domyn/Domyn-Edge-Pretraining-Mix-Stages

Domyn Edge Pretraining Mix Stages

This repository contains the data mixes of pre-training stages 2 to 6 of Domyn Edge Base. The stage 1 mix is domyn/Domyn-Edge-Pretraining-Mix-8.4T.

After the stage 1 mix, Domyn Edge Base was trained in five more stages: specialized pre-training, two context extensions, mid-training and annealing. Together they account for 4.30 trillion tokens. All counts refer to the tokens seen during training, measured with the Domyn Edge tokenizer.

This release contains the text of 4.02 trillion of those tokens. For the remaining documents, it contains upstream record identifiers instead of text. These documents belong to two groups: three upstream datasets whose terms do not allow redistribution, and the stack-edu code, including its fill-in-the-middle version and our version with reasoning traces. You can use the identifiers to rebuild the documents from the original datasets, as described in Documents published as identifiers.

Stages

Each stage is a config. Stages 3, 5 and 6 were trained on a sequence of mixes, and each mix is a subfolder of its stage. Two parts of each stage are in separate folders that are not part of any config. The Wikipedia text is in data/<config>-wiki/, because it has three additional columns (see Wikipedia attribution). The record identifiers of the documents published without text are in data/<config>-ids/ (see Documents published as identifiers). In the tables below, the tokens in this release include the Wikipedia text and exclude the documents published as identifiers.

Config	Mixes	Tokens	In this release
stage2-specialized	specialized-pretraining	1.20T	1.05T (87.56%)
stage3-ctx-ext	long-context-32k-1, mix-64k-100b-pre-1-mid-1-lc-05, mix-128k-100b-pre-1-mid-1-lc-05	298.9B	274.0B (91.68%)
stage4-midtraining	mix-2t-32k-mid-include-gold-node	2.00T	1.95T (97.33%)
stage5-ctx-ext	ctx-extension-64k, ctx-extension-128k-math25-code15, ctx-extension-256k-math25-code15, ctx-extension-512k-math25-code15, ctx-extension-1m-math25-code15	504.6B	489.2B (96.95%)
stage6-annealing	midtrain-annealing-64k-p6, midtrain-annealing-512k-60b	298.4B	260.5B (87.30%)

Mix	Tokens	In this release	Share released
specialized-pretraining	1.20T	1.05T	87.56%
long-context-32k-1	99.6B	91.0B	91.45%
mix-64k-100b-pre-1-mid-1-lc-05	99.7B	91.9B	92.21%
mix-128k-100b-pre-1-mid-1-lc-05	99.7B	91.1B	91.39%
mix-2t-32k-mid-include-gold-node	2.00T	1.95T	97.33%
ctx-extension-64k	240.0B	236.9B	98.73%
ctx-extension-128k-math25-code15	190.1B	181.3B	95.40%
ctx-extension-256k-math25-code15	50.0B	47.6B	95.22%
ctx-extension-512k-math25-code15	14.9B	14.1B	95.04%
ctx-extension-1m-math25-code15	9.7B	9.3B	95.24%
midtrain-annealing-64k-p6	239.0B	201.9B	84.47%
midtrain-annealing-512k-60b	59.4B	58.6B	98.68%

The model checkpoint at the end of each stage is a branch of domyn/Domyn-Edge-Base-checkpoints. The branch name gives the total number of tokens seen up to that checkpoint, stage 1 included.

Config	Checkpoint
stage2-specialized	stage2-specialized-10.1T
stage3-ctx-ext	stage3-ctx-ext-10.3T
stage4-midtraining	stage4-midtraining-12.3T
stage5-ctx-ext	stage5-ctx-ext-12.7T
stage6-annealing	stage6-annealing-13T

Composition

The tables list the sources in this release and their share of the tokens of each stage. The shares of a stage add up to its share released. The remaining documents are published as identifiers (see Documents published as identifiers).

