domyn/Domyn-Edge-Teacher-IF-Mix

Domyn Edge Teacher IF 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 IF, in three configs:

Config	Stage	Train	Test
sft	Supervised fine-tuning, 64k context	2,996,868 (6.32B tokens)	157,177 (331.48M tokens)
rl	GRPO, verifiable instruction-following and tool-call rewards	324,323 (269.89M tokens)	1,691 (1.31M tokens)
rl_lmsys_redacted	GRPO, verifiable instruction-following and tool-call rewards, LMSYS-Chat-1M prompts withheld	5,999 (0.38M tokens)	31 (<0.01M tokens)

sft: supervised fine-tuning mix

Instruction following with verifiable constraints, multi-turn chat, and tool use: single, parallel and irrelevant calls, and single steps of agentic trajectories. Conversations longer than 65,536 tokens were dropped, not truncated; 5% of the mix is held out as the test split. Sources with a weight above 1 appear that many times, as in training; a weight of 0.5 keeps half of a source.

Task	Weight	Sources
agentic	1	ToolMind, Nemotron-SFT-Agentic-v2, Toucan-1.5M, Nemotron-Agentic-v1, Dolci-Instruct-SFT-Tool-Use, xlam-function-calling-60k, ToolACE, Nemotron-Post-Training-Dataset-v1, Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1, LoopTool-23k, xlam-irrelevance-7.5k (Qwen2.5-72B distill)
instruction_following	1	Nemotron-SFT-Instruction-Following-Chat-v2, Nemotron-SFT-Instruction-Following-Chat-v3, Dolci-Think-SFT-7B, IF_multi_constraints_upto5_SFT, Dolci-Instruct-SFT, RECAST-30K, Domyn IF Draft-Verify, ifeval-like-data, SmolTalk2, IF_sft_data_verified, Conifer
table	1	Dolci-Think-SFT-7B

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

Source	License	Weight	Conversations	Tokens
Nemotron-SFT-Instruction-Following-Chat-v2	ODC-BY 1.0	0.5	972,728	1.82B
Nemotron-SFT-Instruction-Following-Chat-v3	CC BY 4.0 (instruction_following)	2	496,968	1.47B
Dolci-Think-SFT-7B	ODC-BY 1.0	1	358,911	781.52M
IF_multi_constraints_upto5_SFT	ODC-BY 1.0	5	208,790	65.50M
ToolMind	Apache 2.0	1	168,136	238.35M
Dolci-Instruct-SFT	ODC-BY 1.0	1	136,800	78.67M
Nemotron-SFT-Agentic-v2	CC BY 4.0, Apache 2.0, MIT (per subset)	1	132,734	554.11M
Toucan-1.5M	Apache 2.0	1	115,285	329.79M
Nemotron-Agentic-v1	CC BY 4.0	1	87,542	351.25M
Dolci-Instruct-SFT-Tool-Use	ODC-BY 1.0	1	79,794	127.74M
RECAST-30K	CC BY 4.0	3	64,365	39.06M
Domyn IF Draft-Verify	Apache 2.0 (responses), ODC-BY 1.0 (prompts)	3	62,634	283.83M
xlam-function-calling-60k	CC BY 4.0	1	58,874	38.04M
ifeval-like-data	Qwen License	1	53,076	11.98M
ToolACE	Apache 2.0	3	31,287	25.67M
SmolTalk2	Apache 2.0 (multi-turn-reasoning-if)	1	28,157	90.51M
Nemotron-Post-Training-Dataset-v1	CC BY 4.0	0.5	28,100	159.34M
Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1	CC BY 4.0	1	20,201	74.41M
LoopTool-23k	Apache 2.0	3	19,194	86.68M
IF_sft_data_verified	ODC-BY 1.0 (as Dolci)	1	14,771	10.21M
Conifer	Apache 2.0	1	13,605	12.23M
xlam-irrelevance-7.5k (Qwen2.5-72B distill)	CC BY 4.0; Qwen License (responses)	1	2,093	1.32M
Total			3,154,045	6.65B

rl: GRPO prompts

Single-turn prompts with verifiable rewards: instruction following graded by IFEval-style constraint checks, structured outputs validated against a schema, citation and free-form formatting checks, and tool calls compared with the reference calls (an empty list means no call should be made). 0.5% of the mix is held out as the test split.

