[ { "task_id": 1, "intent": "Is biophysics mentioned as a related page or branch in the Biology article, the Physics article, both, or neither of them?", "intent_template_id": 1, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "biology" ], "must_exclude": [ "both", "neither", "^[\\s\\S]*\\bnot\\b (in |mentioned in |listed in |on |under )?(the )?biology\\b[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "Biology" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap-4 multiple choice (Biology / Physics / both / neither). must_include the correct option and exclude only the literal choice words 'both'/'neither' -- bare 'physics' cannot be excluded because a correct answer routinely says 'in Biology, not in Physics'. The third exclusion is the mirror near-miss where the negation directly governs 'biology' ('..., not (in/mentioned in/listed in/on/under) (the) Biology'): it catches a wrong answer that says biophysics is NOT in the Biology article, while 'It is not Physics but Biology' still passes because the negation there governs Physics. (Earlier this regex carried a '(?! article'?s)' lookahead that opened a false-accept hole: 'not in the Biology article's related pages' -- a wrong answer -- escaped the exclusion; removed.) scope first_sentence stops a 'Physics.' verdict whose justification later names the Biology article.", "source": "batch_14.json" } } }, { "task_id": 2, "intent": "What is the population difference between the two special administrative regions of China?", "intent_template_id": 2, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 7000000, "rel_tolerance": 0.1 }, "fuzzy_match": "Around 7 million" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Hedged gold ('Around 7 million'); the real difference (Hong Kong ~7.4M minus Macau ~0.68M ~= 6.7M) sits inside a 10% band around 7M, so rel_tolerance 0.1 accepts every plausible correct figure while rejecting 8M / 700k answers. Medium because an answer that reports Hong Kong's raw population (~7.4M) instead of the difference also lands in the band.", "source": "batch_14.json" } } }, { "task_id": 3, "intent": "List all specific animal names found in the 'Natural History' section of the New Zealand page and the 'Indigenous life' section of the Australia page.", "intent_template_id": 3, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "koalas |OR| koala" ], [ "kangaroos |OR| kangaroo" ], [ "wombats |OR| wombat" ], [ "numbats |OR| numbat" ], [ "platypus |OR| platypuses |OR| platypi" ], [ "echidna |OR| short-beaked echidna" ], [ "bats |OR| bat" ], [ "seals |OR| seal" ], [ "dolphins |OR| dolphin" ], [ "whales |OR| whale" ], [ "kiwi |OR| kiwis" ], [ "kakapo |OR| kakapos" ], [ "tuatara |OR| tuataras" ], [ "moa |OR| moas" ], [ "rats |OR| rat" ], [ "dogs |OR| dog" ], [ "cats |OR| cat" ], [ "possums |OR| possum" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "Koalas, kangaroos, wombats, numbats, platypus, short-beaked echidna, bats, seals, dolphins, whales, kiwi, kakapo, tuatara, Moa, rats, dogs, cats, possums." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The intent asks for *all* animal names in two sections, so the set semantics of list_match are exactly right and every one of the 17 gold animals is required. Each item carries singular/plural alternatives because agents freely switch number when re-listing ('koala' vs 'koalas'). 'kakapo' needs no macron variant -- fold_unicode strips the diacritics of 'kakapo' before matching. ordered:false because the intent imposes no order; an agent may report the New Zealand section before the Australia one. Word-boundary matching keeps 'bat' from matching inside 'wombat' and 'cat' from matching inside 'cats' incorrectly. | SPOT-CHECK: gold list omits the short-beaked echidna (named beside platypus). Verifier tolerates extra items, so a correct fuller answer still passes; human should complete the gold. | FIXED (env-verified): added 'short-beaked echidna', named beside platypus in Australia's 'Indigenous life' section. All other gold animals confirmed across the two sections. | NOTE (designer intent): the HF CSV gold for this task adds \"(It is fine if 1/2 animals are missing)\", i.e. some leniency was intended. The deterministic list_match here is kept STRICT (all animals required) per review decision; revisit if it proves too harsh.", "source": "batch_13.json" } } }, { "task_id": 4, "intent": "Go to Black Sabbath's article and click the second hyperlink in the article's content. I need the population of the country that the link leads to. If it's not a country, just go to the first country mentioned on that visited page and give me that country's population.", "intent_template_id": 4, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 57700000, "rel_tolerance": 0.015 }, "fuzzy_match": "Over 57.7 million" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold 'Over 57.7 million' is a hedged population figure (England). number_match parses '57.7 million', '57,690,300' and 'over 57.7m' alike. rel_tolerance 0.015 (+/- ~0.87M) covers the plausible on-page England estimates (57.1M-57.7M) while still rejecting the UK figure (~67.6M), Birmingham (~1.1M) and Italy (~58.9M). No string anchor is needed: the intent asks only for the number.", "source": "batch_15.json" } } }, { "task_id": 5, "intent": "How many total Pro Football Hall of Famers are there in the articles about the American football teams mentioned in Eminem's article?", "intent_template_id": 5, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 13 }, "fuzzy_match": "Thirteen" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold 'Thirteen' is a count; number_match parses both the spelled-out cardinal and the digits. Exact comparison, no tolerance - a count is either right or wrong. Other numbers in the answer (team names, per-article subtotals) are harmless because the verifier only requires 13 to be among the parsed values.", "source": "batch_15.json" } } }, { "task_id": 6, "intent": "Trace this path: start at the \"Territory\" page → the first country mentioned → a country that shares its southern border → the largest city of that country. What is the first word of the second section on that final page (the introductory paragraphs count as one section)?", "intent_template_id": 6, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "mexico" ], "scope": "full", "fuzzy_match": "Mexico" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Final answer is the single word 'Mexico'. Neither 'Mexico' nor a distractor is named in the intent, so plain must_include is safe. Medium: 'Mexico' could also appear as an intermediate country in a wrong answer that narrates the path.", "source": "batch_16.json" } } }, { "task_id": 7, "intent": "Navigate to the article corresponding to the last word in the second paragraph of the 'Hip-hop' article. In the visited page, follow the first hyperlink and tell me the first line from the article's content.", "intent_template_id": 7, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "in politics", "government of a country", "control over" ], "scope": "full", "fuzzy_match": "In politics, a is the government of a country which has control over a geographic area or territory." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a garbled quote ('In politics, a is the government of a country...' -- the word 'state' was dropped/reordered), so it cannot be matched verbatim; flagged for human review. Spec anchors on the three content phrases that survive in both the gold and a clean quote of the same line. scope full because agents often prefix the quote with the article title. | FIXED (env-verified): removed the spurious 'State ' prefix; the article renders 'In politics, a is the government of a country which has control over a geographic area or territory.' NOTE the env HTML drops the word 'state' ('a is' rather than 'a state is') -- an environment bug, but the gold now matches what an agent actually reads. Spec anchors on stable tokens.", "source": "batch_14.json" } } }, { "task_id": 8, "intent": "Recursively follow the first link in the main content of each page, starting from the article on Research. What is the title of the final article whose page does not exist?", "intent_template_id": 8, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "real" ], "scope": "full", "fuzzy_match": "Real" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the single article title 'Real'. Word-boundary matching keeps it from matching 'Reality'/'Realism'/'really', but 'real' is a common English word, so an answer that uses it incidentally could slip through -- hence medium. The starting article 'Research' named in the intent does not contain the token, so no trap-1 handling is needed.", "source": "batch_01.json" } } }, { "task_id": 9, "intent": "Recursively follow the third hyperlink in the main content of each article, starting from the article \"Periodic table.\" Keep traversing until there is no third hyperlink or you encounter a \"page does not exist\" article. List the titles of all the valid pages in the order that you visited.", "intent_template_id": 9, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "periodic table" ], [ "atomic number" ], [ "nucleus |OR| atomic nucleus" ], [ "proton" ], [ "neutron" ], [ "atom" ], [ "solid" ] ], "ordered": true, "forbidden": [] }, "fuzzy_match": "Periodic table, Atomic number, Atomic nucleus, Proton, Neutron, Atom, Solid" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The intent explicitly asks for the titles 'in the order that you visited', so ordered:true is warranted. Word-boundary matching keeps 'atom' from matching inside 'atomic number', so the ordering check on first occurrences is meaningful. 'nucleus' also accepts the full title 'Atomic nucleus'. No forbidden entries: a correct answer may legitimately mention the terminating non-existent page. | SPOT-CHECK: intent asks for page TITLES; gold item 3 is link text \"nucleus\" but the visited page title is \"Atomic nucleus\". Spec still matches via containment; human should fix gold to titles. | FIXED (env-verified): item 3 corrected from link text 'nucleus' to the actual page title 'Atomic nucleus' (Atomic number's third link points to /wiki/atomic nucleus; a separate 'Nucleus' page also exists). Spec already accepts both spellings.", "source": "batch_15.json" } } }, { "task_id": 10, "intent": "Visit the article about the earliest known statistics mentioned in the Statistics article. Then visit the article about the last country mentioned in its main content. What is the name of the first animal mentioned on that country's page?", "intent_template_id": 10, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "kiwi |OR| kiwis" ], "scope": "full", "fuzzy_match": "Kiwi" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single named entity ('Kiwi'); neither the gold nor any competing animal appears in the intent, so plain must_include on the full answer is safe. The |OR| alternative covers the plural; word-boundary matching already keeps 'kiwifruit' from counting as a hit. 'kiwi' is distinctive (not a common word), so the incidental-mention risk of Trap 3 is low.", "source": "batch_02.json" } } }, { "task_id": 11, "intent": "Who is the current mayor of New York City, according to the article on NYC?", "intent_template_id": 11, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "eric adams" ], "scope": "full", "fuzzy_match": "Eric Adams" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single named entity; the gold string does not appear in the intent, so a plain must_include on the full answer is safe. No must_exclude on predecessors (de Blasio/Bloomberg) because a correct verbose answer may legitimately mention who Adams succeeded (guidelines trap 2).", "source": "batch_03.json" } } }, { "task_id": 12, "intent": "Navigate to the article on Canada and find who the head of the military is. Tell me the name of that person.", "intent_template_id": 12, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^[\\s\\S]*(\\bnot\\b|n't|\\bnever\\b)[\\s\\S]{0,25}(mention|name|specif|state|list|provid|give|given|say|said|identif|indicat|avail)[\\s\\S]*$ |OR| ^[\\s\\S]*\\bno\\b[\\s\\S]{0,20}name[\\s\\S]*$ |OR| unnamed" ], "scope": "full", "fuzzy_match": "Name has not been mentioned." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a negative finding ('Name has not been mentioned.'), so the checkable anchor is a negation adjacent to a reporting verb; the regex covers 'not mentioned', 'has not been mentioned', \"doesn't mention\", 'never names', 'does not give/state/specify/list/identify' and 'no name'/'no mention of the name'. Any answer that actually names a person (Mary Simon, Charles III, Wayne Eyre) lacks that pattern and fails.", "source": "batch_14.json" } } }, { "task_id": 13, "intent": "Visit the article on the date of the latest event from the article on the year 1971. Then tell me the year of the first event from the 1900s, as listed in that article.", "intent_template_id": 13, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "1903" ], "scope": "full", "fuzzy_match": "1903" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A year. Distinctive digits; word-boundary matching keeps it from matching 19031 etc. The intent mentions 1971 and '1900s' but not 1903, so requiring 1903 is discriminative.", "source": "batch_16.json" } } }, { "task_id": 14, "intent": "What is the total number of floral and faunal emblems mentioned in the articles for all six Australian states?", "intent_template_id": 14, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "four |OR| 4" ], "scope": "full", "fuzzy_match": "Four" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a bare count. Switched from number_match to string_match ['four |OR| 4']: both accept 'Four' and '4', but string_match refuses the ordinal reading ('the 4th state...') that number_match would parse as 4, closing that false-accept vector. No tolerance-equivalent: the task asks for an exact total. Confidence medium because 4 is a low-discriminative value that still leaks when a wrong answer enumerates a component of 4 ('4 floral and 8 faunal, twelve total') -- that residual cannot be closed without rejecting the bare gold 'Four'. gold_suspect: six states with a floral AND a faunal emblem each would naively give twelve; 'Four' is only right if most state articles omit their emblems in this snapshot -- worth a human check against the served pages. Spec is unaffected either way. | VERIFIED CORRECT (env): only Queensland (floral + fauna) and South Australia (floral + faunal) mention floral/faunal emblems -> 4 total; New South Wales, Victoria, Western Australia and Tasmania articles mention none. Gold 'Four' stands.", "source": "batch_04.json" } } }, { "task_id": 15, "intent": "Find out which U.S. state uses the word borough instead of counties based on the Borough Wiki article.", "intent_template_id": 15, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "alaska" ], "scope": "full", "fuzzy_match": "Alaska" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single named entity. Neither 'Alaska' nor the distractor 'Louisiana' appears in the intent, so plain must_include is safe. No must_exclude: a correct answer may contrast Alaska with Louisiana's parishes. Residual: a self-contradictory wrong answer ('no state uses boroughs, not even Alaska') would still contain the token, but no competent agent produces that when the gold is a clear positive.", "source": "batch_05.json" } } }, { "task_id": 16, "intent": "How many times did you encounter the word Hell in these articles: \"Peru\", \"Pusha T\", and \"Doom (2016 video game)\"?", "intent_template_id": 16, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 10 }, "fuzzy_match": "Ten" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Pure count; gold 'Ten' -> 10 exact. number_match parses both 'ten' and '10'. Per-article breakdowns in the answer add other numbers but the required total must still be present; an off-by-one total has no 10 anywhere.", "source": "batch_06.json" } } }, { "task_id": 17, "intent": "Which specialized tools were mentioned in the approaches section of the Physics article?", "intent_template_id": 17, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "particle accelerators |OR| particle accelerator" ], [ "lasers |OR| laser" ] ], "ordered": false }, "fuzzy_match": "Particle Accelerators, Lasers" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Two specialized tools. Alternatives cover singular/plural because word-boundary matching treats 'accelerator' and 'accelerators' as distinct. Both must appear.", "source": "batch_16.json" } } }, { "task_id": 18, "intent": "How many times is the word \"Ishmael\" mentioned on the page \"Abraham\"?", "intent_template_id": 18, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 5 }, "fuzzy_match": "Five" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold 'Five' is a count; number_match parses both '5' and 'five'. Exact match, no tolerance. Medium because a wrong answer that happens to contain the digit 5 for another reason (e.g. '6 times across 5 sections') would slip through; no string anchor exists since the intent's own words ('Ishmael', 'Abraham') appear in every answer.", "source": "batch_07.json" } } }, { "task_id": 19, "intent": "Trace this path in the exact order: France → Emblem → Democracy → Name → Custom. While visiting these articles, which word appears immediately before either \"Suzuki Ichiro\" or \"Suffrage\"?", "intent_template_id": 19, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "election |OR| name |OR| given name" ], "scope": "first_sentence", "fuzzy_match": "election" }, "revision": 1, "annotation": { "confidence": "low", "subjective": false, "gold_suspect": false, "notes": "Trap 1: the gold token 'Suffrage' is quoted in the intent itself, so a wrong answer that names another word ('Universal') while pointing at 'Suffrage' would satisfy a plain must_include. The must_exclude regex rejects answers that present some other word as the one standing immediately before 'Suffrage'. gold_suspect: the intent asks for the word appearing immediately *before* \"Suzuki Ichiro\" or \"Suffrage\", yet the gold is one of those two anchors, so the gold does not answer the intent as literally phrased -- a human should confirm the intended reading before this spec is trusted. | SPOT-CHECK: gold \"Suffrage\" restates a search string, does not name the word before it (that word is \"election\"). Gold likely wrong. | FIXED (env-verified): old gold 'Suffrage' restated the anchor rather than the word before it. In Democracy '...vote in an election. Suffrage is...' so 'election' precedes Suffrage; in Name '...the given name. \"Suzuki Ichiro\"...' so 'name' precedes Suzuki Ichiro. The intent says 'either ... or', so both are acceptable and the spec accepts either.", "source": "batch_08.json" } } }, { "task_id": 20, "intent": "Which Brazilian president is mentioned in the article on Conservatism?", "intent_template_id": 20, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "bolsonaro" ], "scope": "full", "fuzzy_match": "Jair Bolsonaro" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single named entity; neither the gold nor any competing president appears in the intent, so plain must_include on the surname is safe. Surname alone is required so 'Jair Bolsonaro', 'Bolsonaro' and '**Bolsonaro**' all pass. Word-boundary matching keeps 'Bolsonarism' from matching. Scope kept 'full' because a verbose agent may open with a sentence about the article before naming the president.", "source": "batch_09.json" } } }, { "task_id": 21, "intent": "Go to the page of the first notable Nobel laureate mentioned in the Nobel Prize article. Then visit the last hyperlink in the second section (the introductory paragraph counts as a section). How many countries are mentioned on the final visited page?", "intent_template_id": 21, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 1 }, "fuzzy_match": "One" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold 'One' is a count, so number_match parses both the word and the digit. Full scope on purpose: this is a 3-hop navigation task and agents often narrate the path in the opening sentence, so first_sentence would reject genuinely correct answers. Trap-3 applies: 'one' is low-discriminative, so a wrong answer that happens to say 'one of which' would slip through; the intent names no competing number, so nothing better is available.", "source": "batch_10.json" } } }, { "task_id": 22, "intent": "I heard that possession of illegal drugs is considered a victimless crime. Can you find a reference article to this statement? If yes, then tell me the name of the article.", "intent_template_id": 22, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "victimless crime" ], "must_exclude": [ "could not find |OR| couldn't find |OR| unable to find |OR| did not find |OR| didn't find |OR| no article |OR| no reference article |OR| no such article |OR| does not exist |OR| doesn't exist |OR| found nothing |OR| nothing matching |OR| no results |OR| no matching |OR| no matches |OR| couldn't locate |OR| could not locate |OR| not able to find |OR| wasn't able to find |OR| was not able to find |OR| found no |OR| there is no article |OR| there was no article |OR| there are no articles" ], "scope": "full", "fuzzy_match": "Victimless Crime" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap-1: the gold title 'Victimless Crime' is quoted verbatim in the intent, so a failed agent will echo it ('I could not find an article about victimless crime'). must_include the title, plus a broadened must_exclude of not-found phrasings (found nothing / no results / couldn't locate / there is no article / ...) that cannot occur in an answer that successfully names the article. Adversarial pass widened the not-found net beyond the draft's four phrasings. Residual risk (why medium): an answer that names a DIFFERENT article while restating the question phrase ('the article about victimless crime is Drug Prohibition') still passes -- not separable deterministically without enumerating wrong titles.", "source": "batch_11.json" } } }, { "task_id": 23, "intent": "Which one of these does not have a valid page: Burger King, Wendy's, McDonald's?", "intent_template_id": 23, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "none |OR| ^.