ounce100m-mix-v1

Tokenised, deduplicated English pre-training mix built for project ounce100m — a 90–110M-parameter decoder-only base model trained from scratch on ~1B tokens on 2x Tesla T4. This dataset is the exact input to that run, published so the result can be reproduced.

• Training tokens: 1,109,714,831 across 139 shards
• Held-out validation tokens: 22,934,043 in 12 shards (val/), sampled every 50th merged document — across sources, never whole documents, and excluded from the training shards
• Tokenizer: HuggingFaceTB/SmolLM2-135M:tokenizer.json — vocab 49,152, sha256 9ca9acddb6525a19…
• Built: 2026-09-20T00:30:56Z

Format

Raw binary, no wrapper library. Each shard:

header  : struct '<IIII'  = vocab, n_docs, n_tokens, flags      (16 bytes)
offsets : uint32 x (n_docs + 1)   cumulative token positions, [0] = 0
ids     : uint16 x n_tokens       SmolLM2 token ids

Document i of a shard is ids[offsets[i]:offsets[i+1]]. Concatenating the shards in filename order gives the training stream; a training position is exactly (shard_index, token_offset), which is what lets the run resume at the identical data position after an interruption.

import struct, array
with open(path, "rb") as f:
    vocab, n_docs, n_tokens, flags = struct.unpack("<IIII", f.read(16))
    offs = array.array("I"); offs.fromfile(f, n_docs + 1)
    ids  = array.array("H"); ids.fromfile(f, n_tokens)

Composition

source	tokens staged	docs seen	complete
SimpleStories__default	40,000,251	143,514	True
arxiv_abstracts__default	20,000,259	119,734	True
cosmopedia__auto_math_text	30,000,665	45,300	True
cosmopedia__khanacademy	10,000,346	11,385	True
cosmopedia__openstax	30,000,617	39,763	True
cosmopedia__stanford	60,000,589	67,575	True
cosmopedia__wikihow	30,000,947	34,668	True
finemath__finemath-4plus	160,000,457	111,587	True
finepdfs-edu__eng_Latn	180,001,431	52,277	True
finephrase__faq	30,000,605	91,273	True
finephrase__table	30,007,022	121,346	True
finephrase__tutorial	60,000,409	60,889	True
fineweb-edu__sample-10BT	340,001,020	2,456,651	True
open-web-math__default	70,000,824	38,793	True
wikipedia-monthly__20260101.en	60,000,399	10,448	True

Per-shard contents, source mix and sha256 digests are in manifest.json. Selection rules and the reasoning behind each source are in the project's docs/02-mix-plan.md; licence attribution and the transformations applied to each source are in ATTRIBUTION.md.

Provenance and hygiene

• Built by streaming public datasets on the Hugging Face Hub, quality/English-gated, deduplicated by exact document hash and by source url/id.
• No benchmark content. Overlap was computed mechanically against eval train/validation/dev material only; test splits were never opened during construction. Counts are in audit.json.
• No raw web crawl: web-derived sources here are classifier-scored or curated subsets.
• Trained model: see the ounce100m-* model repos in the same namespace.