README.md

metadata

license: other
language:
  - en
task_categories:
  - text-generation
tags:
  - nanochat
  - nemotron
  - midtrain
  - math
  - reasoning
pretty_name: d24 Midtrain — OLMo-3 Dolmino (10B, whole-doc)
size_categories:
  - 10M<n<100M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet

d24 Midtrain — OLMo-3 Dolmino (10B, whole-doc)

A 9.3B-token reproduction of OLMo 3's Dolma-3 Dolmino mid-train mix, built by taking a fraction of each component of allenai/dolma3_dolmino_mix-100B-1025. No length filter and no chunking — every document is kept whole (some are very long: tens of thousands of tokens). Each component reaches its target, so the realized mix matches OLMo-3's true proportions (the OLMo-3 target % column ≈ the kept share). For training, the Megatron data loader concatenates documents and slices the stream into fixed-length (2048-token) sequences, so whole docs need no pre-chunking.

• 19,403,492 documents, ~9.3B GPT-2 tokens of plain text.
• Each row: component (str) and text (str). Documents are whole (unfiltered, unchunked).

Components (realized ≈ OLMo-3 target)

component	documents	~tokens	OLMo-3 target %
common_crawl-high-quality	1,534,917	~2402M	22.50%
dolmino-math	1,590,032	~1049M	10.70%
cranecode	655,845	~702M	10.00%
stack_edu-fim	873,675	~654M	10.00%
reddit_to_flashcards	8,756,357	~590M	5.90%
cranemath	737,929	~489M	5.63%
dolmino_1-flan	1,297,915	~520M	5.00%
nemotron-synth-qa	890,612	~502M	5.00%
olmocr_science_pdfs	61,883	~552M	5.00%
stem-heavy-crawl	571,094	~522M	5.00%
wiki_to_rcqa	1,371,600	~301M	3.00%
general_reasoning_mix	90,789	~155M	1.87%
qwq-reasoning-traces	15,893	~131M	1.87%
megamatt	322,394	~165M	1.73%
llama_nemotron-reasoning-traces	8,802	~101M	1.25%
openthoughts2-reasoning-traces	23,260	~94M	1.25%
tulu-3-sft	246,103	~116M	1.10%
tinymath-mind	183,429	~83M	0.90%
omr-rewrite-fullthoughts	29,341	~59M	0.85%
code-meta-reasoning	27,363	~51M	0.46%
math-meta-reasoning	16,289	~25M	0.38%
gemini-reasoning-traces	5,345	~15M	0.25%
tinymath-pot	82,526	~24M	0.24%
program_verifiable	10,099	~10M	0.16%

Load

from datasets import load_dataset
ds = load_dataset("sfanm/d24-midtrain-olmo3-10b-wholedoc", split="train")
ds = ds.filter(lambda r: r["component"] == "dolmino-math")

Related OLMo-3 midtrain corpora

• d24-midtrain-olmo3 — ~2.1B, long docs dropped.
• d24-midtrain-olmo3-5b — ~5B, long docs chunked to ≤2048-tok windows.
• this — ~10B, long docs kept whole (no filter/chunk).

License / attribution

Derived from allenai/dolma3_dolmino_mix-100B-1025 (itself a mix of many sources — see its card). Provided as-is for research; refer to the upstream licenses.