*\\b(all|each|every)\\b[^.!?]{0,55}\\b(valid|exist|exists|available|present|active|working|real|have a page|has a page|have pages|has pages|have an article|has an article|have articles)\\b.*$" ], "must_exclude": [ "except", "not all", "none have valid |OR| none has valid |OR| none has a valid |OR| none have a valid |OR| none of them have valid |OR| none of them has valid |OR| none of them has a valid |OR| none of them have a valid |OR| none of the three have valid |OR| none of the three has valid |OR| none of the three has a valid |OR| none of the three have a valid", "does not have a valid |OR| doesn't have a valid |OR| has no valid" ], "scope": "first_sentence", "fuzzy_match": "None, all of them have valid pages" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1: all three brands are named in the intent, so a correct answer necessarily mentions them and must_exclude on a brand is impossible. The discriminating signal is the universal 'all/each/every ... valid|exist|available|...' verdict OR a bare 'none' (the correct one-word answer). Requiring the positive predicate NEAR all/each/every (rather than accepting bare 'all three'/'all of them' tokens) is what rejects the hardest near-miss 'All three were checked, and Wendy's is a red link' -- that sentence names a culprit and never asserts validity, so it matches neither 'none' nor 'all...valid'. The excludes kill the residual holes that still contain 'none' or 'all...valid': 'not all ... have valid'; 'none [of them/of the three] have/has [a] valid' (the all-invalid reading of 'none', which the gold 'None, all of them have valid' does NOT trip because 'all' sits between 'none' and 'have'); 'except'; and a brand that 'does not have a valid / has no valid' page. 'is a red link' / 'is invalid' are deliberately NOT excluded because a correct answer may negate them ('none of them is a red link', 'none of them is invalid'). Residual (medium): a convoluted 'only McDonald's has a valid page' could slip through.", "source": "batch_01.json" } } }, { "task_id": 24, "intent": "Start at the Computer Science article and serially visit the hyperlinks in that article. Stop when you find a hyperlink where the hyperlink text doesn't match the title of the page it leads to, and tell me the text of that hyperlink.", "intent_template_id": 24, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "encoding |OR| encodings" ], "scope": "full", "fuzzy_match": "Encoding" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Single-token hyperlink text. Neither the gold nor any distractor appears in the intent, so plain must_include is safe. The 'encodings' alternative is there only because word-boundary matching would otherwise reject a plural restatement. Residual false-accept (why confidence is medium, not high): the target page is titled 'Character encoding', which every correct answer legitimately quotes, so a comprehension-error answer that reports the *title* 'Character encoding' as the hyperlink text also contains the word 'encoding' and passes. It cannot be excluded without rejecting correct answers that cite the target title, so no deterministic rule separates the two.", "source": "batch_12.json" } } }, { "task_id": 25, "intent": "Does the article on Sociology discuss its relationship with Mathematics? If so, please provide the title of the section.", "intent_template_id": 25, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| does not |OR| doesn't |OR| do not |OR| don't" ], "must_exclude": [ "yes", "does discuss |OR| discusses it |OR| it discusses |OR| is discussed |OR| does cover |OR| is covered |OR| covers it" ], "scope": "first_sentence", "fuzzy_match": "No, it does not." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No, it does not.'). first_sentence scope keeps a justification tail from leaking a 'yes'. The negated-verb alternatives accept 'The article does not discuss ...' which carries no bare 'no' token. The second must_exclude entry rejects the tricky affirmative wrong-answer shapes that still contain a negation token elsewhere -- 'it does not have a dedicated section, but it does discuss ...' and 'No doubt the article discusses it ...' -- by excluding the affirmative-discussion phrases themselves. The excludes are tied to adjacency ('does discuss', 'discusses it', 'is discussed') so contrastive-correct wording like 'never discusses its relationship' or 'is not covered' is NOT tripped ('does not discuss', 'is not covered' keep the negation between the words). Medium confidence: the affirmative-phrase exclude list is judgement-based and a rare wrong answer that affirms with an un-listed verb (e.g. 'No -- actually it IS addressed in ...') would slip through.", "source": "batch_05.json" } } }, { "task_id": 26, "intent": "I've been wanting to learn more about civil wars for a long time. It's fascinating how conflicts can arise between two groups within the same country. Since it's a common topic, there should be plenty of articles on it. I'm especially interested in civil wars where one group wants to separate from, or secede from, the country. Could you search for an article on this and let me know the title?", "intent_template_id": 26, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "secession |OR| secessions" ], "scope": "full", "fuzzy_match": "Secession" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the single article title 'Secession'. The intent uses the verb 'secede', never the noun, so the gold token is not handed to the agent by the question and word-boundary matching keeps 'secede'/'seceding'/'secessionist'/'secessionism' from satisfying it. Scope kept at 'full' deliberately: a correct answer often opens with a preamble ('I found an article that matches. It is titled Secession.') and first_sentence would reject it. Residual risk (hence medium): a wrong answer that names another title but happens to use the noun 'secession' in its justification would slip through.", "source": "batch_13.json" } } }, { "task_id": 27, "intent": "Start at the Computer Science article and serially visit the hyperlinks in that article. Stop when you find a hyperlink where the hyperlink text doesn't match the title of the page it leads to, and tell me the names of all the hyperlinks that you have visited.", "intent_template_id": 27, "sites": [ "wiki" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "^.*(?number binding regex was rejected because the common 'respectively' phrasing ('the Lindy and Kodak cables are 5 m and 1.8 m respectively') decouples names from numbers, so any binding spec would false-reject correct answers - a worse failure than the residual swap false-accept. The swap therefore cannot be added as a negative_example.", "source": "batch_02.json" } } }, { "task_id": 55, "intent": "Which cable is longer: the white Type-C to HDMI Cable or the Type-C USB Charger Cable?", "intent_template_id": 55, "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "hdmi" ], "must_exclude": [ "^(?:(?!\\b(?:than|unlike|versus|vs|over|compared|beat|beats|beaten|rather than|instead of)\\b[-\\s a-z0-9]{0,15}usb charger).)*?\\busb charger cable\\b(?:(?!\\bhdmi\\b|\\bnot\\b|[.;]).)*?\\blonger\\b.*$", "^(?:(?!\\bnot\\b|isn|aren|wasn|weren).)*\\b(?:same length|equal length|equally long|identical (?:in )?length|same size)\\b.*$" ], "scope": "full", "fuzzy_match": "White Type-C to HDMI Cable" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1: both options are named in the intent, so a wrong answer still contains the gold token 'hdmi'. must_include['hdmi'] keeps the answer on-topic; the discrimination is done by two regex exclusions. Exclusion 1 fires only when the USB Charger Cable is the subject asserted to be 'longer' (any words, not just 'is/was', so 'measures 10 ft, longer' is caught). Its leading guard '(?!than|unlike|versus|compared ... usb charger)' and the inner '(?!hdmi)' block the two correct phrasings that mention the USB cable as the loser -- 'HDMI is longer than the USB Charger Cable' and 'the HDMI cable, unlike the USB Charger Cable, is the longer one' -- so they are not false-rejected. Exclusion 2 catches the 'they are the same length' wrong answer (guarded against 'not the same length'). Medium: an exotic wrong answer that asserts USB superiority without the word 'longer'/'same length' (e.g. 'the USB cable wins on length') could still slip through.", "source": "batch_08.json" } } }, { "task_id": 56, "intent": "What is the total cost if I buy all the 100% polyester skirts available in the shop? If there are multiple variants (sizes, colors) of the same item, include only one.", "intent_template_id": 56, "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 9.99 }, "fuzzy_match": "$9.99" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Currency total. Exact comparison (no tolerance) on purpose: an abs_tolerance of 0.01 would accept an off-by-one-cent answer such as $9.98, and every genuinely correct answer reports the price verbatim. Residual leniency: an answer that double-counts variants but still shows the unit price ('two at $9.99, total $19.98') would pass; no deterministic rule separates that from a correct answer that quotes the unit price.", "source": "batch_12.json" } } }, { "task_id": 57, "intent": "What would be the minimum cost to buy almonds and rice crackers?", "intent_template_id": 57, "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 105.46 }, "fuzzy_match": "$105.46" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Price total. Exact comparison with no tolerance: the gold is a precise cent figure the agent reads off two product pages and adds, so an off-by-a-cent answer is a real arithmetic error, not a rounding artifact. number_match ignores the '$' and any per-item prices the agent shows alongside the total, so a breakdown answer still passes.", "source": "batch_13.json" } } }, { "task_id": 58, "intent": "List all the different color options available for the Nautica and GRADO SR325x headphones.", "intent_template_id": 58, "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ "^.*black[^a-z]{0,3}black.*$", "^.*navy[^a-z]{0,3}navy.*$", "^.*navy[^a-z]{0,3}red.*$", "^.*navy[^a-z]{0,3}yellow.*$", "^.*white[^a-z]{0,3}nude.*$", "^.*white[^a-z]{0,3}off[^a-z]{0,3}white.*$" ], "ordered": false, "forbidden": [] }, "fuzzy_match": "black black, navy navy, navy red, navy yellow, white nude, white off white" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Six two-word colour options. Plain containment would break on 'Black/Black' or 'Navy - Red' because the matcher only bridges hyphens and whitespace, so each item is a regex leaf allowing up to three non-letter characters between the two words. Medium confidence: an answer that lists all six plus an invented seventh option would still pass.", "source": "batch_03.json" } } }, { "task_id": 59, "intent": "What is the total price of all the rackets and tennis balls available in the shop? If there are multiple variants of the same item, pick one only.", "intent_template_id": 59, "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "zero |OR| 0 |OR| none |OR| no rackets |OR| no racket |OR| no tennis balls |OR| not available |OR| unavailable |OR| out of stock |OR| nothing |OR| no products |OR| no items |OR| not in stock" ], "must_exclude": [ "^.*\\$\\s*0*[1-9][0-9,]*(\\.[0-9]+)?.*$", "^.*(? place the order' branch, so no extra check is needed. Leading zero survives normalization because matching is textual, not numeric.", "source": "batch_01.json" } } }, { "task_id": 89, "intent": "Look up Organic Diced Pear in the Shop and check the product description. Identify the vitamin mentioned in the description and visit its article on Wiki. Does the fruit pear appear in that article?", "intent_template_id": 89, "sites": [ "wiki", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| doesn't |OR| isn't |OR| does not |OR| never |OR| absent" ], "must_exclude": [ "yes" ], "scope": "full", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold. Switched scope from first_sentence to full: this is a multi-step task where a correct answer often identifies the vitamin in its opening sentence and delivers the 'No' in the next one ('The description mentions Vitamin C. No, the pear does not appear.') -- first_sentence FALSE-REJECTED exactly that shape. Under full scope the negation list (no/not/doesn't/isn't/does not/never/absent) is what carries a correct answer, and must_exclude ['yes'] rejects the affirmative near-miss even when it sneaks in a stray 'not' ('Yes, pear does appear; it is not absent'). must_exclude stays limited to 'yes' per trap 2 -- 'vitamin c' and 'pear' must remain allowed. Residual is only the rare correct answer that uses a 'yes' filler in its justification.", "source": "batch_04.json" } } }, { "task_id": 90, "intent": "Search for \"Vitamin K\" on Wiki and identify all the vegetables listed there as sources of vitamin K. How many of those vegetables are available in the shop?", "intent_template_id": 90, "sites": [ "wiki", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "none |OR| zero |OR| 0 |OR| no vegetables |OR| no vegetable |OR| not one |OR| not a single |OR| none of |OR| not any of |OR| do not carry any |OR| does not carry any |OR| doesn't carry any |OR| don't carry any |OR| do not stock any |OR| does not stock any |OR| doesn't stock any |OR| don't stock any |OR| do not have any |OR| does not have any |OR| doesn't have any" ], "scope": "first_sentence", "fuzzy_match": "None" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold 'None' is again an unparseable zero count, so string_match carries it. Adversarial pass replaced the draft's loose 'any of them / any of those' alternatives (which false-accepted a rhetorical 'Are any of them available? Yes, two.') with negation-anchored forms ('does not stock any', 'does not carry any', 'not any of', 'none of'), which still admit the common correct phrasing 'the shop does not carry any of the vegetables listed'. first_sentence prevents a trailing 'none of the others' clause from rescuing a wrong count. Medium: a single-sentence 'only one ... none of the rest' would still slip through the bare 'none' alternative, but excluding counts would false-reject correct 'none of the four listed' answers.", "source": "batch_11.json" } } }, { "task_id": 91, "intent": "Search News for \"video game\" and visit the sixth search result. Is that news event described in the wiki page for \"Video game\"?", "intent_template_id": 91, "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not described |OR| not mentioned |OR| not covered |OR| not present |OR| not referenced |OR| not discussed |OR| not found |OR| doesn't describe |OR| does not describe |OR| doesn't mention |OR| does not mention |OR| isn't described |OR| isn't mentioned |OR| nowhere |OR| not in the |OR| not on the |OR| no mention" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No'). first_sentence scope keeps the justification (which mentions the wiki page and the news event) from interfering. Bare 'not' was removed because it flips polarity in a wrong answer ('The event is not only described but expanded upon' means it IS described = yes); the alternatives are the specific 'not ' negations that can only express the 'No' verdict, plus bare 'no'. Inherent yes/no crosstalk is unavoidable.", "source": "batch_03.json" } } }, { "task_id": 92, "intent": "Do the following tasks: Search \"crocodile\" on News, visit the first news article, identify the country where that crocodile lives, search that country on Wiki, and tell me the population of that country. If you encounter a redirect link at any point, then please visit that link.", "intent_template_id": 92, "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 27000000, "rel_tolerance": 0.05 }, "fuzzy_match": "27 million people" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Environment-verified: the crocodile in the first news article ('Drunken man is surprised crocodile bit him') is at Broome, Western Australia, so the country is Australia and gold '27 million' matches the Wiki intro ('27 million people live in Australia'). Crucially the SAME article also states 'The population is about 26 million people (2021 census = 25,890,773)', so a correct agent may legitimately quote the census figure (~25.9-26M) instead of 27M. Reviewer widened rel_tolerance from 0.03 to 0.05 (range 25.65M-28.35M) so both readings pass; it still rejects a wrong-country population (Nepal 30M, PNG 10M) or a gross misread. The former negative '25.7 million' was dropped because it is within the article's own census figure and thus not clearly wrong. Medium because the acceptance band spans two figures the article itself gives.", "source": "batch_07.json" } } }, { "task_id": 93, "intent": "Search Canada on Wiki and find the name of the current prime minister. Look up his name on News and click on the first article. When was that article published?", "intent_template_id": 93, "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "april |OR| ^.*\\b2025[-/]0?4[-/]28\\b.*$ |OR| ^.*\\b0?4[-/]28[-/]2025\\b.*$", "28", "2025" ], "scope": "full", "fuzzy_match": "April 28, 2025" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a publication date. Requiring the three components separately accepts 'April 28, 2025', '28 April 2025' and 'Apr 28th, 2025', while the regex alternatives also accept the numeric forms 2025-04-28 / 04/28/2025. Word-boundary matching means '28' cannot come from '2025' or '280'. Any wrong day, month or year is rejected.", "source": "batch_15.json" } } }, { "task_id": 94, "intent": "Which ones of the following sites: Wiki, News, and Shop, have an article or product page on a screwdriver?", "intent_template_id": 94, "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match", "string_match" ], "reference_answers": { "list_match": { "items": [ [ "wiki |OR| wikipedia" ], [ "news |OR| wikinews" ] ], "ordered": false }, "must_exclude": [ "^.*(? ... ' within 22 chars (does not / has no / lacks / nothing / without / no hdmi / missing / ...), plus plain excludes for 'only', 'but not/except/other than/neither/no relevant', and 'not the '. The 'all three'/'all of them'/'all sites' alternatives keep terse correct answers passing; the 22-char gap was tuned so 'without exception' and 'none of the sites are missing it' (correct) do not trip it. Medium because open-ended negation phrasing means a novel wording could still slip.", "source": "batch_11.json" } } }, { "task_id": 96, "intent": "I heard some bad news about biscuits. Can you check and share the title of such a news article with me? If there are none, please place an order for the top search result on \"biscuits\" and share the confirmation code with me.", "intent_template_id": 96, "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "83a1b2e714" ], "scope": "full", "fuzzy_match": "83A1B2E714" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": true, "notes": "The task resolves to the order branch, so the answer must carry the confirmation code. The code is a unique alphanumeric token; normalization lowercases both sides and word-boundary matching prevents a longer code from matching by prefix. | SPOT-CHECK: gold is an order code (took the \"no bad news about biscuits\" branch), but News has a clearly negative biscuit article (Vanuatu banned Fijian biscuits; maker laid off staff, taking legal action). Code also unverifiable. Human: fix gold/intent. | ESCALATED: Order-code gold vs. a plausible negative biscuit article ('Fiji ... ban on Fijian biscuits'). Flipping the task branch is a judgement about whether a trade-ban story counts as 'bad news about biscuits'; needs a human decision. Left flagged, gold unchanged.", "source": "batch_02.json" } } }, { "task_id": 97, "intent": "Can you place an order for the cheapest nuts available and share the confirmation code with me?", "intent_template_id": 97, "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "54f0dc47b1" ], "scope": "full", "fuzzy_match": "54F0DC47B1" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single opaque confirmation code; literal containment is exact and cannot be produced by a wrong trajectory.", "source": "batch_05.json" } } }, { "task_id": 98, "intent": "Which scholars are mentioned in the article on Rayleigh scattering?", "intent_template_id": 98, "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "rayleigh" ], [ "raman" ], [ "krishnan" ] ], "ordered": false }, "fuzzy_match": "Rayleigh, Raman, Krishnan" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Three scholars. 'Rayleigh' also names the article, so it alone is not discriminative; requiring Raman AND Krishnan makes the spec strict.", "source": "batch_16.json" } } }, { "task_id": 99, "intent": "Is it true that Mattel had to recall its products due to safety concerns? If so, please visit the Wiki page for Mattel and find the first product mentioned in that article. Tell me the name of that product and whether it is available in the Shop.", "intent_template_id": 99, "sites": [ "wiki", "news", "webshop" ], "start_url": "__NEWS__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "american girl", "^.*\\bnot\\b[^.]{0,40}\\b(?:availab\\w*|find|found|sold|sell\\w*|listed|carr\\w+|stock\\w*|offered|exist\\w*|have|has|there|in the shop)\\b.*$ |OR| unavailable |OR| no american girl |OR| no results |OR| no such product |OR| out of stock" ], "scope": "full", "fuzzy_match": "American Girl dolls; not available in the shop." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Two-part gold: the product name and the shop-availability verdict. 'barbie' pins the product ('Barbie', 'Barbie dolls', 'Barbie doll' all match). The second entry is a regex that requires a negated availability claim within one clause of 'not' ('not available', 'could not find', 'does not sell/carry/have/stock/list it'), plus plain alternatives for the phrasings that do not use 'not' ('unavailable', 'no Barbie dolls', 'no results'). This is what separates the correct answer from the near-miss that names Barbie but claims it *is* in the Shop, and from an answer that names Barbie and never addresses availability at all. | SPOT-CHECK: gold names \"Barbie dolls\" but the Mattel article lists \"American Girl dolls\" FIRST; the intent asks for the first product mentioned. Human: change gold to American Girl dolls. | FIXED (env-verified): the Mattel article lists 'American Girl dolls' before 'Barbie dolls', and the intent asks for the FIRST product mentioned. Shop search for 'American Girl' returns 0 products, so the 'not available' half stands.", "source": "batch_13.json" } } }, { "task_id": 100, "intent": "I am a huge fan of the Donkey Kong series. Please find related Wiki articles, news, and products from the Shop. If there are multiple entries, pick the most relevant ones. If you cannot find any product or article, mention that in your answer. Share the titles of these items with me using the following format: :