Stage 2: specialized pre-training

Source	Tokens	Share of stage
Dolma 3 mix (6T)	541.7B	45.16%
Dolma 3 Dolmino pool	165.2B	13.77%
FinePhrase	122.2B	10.19%
FineWeb2-HQ	103.2B	8.60%
TxT360-Midas	38.9B	3.24%
FinePDFs	25.5B	2.12%
Nemotron Pretraining Specialized v1 and v1.1	24.1B	2.01%
Common Crawl, Domyn pipeline	10.2B	0.85%
SYNTH	5.1B	0.43%
Swallow Code v2	4.3B	0.36%
daVinci-Dev	3.5B	0.29%
HPLT 3.0	2.2B	0.18%
UltraData-Math	1.5B	0.12%
Issues and Kaggle notebooks	0.8B	0.07%
FineWiki	0.7B	0.06%
OPC annealing corpus	0.6B	0.05%
SmolTalk2	0.5B	0.04%

Stage 3: context extension

Source	Tokens	Share of stage
Dolma 3 mix (6T)	84.6B	28.30%
Dolma 3 Dolmino pool	64.6B	21.61%
TxT360-Midas	46.3B	15.49%
Dolma 3 Longmino pool	23.9B	8.01%
FinePhrase	17.9B	6.00%
FineWeb2-HQ	16.0B	5.35%
FinePDFs	10.5B	3.50%
Nemotron Pretraining Specialized v1 and v1.1	4.5B	1.51%
Common Crawl, Domyn pipeline	1.6B	0.53%
daVinci-Dev	0.9B	0.32%
SYNTH	0.8B	0.27%
UltraData-Math	0.8B	0.25%
Swallow Code v2	0.7B	0.23%
FineWiki	0.5B	0.17%
SmolTalk2	0.2B	0.08%
OPC annealing corpus	0.1B	0.03%
Issues and Kaggle notebooks	<0.1B	0.01%

Stage 4: mid-training

Source	Tokens	Share of stage
Dolma 3 Dolmino pool	1.13T	56.49%
TxT360-Midas	472.6B	23.63%
FinePDFs	196.7B	9.84%
UltraData-Math	47.0B	2.35%
MegaMath-Web-Pro-Max	35.1B	1.76%
FineWiki	31.9B	1.60%
SmolTalk2	13.6B	0.68%
CoderForge-Preview	7.6B	0.38%
Nemotron Terminal Corpus	5.3B	0.26%
Nemotron SFT SWE v3	4.1B	0.20%
SWE-rebench OpenHands trajectories	1.4B	0.07%
Scale-SWE-Distilled	1.3B	0.07%
SWE-chat	<0.1B	<0.01%

Stage 5: context extension

Source	Tokens	Share of stage
TxT360-Midas	142.6B	28.26%
FinePDFs	134.4B	26.64%
Dolma 3 Longmino pool	102.0B	20.22%
Dolma 3 Dolmino pool	88.0B	17.45%
CoderForge-Preview	8.1B	1.60%
Nemotron SFT SWE v3	4.7B	0.93%
Nemotron Pretraining Specialized v1.2	1.9B	0.38%
SWE-rebench OpenHands trajectories	1.8B	0.36%
Nemotron Terminal Corpus	1.6B	0.32%
Scale-SWE-Distilled	1.5B	0.30%
FineWiki	0.9B	0.19%
UltraData-Math	0.9B	0.18%
SmolTalk2	0.3B	0.06%
ToolMind-Web-QA	0.1B	0.03%
MegaMath-Web-Pro-Max	0.1B	0.02%
SWE-chat	0.1B	0.02%

Stage 6: annealing

Source	Tokens	Share of stage
TxT360-Midas	67.5B	22.61%
Dolma 3 Dolmino pool	47.0B	15.75%
daVinci-Dev	40.5B	13.57%
FinePDFs	32.4B	10.84%
CoderForge-Preview	17.4B	5.85%
Nemotron SFT SWE v3	10.3B	3.47%
SmolTalk2	8.2B	2.76%
Domyn agentic coding trajectories	8.2B	2.73%
Nemotron Terminal Corpus	7.0B	2.36%
Dolma 3 Longmino pool	6.0B	2.00%
SWE-rebench OpenHands trajectories	3.5B	1.18%
Scale-SWE-Distilled	3.3B	1.12%
Domyn deep-research trajectories	3.3B	1.10%
UltraData-Math	2.9B	0.97%
Domyn tool-use trajectories	1.0B	0.34%
Nemotron Pretraining Specialized v1.2	0.5B	0.16%
FineWiki	0.4B	0.15%
Hermes agent reasoning traces	0.4B	0.13%
ToolMind-Web-QA	0.4B	0.12%
SWE-chat	0.3B	0.09%

Statistics

	Trained on	In this release
Tokens	4,301,131,591,907	4,020,332,602,246
Documents	3,296,099,695	2,958,468,188

How to use

Each stage is available as a named config:

from datasets import load_dataset

annealing = load_dataset(
    "domyn/Domyn-Edge-Pretraining-Mix-Stages",
    "stage6-annealing",
    split="train",
    streaming=True,
)