Source	License	Weight	Train rows	Train tokens	Test rows	Test tokens
Nemotron-Post-Training-Dataset-v1	CC BY 4.0	1	77,206	10.49M	387	0.05M
Nemotron-RL-Instruction-Following-Structured-Outputs-v2	CC BY 4.0	1, 1.005	65,856	154.68M	337	0.75M
ToolACE	Apache 2.0	8.3138	58,845	4.31M	283	0.02M
Nemotron-RL-instruction_following	ODC-BY 1.0	1	44,027	7.81M	233	0.04M
When2Call	CC BY 4.0	3.0593, 7.5117	30,516	1.08M	200	0.01M
Dolci-RL-Zero-IF-7B	ODC-BY 1.0	1	12,406	2.92M	67	0.01M
Nemotron-RL-Instruction-Following-Citation-Formatting-v1	CC BY 4.0	1	9,479	16.12M	46	0.08M
Nemotron-RL-Agentic-Function-Calling-Pivot-v1	CC BY 4.0	1	9,418	60.76M	46	0.28M
Nemotron-RL-Instruction-Following-Free-Form-Formatting-v1	CC BY 4.0	1	8,992	11.34M	45	0.07M
xlam-irrelevance-7.5k	CC BY 4.0	0.9931	4,821	0.21M	28	<0.01M
ToolRL	Apache 2.0	1	2,757	0.18M	19	<0.01M
Total			324,323	269.89M	1,691	1.31M

rl_lmsys_redacted: prompts from LMSYS-Chat-1M

Instruction-following prompts built on the first user turn of an LMSYS-Chat-1M conversation, with Domyn's constraints appended. The LMSYS-Chat-1M license does not allow redistribution, so the turn is replaced by <LMSYS_PROMPT> in messages and reward, and lmsys_conversation_id names the conversation it comes from.

Source	License	Weight	Train rows	Train tokens	Test rows	Test tokens
LMSYS-Chat-1M	LMSYS-Chat-1M Dataset License Agreement	1.005	5,999	0.38M	31	<0.01M
Total			5,999	0.38M	31	<0.01M

To train on these rows as we did, accept the LMSYS-Chat-1M license on its page and restore the prompts:

import json
from datasets import load_dataset

PLACEHOLDER = "<LMSYS_PROMPT>"
rows = load_dataset("domyn/Domyn-Edge-Teacher-IF-Mix", "rl_lmsys_redacted", split="train")
ids = set(rows["lmsys_conversation_id"])
lmsys = load_dataset("lmsys/lmsys-chat-1m", split="train")
lmsys = lmsys.filter(lambda r: r["conversation_id"] in ids)
prompts = {r["conversation_id"]: r["conversation"][0]["content"].strip() for r in lmsys}

def fill(value, prompt):
    if isinstance(value, str):
        return value.replace(PLACEHOLDER, prompt)
    if isinstance(value, list):
        return [fill(v, prompt) for v in value]
    if isinstance(value, dict):
        return {k: fill(v, prompt) for k, v in value.items()}
    return value

def rehydrate(row):
    prompt = prompts[row["lmsys_conversation_id"]]
    row["messages"] = fill(row["messages"], prompt)
    row["reward"] = json.dumps(fill(json.loads(row["reward"]), prompt), ensure_ascii=False)
    return row

rows = rows.map(rehydrate).remove_columns("lmsys_conversation_id")

Decontamination and filtering

Rows were checked against BFCL v3, BFCL v4 (bfcl-eval 2026.3.23), IFEval, IFBench and IFStruct v1.0. A row is dropped when its system, user and tool messages contain at least half of a benchmark prompt's word 8-grams. 8-grams shared by 8 or more benchmark prompts, such as IFEval's constraint templates, are not counted, and benchmark prompts under 8 words are ignored. Rows are also dropped when their license does not allow redistribution, or when they were generated from prompts under a non-commercial license.