.", "intent_template_id": 100, "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "^.*\\bwiki(?:pedia)?(?:\\s+article)?[^a-z0-9]{0,6}donkey kong\\b.*$" ], [ "no relevant news |OR| no news |OR| none |OR| no relevant articles |OR| no relevant results |OR| no matching news |OR| not find any news |OR| no relevant news articles |OR| nothing relevant |OR| no news found", "^.*\\b(?:no|not|none|nothing|never|without|unable)\\b[^.;,]{0,40}\\bnews\\b.*$" ], [ "tropical freeze |OR| donkey kong surfer |OR| game traveler" ], [ "shop |OR| store |OR| webshop" ] ], "ordered": false }, "fuzzy_match": "Wiki: Donkey Kong, News: No relevant news found, Shop: Nintendo Switch Game Traveler Donkey Kong Country Tropical Freeze Deluxe Travel Case - Donkey Kong Surfer" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Three-part answer in a mandated ': ' format. 'Donkey Kong' also occurs inside the shop product title, so a bare must_include would accept an answer that failed to find the Wiki article; the first item is therefore a regex requiring the Wiki label to be immediately followed by the article title. The News item must assert absence: the literal alternatives cover the fixed phrasings, and a second regex leaf catches any negation word sitting within 40 non-separator chars of the word 'news' (so 'No Donkey Kong news articles found' and 'could not find any relevant news' also pass). The [^.;,] barrier stops that negation from leaking across a field boundary -- a Wiki-field 'No relevant article found' cannot satisfy the News item. The Shop item keys on tokens unique to the gold product.", "source": "batch_12.json" } } }, { "task_id": 101, "intent": "Visit the Wiki article on probability and find the second hyperlink on that page. Search the hyperlink text in the Shop, visit the second search result, and search its brand name on Wiki. If a Wiki article exists, then tell me the name of that article; otherwise, tell me the name of the brand.", "intent_template_id": 101, "sites": [ "wiki", "webshop" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "rusticware |OR| rustic ware" ], "scope": "full", "fuzzy_match": "Rusticware" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single brand/article name that does not appear anywhere in the intent. The spaced spelling is allowed because word-boundary matching would otherwise reject 'Rustic Ware'.", "source": "batch_12.json" } } }, { "task_id": 102, "intent": "Please complete the following steps: (1) Search News for \"Japan\", (2) visit the second search result, (3) visit the Wiki page for the second word found in the news article's body, (4) identify the first country mentioned in that Wiki article, and (5) search that country on News and tell me the title of the first search result.", "intent_template_id": 102, "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "tamiflu", "india" ], "scope": "full", "fuzzy_match": "Generic Tamiflu in India" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a headline. Requiring the two distinctive title words separately rather than the full phrase accepts both a verbatim quote and a light paraphrase ('the first result is about generic Tamiflu in India'), while still being highly discriminative: no other headline in the corpus pairs 'Tamiflu' with 'India'. Neither token occurs in the intent, so trap 1 does not apply.", "source": "batch_04.json" } } }, { "task_id": 103, "intent": "How many people were included in the sample for the latest trial of the Australian bird flu vaccine?", "intent_template_id": 103, "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 400 }, "fuzzy_match": "400" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Sample size gold '400'. Exact comparison (no hedge in the gold), which keeps '4,000' and '40' out while accepting 'four hundred'.", "source": "batch_08.json" } } }, { "task_id": 104, "intent": "List all articles from the 'Related Pages' sections of the Biophysics and Biochemistry pages.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Biophysics in the search box and press Enter.\n2. Scroll to the \"Related Pages\" section and remember the listed pages: \"Bionics, Computational biology, Database of Molecular Motions, Nanotechnology\".\n3. Type Biochemistry in the search box and press Enter.\n4. Scroll to the \"Related Pages\" section and remember the listed pages: \"List of biochemistry topics, Biophysics, Chromatography, Organic chemistry\".\n5. Combine the lists from both pages into a single list. Send a message to the user: \"Bionics, Computational biology, Database of Molecular Motions, Nanotechnology, List of biochemistry topics, Biophysics, Chromatography, Organic chemistry\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "list of biochemistry topics" ], [ "biophysics" ], [ "chromatography" ], [ "organic chemistry" ], [ "bionics" ], [ "computational biology" ], [ "database of molecular motions" ], [ "nanotechnology" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "List of biochemistry topics, Biophysics, Chromatography, Organic chemistry, Bionics, Computational biology, Database of Molecular Motions, Nanotechnology" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Eight-element enumeration. ordered:false because the intent only says 'list all articles' -- it names Biophysics first while the gold happens to lead with the Biochemistry entries, so any order requirement would be arbitrary. 'Biophysics' is a weak item (it also appears in the intent), but the other seven carry all the discrimination: an answer that covers only one of the two pages, or that substitutes a plausible-looking page, is missing at least one required item.", "source": "batch_04.json" } } }, { "task_id": 105, "intent": "Did the Physics article's \"Approaches in Physics\" section mention the names of more scientists than the Chemistry article's \"Atoms\" section?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Physics in the search box and press Enter.\n2. Locate the \"Approaches in Physics\" section and count the number of unique scientists mentioned, which is one. Remember the count.\n3. Type Chemistry in the search box and press Enter.\n4. Locate the \"Atoms\" section and count the number of unique scientists mentioned, which is ten. But no need to count as soon as you find more than one.\n5. Compare the two counts to see if the Physics count is higher than the Chemistry count. As the count in Physics was not higher, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no rule. Not excluding 'physics'/'chemistry' (trap 2) since a correct answer routinely compares both sections.", "source": "batch_06.json" } } }, { "task_id": 106, "intent": "Find the largest city listed on the Wikipedia pages for Canada and Australia. Then tell me the population difference between those two cities.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Canada in the search box and press Enter.\n2. Identify the largest city mentioned in the article (likely Toronto), find its population figure (around 3 million), and remember it.\n3. Type Australia in the search box and press Enter.\n4. Identify the largest city mentioned in the article (likely Sydney), find its population figure (around 5.5 million), and remember it.\n5. Calculate the absolute difference between the two population figures.\n6. Send a message to the user: \"Around 2.5 Million\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 2500000, "rel_tolerance": 0.1 }, "fuzzy_match": "Around 2.5 Million" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Hedged gold ('Around 2.5 Million'), so a relative tolerance is required; 0.1 gives 2.25M-2.75M, which covers the real figures an agent can read off the two articles (Toronto ~2.8-3M, Sydney ~5.4-5.5M). Medium confidence because number_match scans the whole answer: an answer that shows its working could contain a Toronto population inside the tolerance band while stating a wrong difference. Scoping to the first sentence would reject correct answers that show the arithmetic first, so full scope is the safer trade.", "source": "batch_01.json" } } }, { "task_id": 107, "intent": "Visit the pages of the birthplaces of Albert Einstein's first and second sons. Then calculate the population difference between those places.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Albert Einstein in the search box and press Enter.\n2. Find the names of his first and second sons (Hans Albert Einstein and Eduard Tete Einstein) and identify their birthplaces (Bern and Zurich).\n3. Visit the hyperlink of the first birthplace, Bern, and record its population, which is 127,000.\n4. Go back and visit the hyperlink of the second birthplace, Zurich, and record its population, which is 415,000.\n5. Calculate the absolute difference between the population of Zurich and Bern. Send the absolute difference as a message to the user: \"Around 288000\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 288000, "rel_tolerance": 0.02 }, "fuzzy_match": "Around 288000" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a hedged quantity ('Around 288000'), so a small relative tolerance (2%, i.e. +/-5760) covers rounding to '290,000' or slightly different rounded populations, while still rejecting the usual wrong results (the sum 542,000, or a difference computed from the wrong pair of cities). '288 thousand' is parsed correctly by the number parser.", "source": "batch_02.json" } } }, { "task_id": 108, "intent": "Visit the articles for Physics, Chemistry, and Biology to identify which one(s) contain a section that explicitly discusses the discipline's relationship with mathematics.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Physics in the search box and press Enter.\n2. Scan the table of contents or headers for a section discussing a relationship with mathematics. Note that such a section exists, and remember this.\n3. Type Chemistry in the search box and press Enter.\n4. Scan the table of contents or headers for a similar section regarding mathematics. Note that no such section exists, and remember this.\n5. Type Biology in the search box and press Enter.\n6. Scan the table of contents or headers for a similar section regarding mathematics. Note that no such section exists, and remember this.\n7. Conclude that only the Physics article contained the specific section. Send a message to the user: \"Physics\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "physics" ], "must_exclude": [ "both", "all three", "all of them" ], "scope": "first_sentence", "fuzzy_match": "Physics" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Multiple choice among Physics/Chemistry/Biology, all named in the intent; gold is Physics only. Per guideline traps #2/#4 we do NOT must_exclude the competing discipline tokens (a correct answer naturally rules them out: 'Physics, not Chemistry or Biology' — the earlier chemistry/biology exclusion false-rejected exactly that, verified). We also drop 'neither' from must_exclude: a correct answer legitimately says 'neither Chemistry nor Biology has it'. Instead must_include physics and exclude only the multi-select words a wrong answer would use ('both', 'all three', 'all of them'), scoped to the first sentence. Medium: a bare two-item wrong answer 'Physics and Chemistry contain such a section' (no 'both'), or a 'Neither ... has such a section' wrong answer, would still pass; the additional_instructions steer the agent to a terse 'Physics', so those residuals are unlikely, and this is the honest trade for not rejecting contrastive-correct answers.", "source": "batch_16.json" } } }, { "task_id": 109, "intent": "In the article on Language, visit the first hyperlink in the \"Types of Languages\" section. Then check the \"Related Pages\" section of the visited article. Is there any article that appears in both this section and the \"Related Pages\" section of the Language article? If so, list the common pages; otherwise, answer \"None\".", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Language in the search box and press Enter.\n2. Record the list of articles in the \"Related Pages\" section of the Language article: \"Alphabet, Basic English, Sign language, English as a foreign language, English language, Historical linguistics, Language education, Language families and languages, Linguistics, List of languages, Orthography, Phonology, Second language, Semantics, Speech therapy.\"\n3. Navigate to the \"Types of Languages\" section and click the first hyperlink, \"Mathematics\".\n4. On the new page, locate the \"Related Pages\" section and record the articles listed there: \"Timeline of women in mathematics, American Mathematical Society, Society for Industrial and Applied Mathematics, EASIAM, International Congress on Industrial and Applied Mathematics, International Congress of Mathematicians, International Mathematical Olympiad, Mathematics Genealogy Project, Mathematics Subject Classification.\"\n5. Compare the two lists of related pages to find any common articles. As no common articles were found, send a message to the user: \"None\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "none |OR| no common |OR| no article |OR| no articles |OR| no page |OR| no pages |OR| no overlap |OR| no shared |OR| nothing in common |OR| no entries |OR| no entry |OR| no titles |OR| no title |OR| share no |OR| shares no |OR| do not share |OR| does not share |OR| no intersection |OR| empty |OR| neither list |OR| share nothing |OR| shares nothing |OR| have nothing |OR| no matching |OR| not appear in both |OR| none appear in both |OR| no common article" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "None" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the literal fallback answer 'None'. The intent itself dictates the word, so the discriminator is the negative verdict versus naming an actually-common page. Alternatives cover the natural prose forms ('there are no common articles', 'the two sections share nothing', 'neither list has a title in the other'). must_exclude 'yes' in the verdict sentence blocks the near-miss that names a shared page and then denies the 'None' option; a correct answer to 'Is there any article that appears in both?' cannot open its verdict sentence with 'yes'.", "source": "batch_01.json" } } }, { "task_id": 110, "intent": "Visit the first article listed in the \"Related Pages\" sections of the Time and War articles. Then list all the related pages mentioned in each of those visited articles.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Time in the search box and press Enter.\n2. Go to the \"Related Pages\" section and click the first article link, \"Chronology\".\n3. On this new page, record all items listed in its \"Related Pages\" section, which are \"Annals, Calendar\".\n4. Type War in the search box and press Enter.\n5. Go to the \"Related Pages\" section and click the first article link, \"Peace\".\n6. On this new page, record all items listed in its \"Related Pages\" section, which are \"Anti-war movement, Ceasefire, Pacifism, Peace makers, Peace treaty, World peace\".\n7. Combine the lists recorded from both visited articles. Send a message to the user: \"Annals, Calendar, Anti-war movement, Ceasefire, Pacifism, Peace makers, Peace treaty, World peace\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "annals" ], [ "calendar" ], [ "anti-war movement |OR| antiwar movement" ], [ "ceasefire |OR| cease-fire" ], [ "pacifism" ], [ "peace makers |OR| peacemakers" ], [ "peace treaty" ], [ "world peace" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "Annals, Calendar, Anti-war movement, Ceasefire, Pacifism, Peace makers, Peace treaty, World peace" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Eight related-page titles combined from the Chronology and Peace articles. Unordered — the intent only asks to list them. Hyphen/space variants are handled by the matcher; explicit alternatives cover the closed-up spellings 'antiwar', 'peacemakers' and 'cease-fire'. Supersets pass (list_match does not penalise extra items); we do not know a reliable distractor set to forbid.", "source": "batch_05.json" } } }, { "task_id": 111, "intent": "Visit the pages of the people who were murdered in the \"Unnatural causes of death\" section of the \"Death\" article. How many of these visited articles have a \"Legacy\" section?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Death in the search box and press Enter.\n2. Navigate to the \"Unnatural causes of death\" section and identify the people mentioned as murdered.\n3. The first person is Abraham Lincoln. Click on his hyperlink.\n4. On Abraham Lincoln's page, there is a \"Legacy\" section. Remember the Count = 1.\n5. Go back to the previous page on Death and click on the hyperlink of the second person, \"Martin Luther King Jr.\"\n6. On Martin Luther King Jr's page, there is a \"Legacy\" section. Remember the Count = 2.\n7. Go back to the previous page on Death and click on the hyperlink of the third person, \"John F. Kennedy\".\n8. On John F. Kennedy's page, there is a \"Legacy\" section. Remember the Count = 3.\n9. Go back to the previous page on Death, and verify if there is any person left.\n10. As there are no other people left and you found three legacy sections, send a message to the user: \"Three\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "three |OR| 3" ], "must_exclude": [ "^.*(?<!each )(?<!every )(?<!any )(?<!not )(?<!no )(?<![a-z])(?:one|two|1|2) (?:articles? |pages? |people )?(?:out )?of (?:the |these |those )?(?:three|3|them|these|those|articles?|pages?|people|visited)(?![a-z]).*$", "^.*(?<![a-z])(?:only|just) (?:one|two|zero|1|2|0)(?![a-z0-9]).*$", "^.*(?<!each )(?<!every )(?<!any )(?<![a-z])(?:zero|none|one|two|0|1|2) (?:of (?:the )?(?:them|these|those|three|3|articles?|pages?|people) )?(?:articles? |pages? |people )?(?:have|has|contain|contains|include|includes)(?: an?)? legacy.*$", "^.*legacy(?: section)?s?\\b(?:(?!\\bnot\\b|\\ball\\b|\\beach\\b|\\bevery\\b|\\bthree\\b|\\b3\\b).){0,25}?\\b(?:on|in|for)\\b (?:only |just )?(?:zero|none|one|two|0|1|2)(?![a-z0-9]).*$" ], "scope": "full", "fuzzy_match": "Three" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count answer, gold 'Three' (= all three visited people have a Legacy section). Switched from number_match to string_match must_include ['three |OR| 3'] because number_match merges adjacent spelled-out numbers ('Three - each page has one...' would parse 'one three'->4 or drop 3), false-rejecting valid answers. Trap 3 in reverse: the number 3 always appears in a correct answer ('three articles/people visited'), so a wrong 'only two of the three have a Legacy section' also contains 3. The four must_exclude regexes reject a lower count (0/1/2) stated as: 'N of the three/them', 'only/just N', 'N (have|contain) ... legacy', and 'legacy ... on/in N'. Lookbehinds (each/every/any/not/no) protect legitimate 'each one of the three', 'not one of them lacks it'. Medium confidence: an exotic wrong phrasing could still slip through, but all realistic near-misses are covered by the negative_examples.", "source": "batch_02.json" } } }, { "task_id": 112, "intent": "Visit the pages of the hyperlinked geographic regions mentioned in the \"Trees\" section of the Australia page. Among the visited pages, list the titles of the pages that do not have a section on languages.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Australia in the search box and press Enter.\n2. Navigate to the \"Trees\" section and identify any hyperlinked geographic regions mentioned. There will be two: \"Southeast Asia\" and \"Oceania\".\n3. Click on the hyperlink of the first entry: Southeast Asia.\n4. Check the visited article for a section on Languages. As there isn't, remember this fact.\n5. Then go back and click on the hyperlink of Oceania.\n6. Check the visited article for a section on Languages. As there is such a section, remember this fact.\n7. Only Southeast Asia doesn't have a section on languages. Send a message to the user, \"Southeast Asia\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "southeast asia", "^(?:(?!southeast asia[\\s,/&]*(?:and|or|nor|plus|as well as|along with|together with)?[\\s,/&]*oceania)(?!oceania[\\s,/&]*(?:and|or|nor|plus|as well as|along with|together with)?[\\s,/&]*southeast asia).)*$ |OR| ^.*only\\s*(?:the\\s*)?southeast asia(?![^.]*oceania).