Data is stored as Parquet files of about 1 GB each, compressed with Snappy. The files are under data/<config>/<mix>/<group>/, where <group> is a group of sources within the mix. To load a single mix or group, pass its folder to data_files:

mix = load_dataset(
    "domyn/Domyn-Edge-Pretraining-Mix-Stages",
    data_files="data/stage5-ctx-ext/ctx-extension-128k-math25-code15/**/*.parquet",
    split="train",
    streaming=True,
)

The Wikipedia text of each stage is in the folder data/<config>-wiki/ and is loaded in the same way:

wiki = load_dataset(
    "domyn/Domyn-Edge-Pretraining-Mix-Stages",
    data_files="data/stage6-annealing-wiki/**/*.parquet",
    split="train",
    streaming=True,
)

Columns

Column	Type	Description
text	string	The document.
token_length	int64	Length of the document in Domyn Edge tokens.
src_path	string	The upstream corpus file the document came from, so any document can be traced back to its source.

The Wikipedia files have three additional columns, described in Wikipedia attribution.

Wikipedia attribution

Wikipedia text is licensed under CC BY-SA 4.0, which requires attribution to the contributors of each article. The files in data/<config>-wiki/ include three additional columns that identify the source of each document:

Column	Type	Description
source_id	string	The upstream record: enwiki/<page id> for English Wikipedia, <wiki>/<page id> for FineWiki (for example ptwiki/25726), the NVIDIA record uuid for Wiki-Rewrite, or the Dolma 3 Dolmino pool record id for wiki_to_rcqa.
source_url	string	The article URL. Null for Wiki-Rewrite and wiki_to_rcqa.
revision_id	string	The article revision the text was taken from. Null for Wiki-Rewrite and wiki_to_rcqa.

To view a revision, open https://<language>.wikipedia.org/w/index.php?oldid=<revision_id>. The contributors are listed in the article history. Wiki-Rewrite rows are attributed to the NVIDIA record they come from, because NVIDIA does not link its records to the original articles. In the same way, wiki_to_rcqa rows are attributed to the Dolma 3 Dolmino pool record they come from. For 2,496 rows, all three columns are null, because the same text appears in more than one upstream record: 1,837 wiki_to_rcqa rows, 638 FineWiki rows and 21 Wiki-Rewrite rows. Refer to ATTRIBUTION.md for details.

Sources

Source	Link	Notes
Dolma 3 mix (6T)	allenai/dolma3_mix-6T	Replay of the stage 1 web crawl, science PDFs, research papers and Dolma 1.7 Wikipedia text. The Wikipedia text is in the data/<config>-wiki/ folders. The stack-edu subsets are published as identifiers.
Dolma 3 Dolmino pool	allenai/dolma3_dolmino_pool	The wiki_to_rcqa subsets are in the data/<config>-wiki/ folders. The stack-edu fill-in-the-middle subsets are published as identifiers.
Dolma 3 Longmino pool	allenai/dolma3_longmino_pool
Common Crawl, Domyn pipeline	commoncrawl.org	Our FineWeb-style pipeline, run on CC-MAIN-2025-47.
FineWeb2-HQ	epfml/FineWeb2-HQ	High-quality multilingual subsets.
HPLT 3.0	HPLT/HPLT3.0	Lower-quality multilingual subsets.
FinePDFs	HuggingFaceFW/finepdfs	English and the 12 other languages of the mix.
FineWiki	HuggingFaceFW/finewiki	English and the 12 other languages of the mix, in the data/<config>-wiki/ folders.
FinePhrase	HuggingFaceFW/finephrase
SYNTH	PleIAs/SYNTH
TxT360-Midas	IFM/TxT360-Midas
MegaMath-Web-Pro-Max	OctoThinker/MegaMath-Web-Pro-Max
UltraData-Math	openbmb/UltraData-Math	The L3 synthetic subsets.
Nemotron Pretraining Specialized v1	nvidia/Nemotron-Pretraining-Specialized-v1	The Wiki-Rewrite subset is in the data/<config>-wiki/ folders.
Nemotron Pretraining Specialized v1.1	nvidia/Nemotron-Pretraining-Specialized-v1.1
Nemotron Pretraining Specialized v1.2	nvidia/Nemotron-Pretraining-Specialized-v1.2
SmolTalk2	HuggingFaceTB/smoltalk2	The Llama_Nemotron_Post_Training_Dataset_reasoning_r1 and OpenThoughts3_1.2M subsets.
OPC annealing corpus	OpenCoder-LLM/opc-annealing-corpus
Swallow Code v2	tokyotech-llm/swallow-code-v2
Issues and Kaggle notebooks	HuggingFaceTB/issues-kaggle-notebooks
daVinci-Dev	GAIR/daVinci-Dev
CoderForge-Preview	togethercomputer/CoderForge-Preview
Nemotron Terminal Corpus	nvidia/Nemotron-Terminal-Corpus
Nemotron SFT SWE v3	nvidia/Nemotron-SFT-SWE-v3
SWE-rebench OpenHands trajectories	nebius/SWE-rebench-openhands-trajectories
Scale-SWE-Distilled	AweAI-Team/Scale-SWE-Distilled
SWE-chat	SALT-NLP/SWE-chat
ToolMind-Web-QA	Nanbeige/ToolMind-Web-QA
Hermes agent reasoning traces	lambda/hermes-agent-reasoning-traces
Domyn trajectories	this repository	Agentic coding, deep-research and tool-use trajectories generated by us. See Data generated by Domyn.

