Datasets:

domofon
/

Domofon-Cot-Conversations-700k

split string	family string	language string	source_language string	source_dataset string	source_doc_id string	original_document_text string	conversation_id string	conversation_xml string	domain string	system_prompt string	user_instruction string	persona string	generated_at string	source_file string	line_number int64	n_user_turns int64	n_assistant_turns int64	n_advisor_calls int64	n_calc_calls int64	approx_tokens_conversation_xml int64	approx_tokens_original_document_text int64	approx_tokens_total_with_original int64	tool_call_count int64	tool_result_count int64	advisor_tool_call_count int64	calc_tool_call_count int64
train	conv	en	en	fineweb2_en	fw_en_002714	A Reference Resource Life Before the Presidency Herbert Clark Hoover was born on August 10, 1874. For the first nine years of his life, he lived in the small town of West Branch, Iowa, the place of his birth. His Quaker father, Jessie Clark Hoover, a blacksmith and farm equipment salesman, suffered a heart attack and d...	conv_fw_en_002714_1771	<conversation lang="en" source="fw_en_002714" quality="high" domain="general"> <system>Use at least one concrete example in every response.</system> <user>I'm trying to understand if Herbert Hoover was actually a self-made man or just lucky. Was he really a struggling orphan who clawed his way up, or is that just a...	general	Use at least one concrete example in every response.	Tell me if this is even the right question to ask.	skeptical expert — challenges claims, asks for evidence, pushes beyond surface answers	2026-05-28T23:12:51.757586+00:00	CLEAN_conv.jsonl	1	5	5	4	0	7,836	10,336	18,172	4	4	4	0
train	conv	en	en	fineweb2_en	fw_en_032038	Mother and son seek kidney transplants HAMILTON — Arie Pekar was fooling around one day with his mom’s blood pressure machine when he discovered his reading was exceptionally high. His mom, Ilana, has a blood pressure machine because she was then doing home dialysis for polycystic kidney disease (PKD), which causes cys...	conv_fw_en_032038_5580	<conversation lang="en" source="fw_en_032038" quality="high" domain="general"> <system>You are a debate partner. Present the strongest version of opposing viewpoints.</system> <user>I'm looking into organ donation. Tell me the risks before you tell me the benefits.</user> <assistant> <reasoning>The user is as...	general	You are a debate partner. Present the strongest version of opposing viewpoints.	Tell me the risks before you tell me the benefits.	impatient professional — wants quick answers, gets frustrated with long explanations	2026-05-28T23:15:28.926371+00:00	CLEAN_conv.jsonl	2	4	4	3	0	5,393	5,471	10,864	3	3	3	0
train	conv	en	en	fineweb2_en	fw_en_029984	These are a few of my favourite things; three. These are a few of my favourite things; two These are a few of my favourite things; one. Throw out (or donate) the dark colours, lighten up your hair and your nails, and most importantly, the sun is shining, so your smile should be too. Valentines Day is fastly approaching...	conv_fw_en_029984_8453	<conversation lang="en" source="fw_en_029984" quality="high" domain="general"> <system>You are a debate partner. Present the strongest version of opposing viewpoints.</system> <user>I'm thinking about getting professional eyelash extensions. Explain the trade-offs involved.</user> <assistant> <reasoning>The u...	general	You are a debate partner. Present the strongest version of opposing viewpoints.	Explain the trade-offs involved.	devil's advocate — disagrees to test the assistant, says 'but couldn't you argue that...'	2026-05-28T23:20:32.066508+00:00	CLEAN_conv.jsonl	3	4	4	2	0	4,375	5,264	9,639	2	2	2	0
train	conv	en	en	fineweb2_en	fw_en_028848	ATLANTA (AP) — Jay-Z is taking another step into the pro basketball realm: The Grammy-winning rapper and minority owner of the Brooklyn Nets will serve as the executive producer for the coming "NBA 2K13" video game. 2K Sports announced the partnership with Jay-Z on Tuesday. The rapper, who designed the Nets new logo an...	conv_fw_en_028848_8698	<conversation lang="en" source="fw_en_028848" quality="high" domain="general"> <system>End every response with a follow-up question to keep the conversation going.</system> <user>I'm looking into Jay-Z's involvement with the NBA 2K13 video game. Point out anything I might be overlooking.</user> <assistant> <r...	general	End every response with a follow-up question to keep the conversation going.	Point out anything I might be overlooking.	detail-oriented analyst — wants numbers, sources, edge cases	2026-05-28T23:20:32.918793+00:00	CLEAN_conv.jsonl	4	5	5	5	0	4,942	1,996	6,938	5	5	5	0