*$" ], "scope": "full", "fuzzy_match": "Southeast Asia" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Two candidate pages are named in the instructions (Southeast Asia, Oceania) and only Southeast Asia is the answer, so the gold token alone is not enough -- the dominant failure mode is listing BOTH. Second must_include entry is a two-branch regex. Branch 1 (tempered-token) rejects any answer where the two region names are joined as a LIST: adjacent modulo whitespace/comma/slash/&/'and'/'or'/'nor'/'plus'/'as well as'/'along with'/'together with', in either order. Branch 2 rescues verbose-correct answers that recap visiting both by matching 'only ... Southeast Asia' when no 'oceania' follows before the next sentence break. A correct contrast that mentions Oceania as HAVING languages ('Southeast Asia lacks a section, Oceania has one') is NOT rejected because the names are separated by words, not a list connector. This upgrades the earlier spec, which only forbade the literal 'and'/'nor' conjunction and thus falsely accepted the comma/slash/multi-word list forms. Medium confidence: the residual false-reject is the phrasing 'I visited both Southeast Asia and Oceania. Southeast Asia lacks a languages section.' (conjunction present, no 'only' rescue); the additional_instructions tell the agent to answer exactly 'Southeast Asia', so short correct answers dominate.", "source": "batch_03.json" } } }, { "task_id": 113, "intent": "Visit the article on the first rapper mentioned on Kid Cudi's page (excluding Kid Cudi himself). How many studio albums has that rapper released (excluding collaborative albums)?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Kid Cudi in the search box and press Enter.\n2. Scan the text to find the first mention of a rapper other than Kid Cudi, which is Kanye West.\n3. Click on the hyperlink of \"Kanye West\".\n4. Navigate to the Studio Albums subsection in his article, which says he has 10 studio albums. Send this as a message to the user: \"Ten\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "ten |OR| 10" ], "scope": "full", "fuzzy_match": "Ten" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Count gold written as a word. Switched from number_match:{value:10} to string_match ['ten |OR| 10']: album discussions naturally use ordinals ('his 10th studio album'), which number_match parses as 10 and would FALSE-ACCEPT a wrong total ('his 10th album ..., but 11 studio albums in all'). string_match refuses the ordinal on the word boundary before 'th', so that near-miss now fails, while 'Ten', '10' and '**10**' all still pass. Word boundaries also keep '10' from matching inside '100' or a year like '2010', and the rapper's name contributes no digits.", "source": "batch_04.json" } } }, { "task_id": 114, "intent": "What is the last word of the first section of the article on Metallica's drummer?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Metallica in the search box and press Enter.\n2. Identify the drummer (Lars Ulrich) and click the hyperlink to his article.\n3. Read the first section of the article (the introduction/summary before the first section header) and identify the very last word of that section, which is \"Marshall\". Send a message to the user: \"Marshall\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "marshall" ], "scope": "full", "fuzzy_match": "Marshall" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single word answer, a proper noun that does not appear in the intent. Word-boundary matching keeps it from firing on 'Marshalls' or 'Marshall-like' continuations of a different token. Unlike task 226, the last word here is not the article's topic, so a wrong last-word answer is unlikely to mention it incidentally.", "source": "batch_05.json" } } }, { "task_id": 115, "intent": "Visit the article on the first country mentioned in the article on Southeast Asia. If that article has a politics section, go to the article on the first politician mentioned in that section. Otherwise, go to the article on the first island mentioned on that country's page. What are the last three words of the main content of the final visited article (excluding references and \"Related Pages\")?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Southeast Asia in the search box and press Enter.\n2. Find the first country mentioned in the text, which is \"East Timor\" and click the link to visit its page.\n3. Check if this country's page has a \"Politics\" section.\n4. Since it doesn't, click on the hyperlink of the first island mentioned in East Timor's article, which is \"Timor\".\n5. Identify the last three words of the text, \"chain of islands\". Send a message to the user: \"chain of islands\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "chain of islands" ], "scope": "full", "fuzzy_match": "chain of islands" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Verbatim three-word span; word-boundary containment rejects the singular 'chain of island' and the reordered 'island chain'. Nothing in the intent leaks the phrase.", "source": "batch_06.json" } } }, { "task_id": 116, "intent": "Trace this path: start at the \"Territory\" page → the first country mentioned → a country that shares its southern border. What is the last word of the first section (the introduction counts as a section) on that final visited page?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Territory in the search box and press Enter.\n2. Scan the text to find the first country mentioned (United States) and click its link.\n3. On the country's page, find the information regarding its borders and identify the country to the south (Mexico). Click on the hyperlink for Mexico.\n4. Read the first section (the introduction) of this final page and identify the very last word, which is \"world\". Send a message to user: \"world\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "world" ], "scope": "first_sentence", "fuzzy_match": "world" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a single common word, so it can occur incidentally in a justification ('... most populous country in the world'). scope:first_sentence confines the match to the verdict sentence, where a correct answer always names the word and a wrong answer names a different one. Medium per trap 3 (low-discriminative gold).", "source": "batch_07.json" } } }, { "task_id": 117, "intent": "Visit the pages of all the countries in the United Kingdom. Which of these countries have a dedicated section on their culture?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type United Kingdom in the search box and press Enter.\n2. Identify the countries that make up the UK, listed in the article (England, Wales, Scotland, Ireland).\n3. Click on the hyperlink for England. As there is a dedicated section on that country's culture, remember this fact.\n4. Go back to the United Kingdom page, and click on the second hyperlink \"Wales\".\n5. As there is no section on that country's culture, go back to the United Kingdom page.\n6. Click on the third hyperlink, \"Scotland\". As there is a section on that country's culture, remember this fact.\n7. Go back to the United Kingdom page, and click on the last hyperlink, \"Ireland\".\n8. As there is no section on that country's culture, tell the users the countries that you remembered, i.e., send a message to the user, \"England, Scotland\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "england" ], [ "scotland" ] ], "ordered": false, "forbidden": [ "^(?!.*(?:\\bnot\\b|\\bno\\b|\\bneither\\b|\\bnor\\b|\\bexcept\\b|\\bexcluding\\b|\\bexclude[sd]?\\b|\\bunlike\\b|\\bwithout\\b|other than|rather than|\\bbesides\\b|aside from|apart from|\\bsave\\b)[^.]{0,12}(?<![a-z0-9])wales).*(?<![a-z0-9])wales(?![a-z0-9])(?!.{0,30}(?:\\bnot\\b|n't|\\bno\\b|\\bnone\\b|\\black|\\babsent\\b|\\bmissing\\b|\\bwithout\\b|\\bexcept\\b)).*$", "^(?!.*(?:\\bnot\\b|\\bno\\b|\\bneither\\b|\\bnor\\b|\\bexcept\\b|\\bexcluding\\b|\\bexclude[sd]?\\b|\\bunlike\\b|\\bwithout\\b|other than|rather than|\\bbesides\\b|aside from|apart from|\\bsave\\b)[^.]{0,12}(?<![a-z0-9])ireland).*(?<![a-z0-9])ireland(?![a-z0-9])(?!.{0,30}(?:\\bnot\\b|n't|\\bno\\b|\\bnone\\b|\\black|\\babsent\\b|\\bmissing\\b|\\bwithout\\b|\\bexcept\\b)).*$" ] }, "fuzzy_match": "England, Scotland" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Set answer: England + Scotland. The only wrong answers that survive must_include(england, scotland) are ones that ALSO list Wales or Ireland as HAVING a culture section, so the forbidden entries target affirmatively-listed Wales/Ireland. A flat forbidden (trap 2) — or the earlier after-only negation window — false-rejected common contrastive correct answers where the negation precedes the country ('England and Scotland, not Wales or Ireland', 'unlike Wales and Ireland'). The forbidden regexes now suppress firing when a negation/contrast marker appears within ~12 chars BEFORE or ~30 chars AFTER the country, catching both sides. Medium because the proximity windows are heuristics.", "source": "batch_06.json" } } }, { "task_id": 118, "intent": "Visit the article about the earliest known statistics mentioned in the Statistics article. List all the countries mentioned on that page.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Statistics in the search box and press Enter.\n2. Locate the mention of the \"first known statistics\", which is \"Census\". Click on the hyperlink on \"Census\".\n3. Scan the content of the visited page for the names of countries and send those names as a message to the user: \"United States, United Kingdom, Australia, Canada, Japan, New Zealand\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "united states |OR| usa |OR| u.s. |OR| u.s.a." ], [ "united kingdom |OR| uk |OR| u.k. |OR| britain" ], [ "australia" ], [ "canada" ], [ "japan" ], [ "new zealand" ] ], "ordered": false, "forbidden": [ "china", "india", "germany", "france", "russia" ] }, "fuzzy_match": "United States, United Kingdom, Australia, Canada, Japan, New Zealand" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Environment-verified against the Census article: the body names exactly these six countries (US, UK, Australia, Canada, Japan, New Zealand) and no others, so the gold is complete. Six distinct names, no order requested, so ordered:false. Abbreviations are offered for the two multi-word states; bare 'us' is deliberately not an alternative because it would match the pronoun. Reviewer added a 'forbidden' list: the original spec accepted an over-lister that names all six gold countries plus an extra (e.g. '..., New Zealand, China'). The forbidden entries are the most likely erroneous census-country additions and are confirmed absent from the page, so a correct answer never contains them.", "source": "batch_07.json" } } }, { "task_id": 119, "intent": "Navigate to the article on the first specialized tool mentioned in the Physics article. Then navigate to the last hyperlink in the main content of that page (excluding references and related pages). How many related pages does the final visited article have?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Physics in the search box and press Enter.\n2. Scan the text to find the first specialized tool mentioned (you will find it in the Approaches in Physics section) and click its link, which is \"Particle Accelerator\".\n3. Click the very last hyperlink found in the text, which is \"Technology\".\n4. On this final page, locate the \"Related Pages\" section and count the number of articles listed in that section, which is two. Send a message to the user: \"Two\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 2 }, "fuzzy_match": "Two" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold 'Two' (related pages on the final article). number_match accepts both spellings. Full scope because agents commonly recap the navigation path first. Medium: an answer that recaps the path may contain other small integers, so a wrong count paired with an incidental '2' would pass.", "source": "batch_08.json" } } }, { "task_id": 120, "intent": "Recursively follow the first link in the main content of each page, starting from the \"Technology\" page. What is the title of the final article you reach before a \"page does not exist\" message appears?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Technology in the search box and press Enter.\n2. Keep clicking the first hyperlink found in the main body text of the visited page. While clicking the hyperlinks, the order of the pages will be Skill, Experience, Learning, Knowledge, Fact, and Real.\n3. As the page on \"Real\" doesn't exist, report the previous page \"Fact\". Send a message to the user: \"Fact\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*(?<!in )(?<!the )(?<!of )(?<!a )(?<![a-z0-9])fact(?![a-z0-9]).*$" ], "scope": "full", "fuzzy_match": "Fact" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 3: the gold 'Fact' is an ordinary English word whose most likely incidental occurrences are the discourse markers 'in fact', 'the fact (that/is)', 'as a matter of fact' and 'a fact' in an answer that names the WRONG page ('Knowledge, because of the fact that Real does not exist'). The regex leaf requires at least one occurrence of the word 'fact' NOT immediately preceded by 'in '/'the '/'of '/'a ', which rejects all four discourse forms while still accepting a correct answer that ALSO uses one of them elsewhere ('In fact, the final article is Fact.' passes on the clean 'is Fact'). The original spec only guarded 'in ', so 'the fact that ... Knowledge' and 'as a matter of fact ... Knowledge' false-accepted (verified). must_exclude ['real'] is deliberately NOT used: a correct answer very often explains that the next link, 'Real', does not exist. Residual, accepted trade-off: an answer that names the page only as 'the Fact article' (definite article directly before 'Fact', no other mention) would be rejected -- uncommon because the task's send-message instruction is the bare token 'Fact'.", "source": "batch_09.json" } } }, { "task_id": 121, "intent": "Recursively follow the first link in the main content of each page, starting from the \"Technology\" page, until you reach an article that does not exist. List the titles of all the valid pages in the order that you visited.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Technology in the search box and press Enter.\n2. Keep clicking the first hyperlink found in the main body text of the visited page. While clicking the hyperlinks, the order of the pages will be Skill, Experience, Learning, Knowledge, Fact, and Real. Remember the names of every page that you have visited.\n3. As the page on \"Real\" doesn't exist, report all the previous pages that you visited based on your memory. Send a message to the user: \"Technology, Skill, Experience, Learning, Knowledge, Fact\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "technology" ], [ "skill" ], [ "experience" ], [ "learning" ], [ "knowledge" ], [ "fact" ] ], "ordered": true, "forbidden": [ "^.*\\bfact\\b[\\s,]*(?:and\\s+)?\\breal\\b(?![^.]*(?:does not|doesn't|doesnt|do not|don't|not exist|no article|no page|is missing|missing|invalid|non-existent|nonexistent|non existent|no such|dead link|dead end|broken|red link|not found|returns nothing|doesn't have|does not have|has no)).*$" ] }, "fuzzy_match": "Technology, Skill, Experience, Learning, Knowledge, Fact" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The intent demands the visiting order, so ordered: true. 'Real' is the first non-existent article and must NOT be listed as a valid page, but a correct answer legitimately names it as the missing one ('... Fact; the next link, Real, does not exist'). The forbidden regex fires when 'Real' follows 'Fact' as a listed item AND is not accompanied (before the next sentence end) by an existence-negation phrase -- this catches both '... Fact, Real' (Real last) and '... Fact, Real, Truth' (extra pages after Real), which the earlier 'Real at end only' rule missed, while still accepting 'Fact, Real (which does not exist)'.", "source": "batch_08.json" } } }, { "task_id": 122, "intent": "What is the name of the TV character whose singing is compared to death metal?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Death Metal\" and press Enter.\n2. According to the article, the TV character whose singing is compared to Death metal is the Cookie Monster.\n3. Send a message to the user: \"Cookie Monster\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "cookie monster" ], "must_exclude": [ "not cookie monster |OR| not the cookie monster |OR| isn't cookie monster |OR| isn't the cookie monster |OR| rather than cookie monster |OR| rather than the cookie monster |OR| instead of cookie monster |OR| instead of the cookie monster" ], "scope": "full", "fuzzy_match": "Cookie Monster" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Unambiguous named entity, no competing character named in the intent. The exclusions cover the ways a wrong answer could quote the gold token while denying it ('... not Cookie Monster', '... not the Cookie Monster', 'instead of ...'), which no correct answer can contain. The 'the'-variants were added because word-boundary matching treats 'not the cookie monster' as distinct from 'not cookie monster', so the bare form alone left a false-accept hole.", "source": "batch_10.json" } } }, { "task_id": 123, "intent": "What type of material is used to make ceramic glass?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Ceramic glass in the search box and press Enter.\n2. As a valid page doesn't exist, send a message to the user: \"There is no article on ceramic glass\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "ceramic glass", "no article |OR| no such article |OR| does not exist |OR| doesn't exist |OR| no valid page |OR| no page |OR| could not find |OR| couldn't find |OR| no results" ], "scope": "full", "fuzzy_match": "There is no article on ceramic glass" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The correct behaviour is to report that no article exists. Requiring 'ceramic glass' plus a not-found phrase rejects any answer that instead invents a material.", "source": "batch_16.json" } } }, { "task_id": 124, "intent": "What is the total area of El Salvador, according to its article?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type El Salvador in the search box and press Enter.\n2. Scan the infobox and the main text for the area of the country.\n3. As the area has not been mentioned, send a message to the user: \"The total area of El Salvador has not been mentioned in the article.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "not been mentioned |OR| not mentioned |OR| not mention |OR| no mention |OR| doesn't mention |OR| never mentions |OR| never mentioned |OR| not listed |OR| not list |OR| doesn't list |OR| not stated |OR| not state |OR| doesn't state |OR| not given |OR| not give |OR| doesn't give |OR| not provided |OR| not provide |OR| doesn't provide |OR| not specified |OR| not specify |OR| doesn't specify |OR| not included |OR| not include |OR| doesn't include |OR| not contain |OR| doesn't contain |OR| not available |OR| not reported |OR| not report |OR| not found |OR| not find |OR| couldn't find |OR| no information |OR| not present |OR| absent |OR| not say |OR| doesn't say |OR| no total area |OR| no area figure |OR| nowhere |OR| no where" ], "must_exclude": [ "^.*\\d[\\d, ]*(km|square kilomet|sq mi|square mile|hectare|acre).*$" ], "scope": "full", "fuzzy_match": "The total area of El Salvador has not been mentioned in the article." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "The gold is a 'the datum is absent' sentence, so the checkable anchor is the negation phrase, not the exact wording (trap 6). The long alternative list covers the phrasings an agent uses for 'the article does not give it'. The must_exclude regex closes the mixed near-miss where the answer negates a DIFFERENT field but still quotes a real area value ('does not state the population, but the total area is 21,041 km2'): it fires on any number immediately followed by an area unit (km/square kilometre/sq mi/hectare/acre). It does not touch correct answers, which never pair a figure with an area unit (they say the figure is absent), and it tolerates rhetorical numbers with no area unit ('gives the population, about 6.5 million, but no area figure'). Medium confidence: a wrong answer that states a bare area value with no unit and no 'total area' anchor could still slip through, but such an answer must also contain a negation phrase to be accepted at all.", "source": "batch_01.json" } } }, { "task_id": 125, "intent": "What is Thom Yorke's first solo album, according to his article?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Thom Yorke in the search box and press Enter.\n2. Scan the article for the name of his first solo album. As it hasn't been mentioned, send a message to the user: \"The first solo album of Thom Yorke has not been mentioned in his article.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "not mentioned |OR| not been mentioned |OR| not mention |OR| no mention |OR| not named |OR| not name |OR| not stated |OR| not state |OR| not specified |OR| not specify |OR| not listed |OR| not list |OR| not say |OR| not identified |OR| not identify |OR| not appear |OR| not provided |OR| not given |OR| not discussed |OR| not indicated |OR| no reference |OR| no information |OR| no solo album |OR| no first solo album |OR| doesn't mention |OR| doesn't name |OR| doesn't say |OR| doesn't state |OR| doesn't specify |OR| doesn't list |OR| doesn't identify |OR| doesn't provide |OR| isn't mentioned |OR| isn't named |OR| isn't specified |OR| isn't listed |OR| isn't stated |OR| never mention |OR| never mentions |OR| never mentioned |OR| never name |OR| never names |OR| never named |OR| never state |OR| never states |OR| never says |OR| never say |OR| never specifies |OR| never lists |OR| not find any |OR| couldn't find any |OR| unable to find any |OR| no album |OR| no such album |OR| has not been mentioned |OR| hasn't been mentioned |OR| have not been mentioned |OR| was not mentioned |OR| wasn't mentioned |OR| not been named |OR| says nothing |OR| say nothing |OR| said nothing |OR| nothing about |OR| mentions nothing |OR| mention nothing |OR| silent |OR| no album is named |OR| no album is mentioned |OR| not referenced |OR| does not reference |OR| doesn't reference |OR| not include |OR| does not include |OR| doesn't include |OR| not included |OR| not documented |OR| does not cover |OR| doesn't cover |OR| not covered |OR| no details |OR| cannot find |OR| can't find |OR| could not find |OR| couldn't find |OR| not able to find |OR| don't mention |OR| do not mention |OR| left unmentioned" ], "scope": "full", "fuzzy_match": "The first solo album of Thom Yorke has not been mentioned in his article." