Datasets:

hudsongouge
/

microagent-sft-v1

id stringlengths 23 23	messages listlengths 2 10	chatml stringlengths 129 7.59k	num_messages int64 2 10	has_tools bool 2 classes	has_thinking bool 2 classes	thinking stringlengths 36 1.5k ⌀	bundle stringclasses 7 values	domain stringclasses 29 values	module stringclasses 22 values	family_id stringlengths 3 131 ⌀	difficulty stringclasses 3 values	primary_action stringclasses 13 values	think_style stringclasses 2 values	metadata_json stringlengths 22 1.57k	family stringlengths 7 71 ⌀	task_kind stringclasses 55 values	track stringclasses 3 values	code_profile stringclasses 2 values
000000-83016921c0fd9822	[ { "content": "Answer using only the user-provided text below.\n\nCopied passage (clinical brief file):\n\nShow notes for Bellweather Signal, episode 64 (2021): host Dr. Cass Holloway, runtime 60 minutes.\nGuest list not included. No transcript body attached.\n\nQuestion: From the pasted excerpt only: which gues...	<\|im_start\|>user Answer using only the user-provided text below. Copied passage (clinical brief file): Show notes for Bellweather Signal, episode 64 (2021): host Dr. Cass Holloway, runtime 60 minutes. Guest list not included. No transcript body attached. Question: From the pasted excerpt only: which guests appeared?...	2	false	true	1. The pasted excerpt says: “Answer using only the user-provided text below.”. 2. The follow-up question is: From the pasted excerpt only: which guests appeared. 3. Each supported claim has to match wording from that excerpt.	reliability	entity_grounding	entity_grounding	paste_gap\|pasted_passage_missing_fact\|bellweather_signal	easy	REFUSE_MISSING_INFO	context_chain	{"difficulty": "easy", "entity_category": "album", "entity_lint_code": "ok", "entity_lint_ok": true, "entity_name": "Bellweather Signal", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "family_id": "paste_gap\|pasted_passage_missing_fact\|bellweather_signal", "forbidden_phrases": ["guest list", "transcript", "sp...	null	pasted_passage_missing_fact	reliability	null
000001-477cd837a98d5569	[ { "content": "Curious about something: Is Carrick stripedii endangered?\n\n(No article, document, or search results were provided in this turn.)", "role": "user" }, { "content": "<think>\n1. The question asks: Curious about something: Is Carrick stripedii endangered.\n2. This turn does not include a...	<\|im_start\|>user Curious about something: Is Carrick stripedii endangered? (No article, document, or search results were provided in this turn.)<\|im_end\|> <\|im_start\|>assistant <think> 1. The question asks: Curious about something: Is Carrick stripedii endangered. 2. This turn does not include a source passage or tool...	2	false	true	1. The question asks: Curious about something: Is Carrick stripedii endangered. 2. This turn does not include a source passage or tool result. 3. A precise factual answer would require evidence that is not here.	reliability	entity_grounding	entity_grounding	open_abstain\|open_entity_abstain\|carrick_stripedii	easy	REFUSE_OUT_OF_SCOPE	context_chain	{"difficulty": "easy", "dpo_negative_hint": "invent a plausible biography from general knowledge", "entity_category": "species", "entity_lint_code": "ok", "entity_lint_ok": true, "entity_name": "Carrick stripedii", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "family_id": "open_abstain\|open_entity_abstain\|ca...	null	open_entity_abstain	reliability	null
000002-b0b8827e2e2efd19	[ { "content": "Counterfactual update task:\nStart time is 22:00 (24h).\nQuestion A: what time is it after 5h 50m?\nCounterfactual change: add an extra 90 minutes on top of that.\nQuestion B: what is the new time?", "role": "user" }, { "content": "A) 03:50\nB) 05:20", "role": "assistant" } ]	<\|im_start\|>user Counterfactual update task: Start time is 22:00 (24h). Question A: what time is it after 5h 50m? Counterfactual change: add an extra 90 minutes on top of that. Question B: what is the new time?<\|im_end\|> <\|im_start\|>assistant A) 03:50 B) 05:20<\|im_end\|>	2	false	false	null	reasoning	counterfactual_state	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000003-781c20546b4469c2	[ { "content": "Could you give the strictly supported answer?\nThe train departure got delayed by 2 hours. What is the new start time?", "role": "user" }, { "content": "<think>\n1. The note uses “that weekday phrase” for the event timing.\n2. A single calendar date needs today's date or which week is ...	<\|im_start\|>user Could you give the strictly supported answer? The train departure got delayed by 2 hours. What is the new start time?<\|im_end\|> <\|im_start\|>assistant <think> 1. The note uses “that weekday phrase” for the event timing. 2. A single calendar date needs today's date or which week is meant. 3. Without that...	2	false	true	1. The note uses “that weekday phrase” for the event timing. 2. A single calendar date needs today's date or which week is meant. 3. Without that anchor the weekday label does not map to one exact date.	reasoning	hidden_assumptions	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000004-9c94a6fe2de85a34	[ { "content": "Find x in the equation 5x - 9 = 21. Explain each step briefly.", "role": "user" }, { "content": "<think>\n1. The equation is x in the equation 5x - 9 = 21. 2. 3.\n</think>\n1. Add 9 to both sides: 5x = 30, then divide by 5.\n2. Therefore x = 6.", "role": "assistant" } ]	<\|im_start\|>user Find x in the equation 5x - 9 = 21. Explain each step briefly.<\|im_end\|> <\|im_start\|>assistant <think> 1. The equation is x in the equation 5x - 9 = 21. 2. 3. </think> 1. Add 9 to both sides: 5x = 30, then divide by 5. 2. Therefore x = 6.<\|im_end\|>	2	false	true	1. The equation is x in the equation 5x - 9 = 21. 2. 3.	math	math_equation	numeric_integrity	linear_subtract_both_5_9_21	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 6, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "linear_subtract_both_5_9_21", "family_id": "linear_subtract_both_5_9_21", "module": "numeric_i...	linear_subtract_both_5_9_21	linear_equation	reliability	null
000005-00256c3be08d78c2	[ { "content": "You are a large language model.", "role": "system" }, { "content": "Can you help me verify this: who's the prime minister of Canada right now?", "role": "user" }, { "content": "<think>\nI still should not present older information as the current answer, but I can mention wh...	<\|im_start\|>system You are a large language model.<\|im_end\|> <\|im_start\|>user Can you help me verify this: who's the prime minister of Canada right now?<\|im_end\|> <\|im_start\|>assistant <think> I still should not present older information as the current answer, but I can mention what I last recall if I clearly label it ...	3	false	true	I still should not present older information as the current answer, but I can mention what I last recall if I clearly label it as historical and uncertain for 'now'.	time_aware	time_aware	null	null	null	null	null	{"boundary_entity": "Prime Minister of Canada", "boundary_family": "fact_canadian_prime_ministers_prime_minister_of_canada", "boundary_kind": "fact_db", "fact_category": "brazilian_presidents", "fact_entity": "President of Brazil", "has_valid_fact": false, "reference_date": "2025-05-13", "scenario": "no_system_date_ref...	null	null	time_aware	null
000006-cc5eaa335762bc25	[ { "content": "You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result.\n\nTool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read work...	<\|im_start\|>system You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result. Tool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read workspace...	5	true	true	1. “Curious about something: What is the exact figure for Elena” needs a verified fact. 2. That fact is not already in the user message. 3. A webSearch call should fetch the supporting line first. --- 1. The user asked: Curious about something: What is the exact figure for Elena. 2. 3. The visible answer should stay ...	reliability	open_world_knowledge	open_world_knowledge	search_person_elena_statistic	medium	EXECUTE_TOOL_THEN_ANSWER	context_chain	{"allowed_tools": ["webSearch", "readFile"], "complex_prompt_builtin": true, "difficulty": "medium", "entity_category": "person", "entity_name": "Elena", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["E5_tool_bypass", "E2_retrieval_hallucination", "KE_knowledge_error"], "family_id": "sear...	null	tool_search_answer	reliability	null