Documents published as identifiers

For two groups of documents, the folder data/<config>-ids/ of each stage contains the identifier of the upstream record instead of the text. You can use these identifiers to rebuild the documents from the original datasets.

Omitted datasets

Three upstream datasets in the training mixes are distributed under terms that do not allow redistribution, so this release does not contain their text. After you accept the upstream terms, you can rebuild their documents.

Dataset	Revision	Subsets used
AutoMathText-V2	2a8d19c8	reasoning_qa, all ten score buckets from 0-10 to 90-100
Nemotron CC Math v1	397a2502	The 4plus split, with reasoning traces we added
Nemotron Pretraining Code v2	7b1a453d	The synthetic code review, question answering, rewriting, student-teacher and transpilation subsets

Their share of the tokens of each stage is the following:

Config	AutoMathText-V2	Nemotron CC Math v1	Nemotron Pretraining Code v2	Total
stage2-specialized	3.8B (0.32%)	10.2B (0.85%)	42.2B (3.52%)	4.69%
stage3-ctx-ext	2.0B (0.65%)	1.4B (0.46%)	7.2B (2.39%)	3.51%
stage4-midtraining	9.7B (0.49%)	none	none	0.49%
stage5-ctx-ext	8.3B (1.64%)	none	none	1.64%
stage6-annealing	24.4B (8.18%)	none	none	8.18%

Mix	Omitted share
specialized-pretraining	4.69%
long-context-32k-1	3.70%
mix-64k-100b-pre-1-mid-1-lc-05	3.07%
mix-128k-100b-pre-1-mid-1-lc-05	3.76%
mix-2t-32k-mid-include-gold-node	0.49%
ctx-extension-64k	0.64%
ctx-extension-128k-math25-code15	2.47%
ctx-extension-256k-math25-code15	2.74%
ctx-extension-512k-math25-code15	2.92%
ctx-extension-1m-math25-code15	2.31%
midtrain-annealing-64k-p6	10.19%
midtrain-annealing-512k-60b	0.10%

Stack-edu

The stack-edu files of stages 2 and 3 come from the stack-edu subsets of the Dolma 3 mix. They are included in two forms: as published, and in our version that adds a reasoning trace to Python, Java and C++ files. In our version, files longer than 16k tokens have no trace. The stack-edu files of stages 4 to 6 come from the stack-edu fill-in-the-middle subsets of the Dolma 3 Dolmino pool. The Dolmino records contain the files as trained, including the fill-in-the-middle tokens. The records of both datasets include the repository, path and license of each file, which are required to attribute it.

Dataset	Revision	Subsets used
Dolma 3 mix (6T)	689a3ea2	data/stack_edu-<language> for C, C++, C#, Go, Java, JavaScript, Markdown, PHP, Python, Ruby, Rust, Shell, SQL, Swift and TypeScript
Dolma 3 Dolmino pool	091589c5	data/stack_edu_fim-<language>_<number> for the same 15 languages

Config	Stack-edu	With reasoning traces	Fill-in-the-middle	Total
stage2-specialized	51.7B (4.31%)	41.3B (3.44%)	none	7.76%
stage3-ctx-ext	8.0B (2.67%)	6.4B (2.14%)	none	4.81%
stage4-midtraining	none	none	43.7B (2.19%)	2.19%
stage5-ctx-ext	none	none	7.1B (1.41%)	1.41%
stage6-annealing	none	none	13.5B (4.52%)	4.52%

Files and columns

The files are Parquet, under data/<config>-ids/<mix>/<group>/<dataset>/, for example data/stage2-specialized-ids/specialized-pretraining/pretrain_synthetic/Nemotron-CC-Math-v1/. To read them, pass their folder to data_files:

ids = load_dataset(
    "domyn/Domyn-Edge-Pretraining-Mix-Stages",
    data_files="data/stage2-specialized-ids/**/*.parquet",
    split="train",
)