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "The gold is a negative finding: the article never names a first solo album. The checkable anchor is the negation, so must_include lists the realistic ways of expressing 'the article does not mention it' (contractions kept by normalization as \"doesn't\"). The draft's list was expanded to cover common correct phrasings it was missing ('says nothing about', 'is silent on', 'wasn't mentioned', 'does not include the name of'), which were confirmed false-rejects. Any answer that only names an album (The Eraser, Anima, ...) with no negation fails. Downgraded to medium confidence: this is inherently a leaky anchor - a wrong answer that names an album while negating some other detail could false-accept, and an unusual correct phrasing could still fall outside the list. No must_exclude on album titles: a correct answer may add the real album as an aside.", "source": "batch_02.json" } } }, { "task_id": 126, "intent": "What is the name of the latest album by Clipse, according to their article?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Clipse in the search box and press Enter.\n2. As a valid page doesn't exist, send a message to the user: \"There is no article on Clipse\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "clipse", "no article |OR| no wiki article |OR| no such article |OR| no page |OR| no valid page |OR| no results |OR| does not exist |OR| doesn't exist |OR| does not have an article |OR| doesn't have an article |OR| could not find |OR| couldn't find |OR| unable to find |OR| not found |OR| no clipse article |OR| clipse does not have |OR| clipse doesn't have |OR| no matching article |OR| no such page |OR| no wikipedia article |OR| article does not exist |OR| page does not exist |OR| no entry" ], "scope": "full", "fuzzy_match": "There is no article on Clipse" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "The correct answer is an unanswerable-report: no Clipse article exists. Spec = the subject token 'clipse' plus any of the standard 'no such article/page' phrasings. Added 'no clipse article' etc. because the earlier list required 'no article' ADJACENT and would false-reject 'There is no Clipse article' (the subject sits between 'no' and 'article'). The absence phrasings are all about the ARTICLE/PAGE, never a generic 'not available', so a wrong answer that names a hallucinated album and adds 'not available on streaming' is NOT accepted. Contracted forms are spelled with the apostrophe because the matcher only bridges hyphens/whitespace. Medium confidence: a very unusual phrasing of absence ('the wiki has nothing whatsoever on Clipse') could still be rejected.", "source": "batch_03.json" } } }, { "task_id": 127, "intent": "Navigate to the page about the highest point in El Salvador and tell me its height.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type El Salvador in the search box and press Enter.\n2. Scan the article page and infobox to find the name of the highest point (Cerro El Pital).\n3. Click on the hyperlink of Cerro El Pital.\n4. As the article does not exist and the height of the highest point hasn't been mentioned elsewhere, send a message to the user: \"No article on the highest point of El Salvador\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no article |OR| no page |OR| not exist |OR| doesn't exist |OR| not available |OR| unavailable |OR| not found |OR| no height |OR| not mentioned |OR| not listed |OR| not given |OR| unknown |OR| could not find |OR| couldn't find |OR| no information |OR| n/a |OR| missing |OR| no such" ], "must_exclude": [ "^.*\\d[\\d,. ]*(m|meters|metres|meter|ft|feet|foot|km)(?![a-z0-9]).*$" ], "scope": "full", "fuzzy_match": "No article on the highest point of El Salvador" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The correct answer is an unavailability report: the Cerro El Pital page does not exist, so no height can be given. must_include covers the ways an agent phrases 'there is no article / the height is not available'. The must_exclude regex is the other half and is what blocks the hardest near-miss: an answer that admits the article is missing but then supplies a fabricated or externally sourced elevation ('no article for Cerro El Pital, but its height is 2,730 metres'). It only fires on a number immediately followed by a length unit, so an incidental digit (a '404' page, 'the 2 candidate pages') is safe.", "source": "batch_04.json" } } }, { "task_id": 128, "intent": "Find the birth date of the first person mentioned in the scientific method example in the Science article.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Search Science and press Enter.\n2. In the \"Scientific Method\" section's \"An Example\" subsection (as the article is very long, don't scroll, instead click using the element ID), you will find that the first person is Arthur Eddington. Click on his hyperlink and find his birth date.\n3. Then send a message to user, the birth date, \"28 December, 1882\".", "eval": { "eval_types": [ "string_match", "number_match" ], "reference_answers": { "must_include": [ "december |OR| dec |OR| 12" ], "scope": "full", "number_match": { "values": [ 28, 1882 ] }, "fuzzy_match": "28 December 1882" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A date: day and year are checked exactly with number_match (order-independent, so both '28 December 1882' and 'December 28, 1882' pass) and the month with a string anchor. The month alternative accepts 'December', the 'Dec' abbreviation, or the numeric month '12' so a purely numeric date ('28/12/1882', '12/28/1882') is not false-rejected. Adding '12' does not weaken discrimination because number_match still requires BOTH 28 and 1882: a wrong-month answer with the same day/year ('28 November 1882' / '28/11/1882') has no 'december', 'dec' or standalone '12' and is rejected. Extra numbers such as a death year are harmless.", "source": "batch_03.json" } } }, { "task_id": 129, "intent": "According to their respective articles, did Rio de Janeiro host the Summer Olympic Games before Beijing?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Rio de Janeiro in the search box and press Enter.\n2. Find the year it hosted the Summer Olympics, which is 2016, and remember this fact.\n3. Type Beijing in the search box and press Enter.\n4. Find the year it hosted the Summer Olympics, which is 2008, and remember this fact.\n5. Compare the years (2016 is not before 2008). Send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no recipe. The years (2016 vs 2008) are not required because a correct answer may state the verdict alone; first_sentence keeps the year comparison out of scope and blocks 'Yes ... 2016 ... before Beijing' verdicts.", "source": "batch_07.json" } } }, { "task_id": 130, "intent": "Based on the relevant articles, which Australian state uses the Cooktown Orchid as its emblem?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Search \"Australia\" and press Enter.\n2. Click on the hyperlinks of each Australian state one by one.\n3. On each page, scan the \"Symbols\" or \"Emblems\" section (or the infobox) for \"Cooktown Orchid\".\n4. Once you are done visiting a state, go back to the Australia page, and click on the next hyperlink.\n5. Identify that the \"Cooktown Orchid\" is mentioned on the Queensland page as its emblem. Send a message to the user: \"Queensland\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "queensland" ], "must_exclude": [ "^.*\\bnot\\b[^.]{0,14}\\bqueensland\\b.*$", "^.*\\bqueensland\\b[^.]{0,18}\\b(is not|isn't|are not|aren't|does not|doesn't|do not|don't|did not|didn't|was not|wasn't)\\b.*$" ], "scope": "first_sentence", "fuzzy_match": "Queensland" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single named entity that does not appear in the intent. Adversarial pass replaced the draft's blanket must_exclude of the other six states -- that false-rejected a genuinely correct contrastive answer 'It is Queensland, not New South Wales, that uses it' (verified 0.0 under the draft). The two regexes instead forbid only a DENIAL of Queensland ('... not Queensland', 'Queensland does not ...'), so 'Queensland, not Victoria' (correct) passes while 'Victoria uses it, not Queensland' and 'Queensland does not use it' (wrong) are rejected. A bare wrong answer naming only another state fails must_include.", "source": "batch_11.json" } } }, { "task_id": 131, "intent": "How many times in total do you encounter the word \"pigeon\" in these articles: \"Peru\", \"Pusha T\", and \"Doom (2016 video game)\"?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Peru in the search box, press Enter.\n2. Count the occurrences of the word \"pigeon\" in the visited article and remember the count.\n3. Type Pusha T in the search box, press Enter.\n4. Similarly, count the occurrences of the word \"pigeon\" in the visited article, remember the count, and add it to the previous count.\n5. Type Doom (2016 video game) in the search box, press Enter. If you cannot find the article, then first search for Doom, and click on the hyperlink of the Doom (2016 video game).\n6. Similarly, count the occurrences of the word \"pigeon\" in the visited article, remember the count, and add it to the previous count.\n7. As the total count you found is zero for this task, send a message to the user: \"Zero\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "zero |OR| 0 |OR| none |OR| no occurrences |OR| not appear |OR| no mention |OR| no instances" ], "must_exclude": [ "^.*\\b(?:total|sum|altogether|combined|overall)\\b[^a-z0-9]{0,4}(?:count|number|is|of|was|=)?[^a-z0-9]{0,4}(?:[1-9][0-9]*|one|two|three|four|five|six|seven|eight|nine|ten)\\b(?!\\s*(?:articles?|pages?|searches?|results?)).*$" ], "scope": "full", "fuzzy_match": "Zero" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the count zero. number_match on 0 would reject correct answers that never write the digit ('the word appears in none of the three articles'), so string_match with zero-synonyms is used instead. Trap 3 applies: '0' can appear incidentally in a per-article breakdown, so a must_exclude regex kills any answer that states a non-zero aggregate. The leading word is a set (total|sum|altogether|combined|overall) so that a wrong breakdown summed with 'in sum, one occurrence' is also caught, not just literal 'total'; the number class is strictly non-zero, so a correct 'the total is zero' never fires, and the lookahead keeps 'a total of 3 articles' from firing.", "source": "batch_12.json" } } }, { "task_id": 132, "intent": "Which three biogeochemical cycles were mentioned in the article on Biology?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Biology in the search box and press Enter.\n2. Navigate to the section discussing ecosystems or energy/cycles, which is \"Biosphere\".\n3. Identify the specific biogeochemical cycles listed in the text. Send a message to the user: \"nitrogen, carbon, and water\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "nitrogen" ], [ "carbon" ], [ "water" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "nitrogen, carbon, and water" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Three-element enumeration. Order is not requested by the intent, so ordered:false. Only the cycle names are required, so 'the nitrogen cycle' / 'nitrogen' / '**nitrogen**' all match. No forbidden entries: a correct answer could legitimately contrast the three with a cycle it ruled out.", "source": "batch_09.json" } } }, { "task_id": 133, "intent": "How many sections are there in the article on the general theory of relativity (including the introduction section but excluding the Related Pages and See Also sections)?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type General theory of relativity in the search box and press Enter.\n2. Count the Introduction (1) plus the main numbered headers. Ensure \"Related Pages,\" \"See Also,\" and \"References\" are excluded.\n3. The number of sections is three. Send a message to the user: \"Three\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 3 }, "fuzzy_match": "Three" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "A count spelled out in the gold ('Three'); number_match parses both 'three' and '3'. Exact comparison -- a count admits no tolerance. Scope kept at 'full' because a correct answer often states the verdict after naming the section list. Medium confidence for two reasons: the value 3 is low-discriminative (any unrelated gold containing a 3 satisfies this spec, so expect crosstalk), and a verbose answer that enumerates section numbers ('1. Introduction 2. ... 4. ...') would incidentally contain a 3 -- but such an answer that ends on a different total is rare in practice.", "source": "batch_13.json" } } }, { "task_id": 134, "intent": "Visit the first hyperlink in the article \"CryEngine\". What is the name of the country mentioned in the first section of the visited page?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type CryEngine in the search box and press Enter.\n2. Click the very first hyperlink in the main text, which is \"Game engine\".\n3. Read the first section (introduction) for any country names. As none are found, send a message to the user: \"No country mentioned\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no country |OR| no countries |OR| none |OR| ^[\\s\\S]*\\bnot\\b[\\s\\S]{0,30}countr[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "No country mentioned" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Negative finding ('No country mentioned'). Alternatives cover 'no country/countries', a bare 'None', and 'does not mention/name any country'. scope first_sentence prevents a wrong answer that names a country first and then adds 'no other country is mentioned' from passing.", "source": "batch_14.json" } } }, { "task_id": 135, "intent": "Which fast food restaurants were mentioned in the article on hamburgers?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Hamburger in the search box and press Enter.\n2. Scan the entire article for names of major fast food chains.\n3. List the fast food chains found in the text. Send a message to the user: \"Burger King, Wendy's, McDonald's\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "burger king" ], [ "wendy's |OR| wendys |OR| wendy s" ], [ "mcdonald's |OR| mcdonalds |OR| mcdonald s" ] ], "ordered": false }, "fuzzy_match": "Burger King, Wendy's, McDonald's" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Unordered enumeration of three chains; the intent asks 'which ones', not for an order. Apostrophe-less spellings are offered as alternatives because normalization preserves internal apostrophes, so 'Wendys' would not match 'wendy's'. No 'forbidden' list: an answer that adds a further chain the article really does mention should not be punished by a spec written from the gold alone.", "source": "batch_10.json" } } }, { "task_id": 136, "intent": "How many of the hyperlinks from the \"History and geography\" section of the Rio de Janeiro page refer to places (excluding pages that don't exist)?", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type Rio de Janeiro in the search box and press Enter.\n2. Navigate to the \"History and geography\" section.\n3. Visit each hyperlink in this section to see if it refers to a geographical place. Do this by clicking on the hyperlink, checking it, then going back, and clicking on the next hyperlink.\n4. For each visited page, remember if it's a geographic place.\n5. Count the number of geographic places that you have remembered. There should be only one such place, which is Nova Iguaçu. Send a message to the user: \"One\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 1, "scope": "first_sentence" }, "fuzzy_match": "One" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold 'One' is a count; number_match reads both 'One' and '1'. Trap 3: 'one' is a low-discriminative token that can appear incidentally, so the match is restricted to the verdict sentence - an answer that reports a different count in its first sentence ('Two of the links...') fails even if 'one' shows up later in the justification. Medium for the residual case of a first sentence that both states a wrong count and uses 'one' as a pronoun.", "source": "batch_15.json" } } }, { "task_id": 137, "intent": "Visit the article on the last grandmaster mentioned in Magnus Carlsen's article, and tell me the birthplace of that grandmaster.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Magnus Carlsen\" in the search box and press Enter.\n2. Identify the last Grandmaster mentioned in the text, which is Garry Kasparov, and click on his hyperlink.\n3. Identify the place of birth of the grandmaster from his article, which is Baku, Azerbaijan. Send a message to the user \"Baku, Azerbaijan\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "baku" ], "scope": "full", "fuzzy_match": "Baku, Azerbaijan" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Birthplace of Garry Kasparov. Only the city is required: 'Baku' alone is a genuinely correct answer, so demanding 'Azerbaijan' as well would reject it. 'Baku' is unique enough that no wrong birthplace contains it.", "source": "batch_11.json" } } }, { "task_id": 138, "intent": "List all the countries mentioned in the 'Antonio de Cabezón' article, and tell me their names in alphabetical order.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Antonio de Cabezón\" in the search box and press Enter.\n2. Identify names of the countries found: Italy, Spain, Germany, Luxembourg, and Netherlands. Sort these names into alphabetical order.\n3. Send a message to the user \"Germany, Italy, Luxembourg, Netherlands, Spain\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "germany" ], [ "italy" ], [ "luxembourg" ], [ "netherlands" ], [ "spain" ] ], "ordered": true }, "fuzzy_match": "Germany, Italy, Luxembourg, Netherlands, Spain" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Five-country enumeration; the intent explicitly demands alphabetical order, so ordered:true is warranted (trap 5 satisfied). No 'forbidden' entries: a correct answer may legitimately mention a country it ruled out.", "source": "batch_12.json" } } }, { "task_id": 139, "intent": "Start at the \"Valve Corporation\" article and sequentially visit the hyperlinks in that article. Stop when you find a hyperlink where the hyperlink text doesn't match the title of the page it leads to, and tell me the text of that hyperlink.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Valve Corporation\" in the search box and press Enter.\n2. Click the first hyperlink \"video game developer\" and verify whether the page title matches the hyperlink text.\n3. As the texts match, go back and click the second hyperlink \"digital\" and verify whether the page title matches the hyperlink text.\n4. As the texts match, go back and click the third hyperlink \"distributor\".\n5. Observe that the page title (\"distribution\") does not match the hyperlink text (\"distributor\"). Send a message to the user \"distributor\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "distributor" ], "scope": "full", "fuzzy_match": "distributor" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The answer is the hyperlink text 'distributor'. It is not among the anchors named in the additional_instructions ('video game developer', 'digital'), so it is discriminative.", "source": "batch_16.json" } } }, { "task_id": 140, "intent": "Start at the \"Valve Corporation\" article and sequentially visit the hyperlinks in that article. Stop when you find a hyperlink where the hyperlink text doesn't match the title of the page it leads to, and tell me the names of all the hyperlinks that you have visited.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Valve Corporation\" in the search box and press Enter.\n2. Visit the first hyperlink \"video game developer\", confirm the article title matches the hyperlink text, and record the name.\n3. Visit the second hyperlink \"digital\", confirm the article title matches the hyperlink text, and record the name.\n4. Visit the third hyperlink \"distributor\" and observe that the page title does not match the text. Record \"distributor\" and stop the process.\n5. Send a message to the user, the names of the pages that you recorded as visited, \"video game developer, digital, distributor\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "video game developer |OR| videogame developer" ], [ "digital" ], [ "distributor" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "video game developer, digital, distributor" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is the traversal's three hyperlink texts. ordered:false on purpose: the intent asks for 'the names of all the hyperlinks that you have visited', not for them to be reported in order, and agents routinely lead with the stopping link ('I stopped at distributor. The links I visited were video game developer, digital, distributor'), which ordered:true would wrongly reject. Discrimination comes from requiring all three: word-boundary matching means 'distributor' is not satisfied by 'digital distribution', so an agent that stopped one link early or named the wrong stop fails.", "source": "batch_13.json" } } }, { "task_id": 141, "intent": "Start from the article on the country \"Angola\". Sequentially visit the hyperlinks in that article. Stop when you visit a hyperlink that doesn't have a valid page, and tell me the text of that hyperlink.