100B/
  bins/
  supplemental_bins/
  lake/

• bins/ — merged <domain>.bin/.idx, one pair per domain. Point the trainer's --bins_dir here.
• supplemental_bins/ — small SFT/eval supplement bins blended in by weight.
• lake/ — the raw parquet each domain was tokenized from (58 sources). Re-tokenize with your own tokenizer, or just read the source text. Not needed to train on the bins.

Download

pip install huggingface_hub hf_xet
python -m data_prep.prepare_pretrain.hub.download_from_hub \
    --repo_id priyammaz/WrenPretraining100B --local_dir data/wren_100b
# ready-to-train subset only (skip the raw lake):
python -m data_prep.prepare_pretrain.hub.download_from_hub \
    --repo_id priyammaz/WrenPretraining100B --local_dir data/wren_100b --include bins supplemental_bins

Then train against data/wren_100b/bins (and supplemental_bins/).

Reading a bin directly

from data_prep.prepare_pretrain.binned.indexed_dataset import open_indexed
ds = open_indexed("data/wren_100b/bins/fineweb")
print(len(ds), "docs", ds.num_tokens, "tokens")
doc0 = ds.get(0)          # np.uint16 token ids, doc0[0] == 0 (BOS)

Token breakdown

bins/ — 115.91B tokens

domain	tokens
fineweb	41,094,868,387
dclm	27,487,572,843
finepdfs	13,592,939,038
stack_v2_python	4,262,022,220
stack_v2_cpp	3,504,505,050
stack_v2_javascript	2,265,939,959
pes2o	2,136,652,157
finemath_3plus	2,093,666,961
stack_v2_sql	1,651,763,660
stack_v2_java	1,409,478,803
infiniwebmath_3plus	1,261,070,334
stack_v2_c	1,231,624,964
stack_v2_markdown	1,069,447,401
stack_v2_shell	1,048,951,363
finemath_4plus	1,018,958,400
stack_edu_python	987,233,243
stack_v2_typescript	808,167,582
stack_v2_jupyter	768,729,124
stack_edu_cpp	634,537,107
infiniwebmath_4plus	519,143,510
stack_v2_cuda	470,029,442
megamath_web	448,067,474
stack_v2_go	423,449,389
stack_edu_javascript	421,020,588
rstarcoder	381,954,870
stack_edu_java	350,146,548
stack_v2_rust	342,904,612
wiki	294,928,803
stack_edu_sql	288,097,530
github_issues	261,603,925
cosmopedia	260,365,263
stack_v2_c_sharp	256,323,447
stack_edu_shell	244,952,078
stack_edu_markdown	244,424,878
stack_edu_c	223,963,563
stack_edu_typescript	184,374,020
openmathreasoning	163,604,847
openmathinstruct	158,968,750
openthoughts	154,809,570
opencodereasoning	153,656,886
smoltalk	144,372,190
tinygsm_mind	144,346,479
kaggle	130,824,761
twostudents	127,679,373
natural_reasoning	116,964,655
stack_v2_ruby	95,434,840
stack_edu_go	91,158,271
stack_v2_php	86,828,284
stack_edu_rust	80,580,518
stack_v2_swift	72,554,655
stackexchange	67,151,338
stack_edu_c_sharp	64,294,220
arxiv_summarization	52,340,955
stack_edu_ruby	23,416,943
stack_edu_swift	20,640,613
stack_edu_php	20,586,942

supplemental_bins/ — 230.11M tokens

domain	tokens
orca-math-word-problems-200k	67,296,975
Magicoder-Evol-Instruct-110K	62,026,245
Magicoder-OSS-Instruct-75K	39,928,716
long-context	33,255,736
WikiBirdsConversations	17,056,994
WikiBirds	10,540,596

Provenance & licensing

Derived from public datasets (FineWeb / FineWeb-Edu, DCLM, FinePDFs, FineMath, InfiniWebMath, MegaMath, Cosmopedia, The Stack, StackExchange, and others). Each retains its upstream license and terms — consult the original sources before redistribution or commercial use. Released here for research convenience.