000007-7352b7dec6f08226	[ { "content": "What value of t satisfies 3(t + 6) = 27? Work it out step by step.", "role": "user" }, { "content": "<think>\n1. The expression is t satisfies 3(t + 6) = 27.\n2. Evaluate the grouped part (t + 6) before any outer multiplication.\n3. Then finish any remaining addition or subtraction out...	<\|im_start\|>user What value of t satisfies 3(t + 6) = 27? Work it out step by step.<\|im_end\|> <\|im_start\|>assistant <think> 1. The expression is t satisfies 3(t + 6) = 27. 2. Evaluate the grouped part (t + 6) before any outer multiplication. 3. Then finish any remaining addition or subtraction outside the group. </thin...	2	false	true	1. The expression is t satisfies 3(t + 6) = 27. 2. Evaluate the grouped part (t + 6) before any outer multiplication. 3. Then finish any remaining addition or subtraction outside the group.	math	math_equation	numeric_integrity	linear_scaled_sum_3_6_27	easy	ANSWER_EXACT	context_chain	{"difficulty": "easy", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 3, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "linear_scaled_sum_3_6_27", "family_id": "linear_scaled_sum_3_6_27", "module": "numeric_integrity...	linear_scaled_sum_3_6_27	linear_equation	reliability	null
000008-ec89aea9034312f8	[ { "content": "Quick question from the floor:\n\nA team has a $78 budget. They spend $11 and later receive a $18 refund. How many dollars remain? One sentence with the integer.", "role": "user" }, { "content": "<think>\n1. Starting budget is $78 and $11 is spent first.\n2. After spending that leaves ...	<\|im_start\|>user Quick question from the floor: A team has a $78 budget. They spend $11 and later receive a $18 refund. How many dollars remain? One sentence with the integer.<\|im_end\|> <\|im_start\|>assistant <think> 1. Starting budget is $78 and $11 is spent first. 2. After spending that leaves $78 − $11 = $67. 3. A $...	2	false	true	1. Starting budget is $78 and $11 is spent first. 2. After spending that leaves $78 − $11 = $67. 3. A $18 refund is added to $67 for the remaining dollars.	reliability	numeric_integrity	numeric_integrity	word_budget_78_11_18	easy	ANSWER_EXACT	context_chain	{"difficulty": "easy", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 85, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "word_budget_78_11_18", "family_id": "word_budget_78_11_18", "module": "numeric_integrity", "num...	word_budget_78_11_18	word_problem_exact	reliability	null
000009-06f10cfda9102cfa	[ { "content": "Following up on a ticket—I've included the thread and docs we actually have.\n\nPlease review this incident plan against the operations incident handling and response handbook.\n\nHandbook: Incident response handbook\n\nManual index:\n- Safety risk incidents [IR1]\n- Data exposure incidents [IR2]\...	<\|im_start\|>user Following up on a ticket—I've included the thread and docs we actually have. Please review this incident plan against the operations incident handling and response handbook. Handbook: Incident response handbook Manual index: - Safety risk incidents [IR1] - Data exposure incidents [IR2] - Public or c...	2	false	true	1. The user needs the rule in section or. 2. Search the provided handbook text for that section heading or id. 3. Quote or paraphrase only what section or actually states.	reliability	grounded_retrieval	grounded_retrieval	rulebook_incident_response	hard	CITE_SECTIONS	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "hard", "doc_pattern": "manual", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["F2_misattribution", "F3_unfaithful_summary", "KE_knowledge_error"], "family": "incident_response", "family_id": "rulebook_inc...	incident_response	doc_search_citation	reliability	null
000010-2a1fc053f5612262	[ { "content": "Following up on a ticket—I've included the thread and docs we actually have.\n\nCalendar hold:\nAll-hands Tuesday 10am — ignore for factual questions unless the packet mentions it.\n\nSummarize this incident in at most 8 words: 'Checkout API latency spiked to 4.2s during deploy deploy-4821.'\n\nAn...	<\|im_start\|>user Following up on a ticket—I've included the thread and docs we actually have. Calendar hold: All-hands Tuesday 10am — ignore for factual questions unless the packet mentions it. Summarize this incident in at most 8 words: 'Checkout API latency spiked to 4.2s during deploy deploy-4821.' Answer careful...	2	false	true	1. The reply shape is fixed by that instruction, not free-form prose. 2. Satisfy each listed constraint before returning the final formatted answer.	reliability	instruction_lock	instruction_lock	word_limit_summary	medium	FOLLOW_INSTRUCTION	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "medium", "dpo_negative_hint": "ignore format constraints and reply with an essay", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["IFE_instruction_error", "F5_off_topic_drift"], "family": "word_limit_summa...	word_limit_summary	word_limit_label	reliability	null
000011-e4d4df67a1863b21	[ { "content": "You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result.\n\nTool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read work...	<\|im_start\|>system You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result. Tool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read workspace...	5	true	true	1. “Someone on a forum asked: What are the date and location for MV Drift Fairchild” needs a verified fact. 2. That fact is not already in the user message. 3. A webSearch call should fetch the supporting line first. --- 1. The user asked: Someone on a forum asked: What are the date and location for MV Drift Fairchil...	reliability	open_world_knowledge	open_world_knowledge	search_vessel_mv_drift_fairchild_when	easy	EXECUTE_TOOL_THEN_ANSWER	context_chain	{"allowed_tools": ["webSearch", "readFile"], "complex_prompt_builtin": true, "difficulty": "easy", "entity_category": "vessel", "entity_name": "MV Drift Fairchild", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["E5_tool_bypass", "E2_retrieval_hallucination", "KE_knowledge_error"], "family...	null	tool_search_answer	reliability	null
000012-95beae5839ab52da	[ { "content": "DFA acceptance task.\n\nStart state: S0\nAccepting states: S0\nTransition table:\nstate \| a \| b\nS0 \| S1 \| S1\nS1 \| S2 \| S2\nS2 \| S2 \| S2\n\nInput symbols: b b b b b a\n\nQuestion: Is the input accepted?", "role": "user" }, { "content": "<think> goal: - simulate the DFA and compare the...	<\|im_start\|>user DFA acceptance task. Start state: S0 Accepting states: S0 Transition table: state \| a \| b S0 \| S1 \| S1 S1 \| S2 \| S2 S2 \| S2 \| S2 Input symbols: b b b b b a Question: Is the input accepted?<\|im_end\|> <\|im_start\|>assistant <think> goal: - simulate the DFA and compare the final state against the accept...	2	false	true	goal: - simulate the DFA and compare the final state against the accepting set subproblems: - simulate the full path exactly - only after the last symbol compare the ending state against the accepting set state_updates: - start in S0 - read b -> move S0 -> S1 - read b -> move S1 -> S2 - read b -> move S2 -> S2 - read b...	reasoning	state_machine	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000013-c13a609fa76e5ada	[ { "content": "You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestl...	<\|im_start\|>system You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestly t...	3	false	true	1. The question asks: For a trivia night — What's the origin of US10027235B2 as a structure. 2. This turn does not include a source passage or tool result. 3. A precise factual answer would require evidence that is not here.	reliability	open_world_knowledge	open_world_knowledge	abstain_patent_us10027235b2_design	easy	REFUSE_OUT_OF_SCOPE	context_chain	{"complex_prompt_builtin": true, "difficulty": "easy", "dpo_negative_hint": "invent a plausible biography or plot from memory", "entity_category": "patent", "entity_name": "US10027235B2", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["KM_knowledge_missing", "F8_overconfidence_sparse", "C2...	null	no_tools_abstain	reliability	null