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Angola\" in the search box and press Enter.\n2. Begin clicking hyperlinks in the article sequentially to check for page validity. Do this by clicking on a hyperlink, checking, then going back. Remember the pages that you have visited so that you don't revisit them.\n3. Continue until you click the hyperlink with the text \"coastline\".\n4. Identify that this link leads to an invalid page or a \"red link\" (page does not exist).\n5. Send a message to the user \"coastline\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "coastline |OR| coast line" ], "scope": "full", "fuzzy_match": "coastline" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is the literal link text. Word-boundary matching separates it from the near-miss link text 'coast'; the 'coast line' alternative covers the spaced spelling. The token does not appear in the intent, only in the walkthrough instructions, so no trap-1 issue for the answer.", "source": "batch_01.json" } } }, { "task_id": 142, "intent": "Start at the article \"Crayon\" and serially visit the first five hyperlinks in the article's content. Tell me the titles of the articles that don't have a valid page.", "sites": [ "wiki" ], "start_url": "__WIKI__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Crayon\" in the search box and press Enter.\n2. Identify the first five hyperlinks in the article content.\n3. Click on the first hyperlink and identify that it is a valid page. Remember this fact.\n4. Go back and click on the second hyperlink in the \"Crayon\" page and identify that it is a valid page. Remember this fact.\n5. Go back and click on the third hyperlink in the \"Crayon\" page and identify that it is a valid page. Remember this fact.\n6. Go back and click on the fourth hyperlink in the \"Crayon\" page and identify that it is a valid page. Remember this fact.\n7. Go back and click on the fifth hyperlink in the \"Crayon\" page and identify that it is a valid page. Remember this fact.\n8. Confirm that all five pages load correctly and contain content. Send a message to the user: \"All the articles have valid pages\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*\\b(all|every|each)\\b(?:(?!\\bnot\\b|n't|\\bno\\b|except|invalid|missing|broken|fail)[^.!?]){0,60}\\b(valid|exist|exists|work|works|working|load|loads|loaded|fine|good)\\b.*$" ], "must_exclude": [ "except", "other than", "apart from", "does not have a valid |OR| doesn't have a valid |OR| do not have valid |OR| don't have a valid |OR| did not have a valid |OR| didn't have a valid |OR| does not have a page |OR| doesn't have a page |OR| do not have a page |OR| don't have a page |OR| have no valid page |OR| has no valid page", "does not exist |OR| doesn't exist |OR| do not exist |OR| don't exist", "does not load |OR| doesn't load |OR| do not load |OR| don't load |OR| fails to load |OR| failed to load |OR| not found |OR| error page" ], "scope": "full", "fuzzy_match": "All the articles have valid pages." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "The correct answer is a negative-existence claim ('no article lacks a page'), so there is no entity to anchor on. The regex requires the affirmative 'all/every/each ... valid/exists/loads/fine' claim, and — crucially — its lookahead-guarded window forbids any negation token ('not', \"n't\", 'no', 'except', 'invalid', 'missing', 'broken', 'fail') between the quantifier and the validity word. This kills the earlier false accept 'All five links were checked, and Pigment does not load' (a negation sits between 'all' and 'load'). The must_exclude entries catch the 'all valid EXCEPT one' / 'the rest are all valid but X does not have a page / shows an error page' shapes; they are tied to page/valid/load/exist so they do NOT trip a correct 'I don't have any titles to report' or 'none have any page issues'. Medium confidence: a purely negative correct phrasing that never uses all/every/each (e.g. 'none of them is missing a page') is rejected, and a self-contradictory wrong answer ('all five load correctly, but Pigment is broken') that names no excluded phrase would slip through.", "source": "batch_05.json" } } }, { "task_id": 143, "intent": "Which seems to be the deadlier accident (based on the number of casualties) from these two articles: (1) Serb pilots defend colleague in Air India Express disaster, and (2) Molten steel kills 32 workers in Liaoning, China? Mention the name of the article related to the deadlier accident.", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Serb pilots defend colleague in Air India Express disaster\" in the search box and press Enter.\n2. Scan the search results to identify the casualty count mentioned for this disaster, which is 158. Remember the count.\n3. Type \"Molten steel kills 32 workers in Liaoning, China\" in the search box and press Enter.\n4. Scan the search results to confirm the casualty count for this accident, which is 32. Remember the count.\n5. Compare the two casualty counts to determine which is higher. Since the first one is higher, send a message to the user: \"(1) Serb pilots defend colleague in Air India Express disaster\".", "eval": { "eval_types": [ "llm_judge" ], "reference_answers": { "fuzzy_match": "(1) Serb pilots defend colleague in Air India Express disaster" }, "revision": 1, "annotation": { "confidence": "low", "subjective": true, "gold_suspect": false, "notes": "Environment-verified casualties: Air India Express disaster = 158 killed, Molten steel = 32 killed, so gold (1) Air India is correct. But this is trap 1 at its worst -- BOTH option titles are quoted in the intent, and the winner is signalled by relational language ('the latter', 'the former', 'the first one'), not by which title appears first. The earlier deterministic regex used order-of-mention as a proxy for the winner, and the reviewer showed it both false-REJECTS a correct answer ('Between (2) Molten steel ... and (1) Serb pilots ..., the deadlier is the latter') and false-ACCEPTS the mirror-image wrong answer ('Between (1) Serb pilots ... and (2) Molten steel ..., the deadlier is the latter'). must_include(correct title) alone false-accepts 'the deadlier is Molten steel, not Air India'; must_exclude(wrong title) false-rejects the common 'Air India, not Molten steel' form (trap 2). No deterministic combination separates correct from incorrect here, so this task legitimately stays on llm_judge. Examples below are retained for documentation but are not executed for an llm_judge spec.", "source": "batch_07.json" } } }, { "task_id": 144, "intent": "Which cities are mentioned in any one of these two articles: (1) Australia, NATO enhance ties, and (2) Two pieces of suspected plane debris from Maldives taken to Malaysia for analysis?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Australia, NATO enhance ties\" in the search box and press Enter.\n2. Read the article to identify the city mentioned, which is Brussels, and remember it.\n3. Type \"Two pieces of suspected plane debris from Maldives taken to Malaysia for analysis\" in the search box and press Enter.\n4. Read the article to identify any cities mentioned. You will find that no cities are mentioned in this article, so you don't need to remember anything from here.\n5. Send a message to the user with the city you remembered: \"Brussels\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "brussels" ], "must_exclude": [ "kuala lumpur", "canberra", "male" ], "scope": "first_sentence", "fuzzy_match": "Brussels" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the single city Brussels; it does not appear in the intent. The failure mode a plain must_include cannot catch is an over-answer that adds a hallucinated city from the second (city-free) article, so the leading sentence must not name the plausible false positives: Kuala Lumpur (Malaysia), Malé (folds to 'male', Maldives) and Canberra (Australia). first_sentence scope keeps a trailing explanation such as 'the second article mentions no city' from tripping the exclusions. Medium because the exclusion list is a judgement call and because an answer that explicitly rules Kuala Lumpur out inside its first sentence would be rejected.", "source": "batch_13.json" } } }, { "task_id": 145, "intent": "Which countries are mentioned in both of these articles: (1) Kenya deports Somali businessman, and (2) Rwandan genocide suspect arrested in Uganda?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Kenya deports Somali businessman\" in the search box and press Enter.\n2. Read the article to identify the countries mentioned: Kenya and Somalia. Remember the names of these countries.\n3. Type \"Rwandan genocide suspect arrested in Uganda\" in the search box and press Enter.\n4. Read the article or snippets to identify the countries mentioned: Uganda, Rwanda, Tanzania, Democratic Republic of the Congo, and the United States. Remember the names of these countries.\n5. Compare the lists of countries from both articles. As there are no countries that appear in both articles, send a message to the user: \"None\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^(none|no|neither)\\b[\\s\\S]*$ |OR| ^[\\s\\S]*\\bno\\b[\\s\\S]{0,20}countr[\\s\\S]*$ |OR| ^[\\s\\S]*\\bnot\\b[\\s\\S]{0,25}(share|common|overlap|both|appear)[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "None" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is 'None' (empty intersection). Country names cannot be excluded because a correct answer usually enumerates each article's countries to justify the empty overlap, so the anchor is the negation itself: a verdict that opens with none/no/neither, or says 'no countries', or 'do not share/overlap'. scope first_sentence keeps a wrong answer's trailing 'none of the others overlap' from rescuing it.", "source": "batch_14.json" } } }, { "task_id": 146, "intent": "Compare the first two articles that appear when you search for articles on Morocco. Are these two articles related to the same event?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Morocco\" in the search box and press Enter.\n2. Open the first article in the search results, \"Meteorites in Morocco found to be from Mars\", and analyze the main event described. Remember your analysis.\n3. Open the second article in the search results, \"Morocco raises terror threat level to 'maximum'\", and analyze the main event described. Remember your analysis.\n4. Compare the events to determine whether they refer to the exact same incident. Since they don't, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| unrelated |OR| different" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold. first_sentence isolates the verdict. 'not'/'unrelated'/'different' are accepted alternatives because a correct answer often states the verdict as 'They are not about the same event.' rather than a bare 'No'. must_exclude 'yes' in the same scope kills the affirmative near-miss whose tail mentions the differing topics.", "source": "batch_10.json" } } }, { "task_id": 147, "intent": "Is there any news article regarding San Diego that was published in 2024?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"San Diego\" in the search box and press Enter.\n2. Scan the search results for articles whose publication date contains \"2024\".\n3. As the first search result is from 2024, send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "no" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no recipe from the guidelines: verdict in the leading sentence, 'no' forbidden there. Word-boundary matching keeps 'no' from firing on 'none of the newer ones' appearing later in a justification, which first_sentence excludes anyway. Cross-task leniency against other yes-golds is inherent to a bare 'Yes' gold and cannot be tightened without rejecting the gold itself.", "source": "batch_11.json" } } }, { "task_id": 148, "intent": "Were these two articles published on the same day: (a) Former guitarist of pioneering death metal band Possessed dies in shootout with San Francisco police, and (b) Seven civilians killed in Iraq?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Former guitarist of pioneering death metal band Possessed dies in shootout with San Francisco police\" in the search box and press Enter.\n2. Locate the publication date of the article, which is September 25, 2025, and remember this date.\n3. Type \"Seven civilians killed in Iraq\" in the search box and press Enter.\n4. Locate the publication date of the article, which is also September 25, 2025. Since this date matches the date you remembered, send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "^(?:no|nope)\\b.*$" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/No gold. The must_exclude is anchored (^no|^nope) rather than a bare 'no' token: the wrong answer always opens with the verdict 'No, ...', while a correct 'Yes' answer can legitimately contain a mid-sentence 'no' in the same first sentence ('Yes, both share the same date, so there is no difference'). A bare must_exclude ['no'] false-rejects that correct phrasing; the anchor fixes it without letting any wrong answer through, since a wrong answer that fails to open with no/nope has no 'yes' token and is already caught by must_include.", "source": "batch_12.json" } } }, { "task_id": 149, "intent": "Navigate to the 23rd article in the search results for \"crime patrol\". Then take the first two words of the title of the visited article, search those two words, and report the title of the last article in the search results.", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"crime patrol\" in the search box and press Enter.\n2. Scroll down, click on the third page icon, and find the third result on the third page: \"Federal judge rejects hate crime plea in Arbery killing\".\n3. Identify the first two words of the title, \"Federal judge\".\n4. Type those two words into the search box and press Enter.\n5. Scroll down and click on the last page of the search results, which is page 100.\n6. Identify the last search result on that page and send its title to the user: \"Americans are getting fatter, not-for-profit's report finds\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "americans are getting fatter" ], "scope": "full", "fuzzy_match": "Americans are getting fatter, not-for-profit's report finds" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Article title. The leading clause is highly distinctive and unaffected by how the agent renders the 'not-for-profit's' half (hyphens, apostrophe, 'nonprofit'), so only that clause is required.", "source": "batch_05.json" } } }, { "task_id": 150, "intent": "Follow this navigation path: Search \"Utah\", visit the fourth search result, identify the name of the first city in the body/content of the article, and search its name. Visit the fourth search result and tell me the name of the last state mentioned in the visited article. If you encounter a redirect link at any point, then visit that link. If any of the articles are not found, then report the task as incomplete.", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Utah\" in the search box and press Enter.\n2. Navigate to and click on the fourth search result, which is \"U.S. state of Utah begins to accept evacuees from Hurricane Katrina\".\n3. Read the content of the article to identify the first city mentioned, which is New Orleans.\n4. Type the name of that city into the search box and press Enter.\n5. Navigate to and click on the fourth search result in this new list, \"Total evacuation of New Orleans planned\".\n6. Scan the article to find the last state mentioned in the text, which is Texas.\n7. Send a message to the user: \"Texas\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "texas" ], "scope": "first_sentence", "fuzzy_match": "Texas" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a single state name that does not appear in the intent (the intent names Utah, the additional_instructions also mention New Orleans). first_sentence scope keeps the verdict from being satisfied by a trailing recap of the navigation path. No must_exclude on 'Louisiana': a correct answer may legitimately mention it as the state New Orleans is in.", "source": "batch_15.json" } } }, { "task_id": 151, "intent": "How many articles did you find on India winning the 2025 Champions Trophy?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"India winning the 2025 Champions Trophy\" in the search box and press Enter.\n2. Count the number of relevant articles in the search results. Then send the number as a message to the user: \"One\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 1, "scope": "full" }, "fuzzy_match": "One" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "A count ('One'). number_match parses 'one' and '1' and just requires 1 to be PRESENT, so the incidental '2025' from the intent (2025 != 1) does not cause a false accept. Switched scope from first_sentence to full: first_sentence false-rejected a natural correct answer whose count lands in the second sentence ('I searched ... I found one.') because the first sentence carried only '2025' — verified. Medium because 1 is a low-discriminative value: a wrong answer that mentions a stray '1' (e.g. 'article 1') alongside a different total could false-accept under either scope.", "source": "batch_16.json" } } }, { "task_id": 152, "intent": "How many suspects have been arrested so far in connection with Pheobe Bishop's murder?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Pheobe Bishop murder suspects arrested\" in the search box and press Enter.\n2. Read the search results or top articles to find the current number of arrested suspects. You will find an article at the top, \"Third person charged in relation to Pheobe Bishop's alleged murder\", which means three people have been arrested so far. Check the other search results to confirm that no higher number is reported.\n3. Send a message to the user: \"Three\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 3 }, "fuzzy_match": "Three" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Spelled-out count; number_match parses both 'Three' and '3'. Exact comparison, no tolerance. Ordinal words like 'third' are not parsed as cardinals, so an answer that cites the 'Third person charged' headline while giving a different count still fails. Medium because a stray literal '3' anywhere in the answer would satisfy it; first_sentence scope was rejected because agents often state the evidence before the count.", "source": "batch_01.json" } } }, { "task_id": 153, "intent": "I heard workers died in an accident involving molten steel in China. Can you tell me when the last similar major incident happened in China?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"China molten steel accident workers died\" in the search box and press Enter.\n2. Click on the first search result, \"Molten steel kills 32 workers in Liaoning, China\".\n3. Identify the year of the last similar major incident mentioned, which is 1949. Send a message to the user: \"1949\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "1949" ], "scope": "full", "fuzzy_match": "1949" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Year answer: literal digits, no magnitude ambiguity, and 1949 appears nowhere in the intent. Word-boundary matching keeps it from matching inside a longer number.", "source": "batch_14.json" } } }, { "task_id": 154, "intent": "I heard Utah refused radioactive waste from a Maine plant. When did that plant stop producing electricity?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Utah refuses radioactive waste Maine plant\" in the search box and press Enter.\n2. Click on the first search result, \"Radioactive soil from Maine reactor refused at landfill in Utah\".\n3. Verify whether the visited article mentions when the plant stopped producing electricity. Since this information is present, send a message to the user: \"1997\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "1997" ], "scope": "full", "fuzzy_match": "1997" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a year: literal digits with no magnitude ambiguity, so string_match with must_include is the right verifier (number_match would be equivalent but pointlessly parses every other figure in the answer). Full scope is safe because no competing year is named in the intent, and word-boundary matching stops '1997' from matching inside a longer digit run.", "source": "batch_15.json" } } }, { "task_id": 155, "intent": "I heard Greta Thunberg was named Time magazine's Person of the Year. What is the first article about this?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Greta Thunberg Time magazine Person of the Year\" in the search box and press Enter.\n2. Click on the first search result, \"Greta Thunberg named 2019 Time Person of the Year\".