domyn/Domyn-Edge-SFT-Mix

Domyn Edge SFT Mix

Phased release. Domyn Edge is a rolling release in three bundles. This dataset is part of the second bundle. It is the supervised fine-tuning mix of domyn/Domyn-Edge-SFT, released in the first bundle. For an overview, see the launch blog post.

Domyn Edge SFT was built in three supervised stages, each trained for one epoch from the previous stage's weights. Each stage has its own mix, released here as a separate config:

Config	Stage	Sequence length	Steps	Trained on (train split)	Released: train	Released: test
64k	64k base	65,536	14,195	26,168,884 conversations, 228.1B tokens	24,831,170 (211.00B tokens)	1,302,432 (11.05B tokens)
128k	128k context extension	131,072	2,022	2,596,448 conversations, 29.8B tokens	2,501,799 (29.58B tokens)	131,352 (1.55B tokens)
512k	512k context extension	524,288	3,592	3,239,400 conversations, 57.3B tokens	3,126,965 (57.00B tokens)	162,702 (2.81B tokens)

The released configs leave out the data whose license we could not confirm. See Differences from the training mix; the withheld public data can be rebuilt from its source as described in Rebuilding the withheld rows.

How the mixes were built

64k base. A weighted mix of public SFT datasets covering math, code, multilingual chat, STEM, software engineering, agentic tool use, reasoning, instruction following, finance and safety, plus our internal data. Conversations longer than 65,536 tokens were dropped.

128k and 512k context extension. Each extension stage trains on a "context extension" mix: every available conversation longer than the previous context length, joined with a downsampled copy of the shorter mix so that short-context skills are preserved.

• 128k = the 64k mix downsampled to 10% (24.5B tokens, conversations up to 65,535 tokens), joined with every conversation of 65,536 to 131,072 tokens from 13 datasets (5.4B tokens).
• 512k = the 64k mix downsampled to about 12%, so that its token count roughly matches the long half, joined with every conversation of 65,536 to 524,288 tokens, including our long-context QA data.

Each stage shifts the mix toward long context:

Task	64k base	128k c.ext	512k c.ext
Long context	1.2%	14.9%	43.0%
Math	31.7%	24.2%	14.4%
Code	20.8%	15.9%	7.4%
Multilingual	12.2%	9.4%	5.6%
STEM	10.9%	8.3%	4.4%
Software engineering	3.6%	7.6%	8.3%
Agentic	3.9%	7.9%	7.6%
Reasoning	4.2%	3.3%	4.0%
Instruction following	4.4%	3.3%	2.1%
Finance	4.2%	3.2%	2.0%
Deep research	1.3%	1.0%	0.6%
Other (mixed, chat, safety)	1.4%	1.0%	0.6%

Sources

Released conversations and tokens per source, train and test splits together. Licenses are short labels of those published by the upstream source; the full terms and notes are in ATTRIBUTION.md.