000014-516ba5ad4f701990	[ { "content": "Stay inside what the user pasted — below.\n\nPasted author's note:\n\nAster Census uses \"Little Springs Parish University\" as a made-up university; the note says it is not real.\n\nQuestion: Wasn't the actual Little Springs Parish University chartered in the 1800s — which building is oldest?",...	<\|im_start\|>user Stay inside what the user pasted — below. Pasted author's note: Aster Census uses "Little Springs Parish University" as a made-up university; the note says it is not real. Question: Wasn't the actual Little Springs Parish University chartered in the 1800s — which building is oldest?<\|im_end\|> <\|im...	2	false	true	1. Wasn't the actual Little Springs Parish University chartered in the 1800s — which building is oldest sets the target conclusion. 2. Only the facts written in this turn can support the reply. 3. Line up those facts in the order the materials present them.	reliability	entity_grounding	entity_grounding	fiction_entity\|fiction_entity_correction\|little_springs_parish_university	easy	CORRECT_FALSE_PREMISE	context_chain	{"difficulty": "easy", "entity_lint_code": "ok", "entity_lint_ok": true, "entity_name": "Little Springs Parish University", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "family_id": "fiction_entity\|fiction_entity_correction\|little_springs_parish_university", "module": "entity_grounding", "primary_action": "C...	null	fiction_entity_correction	reliability	null
000015-cc92603389b80ca6	[ { "content": "Please review this Signal Forge procedure under the station control and manufacturing procedures manual.\n\nDevice: Signal Forge\n\nDocument environment: Manufacturing station procedure handbook\nSearch patterns supported here: keyword_search, section_lookup, precondition_scan.\nManual index:\n- O...	<\|im_start\|>user Please review this Signal Forge procedure under the station control and manufacturing procedures manual. Device: Signal Forge Document environment: Manufacturing station procedure handbook Search patterns supported here: keyword_search, section_lookup, precondition_scan. Manual index: - Opening the h...	2	false	true	1. The visible terms are 1, 2025, 2, 2. 2. Several rules could continue that pattern. 3. Without the generating rule, the next term is not forced to one value.	reliability	grounded_retrieval	grounded_retrieval	rulebook_device_control	medium	CITE_SECTIONS	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "medium", "doc_pattern": "manual", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["F2_misattribution", "F3_unfaithful_summary", "KE_knowledge_error"], "family": "device_control", "family_id": "rulebook_devi...	device_control	doc_search_citation	reliability	null
000016-aeafe0264cd1b848	[ { "content": "Pasting internal notes; a teammate may have mixed up figures.\n\nSlack (#support):\n@alex: Is the cap 1200 or 1500?\n@sam: check SW7 in the handbook before you answer the ticket.\n\nI need a careful handbook check for this incident plan.\n\nHandbook: Incident response handbook\n\nManual index:\n- ...	<\|im_start\|>user Pasting internal notes; a teammate may have mixed up figures. Slack (#support): @alex: Is the cap 1200 or 1500? @sam: check SW7 in the handbook before you answer the ticket. I need a careful handbook check for this incident plan. Handbook: Incident response handbook Manual index: - Safety risk inci...	2	false	true	1. The user needs the rule in section or. 2. Quote or paraphrase only what section or actually states.	reliability	grounded_retrieval	grounded_retrieval	rulebook_incident_response	hard	CITE_SECTIONS	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "hard", "doc_pattern": "manual", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["F2_misattribution", "F3_unfaithful_summary", "KE_knowledge_error"], "family": "incident_response", "family_id": "rulebook_inc...	incident_response	doc_search_citation	reliability	null
000017-c96fc8f06fb8a971	[ { "content": "Convert the sentence to passive voice: A manager should dismiss a customer.", "role": "user" }, { "content": "A customer should be dismissed by a manager.", "role": "assistant" } ]	<\|im_start\|>user Convert the sentence to passive voice: A manager should dismiss a customer.<\|im_end\|> <\|im_start\|>assistant A customer should be dismissed by a manager.<\|im_end\|>	2	false	false	null	rewrite	rewrite_voice	null	null	medium	null	null	{"difficulty": "medium"}	null	null	null	null
000018-01e0bed392cafe17	[ { "content": "You are an artificial intelligence.\nCurrent date: 2062/05/10", "role": "system" }, { "content": "I'm trying to remember the latest iOS version on 2139-08-17.", "role": "user" }, { "content": "<think>\nThe user is asking about a date well beyond my knowledge cutoff. I shoul...	<\|im_start\|>system You are an artificial intelligence. Current date: 2062/05/10<\|im_end\|> <\|im_start\|>user I'm trying to remember the latest iOS version on 2139-08-17.<\|im_end\|> <\|im_start\|>assistant <think> The user is asking about a date well beyond my knowledge cutoff. I should accept the question but say I can't be...	3	false	true	The user is asking about a date well beyond my knowledge cutoff. I should accept the question but say I can't be sure rather than guessing or rejecting the premise.	time_aware	time_aware	null	null	null	null	null	{"fact_category": "ios_versions", "fact_entity": "latest iOS version", "has_valid_fact": true, "query_date": "2139-08-17", "question_granularity": "full_date", "reference_date": "2016-07-20", "scenario": "far_future", "system_date": "2062-05-10", "temporal_relationship": "far_future", "track": "time_aware"}	null	null	time_aware	null
000019-bc8542cd87336876	[ { "content": "I'm writing a blog post — How does Study CARRICK-12F compare to SJ Archive?\n\n(I wasn't given excerpts, files, or search evidence to use.)", "role": "user" }, { "content": "<think>\n1. I'm writing a blog post — How does Study CARRICK-12F compare to SJ Archive asks for one exact value....	<\|im_start\|>user I'm writing a blog post — How does Study CARRICK-12F compare to SJ Archive? (I wasn't given excerpts, files, or search evidence to use.)<\|im_end\|> <\|im_start\|>assistant <think> 1. I'm writing a blog post — How does Study CARRICK-12F compare to SJ Archive asks for one exact value. 2. The stated lines d...	2	false	true	1. I'm writing a blog post — How does Study CARRICK-12F compare to SJ Archive asks for one exact value. 2. The stated lines do not pin every quantity that formula needs. 3. Naming what was never given is safer than inventing a number.	reliability	entity_grounding	entity_grounding	compare_study_carrick_12f_sj_archive\|compare_entities_abstain\|study_carrick_12f	medium	REFUSE_OUT_OF_SCOPE	context_chain	{"difficulty": "medium", "dpo_negative_hint": "invent a comparison from memory", "entity_category": "clinical_trial", "entity_lint_code": "ok", "entity_lint_ok": true, "entity_name": "Study CARRICK-12F", "entity_name_b": "SJ Archive", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "family_id": "compare_study_c...	null	compare_entities_abstain	reliability	null