\n3. Identify the publication date of the visited article. Send a message to the user: \"December 18, 2024\".", "eval": { "eval_types": [ "string_match", "number_match" ], "reference_answers": { "must_include": [ "december |OR| dec |OR| 12" ], "number_match": { "values": [ 18, 2024 ] }, "scope": "full", "fuzzy_match": "December 18, 2024" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a publication date but the intent literally asks 'What is the first article about this?' — the date only answers the additional_instructions, so the gold does not answer the stated intent (flagged for human review; spec still follows the gold). Day+year required exactly and the month via december/dec/12, so 'December 18, 2019', 'November 18, 2024' and a title-only answer all fail while 2024-12-18 and 12/18/2024 pass. The article title contains '2019', which is harmless here. | SPOT-CHECK: gold is a publication date but the intent primarily asks WHICH article; only additional_instructions reveal a date was wanted. | GOLD DEFENSIBLE: the additional_instructions direct the agent to report the publication date, which the gold gives. The intent alone reads as 'which article', but agents receive intent + instructions together. Left unchanged.", "source": "batch_06.json" } } }, { "task_id": 156, "intent": "How many articles did you find on Ken Lay (articles that consist only of redirect pages should be excluded)?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Ken Lay\" in the search box and press Enter.\n2. Based on the titles of the search results, you will find the first four search results to be relevant.\n3. Visit the first search result, \"Ken Lay dies in Aspen\". As this is a redirect link, go back.\n4. Visit the second search result, \"Ken Lay dies of heart attack\". As this is a redirect link, go back.\n5. Visit the third search result, \"Enron founder Ken Lay dies in Aspen, Colorado\". This is a valid search result; remember the title of the article.\n6. Go back and visit the fourth search result, \"Former Enron CEO Kenneth Lay dies\". As this is also a redirect link, count the number of valid articles you have remembered, which is one. Send a message to the user: \"One\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "one |OR| 1" ], "must_exclude": [ "^.*(?<![a-z])(?:zero|two|three|four|five|0|2|3|4|5) (?:valid|non[- ]redirect|non redirect|real|actual|genuine|legitimate|proper|distinct|unique|standalone|non[- ]duplicate) articles?(?![a-z]).*$", "^.*(?:there are|there is|leaving|that leaves|leaves|so there are|so we have|so i have|answer is|count is|total of|end up with|end with|arrive at) (?:only |just |a |the )?(?:zero|two|0|2) (?!redirect)articles?(?![a-z]).*$", "^.*(?<![a-z])(?:zero|two|0|2) (?:of (?:the )?(?:them|four|4|results|search results|these|those) )?articles? (?:remain|remains|are left|is left|qualify|qualifies|count|counts|are valid|were valid|are real)(?![a-z]).*$", "^.*(?<![a-z])(?:zero|two|0|2) (?:of (?:the )?(?:them|four|4|results|search results|these|those) )?(?:are|were) (?:valid|real|non[- ]redirect|actual|genuine|legitimate|standalone|non[- ]duplicate)(?![a-z]).*$", "^.*(?<![a-z])(?:only|just) (?:zero|two|0|2)(?![a-z0-9]).*$" ], "scope": "full", "fuzzy_match": "One" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count answer, gold 'One' (one non-redirect article). Switched from number_match to string_match must_include ['one |OR| 1'] because number_match merges adjacent spelled-out numbers ('Just one. Three of the four...' parses 'one three'->4 and loses the 1), false-rejecting valid answers. Trap 3: 'One' is low-discriminative and a correct answer also cites the four results and three redirects, so full scope is required. The dominant wrong answer 'Two' fails must_include outright; the must_exclude regexes then catch the residual case where a wrong final count of two/zero co-occurs with an incidental 'one' ('...two valid articles... one of the four is...'). The exclude targets 'two' only in a VALIDITY context (valid/real/remain/count) so it never fires on the legitimate 'the first two results are redirects'. Medium confidence: the intermediate counts 2/3/4 all appear legitimately, so completeness is impossible; realistic near-misses are covered by negative_examples.", "source": "batch_02.json" } } }, { "task_id": 157, "intent": "Can you tell me the name of the article on the Indian prime minister testing positive for COVID-19?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Indian prime minister testing positive for COVID-19\" in the search box and press Enter.\n2. Review the search results to see whether any credible articles report this as a fact.\n3. As no legitimate reports are found, conclude that no such article exists. Send a message to the user: \"No such article exists\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*\\bno\\b[^.;]{0,25}\\b(?:article|articles|news|report|reports|story|stories|coverage|record|records|result|results|match|matches|source|sources)\\b.*$ |OR| ^.*\\b(?:not|never)\\s+exist\\w*\\b.*$ |OR| ^.*\\b(?:could not|couldn't|couldnt|can not|cannot|can't|cant|did not|didn't|didnt|do not|don't|dont|unable to|failed to|was not able to)\\s+(?:find|locate|identify)\\b[^.;]{0,30}\\b(?:article|articles|news|report|story|coverage|such|any|it|one|match|record|anything|result|results|source|evidence)\\b.*$ |OR| no such |OR| nothing relevant |OR| nothing found |OR| no legitimate |OR| no credible |OR| no evidence |OR| does not exist |OR| doesn't exist" ], "scope": "full", "fuzzy_match": "No such article exists" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trick question: the gold is a 'nothing found' verdict, so the anchor is a negative-existence phrase. Rather than a flat list containing bare 'not find' (which the draft matched inside 'could not find the exact date' next to a hallucinated title -- a false accept), the anchors are regexes that require the negation to be ABOUT an article/report/search: 'no ... article/news/report/...', 'not/never exist...', or 'could not find/locate ... article/such/any/it/coverage/...'. 'yes' is deliberately NOT excluded (the intent opens 'Can you tell me ...', so 'Yes, but no such article exists' is correct). Medium: a wrong answer that names a title AND asserts non-existence of the article in the same breath could still pass.", "source": "batch_08.json" } } }, { "task_id": 158, "intent": "Follow this navigation path: search for \"England,\" visit the second article in the results, then take the first three words of its main content and search them. Visit the fifth search result. What is the first sentence of that article's main content? If you encounter a redirect link at any point, follow the redirect. If any of the articles are not found, report the task as incomplete.", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"England\" in the search box and press Enter.\n2. Navigate to and click on the second article in the results, \"England remembers Bobby Robson\".\n3. Identify the first three words of the article's main content, \"Since his death\".\n4. Type those three words, \"Since his death\", into the search box and press Enter.\n5. Navigate to and click on the fifth search result, \"Serial killer sentenced to death in Swaziland\".\n6. Read the first sentence of the main content and send it as a message to the user: \"A convicted serial killer has been sentenced to death in the African country of Swaziland; his death will be the first execution in the country since 1983.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "swaziland", "1983", "first execution |OR| execution" ], "scope": "full", "fuzzy_match": "A convicted serial killer has been sentenced to death in the African country of Swaziland; his death will be the first execution in the country since 1983." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": true, "notes": "GOLD/ENV MISMATCH (spot-check): the environment's rendered HTML has dropped the phrases 'serial killer' and 'sentenced to death' from the first sentence, which actually reads 'A convicted has been by in the African country of Swaziland; his death will be the first execution in the country since 1983.' Requiring those gold phrases would false-reject a faithful agent, so the spec anchors on tokens that DO render: 'swaziland' + '1983' + 'first execution'. These still reject the bare article title 'Serial killer sentenced to death in Swaziland' (which lacks both 1983 and 'first execution'). Human review: fix the environment HTML or the gold. | ESCALATED: Gold quotes phrases the environment's HTML has dropped ('serial killer', 'sentenced to death'). This is an ENVIRONMENT bug, not a gold error; fixing the rendered article is the real remedy. Spec already anchors on tokens that render. Left flagged, gold unchanged.", "source": "batch_09.json" } } }, { "task_id": 159, "intent": "How many external links are provided in the article \"World Health Organization declares COVID-19 pandemic\"?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"World Health Organization declares COVID-19 pandemic\" in the search box and press Enter.\n2. Open the specific article matching this title.\n3. Count the number of external hyperlinks embedded in the article body. Then send it as a message to the user: \"Six\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 6 }, "fuzzy_match": "Six" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "A count spelled out in the gold; number_match parses both 'Six' and '6'. Scope kept at 'full' so a leading preamble sentence does not hide the count (first_sentence would both hide the count behind a preamble and still admit an incidental '6', so it is strictly worse). Medium confidence per guidelines trap 3: the answer often quotes the article title 'COVID-19' (a 19) and a small integer like 6 can appear incidentally, so an answer that said '6' of something else would be accepted; no stricter deterministic anchor is available for a bare count.", "source": "batch_03.json" } } }, { "task_id": 160, "intent": "Does Stephen Hawking appear in the search results when you search \"Stephen Curry\"?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Stephen Curry\" in the search box and press Enter.\n2. Scan the search result titles for the name \"Stephen Hawking\". The name should appear in the first search result. Send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "no" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no rule. first_sentence isolates the verdict so a justification like 'Yes. Stephen Curry himself is not in the first result.' is not rejected by the must_exclude. Word-boundary matching means the excluded 'no' cannot fire on 'not', 'none' or 'Nobel'.", "source": "batch_13.json" } } }, { "task_id": 161, "intent": "There is an article called \"2007 European football friendly match: England vs. Germany\". Do you think the publication year of the article is consistent with the year when the match took place?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"2007 European football friendly match: England vs. Germany\" in the search box and press Enter.\n2. Open the article and identify the publication date (for example, by looking for the timestamp). It should be December 18, 2024.\n3. Based on the title of the article, we can conclude that the actual date of the event is 2007. As the actual date is inconsistent with the publication date, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| inconsistent |OR| mismatch |OR| discrepancy" ], "must_exclude": [ "yes", "^[\\s\\S]*\\bno (mismatch|discrepancy|inconsistency|inconsistencies)\\b[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No'). first_sentence + exclude 'yes' is the standard recipe; the include list also accepts the common phrasing 'the publication year is not consistent with the match year' that never uses a bare 'no'. The positive-signal words (mismatch/discrepancy/inconsistent) can be negated to mean the OPPOSITE ('there is no discrepancy' = consistent = wrong), so the second must_exclude rejects a first sentence that says 'no discrepancy/mismatch/inconsistency'. A correct answer never asserts an absent discrepancy, so this only closes a false-accept hole; plainly-wrong 'the years are consistent' already fails must_include (none of the tokens appear).", "source": "batch_14.json" } } }, { "task_id": 162, "intent": "Did the 2024 article, \"Indian flooding displaces thousands\", get published before the 2024 article \"Satellite radio approved in Canada\"?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Indian flooding displaces thousands 2024\" in the search box and press Enter.\n2. Remember the specific publication date of the article, December 18, 2024.\n3. Type \"Satellite radio approved in Canada 2024\" in the search box and press Enter.\n4. Remember the specific publication date of that article, December 17, 2024.\n5. Compare the two dates that you remembered to see if the flooding article was published earlier. Send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Straight yes/no gold. first_sentence scope isolates the verdict from the date justification that follows, and word-boundary matching keeps 'no' from firing on 'not'/'north'. A correct answer may quote both December dates afterwards without affecting the match.", "source": "batch_15.json" } } }, { "task_id": 163, "intent": "Based on the article, \"Allahabad High Court grants bail to accused linked to Umesh Pal murder case\", when was Umesh Pal shot?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Allahabad High Court grants bail to accused linked to Umesh Pal murder case\" in the search box and press Enter.\n2. Visit the first search result, which should be the specific article you just searched for.\n3. Read the text to find the date mentioned for when Umesh Pal was shot. Then send it as a message to the user: \"February 24, 2023\".", "eval": { "eval_types": [ "string_match", "number_match" ], "reference_answers": { "must_include": [ "february |OR| feb" ], "number_match": { "values": [ 24, 2023 ] }, "fuzzy_match": "February 24, 2023" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A date splits cleanly into a month token plus two exact numbers, which also makes the spec order-independent ('February 24, 2023' and '24 February 2023' both pass). A single must_include of 'february 24, 2023' would demand the literal comma, so it is avoided.", "source": "batch_07.json" } } }, { "task_id": 164, "intent": "Navigate to the article \"Hurricane Emily is second strike for Jamaica\", find the name of the last country mentioned, search for the name of that country, and tell me the title of the last search result.", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Hurricane Emily is second strike for Jamaica\" in the search box and press Enter.\n2. Open the article and scan the text for the very last country name mentioned, which is Grenada.\n3. Type \"Grenada\" in the search box and press Enter.\n4. Click on the last page (if there are multiple pages in the search results).\n5. Scroll to the last result on the last page. Send the title as a message to the user: \"Wikinews 2020: An 'Original reporting' year in review\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "wikinews 2020", "original reporting", "year in review" ], "scope": "full", "fuzzy_match": "Wikinews 2020: An 'Original reporting' year in review" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The gold is an article title. Matching it as one literal would be brittle (the internal quotes and colon would have to survive verbatim), so it is split into three distinctive fragments that together identify only this title. 'wikinews 2020' also rejects the neighbouring year-in-review titles, which is the hardest near-miss here.", "source": "batch_10.json" } } }, { "task_id": 165, "intent": "Based on these two articles: (1) \"At least eight dead after drone strikes in Pakistan\" and (2) \"Tornado touches down in Joplin, Missouri\", which appears to describe the deadlier incident?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"At least eight dead after drone strikes in Pakistan\" in the search box and press Enter.\n2. Identify the casualty count mentioned in the article, which is eight. Remember the count.\n3. Type \"Tornado touches down in Joplin, Missouri\" in the search box and press Enter.\n4. Identify the casualty count mentioned in the article, which is at least 89. Remember the count.\n5. Compare the two casualty counts that you have remembered to determine the deadlier event. Send a message to the user: \"(2) Tornado touches down in Joplin, Missouri\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "tornado |OR| joplin" ], "must_exclude": [ "^.*(?:\\bdrone\\b|\\bpakistan\\b|\\(1\\))(?:(?!\\btornado\\b|\\bjoplin\\b|\\bnot\\b|\\bnor\\b|\\bless\\b|\\bfewer\\b|[.;]).)*?\\b(?:deadlier|deadliest|more lethal|most lethal|more fatal|more deadly|most deadly|more casualties|more deaths|more fatalities|higher (?:death|casualty|fatality)|greater (?:death|casualty)|bloodier)\\b.*$", "^.*\\b(?:deadlier|deadliest|more lethal|most lethal|more fatal|more deadly|most deadly|more casualties|more deaths|more fatalities|higher (?:death|casualty|fatality)|greater (?:death|casualty)|bloodier)\\b(?:(?!\\btornado\\b|\\bjoplin\\b|[.;]).)*?\\b(?:is|was|were|being|appears? to be|seems? to be)\\s+(?:(?!\\bnot\\b)[a-z0-9 ()]){0,18}?(?:\\(1\\)|\\bdrone\\b|\\bpakistan\\b).*$" ], "scope": "full", "fuzzy_match": "(2) Tornado touches down in Joplin, Missouri" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1: both article titles are quoted in the intent, so a wrong answer also contains 'tornado'/'joplin'. must_include['tornado |OR| joplin'] rejects most wrong answers outright (an agent picking incident (1) usually names only the drone strikes). The two regex exclusions catch the remaining comparative wrong answers that assert the Pakistan drone strikes are the deadlier one, in either word order, using a broadened synonym set (deadlier/deadliest/more lethal/more deaths/higher death toll/bloodier/...). Both exclusions block the competing terms 'tornado'/'joplin' and 'not' between the drone mention and the superlative, so correct answers that justify the verdict by citing the 8-vs-89 counts ('the drone strike killed 8 and the tornado was deadlier') and the negated form ('the deadlier one is not the drone strikes but the tornado') are not false-rejected. Medium: an assertion of drone superiority phrased without any covered synonym could slip through.", "source": "batch_08.json" } } }, { "task_id": 166, "intent": "How many people were killed in the 2024 Joplin tornado?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"2024 Joplin tornado\" in the search box and press Enter.\n2. Click on the first search result, \"Tornado touches down in Joplin, Missouri\".\n3. Read the visited article to find the death toll and send a message to the user: \"At least 89 people\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 89 }, "fuzzy_match": "At least 89 people" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Death toll ('At least 89 people'). number_match requires 89 regardless of the 'at least' framing.", "source": "batch_16.json" } } }, { "task_id": 167, "intent": "Based on the article, \"Pilots spot 'UFOs' near the Channel Islands\", is it true that the UFOs are said to have come from Neptune?", "sites": [ "news" ], "start_url": "__NEWS__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Pilots spot 'UFOs' near the Channel Islands\" in the search box and press Enter.\n2. Open the article and read the details regarding the origin of the UFOs.\n3. Check if there is any claim that they came from Neptune. As there are no such claims, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| false |OR| does not |OR| doesn't |OR| no claim |OR| no such claim |OR| incorrect |OR| untrue" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No'). first_sentence scope with a negation and must_exclude 'yes' follows the yes/no recipe; the negated-verb alternatives accept 'The article makes no such claim'.", "source": "batch_16.json" } } }, { "task_id": 168, "intent": "What unique sizes are available for all the men's and women's jeans?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"jeans\" in the search box and press Enter.\n2. Based on the titles, open each product page for men's and women's jeans. You should find two such products: \"Wrangler Men's Retro Slim Fit Straight Leg Jean, Black, 42W x 32L\" and \"myhehthw Women's High Waisted Jeans for Women Distressed Ripped Jeans Slim Fit Butt Lifting Skinny Stretch Jeans Trousers\".\n3. Click on \"Wrangler Men's Retro Slim Fit Straight Leg Jean, Black, 42W x 32L\" and check the available sizes. There should be no sizes available.\n4. Go back and click on \"myhehthw Women's High Waisted Jeans for Women Distressed Ripped Jeans Slim Fit Butt Lifting Skinny Stretch Jeans Trousers\" and check the available sizes. Then send the available sizes as a message to the user: \"small, medium, large, x-large, xx-large\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "small" ], [ "medium" ], [ "^.*(?<![a-z0-9-])(?<!x )(?<!extra )large(?![a-z0-9]).*$" ], [ "x-large |OR| xl |OR| extra large |OR| extra-large" ], [ "xx-large |OR| xxl |OR| 2xl |OR| extra extra large" ] ], "ordered": false }, "fuzzy_match": "small, medium, large, x-large, xx-large" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Unordered set of five sizes. Plain 'large' is a regex because the substring 'large' also sits inside 'x-large'/'x large'/'extra large'/'extra-large' (the word-boundary guard only blocks letters/digits, not the hyphen). Adversarial review: the earlier regex still matched 'large' inside 'extra large', so an answer that omitted the plain-large size but named x-large as 'extra large' and xx-large as 'extra extra large' falsely passed. Added a (?<!extra ) guard (the hyphen guard already blocks 'extra-large'). No 'forbidden' on the men's 42W x 32L jean: a correct answer may mention it has no size options.", "source": "batch_06.json" } } }, { "task_id": 169, "intent": "What colors are available for the Men's Running Tennis shoes?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Men's Running Tennis shoes\" in the search box and press Enter.\n2. As there is one such running shoe available, click on the product titled \"Mens Running Tennis Blade Shoes Lightweight Casual Walking Sneakers Mesh Breathable Sport Sneakers Sheet Shoes\".