Source	License	64k conversations	64k tokens	128k conversations	128k tokens	512k conversations	512k tokens
INTELLECT-3-SFT	CC BY 4.0, MIT, Apache 2.0 (per subset)	4,861,029	53.94B	486,209	5.35B	579,663	6.46B
Nemotron-SFT-Science-v2	CC BY-SA 4.0	2,784,704	11.78B	283,558	1.54B	331,948	1.40B
Nemotron-SFT-Multilingual-v1	CC BY 4.0, CC BY-SA 4.0	2,563,401	23.54B	258,029	2.49B	305,563	2.80B
Nemotron-SFT-Math-v3	CC BY 4.0, CC BY-SA 4.0	2,482,941	31.93B	252,818	3.32B	295,909	3.80B
Dolci-Think-SFT-7B	ODC-BY 1.0	1,986,352	18.71B	198,630	1.87B	236,763	2.23B
Nemotron-SFT-Instruction-Following-Chat-v2	ODC-BY 1.0	1,943,201	3.63B	194,303	360.6M	231,640	433.3M
Dolci-Instruct-SFT	ODC-BY 1.0	2,055,213	1.66B	—	—	—	—
Domyn internal data	Apache 2.0	1,261,230	3.15B	126,126	302.3M	150,342	374.6M
Nemotron-SFT-Agentic-v2	CC BY 4.0	968,495	5.18B	98,202	443.2M	118,139	859.2M
Nemotron-SFT-Instruction-Following-Chat-v3	CC BY 4.0 and ODC-BY 1.0	749,111	3.83B	75,032	391.4M	89,614	479.9M
GLM-5.1-Reasoning-1M-Cleaned	Apache 2.0	726,908	3.77B	72,804	387.2M	86,645	449.2M
Nemotron-SFT-Competitive-Programming-v2	CC BY 4.0, ODC-BY 1.0	692,261	14.01B	84,105	2.69B	82,512	1.67B
Nemotron-SFT-OpenCode-v1	CC BY 4.0	449,123	6.79B	45,030	277.4M	53,794	829.8M
Nemotron-SFT-Math-v4	CC BY 4.0, CC BY-SA 4.0	371,537	3.83B	38,133	465.2M	44,287	457.4M
Nemotron-SFT-Multilingual-v2	CC BY 4.0, CC BY-SA 4.0	367,828	3.77B	37,002	396.3M	43,843	447.4M
ToolMind	Apache 2.0	337,881	820.1M	33,788	45.1M	40,275	97.8M
Nemotron-SpecializedDomains-Finance-v1	CC BY 4.0	326,698	10.15B	32,670	1.02B	38,943	1.21B
Nemotron-SFT-SWE-v2	CC BY 4.0	251,494	3.33B	26,436	372.7M	34,732	766.9M
Nemotron-Terminal-Corpus	CC BY 4.0	—	—	110,548	1.42B	175,354	3.15B
Nemotron-SFT-ARC-AGI-v1	CC BY 4.0	211,524	6.24B	23,312	780.5M	42,755	2.07B
Nemotron-Agentic-v1	CC BY 4.0	202,001	1.05B	20,235	102.2M	24,169	125.7M
Agentic-Long-Context-Understanding-QA	MIT	63,066	2.97B	50,644	4.68B	52,003	4.75B
Nemotron-SFT-Safety-v2	CC BY 4.0	130,338	111.6M	13,034	11.0M	15,536	13.2M
Scale-SWE-Distilled	CC BY 4.0	54,255	2.32B	22,708	1.27B	36,428	2.41B
daVinci-Dev	CC BY 4.0	62,696	2.16B	15,961	581.9M	26,314	1.43B
dolma3_longmino_pool	ODC-BY 1.0 (documents), Apache 2.0 (QA)	—	—	—	—	83,961	14.28B
Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1	CC BY 4.0	57,080	250.5M	5,672	26.0M	6,712	29.5M
Nemotron-RL-Agentic-SWE-Pivot-v1	CC BY 4.0	48,152	1.16B	4,913	108.2M	6,445	190.5M
Nemotron-SFT-Safety-v1	CC BY 4.0	45,134	29.5M	4,514	2.9M	5,380	3.5M
Nemotron-Math-Proofs-v2	CC BY 4.0	40,238	1.63B	5,792	371.2M	4,796	193.4M
long-context-synth	Apache 2.0	—	—	—	—	29,578	6.29B
smoltalk2	Apache 2.0	28,158	90.5M	—	—	—	—
Toucan-1.5M	Apache 2.0	—	—	11,562	18.2M	13,952	58.5M
Nemotron-SFT-SWE-v3	CC BY 4.0	5,643	142.0M	790	33.1M	968	41.2M
SWE-Next-SFT-Trajectories	MIT	3,692	63.9M	369	5.2M	440	7.8M
Nemotron-SFT-CUDA-v1	CC BY 4.0	2,218	21.4M	222	0.8M	264	2.5M
Total		26,133,602	222.05B	2,633,151	31.12B	3,289,667	59.81B

Differences from the training mix

The release differs from the data Domyn Edge SFT was trained on in two ways:

• Public data not redistributed. Some public data used in training is not redistributed here because its terms do not allow it. This includes IF_sft_data_verified (no published license) and ifeval-like-data (Qwen license), about 2.2k and 1.2k conversations of the 64k training mix.
• Withheld prompts. In Nemotron-SFT-Instruction-Following-Chat-v3, the seed prompts taken from LMSYS-Chat-1M are not redistributable. As in the upstream release, the first system and user messages of those rows are set to null; NVIDIA's prepare_chat_prompts.py restores them for users with access to LMSYS-Chat-1M. The rest of each conversation is kept.

Rebuilding the withheld rows

Part of the public data left out of this release can be downloaded from its source and added back.

IF_sft_data_verified and ifeval-like-data. These appear only in the 64k training mix. To approximate the training mix, add about 2,200 conversations of IF_sft_data_verified and 1,200 of ifeval-like-data, sampled at random (95% to train, 5% to test), keeping conversations of at most 65,536 tokens.

LMSYS-seeded prompts of Nemotron-SFT-Instruction-Following-Chat-v3. Use NVIDIA's prepare_chat_prompts.py from the upstream repository, which restores the null messages for users with access to LMSYS-Chat-1M.

How to use

from datasets import load_dataset

mix_128k = load_dataset("domyn/Domyn-Edge-SFT-Mix", "128k", split="train", streaming=True)

Columns

Column	Type	Description
messages	list	The conversation. Each message has role, content, reasoning_content (assistant reasoning, if any), tool_calls (each with id, type and function {name, arguments as a JSON string}), and name and tool_call_id for tool results.
tools	string	JSON-encoded list of the tool schemas available in the conversation, or null.
task	string	Task family (math, code, stem, agentic, ...).
source	string	Upstream dataset.
license	string	License of the upstream record.
metadata	string	JSON-encoded upstream metadata (for example topic or generation model), or null.
uuid	string	Record identifier.

License

Our contributions to this dataset, including the curation and the mix composition, are released under the Apache License 2.0. Upstream datasets keep their original license terms. Read ATTRIBUTION.md before redistributing any part of this data.

Citation

@misc{domyn_edge_sft_mix_2026,
  title     = {Domyn Edge SFT Mix},
  author    = {Domyn},
  year      = {2026},
  note      = {Technical report, forthcoming in November 2026.},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/domyn/Domyn-Edge-SFT-Mix}
}