000020-88f34f6e33ef9261	[ { "content": "You are a digital assistant.\n\nI'm trying to confirm something: who currently serves as the in charge at Alphabet?", "role": "user" }, { "content": "<think>\n1. I'm trying to confirm something: who currently serves as the in charge at Alphabet asks for one exact value.\n2. The stated ...	<\|im_start\|>user You are a digital assistant. I'm trying to confirm something: who currently serves as the in charge at Alphabet?<\|im_end\|> <\|im_start\|>assistant <think> 1. I'm trying to confirm something: who currently serves as the in charge at Alphabet asks for one exact value. 2. The stated lines do not pin every ...	2	false	true	1. I'm trying to confirm something: who currently serves as the in charge at Alphabet asks for one exact value. 2. The stated lines do not pin every quantity that formula needs. 3. Naming what was never given is safer than inventing a number.	reliability	temporal_boundary	temporal_boundary	fact_alphabet_ceos_ceo_of_alphabet	easy	REFUSE_MISSING_INFO	context_chain	{"boundary_entity": "CEO of Alphabet", "boundary_family": "fact_alphabet_ceos_ceo_of_alphabet", "boundary_kind": "fact_db", "complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "easy", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "fact_category": "argentine_presidents", "fact_en...	fact_alphabet_ceos_ceo_of_alphabet	null	reliability	null
000021-b538742b13980df2	[ { "content": "You are an AI assistant designed for factual responses.", "role": "system" }, { "content": "For a timeline I'm putting together, who was the NATO Secretary General today?", "role": "user" }, { "content": "<think>\nThis question is anchored to a phrase like 'last year' or 'n...	<\|im_start\|>system You are an AI assistant designed for factual responses.<\|im_end\|> <\|im_start\|>user For a timeline I'm putting together, who was the NATO Secretary General today?<\|im_end\|> <\|im_start\|>assistant <think> This question is anchored to a phrase like 'last year' or 'next month', but no current date was pro...	5	false	true	This question is anchored to a phrase like 'last year' or 'next month', but no current date was provided. I should ask the user what date I should count from. --- The user asked about March 2017, but my cutoff is March 13, 2017, so I only know part of that month. I should answer conservatively rather than treat the w...	time_aware	time_aware	null	null	null	null	null	{"fact_category": "brazilian_presidents", "fact_entity": "President of Brazil", "has_valid_fact": true, "query_date": "2017-03-13", "reference_date": "2017-03-13", "scenario": "no_system_date_ask_date", "temporal_relationship": "present", "track": "time_aware"}	null	null	time_aware	null
000022-226a2fb92c38aca8	[ { "content": "Rephrase more concise: The manager sends the contract.", "role": "user" }, { "content": "Manager sends contract.", "role": "assistant" } ]	<\|im_start\|>user Rephrase more concise: The manager sends the contract.<\|im_end\|> <\|im_start\|>assistant Manager sends contract.<\|im_end\|>	2	false	false	null	rewrite	rewrite_paraphrase	null	null	medium	null	null	{"difficulty": "medium"}	null	null	null	null
000023-86c6ac795d1357ac	[ { "content": "Following up on a ticket—I've included the thread and docs we actually have.\n\nAnswer carefully.\n\nIs 91 divisible by 7? Answer with only Yes or No.", "role": "user" }, { "content": "<think>\n1. The yes/no question is: Is 91 divisible by 7.\n2. Check the stated fact against what the ...	<\|im_start\|>user Following up on a ticket—I've included the thread and docs we actually have. Answer carefully. Is 91 divisible by 7? Answer with only Yes or No.<\|im_end\|> <\|im_start\|>assistant <think> 1. The yes/no question is: Is 91 divisible by 7. 2. Check the stated fact against what the question asks. 3. Reply w...	2	false	true	1. The yes/no question is: Is 91 divisible by 7. 2. Check the stated fact against what the question asks. 3. Reply with only Yes or No as instructed.	reliability	instruction_lock	instruction_lock	yes_no_divisible	medium	FOLLOW_INSTRUCTION	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "medium", "dpo_negative_hint": "ignore format constraints and reply with an essay", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["IFE_instruction_error", "F5_off_topic_drift"], "family": "yes_no_divisible...	yes_no_divisible	yes_no_only	reliability	null
000024-65f97c463fb03864	[ { "content": "DFA acceptance task.\n\nStart state: S2\nAccepting states: S0, S2\nTransition table:\nstate \| 0 \| 1 \| x\nS0 \| S0 \| S2 \| S2\nS1 \| S1 \| S1 \| S0\nS2 \| S2 \| S1 \| S0\nS3 \| S0 \| S3 \| S3\n\nInput symbols: 0 1 x 1 0 1 1 x 0 1\n\nQuestion: Is the input accepted?", "role": "user" }, { "content":...	<\|im_start\|>user DFA acceptance task. Start state: S2 Accepting states: S0, S2 Transition table: state \| 0 \| 1 \| x S0 \| S0 \| S2 \| S2 S1 \| S1 \| S1 \| S0 S2 \| S2 \| S1 \| S0 S3 \| S0 \| S3 \| S3 Input symbols: 0 1 x 1 0 1 1 x 0 1 Question: Is the input accepted?<\|im_end\|> <\|im_start\|>assistant <think> goal: - simulate the D...	2	false	true	goal: - simulate the DFA and compare the final state against the accepting set subproblems: - simulate the full path exactly - only after the last symbol compare the ending state against the accepting set state_updates: - start in S2 - read 0 -> move S2 -> S2 - read 1 -> move S2 -> S1 - read x -> move S1 -> S0 - read 1...	reasoning	state_machine	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000025-c267bbdf3394d56e	[ { "content": "This is an exact symbolic execution problem.\nState-visit counting task.\n\nStart state: S4\nTransition table:\nstate \| a \| b \| c\nS0 \| S5 \| S1 \| S2\nS1 \| S4 \| S0 \| S5\nS2 \| S2 \| S6 \| S3\nS3 \| S6 \| S2 \| S3\nS4 \| S0 \| S6 \| S6\nS5 \| S6 \| S4 \| S2\nS6 \| S6 \| S6 \| S4\n\nInput symbols: b b b b c b a b\n...	<\|im_start\|>user This is an exact symbolic execution problem. State-visit counting task. Start state: S4 Transition table: state \| a \| b \| c S0 \| S5 \| S1 \| S2 S1 \| S4 \| S0 \| S5 S2 \| S2 \| S6 \| S3 S3 \| S6 \| S2 \| S3 S4 \| S0 \| S6 \| S6 S5 \| S6 \| S4 \| S2 S6 \| S6 \| S6 \| S4 Input symbols: b b b b c b a b Question: How many ...	2	false	true	goal: - count every visit to S4, including the start if applicable subproblems: - simulate the path exactly - count each occurrence of S4 in that path state_updates: - start in S4 - read b -> move S4 -> S6 - read b -> move S6 -> S6 - read b -> move S6 -> S6 - read b -> move S6 -> S6 - read c -> move S6 -> S4 - read b -...	reasoning	state_machine	null	null	medium	null	null	{"difficulty": "medium"}	null	null	null	null
000026-92712fb72431bb2d	[ { "content": "If the setup is incomplete or inconsistent, say that directly: Find two U.S. coins summing to 30 cents where the statement says one coin is not a nickel.", "role": "user" }, { "content": "<think>\n1. The claim to evaluate is: If the setup is incomplete or inconsistent, say that directl...	<\|im_start\|>user If the setup is incomplete or inconsistent, say that directly: Find two U.S. coins summing to 30 cents where the statement says one coin is not a nickel.<\|im_end\|> <\|im_start\|>assistant <think> 1. The claim to evaluate is: If the setup is incomplete or inconsistent, say that directly: Find two U.S. coi...	2	false	true	1. The claim to evaluate is: If the setup is incomplete or inconsistent, say that directly: Find two U.S. coins summing to 30 cents where the stateme. 2. Check whether the stated rule matches how the terms are actually defined. 3. If a counterexample or category mistake appears, the strict answer follows from that.	reliability	epistemic_core	epistemic_core	coin_semantics	easy	ANSWER_EXACT	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "easy", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": [], "family": "coin_semantics", "family_id": "coin_semantics", "module": "epistemic_core", "outcome_type": "exact", "preserve_think": true, "primary_act...	coin_semantics	null	reliability	null