train	conv	en	en	fineweb2_en	fw_en_016394	When I first watched this clip I had flashbacks to my one and only trip to African Lion Safari located here in Southern Ontario, just west of Hamilton, ON. Our car seemed to be in peril throughout the whole drive through the game reserve. A group of zebra rubbed up against the door, an ostrich pecked at the window, and...	conv_fw_en_016394_9410	<conversation lang="en" source="fw_en_016394" quality="high" domain="general"> <system>If the user asks a yes-or-no question, answer yes or no first, then explain.</system> <user>I just saw a video of monkeys destroying a car! Was it actually a real safari park?</user> <assistant> <reasoning>The user is askin...	general	If the user asks a yes-or-no question, answer yes or no first, then explain.		enthusiastic learner — excited about the topic, asks 'what else?' and 'that's cool, but why?'	2026-05-28T23:20:34.064414+00:00	CLEAN_conv.jsonl	5	5	5	5	0	4,293	1,499	5,792	5	5	5	0
train	conv	en	en	markdown_corpus	wiki_012890	Tom Clancy's Splinter Cell: Essentials is an action-adventure video game with heavy emphasis on stealth. It is part of the Splinter Cell series and was released for the PlayStation Portable handheld system. It was developed by Ubisoft Montreal and published by Ubisoft on March 21, 2006. It is the fourth entry in the se...	conv_wiki_012890_1774	<conversation lang="en" source="wiki_012890" quality="high" domain="reference"> <system>Include one specific number or statistic in every response.</system> <user>I'm looking into the release of Splinter Cell: Essentials. Can you compare the pros and cons of the title?</user> <assistant> <reasoning>The user w...	reference	Include one specific number or statistic in every response.	Compare the pros and cons.	detail-oriented analyst — wants numbers, sources, edge cases	2026-05-28T23:20:34.194106+00:00	CLEAN_conv.jsonl	6	5	5	4	1	4,079	3,493	7,572	5	5	4	1
train	conv	en	en	cc_raw	cc_050445	Sunday, September 28, 2008 A simple wish... ... that this year should be a sweet one filled with health, happiness and prosperity. Before I sign off for Rosh Hashanah, I'd like to share one of the most important lesson I've learned from my bees: The best way to ensure a sweet future is to resist the urge to take every ...	conv_cc_050445_2847	<conversation lang="en" source="cc_050445" quality="high" domain="general"> <system>You are a concise technical advisor. Prioritize accuracy over friendliness.</system> <user>I'm hearing about issues with Joseph's Tomb. How do I actually get access to it as a visitor?</user> <assistant> <reasoning>The user is...	general	You are a concise technical advisor. Prioritize accuracy over friendliness.		practical doer — only cares about actionable advice, asks 'how do I actually do this?'	2026-05-28T23:22:44.026200+00:00	CLEAN_conv.jsonl	7	5	5	4	0	4,559	55,882	60,441	4	4	4	0
train	conv	en	en	fineweb2_en	fw_en_022918	The few West Indian novels I studied in school have a special place in my heart, because they were some of the only books that I could see myself reflected in. I love “Harry Potter” and “The Hunger Games” as much as anyone, but none of them remind of me running barefoot through the streets of Tobago, during school vaca...	conv_fw_en_022918_2466	<conversation lang="en" source="fw_en_022918" quality="high" domain="general"> <system>You are a helpful assistant.</system> <user>I've been looking into the cultural identity of the Caribbean. What is the single most important thing I should know?</user> <assistant> <reasoning>The user is asking for a high-l...	general	You are a helpful assistant.	What is the single most important thing I should know?	skeptical expert — challenges claims, asks for evidence, pushes beyond surface answers	2026-05-28T23:23:32.842501+00:00	CLEAN_conv.jsonl	8	4	4	3	0	4,220	6,199	10,419	3	3	3	0
train	conv	en	en	fineweb2_en	fw_en_029374	Perhaps the most interesting "sidebar" story to emerge from the Red Wings' 4-3 OT loss to the St. Louis blues involved Red Wings coach Mike Babcock's declaration that his team's personnel changes (see; no Nicklas Lidstrom, Brad Stuart or Brian Rafalski, with no real replacements found) and injuries on defense (see: Bre...	conv_fw_en_029374_9890	<conversation lang="en" source="fw_en_029374" quality="high" domain="general"> <system>When you are uncertain, say so explicitly and estimate your confidence as a percentage.</system> <user>I'm trying to understand why the Detroit Red Wings are playing differently lately. Focus on what matters most and ignore the r...	general	When you are uncertain, say so explicitly and estimate your confidence as a percentage.	Focus on what matters most and ignore the rest.	confused beginner — needs things explained multiple ways, asks 'what do you mean by...'	2026-05-28T23:24:32.943361+00:00	CLEAN_conv.jsonl	9	4	4	3	0	3,954	1,574	5,528	3	3	3	0
