domyn/Domyn-Edge-Teacher-STEM-Mix

Domyn Edge Teacher STEM Mix

Phased release. Domyn Edge is a rolling release in three bundles. This dataset is part of the second bundle, together with the domain teachers and their post-training data. For an overview, see the launch blog post.

The training data of Domyn Edge Teacher STEM, in two configs:

Config	Stage	Train	Test
sft	Supervised fine-tuning, 64k context	5,125,269 (28.13B tokens)	269,426 (1.48B tokens)
rl	GRPO, verifiable multiple-choice questions	47,089 (9.64M tokens)	2,464 (0.51M tokens)

sft: supervised fine-tuning mix

The STEM slice of the Domyn Edge SFT corpus plus general-reasoning QA and long-document QA. Conversations longer than 65,536 tokens were dropped, not truncated; 5% of the mix is held out as the test split. Conversations whose prompts overlap with the final evaluation benchmarks are excluded, to keep the mix usable for evaluation.

Task	Weight	Sources
STEM	1.0	Nemotron-SFT-Science-v2, INTELLECT-3-SFT (openreasoning_science), Dolci-Think-SFT-7B, GLM-5.1-Reasoning-1M-Cleaned (PhD science)
Reasoning	1.0	GLM-5.1-Reasoning-1M-Cleaned (general-domain reasoning QA)
Chat	1.0	Dolci-Think-SFT-7B (general QA)
Mixed	0.5	Dolci-Think-SFT-7B (mixed-domain subsets)
Long context	1.0	Agentic-Long-Context-Understanding-QA

Released conversations and tokens per source, train and test splits together (licenses in short; full terms in ATTRIBUTION.md):

Source	License	Conversations	Tokens
Nemotron-SFT-Science-v2	CC BY-SA 4.0	2,784,704	11.78B
INTELLECT-3-SFT	CC BY 4.0, MIT, Apache 2.0 (per subset)	1,588,967	9.45B
GLM-5.1-Reasoning-1M-Cleaned	Apache 2.0	630,491	2.91B
Dolci-Think-SFT-7B	ODC-BY 1.0	327,467	2.50B
Agentic-Long-Context-Understanding-QA	MIT	63,066	2.97B
Total		5,394,695	29.60B

rl: GRPO questions

Multiple-choice and true/false science questions with a single correct option. The reward is 1 when the option extracted from the response matches reward.ground_truth.answer, else 0. Every prompt states its answer format (for example \boxed{} or "Answer: X"). Questions that overlap with the final evaluation benchmarks are excluded, to keep the config usable for evaluation.

Source	License	Train rows	Train tokens	Test rows	Test tokens
Nemotron-3-Nano-RL-Training-Blend, multiple-choice slice (from Nemotron-RL-knowledge-mcqa)	ODC-BY 1.0 (blend); CC BY 4.0 (component)	18,673	5.44M	990	0.29M
WebInstruct-verified, multiple-choice and Boolean questions	Apache 2.0	28,416	4.20M	1,474	0.22M
Total		47,089	9.64M	2,464	0.51M

How to use

from datasets import load_dataset

sft = load_dataset("domyn/Domyn-Edge-Teacher-STEM-Mix", "sft", split="train", streaming=True)
rl = load_dataset("domyn/Domyn-Edge-Teacher-STEM-Mix", "rl", split="train")

Columns

sft 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.
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.

rl columns

The rl rows use the same canonical form as the sft rows, plus the reward data and the prompt length.

Column	Type	Description
messages	list	The prompt: role and content of each message (the other message fields are null).
tools	string	Always null in this config.
task	string	Task family.
source	string	Upstream dataset.
license	string	License of the upstream record.
metadata	string	JSON-encoded upstream metadata.
reward	string	JSON-encoded reward data: ground_truth and the scoring style. Here ground_truth.answer is the correct option letter (WebInstruct rows also give the question type).
token_length	int	Prompt length in tokens.
uuid	string	Record identifier.

To train with NeMo-Gym, as we did, map each row to the mcqa_simple_agent input: responses_create_params.input = messages; expected_answer = reward.ground_truth.answer; agent_ref = metadata.agent_ref.

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_teacher_stem_mix_2026,
  title     = {Domyn Edge Teacher STEM Mix},
  author    = {Domyn},
  year      = {2026},
  note      = {Technical report, forthcoming in November 2026.},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/domyn/Domyn-Edge-Teacher-STEM-Mix}
}