Hugging Face's logo

Datasets:
juliannunezb
/
crisp-v8

Tasks:
Text Generation
Languages:
English
Size:
1B<n<10B
Tags:
pretraining
llm
single-digit-tokenizer
nested-sampling
math
code
License:
Dataset card Files
xet
 Community

Access to crisp-v8 is granted manually

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

crisp-v8 is a derived pretraining corpus. About half of its tokens come from NVIDIA Nemotron-CC-v2.1, which is governed by the NVIDIA Data Agreement for Model Training and cannot be redistributed. The token shards are gated for that reason. The dataset card, the mixture specification and the tokenizer are open.

Log in or Sign Up to review the conditions and access this dataset content.

crisp-v8

An 8B-token, pre-tokenized English pretraining corpus with a single-digit tokenizer, in two nested-sampled phases (stable + anneal). Built to pretrain crisp3b-v8, a 2.67B model, on one GPU.

The shards are gated (see Licensing and access); the tokenizer and this card are open.

What is in it

Tokens 8.0B: final_stable/ 5.60B (6 shards) + final_anneal/ 2.40B (3 shards) + val.bin 85.5M held-out
Format uint16 token ids, documents separated by `<
Tokenizer tokenizer/crisp3b_v8.json — byte-level BPE, vocab 49,152 (48,896 merges + 16 named specials + 240 reserved), every digit is a separate token (12345 → 5 tokens), trained on a 14.6 GB sample of all seven sources. Ids ≥ 48896 never appear in the shards except the separator.
Files final_stable/train_0000..0005.bin, final_anneal/train_0000..0002.bin, val.bin, manifest.json, tokenizer/{crisp3b_v8.json, crisp3b_v8_meta.json, special_tokens_v8.py}

Both phases are nested-sampled: shards are round-robins of document-aligned blocks across sources, so any prefix (the first shard, the first 2B tokens, …) preserves the mixture.

Composition

Source HF dataset Subset Stable (5.6B) Anneal (2.4B)
Nemotron-CC-v2.1 — organic web nvidia/Nemotron-CC-v2.1 High-Quality/ 24% 13%
Nemotron-CC-v2.1 — synthetic rephrased web nvidia/Nemotron-CC-v2.1 High-Quality-Synthetic/ 20% 15%
Nemotron-CC-v2.1 — diverse QA nvidia/Nemotron-CC-v2.1 High-Quality-DQA/ — 25%
FineWeb-Edu HuggingFaceFW/fineweb-edu sample-100BT 18% —
FineMath HuggingFaceTB/finemath finemath-4plus 18% 25%
StarCoderData bigcode/starcoderdata python, cpp, c, java, javascript, go, rust, sql, markdown, shell 13% 15%
peS2o allenai/peS2o v2 7% 7%

Design: the stable phase is ≈42% real web + 20% synthetic web (≈one third synthetic within web, following the BeyondWeb/WRAP results) + 18% math + 13% code + 7% papers. The anneal phase shifts toward math, QA and code (≈40% synthetic, real web kept as an anchor) for the LR-decay stage of a WSD schedule. Nemotron-CC-v2.1 supplies ≈44% of the stable tokens and ≈53% of the anneal tokens.

Pipeline

  1. Download + clean per source (whitespace/boilerplate normalization; code files kept as-is).
  2. MinHash near-dedup across all sources — 316,043 documents removed.
  3. 13-gram decontamination against WikiText-103 (test), HellaSwag (val), GSM8K (test), ARC-Challenge (test) and OpenBookQA (test); overlapping documents dropped.
  4. Tokenize with the v8 tokenizer; any literal special-token string in the text is stripped so only BPE ids and the separator appear.
  5. Nested interleave into the two phases with disjoint document cursors (no document appears in both).
  6. Verify: every shard checked for out-of-vocab ids, separator density and mixture proportions; every shard decoded and spot-read.

Pipeline code: code/data_pipeline/ in the model repo (prepare_v8.py, train_tokenizer_v8.py, special_tokens_v8.py, run_v8_hetzner.sh). Built in ≈6 hours on a 16-core Hetzner CPU box.

Licensing and access

This is a derived dataset; each upstream source keeps its own license and you must comply with all of them:

  • Nemotron-CC-v2.1 (NVIDIA) — governed by the NVIDIA Data Agreement for Model Training, which permits training models on the data but not redistributing it. The token shards are gated for that reason. The card, the mixture specification and the tokenizer are open.
  • FineWeb-Edu — ODC-By 1.0 (with the Common Crawl terms of use).
  • FineMath — ODC-By 1.0.
  • StarCoderData — permissively licensed source code only (see the BigCode terms); attribution requirements of the original repositories apply.
  • peS2o — ODC-By 1.0.

The tokenizer, the mixture specification, the manifest and this card are released without restriction. If you are a source owner and want something changed, open a discussion in this repo.

Loading 

import numpy as np
from huggingface_hub import snapshot_download   # requires approved access + a token

path = snapshot_download("juliannunezb/crisp-v8", repo_type="dataset",
                         allow_patterns=["final_stable/train_0000.bin", "tokenizer/*"])
data = np.memmap(f"{path}/final_stable/train_0000.bin", dtype=np.uint16, mode="r")
# documents are separated by 48896; ids are < 49152

from tokenizers import Tokenizer
tok = Tokenizer.from_file(f"{path}/tokenizer/crisp3b_v8.json")
print(tok.decode([int(t) for t in data[:200] if t != 48896]))

Citation 

@misc{nunez_crisp_v8_2026,
  author = {Julián Núñez Barrero},
  title  = {crisp-v8: an 8B-token single-digit-tokenizer pretraining corpus in two nested-sampled phases},
  year   = {2026},
  url    = {https://huggingface.co/datasets/juliannunezb/crisp-v8}
}
Downloads last month
7
Total file size:
16.2 GB

Models trained or fine-tuned on juliannunezb/crisp-v8

Collection including juliannunezb/crisp-v8