The files have the following columns:

Column	Type	Description
source_dataset	string	The upstream dataset, for example nvidia/Nemotron-CC-Math-v1.
source_subset	string	The upstream folder that contains the record: reasoning_qa/<score bucket>, 4plus, Synthetic-Code/synthetic-<subset>, data/stack_edu-<language>, or data/stack_edu_fim-<language>_<number>.
source_id	string	The upstream record: its id for AutoMathText-V2 and Nemotron CC Math v1, its nemotron_pretraining_id for Nemotron Pretraining Code v2, and the Software Heritage blob id of the file for stack-edu.
reasoning_trace	string	Our reasoning trace, for Nemotron CC Math v1 documents and for the stack-edu files of our version with reasoning traces, up to 16k tokens. Null otherwise.
token_length	int64	Length of the document as trained, in Domyn Edge tokens.
src_path	string	The upstream corpus file the document came from, as in the other configs.

Each row is one document of the mix, so a record that was repeated in a mix appears once for each repetition. Documents with no text are not listed.

Rebuilding the documents

To rebuild the documents of a stage:

1. Accept the terms of each upstream dataset on its Hugging Face page. Nemotron Pretraining Code v2 requires approval by NVIDIA.
2. Download the folders named in source_subset, at the revisions listed in the tables above.
3. Keep one upstream record per identifier. The Dolma 3 mix repeats most stack-edu files up to five times, with the same text. In the Dolma 3 mix, the blob id is the blob_id field of the Python files and the id field of the files of the other languages. In the Dolma 3 Dolmino pool, it is the id field.
4. Join the identifier files with the upstream records on source_id. The upstream text is in the content column for Nemotron Pretraining Code v2 and in the text column for the other datasets.
5. Where reasoning_trace is not null, add it after the upstream text: text + "\n<think>" + reasoning_trace + "</think>".

The following example rebuilds the Nemotron CC Math v1 documents of stage 2 with polars:

import polars as pl
from huggingface_hub import snapshot_download

ids_dir = snapshot_download(
    "domyn/Domyn-Edge-Pretraining-Mix-Stages",
    repo_type="dataset",
    allow_patterns="data/stage2-specialized-ids/**",
)
math_dir = snapshot_download(
    "nvidia/Nemotron-CC-Math-v1",
    repo_type="dataset",
    revision="397a2502f2028c659ba411a6c4935b464a7f03aa",
    allow_patterns="4plus/*",
)

ids = pl.scan_parquet(f"{ids_dir}/data/stage2-specialized-ids/**/*.parquet").filter(
    pl.col("source_dataset") == "nvidia/Nemotron-CC-Math-v1"
)
upstream = (
    pl.scan_parquet(f"{math_dir}/4plus/*.parquet")
    .select(pl.col("id").alias("source_id"), "text")
    .unique("source_id")
)
documents = ids.join(upstream, on="source_id").select(
    (pl.col("text") + "\n<think>" + pl.col("reasoning_trace") + "</think>").alias("text"),
    "token_length",
    "src_path",
)

Data generated by Domyn

Stage 6 includes three sets of trajectories that we generated:

Source	Content
Domyn agentic coding trajectories	Multi-turn trajectories of a coding agent working on software tasks.
Domyn deep-research trajectories	Multi-step research tasks solved with search and browsing tools.
Domyn tool-use trajectories	Multi-turn conversations in which a model calls tools to complete a task.

These sets are released under the Apache License 2.0.

License

Our contributions to this dataset, including the curation and the data we generated, are released under the Apache License 2.0. Upstream corpora maintain their original license terms.

Read ATTRIBUTION.md before redistributing any part of this data. It lists every upstream dataset with its publisher, repository and license. The LICENSES/ folder contains the full text of each license: ODC-BY 1.0, CC BY 4.0, CC BY 2.0, CC BY-SA 4.0, CC0 1.0, GFDL 1.3, Apache 2.0, MIT, BSD-2-Clause and BSD-3-Clause.

Citation

The technical report is forthcoming in November 2026. Until it is published, please cite the dataset:

@misc{domyn_edge_pretraining_mix_2026,
  title     = {Domyn Edge Pretraining Mix},
  author    = {Domyn},
  year      = {2026},
  note      = {Technical report, forthcoming in November 2026.},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/domyn/Domyn-Edge-Pretraining-Mix-Stages}
}