000027-645c098d7fdd8c8c	[ { "content": "Manufacturing QA log is noisy; ignore unrelated sensor alerts unless the question asks.\n\nI'm checking a sterilization equipment and clinical procedures run for Cinder Loom.\n\nDevice: Cinder Loom\n\nDocument environment: Sterilization unit procedure handbook\nSearch patterns supported here: keyw...	<\|im_start\|>user Manufacturing QA log is noisy; ignore unrelated sensor alerts unless the question asks. I'm checking a sterilization equipment and clinical procedures run for Cinder Loom. Device: Cinder Loom Document environment: Sterilization unit procedure handbook Search patterns supported here: keyword_search, ...	2	false	true	1. The visible terms are 1, 2025, 2, 3. 2. Several rules could continue that pattern. 3. Without the generating rule, the next term is not forced to one value.	reliability	grounded_retrieval	grounded_retrieval	rulebook_device_control	medium	CITE_SECTIONS	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "medium", "doc_pattern": "manual", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["F2_misattribution", "F3_unfaithful_summary", "KE_knowledge_error"], "family": "device_control", "family_id": "rulebook_devi...	device_control	doc_search_citation	reliability	null
000028-490d5bc927296ded	[ { "content": "DFA acceptance task.\n\nStart state: S3\nAccepting states: S2, S3\nTransition table:\nstate \| 0 \| 1 \| x\nS0 \| S0 \| S1 \| S2\nS1 \| S0 \| S2 \| S1\nS2 \| S0 \| S2 \| S3\nS3 \| S2 \| S1 \| S3\n\nInput symbols: x 1 1 1 0 x\n\nQuestion: Is the input accepted?", "role": "user" }, { "content": "<think...	<\|im_start\|>user DFA acceptance task. Start state: S3 Accepting states: S2, S3 Transition table: state \| 0 \| 1 \| x S0 \| S0 \| S1 \| S2 S1 \| S0 \| S2 \| S1 S2 \| S0 \| S2 \| S3 S3 \| S2 \| S1 \| S3 Input symbols: x 1 1 1 0 x Question: Is the input accepted?<\|im_end\|> <\|im_start\|>assistant <think> goal: - simulate the DFA and c...	2	false	true	goal: - simulate the DFA and compare the final state against the accepting set subproblems: - simulate the full path exactly - only after the last symbol compare the ending state against the accepting set state_updates: - start in S3 - read x -> move S3 -> S3 - read 1 -> move S3 -> S1 - read 1 -> move S1 -> S2 - read 1...	reasoning	state_machine	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000029-a8c8a6c15ed30813	[ { "content": "You are a helpful assistant.\n\nTool contract: issue tool calls with explicit custom `<CALL ...>` tags. Registry purpose: Search indexed corpus and read workspace files for grounded answers.. The available tool names and argument names below are the only valid ones for this sample.\n\nTool: webSea...	<\|im_start\|>system You are a helpful assistant. Tool contract: issue tool calls with explicit custom `<CALL ...>` tags. Registry purpose: Search indexed corpus and read workspace files for grounded answers.. The available tool names and argument names below are the only valid ones for this sample. Tool: webSearch D...	5	true	true	The safe sequence is webSearch first, inspect the result, then answer only from that text. --- 1. “Podcast research — Who founded Series Drift-9286 and when” needs a verified fact. 2. That fact is not already in the user message. 3. A webSearch call should fetch the supporting line first.	reliability	tool_epistemics	tool_epistemics	proc_product_series_drift_9286_founding	easy	EXECUTE_TOOL_THEN_ANSWER	context_chain	{"allowed_tools": ["webSearch", "readFile"], "complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "easy", "dpo_negative_hint": "answer from memory without tool call", "entity_category": "product", "entity_name": "Series Drift-9286", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "...	proc_product_series_drift_9286_founding	search_then_answer	reliability	null
000030-4bd8f301bcf9e387	[ { "content": "Compute the value of -8 + -12 − 5 step by step.", "role": "user" }, { "content": "<think>\n1. The expression is the value of -8 + -12 − 5 step by step.\n2. Multiplication comes before addition or subtraction here.\n3. Reduce the operations in that order until one integer remains.\n</th...	<\|im_start\|>user Compute the value of -8 + -12 − 5 step by step.<\|im_end\|> <\|im_start\|>assistant <think> 1. The expression is the value of -8 + -12 − 5 step by step. 2. Multiplication comes before addition or subtraction here. 3. Reduce the operations in that order until one integer remains. </think> 1. Evaluate left t...	2	false	true	1. The expression is the value of -8 + -12 − 5 step by step. 2. Multiplication comes before addition or subtraction here. 3. Reduce the operations in that order until one integer remains.	math	math_expr	numeric_integrity	expr_chain_-8_-12_5	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": -25, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "expr_chain_-8_-12_5", "family_id": "expr_chain_-8_-12_5", "module": "numeric_integrity", "nu...	expr_chain_-8_-12_5	expression_eval	reliability	null
000031-4a572b2e92b21980	[ { "content": "Thread excerpt:\n- Ava signs messages as 'T'.\n- Jade signs messages as 'Q'.\n- Kai writes: \"T told me that Q said 'Owen moved the shipment after Nora approved it.'\"\n- Ava replies: \"I only said that Jade believed Owen had moved it. I did not say any approval was confirmed.\"\n- Jade replies: \...	<\|im_start\|>user Thread excerpt: - Ava signs messages as 'T'. - Jade signs messages as 'Q'. - Kai writes: "T told me that Q said 'Owen moved the shipment after Nora approved it.'" - Ava replies: "I only said that Jade believed Owen had moved it. I did not say any approval was confirmed." - Jade replies: "I never said I...	2	false	true	goal: - separate alias identity, reported belief, and event truth across a repaired multi-message chain subproblems: - identify the candidate referents, groups, or readings that are still live after the first pass - use later clauses or repairs to eliminate unsupported readings one by one - state the strongest exact co...	reasoning	reference_resolution	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000032-72b5c88e63976d53	[ { "content": "Please review this lab procedure distribution and confirmation tracking ledger carefully.\n\nHandbook: Procedure distribution handbook\n\nManual index:\n- Hold blocks knowledge transfer [KL1]\n- Publishing the bulletin [KL2]\n- Opening the portal bulletin [KL3]\n- Direct briefing transfer [KL4]\n-...	<\|im_start\|>user Please review this lab procedure distribution and confirmation tracking ledger carefully. Handbook: Procedure distribution handbook Manual index: - Hold blocks knowledge transfer [KL1] - Publishing the bulletin [KL2] - Opening the portal bulletin [KL3] - Direct briefing transfer [KL4] - Forwarded com...	2	false	true	goal: - determine the resulting state by following the stated workflow or simulator steps exactly constraint_focus: - use only the declared workflow state, handbook order, and tool contract for this sample chosen_method: - Replay the listed steps from the stated state. pitfall_to_avoid: - Infer the result without repla...	reasoning	rulebook_world	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000033-e6ad8e127487f7e2	[ { "content": "You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result.\n\nTool contract: issue tool calls with explicit custom `<CALL ...>` tags. Registry purpose: Search indexed corpus and read workspace files for grou...	<\|im_start\|>system You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result. Tool contract: issue tool calls with explicit custom `<CALL ...>` tags. Registry purpose: Search indexed corpus and read workspace files for grounded ...	5	true	true	1. “For a trivia night — Name all subsidiaries under The Patrol Desk.” needs a verified fact. 2. That fact is not already in the user message. 3. A webSearch call should fetch the supporting line first. --- 1. webSearch returned no line that answers “For a trivia night — Name all subsidiaries under The Patrol Desk.”....	reliability	open_world_knowledge	open_world_knowledge	empty_podcast_the_patrol_desk	easy	REFUSE_MISSING_INFO	context_chain	{"allowed_tools": ["webSearch", "readFile"], "complex_prompt_builtin": true, "difficulty": "easy", "entity_name": "The Patrol Desk", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expect_empty_result": true, "failure_mode_ids": ["E2_retrieval_hallucination", "KM_knowledge_missing", "F8_overconfidence_sparse"]...	null	tool_empty_refuse	reliability	null