train	conv	en	en	fineweb2_en	fw_en_029909	I don’t feel like taking a shower this morning. I took one yesterday and I’m feeling LAZY this morning. So, I am trying out some “hair powder” (fancy word for dry shampoo) that I got in a Birchbox a couple months ago. It’s called Lulu Organics Hair Powder and I’ll be honest – it smells a bit “old lady” but we shall see...	conv_fw_en_029909_2913	<conversation lang="en" source="fw_en_029909" quality="high" domain="general"> <system>Structure every response as: acknowledge the question, explain the core idea, suggest a next step.</system> <user>I'm looking for information on Lulu Organics Hair Powder. What exactly is it and how is it applied?</user> <assis...	general	Structure every response as: acknowledge the question, explain the core idea, suggest a next step.		detail-oriented analyst — wants numbers, sources, edge cases	2026-05-28T23:24:34.027836+00:00	CLEAN_conv.jsonl	10	6	6	5	0	5,388	691	6,079	5	5	5	0
train	conv	en	en	fineweb2_en	fw_en_036270	Full Show Index Advertise With Us Write For Us RNO Roundtable: ‘Survivor: One World’ Pre-Show PredictionsPage 5 View Printable version of this article Sarah: Her name is Kourtney Moon, which immediately makes me think of Daphne Moon from Frasier, and it's entirely appropriate since she's been cast as the kooky-girl (wh...	conv_fw_en_036270_9795	<conversation lang="en" source="fw_en_036270" quality="high" domain="general"> <system>Explain your reasoning step by step. Show your work.</system> <user>I'm reading some old predictions for Survivor: One World. Who is Kourtney Moon and what do people think of her?</user> <assistant> <reasoning>The user is a...	general	Explain your reasoning step by step. Show your work.		devil's advocate — disagrees to test the assistant, says 'but couldn't you argue that...'	2026-05-28T23:25:40.013129+00:00	CLEAN_conv.jsonl	11	4	4	3	0	6,523	10,807	17,330	3	3	3	0
train	conv	en	en	fineweb2_en	fw_en_021261	Our website uses anonymous cookies to generate statistics. If you have any concerns about this, please click here to find out more. This site uses Google Analytics, a web analytics service provided by Google, Inc. Google Analytics sets a number of ANONYMOUS cookies in order to evaluate your use of the site and compile ...	conv_fw_en_021261_3298	<conversation lang="en" source="fw_en_021261" quality="high" domain="general"> <system>Always respond in exactly three sentences. No more, no less.</system> <user>I see your site uses cookies for stats. How do I actually stop my data from being tracked?</user> <assistant> <reasoning>The user is asking for the...	general	Always respond in exactly three sentences. No more, no less.		practical doer — only cares about actionable advice, asks 'how do I actually do this?'	2026-05-28T23:25:49.153717+00:00	CLEAN_conv.jsonl	12	5	5	4	0	4,037	826	4,863	4	4	4	0
train	conv	en	en	cc_raw	cc_052460	Thorsten Heins is now a Canadian citizen BlackBerry CEO Thorsten Heins celebrated the release of the new CrackBerry in a better way than most of us - by becoming a Canadian citizen. In a small ceremony that took place in Kitchener yesterday, Heins was one of many who were granted Canadian citizenship. He is now a dual ...	conv_cc_052460_5558	<conversation lang="en" source="cc_052460" quality="high" domain="general"> <system>Begin every response by restating the user's question in your own words.</system> <user>I'm looking for some specific details about Thorsten Heins. Skip the background, I already know the basics.</user> <assistant> <reasoning>...	general	Begin every response by restating the user's question in your own words.	Skip the background, I already know the basics.	confused beginner — needs things explained multiple ways, asks 'what do you mean by...'	2026-05-28T23:29:09.654259+00:00	CLEAN_conv.jsonl	14	4	4	3	0	3,209	790	3,999	3	3	3	0
train	conv	en	en	fineweb2_en	fw_en_013185	I usually trade in 3-minute bits, and confidence, or the illusion thereof, is pretty important. In my case, my sights are set on a bunch of laughs throughout, but I'd would think especially if you're building one long slow burn into a huge punchline at the end, you can't waver in what you're doing. An example: When the...	conv_fw_en_013185_2791	<conversation lang="en" source="fw_en_013185" quality="high" domain="general"> <system>You are a concise technical advisor. Prioritize accuracy over friendliness.</system> <user>I'm getting into stand-up comedy! Is it better to do a bunch of short jokes or one long story?</user> <assistant> <reasoning>The use...	general	You are a concise technical advisor. Prioritize accuracy over friendliness.		enthusiastic learner — excited about the topic, asks 'what else?' and 'that's cool, but why?'	2026-05-28T23:29:15.178102+00:00	CLEAN_conv.jsonl	15	4	4	3	0	4,183	2,179	6,362	3	3	3	0

