ouro-stage4

Tokenized training data for mid-training ByteDance/Ouro-1.4B, built to the Stage 4 recipe of Ouro: Looped Language Models (arXiv:2510.25741, Section 4.2) at about 1/25 of the paper's scale. Everything here is token ids for the Ouro tokenizer (SmolLM family, vocabulary 49,152); no raw text is redistributed.

The paper's Stage 4 trains on 90B tokens of open-source SFT data beside 30B replayed from Stage 1 and 180B from Stage 2, 300B effective, at 32K sequence length. ouro-stage4-v2 is that mixture at 12.30B tokens of content, packed into 411,094 rows of 32,768: one epoch is 51,386 steps on eight ranks at one row each.

Layout

ouro-stage4-v2/          the packed pool a training config reads

tokens.npy is two-dimensional, one row per training sequence, each holding whole documents from every source shuffled together; segment_offsets.npy marks where each document begins in the flattened rows, so that no document crosses a row and a row's document boundaries can be handed to varlen attention (cu_seq_lens) with position_ids restarting at each; segment_prompts.npy says how many leading tokens of each document carry no label. A row its documents leave short ends in a padding segment of <|endoftext|> that labels none of itself. Padding is 8.7% of the rows. 15,948,918 documents sit in 16,356,834 segments.

Composition

Slot	Source	Origin	Tokens (M)	Budget (M)	Share of content (%)	Decontaminated docs
SFT	UltraData	openbmb/UltraData-SFT-2605, 5.7% subsample	3689.9	-	30.00	-
Stage 2	nemotron_cc_high	data.commoncrawl.org Nemotron-CC mirror, quality=high	4907.6	4907.6	39.90	38,467
Stage 2	nemotron_cc_math	nvidia/Nemotron-CC-Math-v1, 4plus	1107.0	1107.0	9.00	76,781
Stage 1	nemotron_cc	data.commoncrawl.org Nemotron-CC mirror, every quality bucket	902.8	902.8	7.34	4,649
Stage 2	nemotron_sft_general	nvidia/Nemotron-Pretraining-SFT-v1, Nemotron-SFT-General	457.6	457.6	3.72	65,431
Stage 2	megamath	LLM360/MegaMath, megamath-web-pro	339.5	339.5	2.76	29,830
Stage 2	nemotron_synthetic_code	nvidia/Nemotron-Pretraining-Code-v1, Synthetic-Code	280.4	280.4	2.28	15,301
Stage 2	nemotron_sft_code	nvidia/Nemotron-Pretraining-SFT-v1, Nemotron-SFT-Code	250.9	250.9	2.04	110,041
Stage 1	mapcc	m-a-p/MAP-CC, first part of each sub-corpus	159.9	159.9	1.30	334
Stage 1	opencoder_pretrain	OpenCoder-LLM/opc-fineweb-code-corpus	92.2	92.2	0.75	579
Stage 1	megamath_web	LLM360/MegaMath, megamath-web	50.4	50.4	0.41	1,324
Stage 2	opc_annealing	OpenCoder-LLM/opc-annealing-corpus	36.9	36.9	0.30	13,611
Stage 1	ultrafineweb_zh	openbmb/Ultra-FineWeb, data/ultrafineweb_zh	24.6	24.6	0.20	9

Shares of content are SFT 30.00%, Stage 1 replay 10.00%, Stage 2 replay 60.00%, the paper's 30 / 10 / 60. Every replay source is drawn to its budget.

How it was built

The SFT slot is a 5.7% subsample of openbmb/UltraData-SFT-2605 rendered through the Ouro tokenizer's ChatML template. A prompt ends at <|im_start|>assistant\n, or at <|im_start|>assistant\n<think>\n on a think row, so the opener that selects the mode is conditioned on rather than learned; the completion holds the reasoning inside </think> and then the answer, ending at <|im_end|>\n. Conversations longer than 32,768 tokens are dropped (46,358 of 862,714).

Replay text is raw, not ChatML. A source's budget is split over its shards and each shard is read from its head until it has given its quota. The two Common Crawl mirror slots draw 1,024 shards uniformly over the mirror's paths. Every document is dropped whose 13-token windows meet a test item of MMLU, BBH, GSM8K, HumanEval+ or MBPP+ (22,414 items, 1,772,051 distinct windows, matched by 64-bit hashes).

The pool sizes each replay source's budget against the SFT tokens (SFT is 30% of the mixture; a source's budget is its stage's share times its share of the stage, over 0.30), draws a seeded random subset of its documents up to that budget, splits replay documents longer than a row into row-sized pieces, shuffles every source's documents together, and packs them whole into rows.

OpenCoder-LLM/RefineCode, the corpus the paper's Stage 1 names as "OpenCoder-pretrain", is not published. opc-fineweb-code-corpus, the code-bearing web text of the same pipeline, stands in for it.

Licensing

Derived from corpora under their own terms, several restrictive: openbmb/UltraData-SFT-2605 is gated upstream, nvidia/Nemotron-CC-Math-v1 and the two nvidia/Nemotron-Pretraining-* corpora are under the NVIDIA Open Data License Agreement, and the Nemotron-CC mirror is subject to Common Crawl's terms. Check each upstream licence before using or redistributing this.