Config	Reason	Rows	Sources
rl	no license on the dataset card	60,882	swiss-ai/if-rl-singleturn-hard-prompts (60,882)
rl	overlaps BFCL v3	18	ToolACE (8), ToolRL (6), Nemotron-RL-Agentic-Function-Calling-Pivot-v1 (3), Nemotron-Post-Training-Dataset-v1 (1)
rl	overlaps IFBench	1	Dolci-RL-Zero-IF-7B (1)
sft	no license on the dataset card (am_if subset, built from a-m-team distillations, which carry none either)	50,841	PrimeIntellect/INTELLECT-3-SFT (50,841)
sft	overlaps BFCL v3	789	xlam-function-calling-60k (280), ToolMind (278), Nemotron-SFT-Agentic-v2 (139), Nemotron-SFT-Instruction-Following-Chat-v2 (26), Toucan-1.5M (25), Nemotron-Agentic-v1 (23), IF_multi_constraints_upto5_SFT (5), LoopTool-23k (3), ToolACE (3), Nemotron-SFT-Instruction-Following-Chat-v3 (2), Dolci-Instruct-SFT (2), Dolci-Think-SFT-7B (2), ifeval-like-data (1)
sft	overlaps IFBench	155	Nemotron-SFT-Instruction-Following-Chat-v2 (143), Dolci-Instruct-SFT (5), Dolci-Think-SFT-7B (3), IF_sft_data_verified (2), Nemotron-SFT-Instruction-Following-Chat-v3 (2)
sft	overlaps IFEval	28	IF_sft_data_verified (16), Dolci-Instruct-SFT (5), Dolci-Think-SFT-7B (5), ifeval-like-data (1), Nemotron-SFT-Instruction-Following-Chat-v2 (1)
sft	prompt from Salesforce/APIGen-MT-5k (cc-by-nc-4.0)	661	ToolMind (661)

How to use

from datasets import load_dataset

sft = load_dataset("domyn/Domyn-Edge-Teacher-IF-Mix", "sft", split="train", streaming=True)
rl = load_dataset("domyn/Domyn-Edge-Teacher-IF-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}), name and tool_call_id for tool results, and language, reasoning_language, token_length and reasoning_token_length annotations.
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 subset or generation model).
token_length	int	Conversation length in tokens.
uuid	string	Record identifier.

rl columns

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

Column	Type	Description
messages	list	The prompt: system and user messages, or the dialogue up to the turn to answer.
tools	string	JSON-encoded tool schemas for tool-call prompts, else null.
task	string	Task family.
source	string	Upstream dataset.
license	string	License of the upstream record.
metadata	string	JSON-encoded upstream metadata, including the NeMo-Gym agent_ref.
reward	string	JSON-encoded reward data: the scoring style and its ground_truth (see below).
token_length	int	Prompt length in tokens.
uuid	string	Record identifier.

rl_lmsys_redacted has the rl columns plus lmsys_conversation_id.

reward.style	reward.ground_truth
[if]	constraints: the checks, each an instruction id (checker) and its JSON-encoded arguments (function); prompt.
[structured_output]	schema_str and schema_type (JSON, YAML, XML, TOML or CSV); some rows add code_block or a tool-call response_mode.
[citation_formatting], [freeform_formatting]	verifier: the pattern the response must follow.
[fc]	answer: the JSON-encoded reference calls (an empty list for no call), or the reference action for the pivot rows.

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_if_mix_2026,
  title     = {Domyn Edge Teacher IF Mix},
  author    = {Domyn},
  year      = {2026},
  note      = {Technical report, forthcoming in November 2026.},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/domyn/Domyn-Edge-Teacher-IF-Mix}
}