End of preview. Expand in Data Studio

Domofon-Cot-Conversations-700k

Synthetic XML conversation data for training small language models on reasoning, instruction following, XML formatting, and tool-use traces.

Repository: domofon/Domofon-Cot-Conversations-700k

What is inside

The dataset contains cleaned generated XML conversations from six families:

conv: multi-turn factual conversations with tool-use traces.
instruct: text-processing instructions, including deterministic count tool calls.
ds: DeepSeek/Gemma-generated repair and XML/text transformation tasks.
dschat: chat-style assistant behavior tasks.
mm: MiniMax-generated text-processing tasks, aggressively filtered.
mimo: Xiaomi MiMo-generated repair/XML/text tasks.

Rows include the generated conversation_xml and an original_document_text column recovered from the local source corpora by source_doc_id.

The rejected split is included for audit/research. It contains rows rejected by the cleaner with reject_reason and reject_detail.

Tool-use idea

This dataset is intended for pretraining or SFT of SLMs that need to learn tool use and XML-style interaction. In the conv family, the experimental Advisor tool represents a small model calling a larger LLM or external advisor to answer questions about context. Other tools include WebSearch, calc, and count.

Splits and shard configs

The default Hugging Face config keeps two aggregate splits:

train: cleaned retained rows.
rejected: rejected/audit rows.

Additional configs expose data-type shards directly:

by_family: family/data-type splits: train_conv/rejected_conv, train_instruct/rejected_instruct, train_ds/rejected_ds, train_dschat/rejected_dschat, train_mm/rejected_mm, train_mimo/rejected_mimo.
by_family_language: language + family/data-type splits: train_en_conv/rejected_en_conv, train_en_instruct/rejected_en_instruct, train_en_ds/rejected_en_ds, train_en_dschat/rejected_en_dschat, train_en_mm/rejected_en_mm, train_en_mimo/rejected_en_mimo, train_ru_conv/rejected_ru_conv, train_ru_instruct/rejected_ru_instruct, train_ru_ds/rejected_ru_ds, train_ru_dschat/rejected_ru_dschat, train_ru_mm/rejected_ru_mm, train_ru_mimo/rejected_ru_mimo.

Parquet shards are physically split by language and family in the filename, for example train-en-conv-00000.parquet or rejected-ru-instruct-00000.parquet. Use the language and family columns to filter.

Metrics

Token counts are approximate, using the requested heuristic:

English/non-Cyrillic character = 1 token
Russian/Cyrillic character = 1.8 tokens

Metric	Count
Total rows	908,670
Train rows	768,359
Rejected rows	140,311
Approx train tokens in `conversation_xml`	4,921,333,331
Approx rejected tokens in `conversation_xml`	644,341,246
Approx total tokens in `conversation_xml`	5,565,674,577
Approx total tokens including `original_document_text`	9,751,251,367

By family

Family	Rows	Approx conversation XML tokens
conv	193,607	2,497,197,156
ds	251,220	671,511,925
dschat	72,807	140,357,015
instruct	348,080	2,142,207,736
mimo	22,849	63,598,674
mm	20,107	50,802,071

By language

Language	Rows	Approx conversation XML tokens	Approx tokens incl. original text
en	708,307	4,118,087,295	6,784,171,884
ru	200,363	1,447,587,282	2,967,079,483

Top rejected reasons

truncated_or_empty_response: 59,032
bloom_ladder_missing_question: 28,747
hidden_context_missing_source: 18,115
web_boilerplate_or_source_garbage: 13,140
uncertain_tool_result_assertive_response: 8,463
count_mismatch: 5,591
inner_xml_invalid: 1,441
assistant_missing_reasoning: 1,425
assistant_missing_response: 1,266
mixed_cjk_contamination: 1,228
malformed_role_order: 669
unpaired_tool_result: 344
bloom_ladder_missing_level: 275
dangling_final_user: 243
dangling_or_repeated_user: 175
pending_tool_call: 98
calc_mismatch: 31
bloom_ladder_incomplete: 21
inner_xml_missing: 4
invalid_tool_name: 2

Columns

Important columns:

conversation_xml: generated XML conversation, including reasoning/tool calls when present.
original_document_text: source text/document used to make the conversation.
family: one of conv, instruct, ds, dschat, mm, mimo.
language: normalized en or ru.
source_dataset, source_doc_id: provenance into the source corpus.
reject_reason, reject_detail: populated for the rejected split.
approx_tokens_*: approximate token estimates using the heuristic above.

Caveats

This is synthetic data and should be inspected before training production systems. Source corpora are mixed web/wiki/common-crawl style text; verify license and safety requirements for your use case.

Downloads last month: 115