000034-6561130bc63757e7	[ { "content": "I'm checking whether this is actually answerable: Average rating: 91. What is the overall total?", "role": "user" }, { "content": "<think>\n1. An average is stated without the full set of values behind it.\n2. Means need either every value or the count of values.\n</think>\nIt cannot b...	<\|im_start\|>user I'm checking whether this is actually answerable: Average rating: 91. What is the overall total?<\|im_end\|> <\|im_start\|>assistant <think> 1. An average is stated without the full set of values behind it. 2. Means need either every value or the count of values. </think> It cannot be determined exactly fr...	2	false	true	1. An average is stated without the full set of values behind it. 2. Means need either every value or the count of values.	reliability	epistemic_core	epistemic_core	average_missing_denominator	easy	REFUSE_MISSING_INFO	context_chain	{"complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "easy", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["KM_knowledge_missing", "F8_overconfidence_sparse"], "family": "average_missing_denominator", "family_id": "average_missing_denominator", "module": "ep...	average_missing_denominator	null	reliability	null
000035-bd69308e00407ce7	[ { "content": "A token code is valid only when its remainder modulo 7 is 3.\nThe token 52 has remainder 3 modulo 7.\n\nIs token 52 valid?", "role": "user" }, { "content": "<think>\n1. The latest user message pushes: “Is token 52 valid?”.\n2. Earlier turns or materials may include a specific anchored ...	<\|im_start\|>user A token code is valid only when its remainder modulo 7 is 3. The token 52 has remainder 3 modulo 7. Is token 52 valid?<\|im_end\|> <\|im_start\|>assistant <think> 1. The latest user message pushes: “Is token 52 valid?”. 2. Earlier turns or materials may include a specific anchored fact. 3. The answer shou...	10	false	true	1. The latest user message pushes: “Is token 52 valid?”. 2. Earlier turns or materials may include a specific anchored fact. 3. The answer should follow that anchored fact, not agree for politeness. --- 1. The latest user message pushes: “Doesn't that show you were wrong?”. 2. Earlier turns or materials may include a...	reliability	logic	sycophancy_pressure	logic_modulo	hard	HOLD_UNDER_REBUTTAL	context_chain	{"anchor_fact": "token 52 is valid", "cluster_id": "truth_decay_00047bc7", "cluster_type": "truth_decay", "complex_prompt_applied": true, "complex_prompt_layout": "enriched", "difficulty": "hard", "domain": "logic", "dpo_negative_hint": "apologize and adopt the user's false correction", "epistemic_lint_code": "ok", "ep...	logic_modulo	truth_decay	reliability	null
000036-286ea9b04f05ebae	[ { "content": "You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestl...	<\|im_start\|>system You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestly t...	3	false	true	1. The question asks: (Nothing was pasted and no tool output is available here.). 2. This turn does not include a source passage or tool result. 3. A precise factual answer would require evidence that is not here.	reliability	open_world_knowledge	open_world_knowledge	abstain_character_pilot_rosa_osgood_tell_about	hard	REFUSE_OUT_OF_SCOPE	context_chain	{"complex_prompt_builtin": true, "difficulty": "hard", "dpo_negative_hint": "invent a plausible biography or plot from memory", "entity_category": "character", "entity_name": "Pilot Rosa Osgood", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["KM_knowledge_missing", "F8_overconfidence_spar...	null	no_tools_abstain	reliability	null
000037-bd6e68d74c12b1b1	[ { "content": "A warehouse shelf starts with 105 boxes. A pickup removes 32 boxes and a restock adds 22 boxes. How many boxes are on the shelf now? One sentence with the integer.", "role": "user" }, { "content": "<think>\n1. The shelf starts with 105 boxes, then loses 32 in a pickup.\n2. After the pi...	<\|im_start\|>user A warehouse shelf starts with 105 boxes. A pickup removes 32 boxes and a restock adds 22 boxes. How many boxes are on the shelf now? One sentence with the integer.<\|im_end\|> <\|im_start\|>assistant <think> 1. The shelf starts with 105 boxes, then loses 32 in a pickup. 2. After the pickup that is 105 − 32...	2	false	true	1. The shelf starts with 105 boxes, then loses 32 in a pickup. 2. After the pickup that is 105 − 32 = 73 boxes. 3. A restock adds 22 to those 73 boxes for the final on-shelf count.	math	math_word	numeric_integrity	word_inventory_105_32_22	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 95, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "word_inventory_105_32_22", "family_id": "word_inventory_105_32_22", "module": "numeric_integr...	word_inventory_105_32_22	word_problem_exact	reliability	null
000038-a610defe777588b8	[ { "content": "Fact-check desk (legal_discovery)—two sources may conflict; say which is authoritative.\n\nYou may cite ONLY the sources listed in the bibliography below. Do not invent DOIs, titles, or authors. If nothing matches, say you cannot verify a match in the provided bibliography.\n\n[1] Scalable Halluci...	<\|im_start\|>user Fact-check desk (legal_discovery)—two sources may conflict; say which is authoritative. You may cite ONLY the sources listed in the bibliography below. Do not invent DOIs, titles, or authors. If nothing matches, say you cannot verify a match in the provided bibliography. [1] Scalable Hallucination Au...	2	false	true	1. The user supplied a bounded bibliography or excerpt block. 2. 3.	reliability	grounded_citation	grounded_citation	cite_from_allowlist	hard	ANSWER_EXACT	context_chain	{"allowed_dois": ["10.5555/reli.2022.002", "10.5555/reli.2022.009", "10.5555/reli.2023.001", "10.5555/reli.2023.003", "10.5555/reli.2023.006", "10.5555/reli.2023.010", "10.5555/reli.2024.004", "10.5555/reli.2024.007", "10.5555/reli.2024.008", "10.5555/reli.2025.005"], "citation_lint_code": "ok", "citation_lint_ok": tru...	cite_from_allowlist	cite_from_allowlist	reliability	null
000039-de64189931a711f7	[ { "content": "Trying to settle this without guessing.\nSuppose both of us are teachers. What are we?", "role": "user" }, { "content": "<think>\n1. Suppose both of us are teachers. What are we sets the target conclusion.\n2. Only the facts written in this turn can support the reply.\n3. Line up those...	<\|im_start\|>user Trying to settle this without guessing. Suppose both of us are teachers. What are we?<\|im_end\|> <\|im_start\|>assistant <think> 1. Suppose both of us are teachers. What are we sets the target conclusion. 2. Only the facts written in this turn can support the reply. 3. Line up those facts in the order the...	2	false	true	1. Suppose both of us are teachers. What are we sets the target conclusion. 2. Only the facts written in this turn can support the reply. 3. Line up those facts in the order the materials present them.	reasoning	hidden_assumptions	null	null	easy	null	null	{"difficulty": "easy"}	null	null	null	null
000040-3a562c619d1e6d46	[ { "content": "From the whiteboard:\n\nA bakery bakes 114 muffins, sells 19, then bakes 18 more. How many muffins are available now? One sentence with the integer.", "role": "user" }, { "content": "<think>\n1. The bakery starts with 114 muffins and sells 19.\n2. That leaves 114 − 19 = 95 muffins stil...	<\|im_start\|>user From the whiteboard: A bakery bakes 114 muffins, sells 19, then bakes 18 more. How many muffins are available now? One sentence with the integer.<\|im_end\|> <\|im_start\|>assistant <think> 1. The bakery starts with 114 muffins and sells 19. 2. That leaves 114 − 19 = 95 muffins still available. 3. A secon...	2	false	true	1. The bakery starts with 114 muffins and sells 19. 2. That leaves 114 − 19 = 95 muffins still available. 3. A second batch of 18 muffins is added to that remaining amount.	math	math_word	numeric_integrity	word_bakery_114_19_18	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 113, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "word_bakery_114_19_18", "family_id": "word_bakery_114_19_18", "module": "numeric_integrity",...	word_bakery_114_19_18	word_problem_exact	reliability	null
000041-1fafb76fcc751482	[ { "content": "What does -9 × 6 + 3 equal? Work through the order of operations.", "role": "user" }, { "content": "<think>\n1. This is 9 × 6.\n2. Multiply the two factors directly.\n3. The product is one integer once the multiplication is finished.\n</think>\nMultiply first: -9 × 6 = -54. Then add 3 ...	<\|im_start\|>user What does -9 × 6 + 3 equal? Work through the order of operations.<\|im_end\|> <\|im_start\|>assistant <think> 1. This is 9 × 6. 2. Multiply the two factors directly. 3. The product is one integer once the multiplication is finished. </think> Multiply first: -9 × 6 = -54. Then add 3 to get -51. The expressi...	2	false	true	1. This is 9 × 6. 2. Multiply the two factors directly. 3. The product is one integer once the multiplication is finished.	math	math_expr	numeric_integrity	expr_mul_add_-9_6_3	easy	ANSWER_EXACT	context_chain	{"difficulty": "easy", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": -51, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "expr_mul_add_-9_6_3", "family_id": "expr_mul_add_-9_6_3", "module": "numeric_integrity", "nume...	expr_mul_add_-9_6_3	expression_eval	reliability	null
000042-b6ce644a732a37d6	[ { "content": "Treat the following as the entire evidence set.\n\nMemo excerpt (local archive archive):\n\nAward GR-2012-212 funds the 2020 Field symposium (2012). PI Dr. Ulrich; approved amount $1,588,800.\nThe abstract does not list co-investigators, subcontractors, or final reports.\n\nQuestion: Based on the ...	<\|im_start\|>user Treat the following as the entire evidence set. Memo excerpt (local archive archive): Award GR-2012-212 funds the 2020 Field symposium (2012). PI Dr. Ulrich; approved amount $1,588,800. The abstract does not list co-investigators, subcontractors, or final reports. Question: Based on the excerpt, his...	2	false	true	1. The stated lines do not pin every quantity that formula needs. 2. Naming what was never given is safer than inventing a number.	reliability	entity_grounding	entity_grounding	partial_snippet\|partial_snippet_extrapolation_refuse\|the_2020_field_symposium	medium	REFUSE_MISSING_INFO	context_chain	{"asked_extrapolation": "historical funding sources", "difficulty": "medium", "entity_category": "event", "entity_lint_code": "ok", "entity_lint_ok": true, "entity_name": "the 2020 Field symposium", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "family_id": "partial_snippet\|partial_snippet_extrapolation_refus...	null	partial_snippet_extrapolation_refuse	reliability	null
000043-62cae3f71a02e8f0	[ { "content": "Can you sanity-check this?\n\nA team has a $647 budget. They spend $418 and later receive a $65 refund. How many dollars remain? One sentence with the integer.", "role": "user" }, { "content": "<think>\n1. Starting budget is $647 and $418 is spent first.\n2. After spending that leaves ...	<\|im_start\|>user Can you sanity-check this? A team has a $647 budget. They spend $418 and later receive a $65 refund. How many dollars remain? One sentence with the integer.<\|im_end\|> <\|im_start\|>assistant <think> 1. Starting budget is $647 and $418 is spent first. 2. After spending that leaves $647 − $418 = $229. 3. ...	2	false	true	1. Starting budget is $647 and $418 is spent first. 2. After spending that leaves $647 − $418 = $229. 3. A $65 refund is added to $229 for the remaining dollars.	math	math_word	numeric_integrity	word_budget_647_418_65	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 294, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "word_budget_647_418_65", "family_id": "word_budget_647_418_65", "module": "numeric_integrity...	word_budget_647_418_65	word_problem_exact	reliability	null
000044-455db0671a49e912	[ { "content": "You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestl...	<\|im_start\|>system You are a helpful assistant. You do not have web search, file access, or other tools in this chat. Answer only when the fact is obvious, certain, or clearly inferable from what the user said. For specific people, places, media, statistics, or events you cannot verify from memory alone, say honestly t...	3	false	true	1. The stated lines do not pin every quantity that formula needs. 2. Naming what was never given is safer than inventing a number.	reliability	open_world_knowledge	open_world_knowledge	ow_compare_pilot_alden_pruitt_northcott_sparrow_iverson_2025_v1	medium	REFUSE_OUT_OF_SCOPE	context_chain	{"complex_prompt_builtin": true, "difficulty": "medium", "entity_category": "character", "entity_name": "Pilot Alden Pruitt-Northcott", "entity_name_b": "sparrow_iverson_2025_v1", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["KM_knowledge_missing", "F8_overconfidence_sparse", "C2_confabu...	null	no_tools_compare_abstain	reliability	null
000045-9570334b9fe29355	[ { "content": "You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result.\n\nTool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read work...	<\|im_start\|>system You are a helpful assistant. Use tools when you need time-sensitive or specific facts. Do not invent details that are not supported by a tool result. Tool contract: issue tool calls as JSON objects with top-level `tool` and `arguments` keys. Registry purpose: Search indexed corpus and read workspace...	5	true	true	1. “Random one for you: Who designed Landing Watchtower and when was it built” needs a verified fact. 2. That fact is not already in the user message. 3. A webSearch call should fetch the supporting line first. --- 1. The user asked: Random one for you: Who designed Landing Watchtower and when was it built. 2. 3. The...	reliability	open_world_knowledge	open_world_knowledge	search_landmark_landing_watchtower_design	medium	EXECUTE_TOOL_THEN_ANSWER	context_chain	{"allowed_tools": ["webSearch", "readFile"], "complex_prompt_builtin": true, "difficulty": "medium", "entity_category": "landmark", "entity_name": "Landing Watchtower", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "failure_mode_ids": ["E5_tool_bypass", "E2_retrieval_hallucination", "KE_knowledge_error"], "fa...	null	tool_search_answer	reliability	null
000046-691bd957f76e3798	[ { "content": "Compute the value of -7 × -14 + 1 step by step.", "role": "user" }, { "content": "<think>\n1. The expression is the value of -7 × -14 + 1 step by step.\n2. Multiplication comes before addition or subtraction here.\n3. Reduce the operations in that order until one integer remains.\n</th...	<\|im_start\|>user Compute the value of -7 × -14 + 1 step by step.<\|im_end\|> <\|im_start\|>assistant <think> 1. The expression is the value of -7 × -14 + 1 step by step. 2. Multiplication comes before addition or subtraction here. 3. Reduce the operations in that order until one integer remains. </think> 1. Multiply first:...	2	false	true	1. The expression is the value of -7 × -14 + 1 step by step. 2. Multiplication comes before addition or subtraction here. 3. Reduce the operations in that order until one integer remains.	math	math_expr	numeric_integrity	expr_mul_add_-7_-14_1	medium	ANSWER_EXACT	context_chain	{"difficulty": "medium", "dpo_negative_hint": "state a nearby wrong integer confidently", "epistemic_lint_code": "ok", "epistemic_lint_ok": true, "expected_value": 99, "failure_mode_ids": ["F_num_arithmetic_drift"], "family": "expr_mul_add_-7_-14_1", "family_id": "expr_mul_add_-7_-14_1", "module": "numeric_integrity", ...	expr_mul_add_-7_-14_1	expression_eval	reliability	null

End of preview. Expand in Data Studio

microagent-sft-v1

6,884 procedurally generated SFT examples for training small reasoning agents. Built for models that need to think before answering, use tools honestly, and refuse when evidence is missing—without memorizing frontier-scale code dumps.

What’s in the mix

Track	Examples	What it teaches
Reliability & agentic basics	3,964	Anti-hallucination, grounded retrieval/citation, tool epistemics, instruction lock, sycophancy resistance, entity grounding
Logic	1,540	State machines, rulebooks, hidden assumptions, constraint grids, counterfactuals
Math	680	Linear equations, expression eval, and word problems (~26%) with visible steps when requested
Rewrite	240	Voice conversion & concision
Time-aware	200	Temporal boundaries & date-sensitive answers
Religion / scripture agentic	200	Bible study, tool-backed search, memory-boundary refusals
Easy code	60	Compact synthetic-language tasks (engine + modes)

Quality highlights (automated audit, seed 99):

Zero exact chatml duplicates
No single think-opener phrase above 2% of the dataset
1,541 tool workflows with mixed surface formats (json, xml, markdown, shell, line, tags, yamlish—not XML-only)
6,404 samples with <think> reasoning blocks
Math track includes 178 word-problem rows (inventory, queue, budget, distance, etc.)

Generation seed: 99 · Difficulty: medium · Builder: generators/build_tiny_model_bundle.py · Run: run-20260616-123408

Message format

Each row has a messages column: an ordered list of {role, content} dicts compatible with OpenAI-style trainers.

from datasets import load_dataset

ds = load_dataset("hudsongouge/microagent-sft-v1", split="train")
row = ds[0]
print(row["messages"])

Roles: system, user, assistant, tool

Reasoning: Thinking stays inside assistant content as <think>…</think>. A separate thinking column concatenates extracted think text for analysis or masking.

Tools: Tool calls use varied surface syntax per sample (JSON objects, XML tags, markdown fences, shell-style flags, line records, etc.). System prompts declare the contract for that row. Multi-turn clusters preserve full conversation history.

ChatML: Original chatml string is included for trainers that consume ChatML directly (<|im_start|>role\n…<|im_end|>).

Columns

Column	Description
`messages`	Training conversation (primary field)
`chatml`	Source ChatML serialization
`thinking`	Extracted think text (joined across assistant turns)
`has_tools` / `has_thinking`	Boolean filters
`bundle`, `module`, `family_id`, `difficulty`	Track metadata for stratified sampling
`metadata_json`	Full generator metadata as JSON string (includes `tool_trace_format` when tools are used)

Loading

from datasets import load_dataset

ds = load_dataset("hudsongouge/microagent-sft-v1", split="train")

tools = ds.filter(lambda x: x["has_tools"])
math = ds.filter(lambda x: x["bundle"] == "math")
word_math = math.filter(lambda x: x.get("domain") == "math_word")

import pandas as pd
df = pd.read_parquet("hf://datasets/hudsongouge/microagent-sft-v1/train.parquet")

Intended use

Supervised fine-tuning of small models for:

Epistemic humility (abstain / qualify instead of confabulate)
Faithful chain-of-thought aligned with final answers
Tool-calling discipline (call before claiming; respect empty results)
Structured reasoning over rules, state, and documents

Not intended as a benchmark or for training models to quote scripture from memory without sources.

Citation

If you use this dataset, please cite the Procedural-Data project and note version microagent-sft-v1.

Downloads last month: 101