\n3. There are three colors available on the product page. Send the colors as a message to the user: \"black, grey, khaki\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "black" ], [ "grey |OR| gray" ], [ "khaki" ] ], "ordered": false, "forbidden": [ "white", "blue", "red", "green", "brown" ] }, "fuzzy_match": "black, grey, khaki" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Environment-verified against the product's Color option list, which is exactly Black / Grey / Khaki -- no other colour is offered, so the gold is complete. Three colour names, unordered. 'gray' is offered as the US spelling of 'grey'. All three must appear, so dropping or substituting one scores 0. Reviewer added a 'forbidden' list of common colours the product does not offer: the original spec accepted an over-lister that gives the three correct colours plus an extra ('black, grey, khaki, and white'). A correct answer never names a colour that is not an available option.", "source": "batch_07.json" } } }, { "task_id": 170, "intent": "Are the FIN86 Men's Elastic High Waisted Shorts made of the same material as the Yinimo Men's Gym Shorts? If not, how do they differ?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"FIN86 Men's Elastic High Waisted Shorts\" in the search box and press Enter.\n2. Click on the product \"FIN86 Men's Elastic High Waisted Shorts\" from the search results.\n3. Note the material listed in the description, which is Spandex. Remember this fact.\n4. Type \"Yinimo Men's Gym Shorts\" in the search box and press Enter.\n5. Click on the product \"Yinimo Men's Gym Shorts\" from the search results.\n6. Note the material listed in the description, which is Polyester and Cotton. Remember this fact.\n7. Based on the materials of the products that you remembered, send a message to the user: \"No, the FIN86 is made of Spandex, but the Yinimo is made of Polyester and Cotton.\"", "eval": { "eval_types": [ "string_match", "list_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| differ |OR| differs |OR| different |OR| distinct |OR| unlike |OR| dissimilar" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "list_match": { "items": [ [ "spandex" ], [ "polyester |OR| poly cotton" ], [ "cotton" ] ], "ordered": false }, "fuzzy_match": "No, the FIN86 is made of Spandex but the Yinimo is made of Polyester and Cotton." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Two-part answer: a negative verdict plus the two materials. string_match checks the verdict in the leading sentence only; adversarial pass added 'distinct/differs/unlike/dissimilar' so a correct 'They use distinct fabrics: ...' is not false-rejected, and added a 'poly cotton' alternative so 'a poly-cotton blend' still satisfies the polyester item. must_exclude 'yes' catches 'Yes, same material' and 'Yes and no' hedges. list_match runs over the full answer for spandex + polyester + cotton. Medium (was high in draft): the list_match cannot check attribution, so a swapped 'No, the FIN86 is polyester/cotton but the Yinimo is spandex' still names all three materials and would pass -- a plausible agent error not separable without fragile proximity checks.", "source": "batch_11.json" } } }, { "task_id": 171, "intent": "Does the longest variant of the Lindy DisplayPort to HDMI cable exceed the length of the longest variant of the Kodak HDMI Mini Cable?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Lindy DisplayPort to HDMI cable\" in the search box and press Enter.\n2. Click on the product \"LINDY 36920 0.5m DisplayPort to HDMI 10.2G Cable, Black\" from the search results.\n3. Identify the longest length variant available, which is 5 m. Remember this fact.\n4. Type \"Kodak HDMI Mini Cable\" in the search box and press Enter.\n5. Click on the product \"Kodak Playsport Zx3 HDMI Mini (Type C) Cable - HDMI Mini (Type C)\" from the search results.\n6. Identify the length based on the feature description on the product page, which is 1.8 m. Remember this fact.\n7. Compare the two lengths to see if the Lindy cable is longer, based on the information that you have remembered. Since it is, send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "^(?:no|nope)\\b.*$" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/No gold. must_exclude is anchored (^no|^nope) rather than a bare 'no' so that a correct 'Yes' answer can carry a mid-sentence 'no' in the same first sentence ('Yes, the Lindy has a 5 m variant, no shorter than the Kodak cable') without false-rejecting. Every wrong answer opens with 'No, ...'; one that does not open with no/nope lacks a 'yes' token and is already excluded by must_include.", "source": "batch_12.json" } } }, { "task_id": 172, "intent": "What is the total price of all the toy cars available in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"toy cars\" in the search box and press Enter.\n2. Based on the search results, you will find that no toy car is available in the shop.\n3. Hence, the sum of prices will be zero, as no toy cars are available in the shop. Send a message to the user: \"Zero (no toy cars are available in the shop)\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "zero |OR| 0 |OR| none |OR| nothing |OR| no toy cars |OR| no toy car" ], "scope": "full", "fuzzy_match": "Zero (no toy cars are available in the shop)" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Total price is zero because the shop stocks no toy cars. string_match rather than number_match so that '$0.00', 'Zero' and 'nothing' all pass without demanding the parenthetical justification of the gold (number_match would false-reject 'None'/'nothing', which carry no parseable digit). Digit-boundary matching means real totals like $130.00 or $45.97 do not match the '0' alternative (verified with the oracle). Medium: a hypothetical wrong sub-dollar total of $0.99 would match '0', but no set of toy cars can produce that.", "source": "batch_11.json" } } }, { "task_id": 173, "intent": "How many grams of protein are in a standard serving of Schoolyard cheese puffs?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Schoolyard cheese puffs\" in the search box and press Enter.\n2. Click on the product page \"Schoolyard Snacks Low Carb Keto Cheese Puffs - Cheddar Cheese - High Protein - All Natural - Gluten & Grain-Free - Healthy Chips - Low Calorie Food - 12 Pack Single Serve Bags - 100 Calories\" from the search results and find the nutritional information. In the features, you should find the amount per serving, which is 14 grams.\n3. Send a message to the user: \"14 grams\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 14 }, "fuzzy_match": "14 grams" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Straight quantity from a product's nutrition information; exact match, no tolerance. Quoting the long product title brings in 12 (pack size) and 100 (calories), which is harmless because number_match only requires 14 to be present and 14 appears nowhere else in the title. Spelled-out 'fourteen' is parsed too.", "source": "batch_02.json" } } }, { "task_id": 174, "intent": "Is it recommended to dry clean the Burband Women's High Waist Bike Shorts?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Burband Women's High Waist Bike Shorts\" in the search box and press Enter.\n2. Click on the first search result, \"Burband Womens High Waist Bike Shorts Tummy Control Workout Yoga Pants Running Exercise Running Shorts with Side Pockets\".\n3. Read the \"Care Instructions\" in the product description and features.\n4. Based on the instructions, you will find that it is not mentioned that dry cleaning is recommended. Send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not |OR| never |OR| isn't |OR| doesn't |OR| don't" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold 'No'. The must_include is widened past the bare 'no' because the natural correct phrasings are negations rather than a leading 'No' ('Dry cleaning is not recommended', 'the care instructions don't mention dry cleaning'). must_exclude['yes'] under first_sentence is what rejects the affirmative near-miss that also contains a 'not' ('Yes — it says dry clean only, so it is not machine washable').", "source": "batch_13.json" } } }, { "task_id": 175, "intent": "Does the description of the XXBR Summer T-shirts mention the material type?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"XXBR Summer T-shirts\" in the search box and press Enter.\n2. Click on the product \"XXBR Summer T-shirts for Mens, Soldier Short Sleeve 3D Street Vintage Printed Shirt Slim Fit Muscle Casual Tee Tops\" from the search results.\n3. Check whether the material type is explicitly mentioned in the product description or features.\n4. Since it is mentioned, send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes |OR| does mention |OR| is mentioned |OR| does specify |OR| is specified |OR| does state |OR| is stated |OR| does list |OR| is listed" ], "must_exclude": [ "^no\\b[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('Yes'). Plain must_exclude ['no'] would reject correct answers such as 'Yes, cotton blend, no polyester', so the exclusion is a regex for a first sentence that *opens* with 'No'. Affirmative paraphrases are matched contiguously ('is mentioned' cannot match 'is not mentioned').", "source": "batch_14.json" } } }, { "task_id": 176, "intent": "What is the price of the most expensive bookshelf?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"bookshelf\" in the search box and press Enter.\n2. Check the prices of the bookshelves in the search results. If there are multiple pages, visit the other pages as well.\n3. Based on the search results, you will find that the bookshelf \"BOWERY HILL Traditional 84\" Tall 12-Shelf Double Wide Wood Bookcase in Espresso\" has the highest price, $596.71.\n4. Send this price as a message to the user: \"$596.71\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 596.71, "abs_tolerance": 0.01 }, "fuzzy_match": "$596.71" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A price. number_match ignores the currency symbol and any other numbers in the product name (84\", 12-shelf). abs_tolerance 0.01 is the standard currency slack: it absorbs a cent-level rounding of the same price ($596.70) but not a different one -- the transposed near-misses $596.17 and $569.71 both differ by far more than a cent and are rejected.", "source": "batch_03.json" } } }, { "task_id": 177, "intent": "What's the warranty period of the Samsung remote for the UN85S9VF model?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Samsung remote UN85S9VF\" in the search box and press Enter.\n2. Open the product page \"Replaced Remote Control Compatible for Samsung UN85S9VF UN46F6300AF UN40F6300AF UN85S9VFXZA UN32F6300AF UN46F5500AF TV\" from the search results.\n3. Check the warranty period in the product description and features.\n4. In both places, you will find a 30-day quality warranty. Send a message to the user: \"30 days\".", "eval": { "eval_types": [ "string_match", "number_match" ], "reference_answers": { "must_include": [ "day |OR| days" ], "number_match": { "value": 30 }, "scope": "full", "fuzzy_match": "30 days" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Duration gold: the magnitude and the unit both matter, so number_match pins 30 and string_match pins the unit. This rejects '1 year (365 days)' (right unit, wrong number) and '30-month warranty' (right number, wrong unit). '30-day' satisfies both entries because the containment pattern treats the hyphen as a separator and number parsing ignores it. The model code UN85S9VF contributes no numbers -- its digits are preceded by word characters, which the number regex refuses.", "source": "batch_04.json" } } }, { "task_id": 178, "intent": "Does the Samsung remote for the UN85S9VFXZA model contain batteries?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Samsung remote UN85S9VF\" in the search box and press Enter.\n2. Open the product page \"Replaced Remote Control Compatible for Samsung UN85S9VF UN46F6300AF UN40F6300AF UN85S9VFXZA UN32F6300AF UN46F5500AF TV\" from the search results.\n3. Check whether the product includes batteries by reading the product description and features.\n4. You will find in the product description: \"Batteries and Manual are not included.\" Send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Straight yes/no gold ('Batteries and Manual are not included'). first_sentence scope is essential here: the justification quotes the product description, and the hardest near-miss is a wrong 'Yes' verdict followed by that same quotation, which the exclusion catches inside the verdict sentence.", "source": "batch_15.json" } } }, { "task_id": 179, "intent": "What's the material of the Yinimo Men's Gym Shorts?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Yinimo Men's Gym Shorts\" in the search box and press Enter.\n2. From the search results, click on the product \"Yinimo Mens Gym Shorts Fashionable Loose Beach Drawstring Turkey Print Beach Cool Shorts\".\n3. In the features, you will find the material composition. Send it as a message to the user: \"Polyester and Cotton\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "polyester" ], [ "cotton" ] ], "ordered": false, "forbidden": [ "^(?:(?!\\bnot\\b|\\bno\\b|without|excluding|rather than|instead of|besides|aside from|other than).)*\\b(?:spandex|elastane)\\b.*$" ] }, "fuzzy_match": "Polyester and Cotton" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Two-element material list (95% polyester / 5% cotton); intent imposes no order so ordered stays false. Both fibres must be named, which rejects '100% polyester' and 'polyester and spandex'. Added a guarded forbidden on 'spandex'/'elastane' -- the most common blend component an agent hallucinates as a third material -- so 'polyester, cotton, and spandex' is rejected as an extra-item false accept; the leading '(?!not|no|without|excluding|...)' guard still accepts a correct contrastive answer such as 'polyester and cotton, not spandex'.", "source": "batch_08.json" } } }, { "task_id": 180, "intent": "Is the Goya Foods wafer marketed as a Latino snack?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Goya Foods wafer\" in the search box and press Enter.\n2. Click on the product \"Goya Foods Chocolate & Vanilla Wafers, 4.9 Ounce (Pack of 24)\" from the search results.\n3. Read the product description. You will find that the product has been described as a \"Latino snack\". Send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes |OR| indeed |OR| correct |OR| marketed as a latino |OR| true |OR| affirmative" ], "must_exclude": [ "no", "not" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('Yes'). Mirror of the yes/no recipe, excluding both 'no' and 'not' (a correct affirmative first sentence needs neither). 'it is' was dropped as an alternative because it is a substring of 'it is not'.", "source": "batch_16.json" } } }, { "task_id": 181, "intent": "I'm looking to buy a jute pouf. Can you please search the shop and let me know how many such products are available?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"jute pouf\" in the search box and press Enter.\n2. Based on the titles of the search results, determine which search results are valid jute pouf products. Visit the other pages if there are multiple pages of search results. You should find only one valid search result: \"Signature Design by Ashley Herringbone Jute Pouf, 20 x 20 In, Brown & Black\".\n3. Send a message to the user: \"One\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 1 }, "fuzzy_match": "One" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count answer; number_match accepts 'One', 'one', '1' and 'a single product' is the only phrasing it would miss. Exact match, no tolerance. Medium confidence: '1' is low-discriminative, so an enumerated answer ('1. Signature Design ... 2. ...') could satisfy it incidentally. The dimensions in the product title (20 x 20 In) are harmless because they parse to 20, not 1.", "source": "batch_01.json" } } }, { "task_id": 182, "intent": "How many socks are available in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"socks\" in the search box and press Enter.\n2. Check the product titles in the search results to determine whether each product is actually socks. Visit the other pages too, if it is a multi-page search result.\n3. You will find five sock products. Send a message to the user: \"Five\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "five |OR| 5" ], "must_exclude": [ "^.*(?<![a-z])(?:one|two|three|four|six|seven|eight|nine|ten|1|2|3|4|6|7|8|9|10) (?:(?!\\bnot\\b|\\bpairs?\\b|\\bpages?\\b|\\bfor\\b).){0,25}?\\b(?:are|is|were|count as|counts as|qualify as|qualifies as)\\b (?:(?!\\bnot\\b).){0,15}?socks?(?![a-z])(?!\\s+for\\b).*$", "^.*(?<!not )(?<![a-z])(?:one|two|three|four|six|seven|eight|nine|ten|1|2|3|4|6|7|8|9|10) (?:actual |actually |real |genuine |true |proper )?socks?(?![a-z])(?!\\s+for\\b).*$" ], "scope": "full", "fuzzy_match": "Five" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count answer, gold 'Five'. Switched from number_match to string_match must_include ['five |OR| 5'] because number_match merges adjacent spelled-out numbers ('There are five: three pairs...' parses 'five three'->8 and loses the 5), false-rejecting valid answers. Trap 3: a wrong answer often keeps the gold number ('five results, but only four are socks'). The first must_exclude regex rejects a wrong count stated affirmatively as socks ('four are/qualify as/count as socks'); it forbids 'not'/'pair'/'page' in the gap so it never fires on the correct 'the other four results are not socks' or 'three pairs of socks'. The second rejects a bare 'N sock(s)' count ('four sock products'), excluding 'socks for ...'. Medium confidence because of the reasoned must_exclude; realistic near-misses are in the negative_examples.", "source": "batch_02.json" } } }, { "task_id": 183, "intent": "How many search results appear when you search the word \"cat\" in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"cat\" in the search box and press Enter.\n2. Check the number of search results displayed. It should say four. Send a message to the user: \"Four\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 4 }, "fuzzy_match": "Four" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "A small count spelled out in the gold; number_match handles 'Four' and '4' alike. Medium confidence per guidelines trap 3: a small integer can occur incidentally in a verbose answer, so an answer naming '4' of something other than the result count would be accepted; nothing stricter is possible for a bare count.", "source": "batch_03.json" } } }, { "task_id": 184, "intent": "Is there any red HDMI cable in the shop? If yes, order a single unit and tell me the confirmation code. Otherwise, say that none are available.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"red HDMI cable\" in the search box and press Enter.\n2. Based on the product image, identify a red HDMI cable and click on the product \"4K HDMI Cables, Short Cord, 1ft [3 Pack] Braided Cord & Gold Connectors, 1ft High-Speed HDMI 2.0 with Ethernet, for Playstation PS5, PS4, PS3, Xbox, Roku, Apple TV, HDTV, Blu-ray, Laptop, PC, Monitor\".\n3. Select the color option \"red-braided\" and the \"single pack\" option.\n4. Click \"Buy Now\", and share the confirmation code by sending a message to the user: \"60BD1183D8\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "60bd1183d8" ], "scope": "full", "fuzzy_match": "60BD1183D8" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Opaque confirmation code. The 'otherwise say none are available' branch is the wrong branch here, and any answer taking it necessarily lacks the code.", "source": "batch_05.json" } } }, { "task_id": 185, "intent": "How many rugs are available to buy?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"rugs\" in the search box and press Enter.\n2. Based on the product titles, verify which items are actually rugs. There should be two such products. Send a message to the user: \"Two\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "two |OR| 2" ], "scope": "full", "fuzzy_match": "Two" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Bare count gold. Switched from number_match:{value:2} to string_match ['two |OR| 2']: string_match refuses the ordinal reading ('the 2nd result is a rug, 3 rugs in total') that number_match parses as 2 and would FALSE-ACCEPT, while still matching 'Two', '2' and '**2**'. Medium per trap 3: 2 remains a low-discriminative value, so a wrong answer that enumerates its items with digit markers ('three rugs: 1. ... 2. ... 3. ...') still leaks a standalone 2 -- that residual is unavoidable without rejecting the bare gold 'Two'. Adding a 'rug' token is impossible for the same reason (the gold contains no such word). full scope avoids rejecting answers that open with a preamble sentence before the count.", "source": "batch_04.json" } } }, { "task_id": 186, "intent": "Is there any bikini with a wrap skirt available in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"bikini wrap skirt\" in the search box and press Enter.\n2. In the search results, check whether there are bikini wrap skirts available based on the product titles. You should find at least one such product. Then send a message to the user: \"Yes\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "no" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no rule: verdict scoped to the leading sentence so a justification like 'not available in every size' cannot flip it. Word-boundary matching means excluding 'no' does not exclude 'not' or 'none of the other results'.", "source": "batch_01.json" } } }, { "task_id": 187, "intent": "Can I purchase the Dota 2 video game from this shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Dota 2\" in the search box and press Enter.\n2. Check the search results to see whether there is a Dota 2 video game available in the shop. As there isn't, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no recipe. 'no' matches on word boundaries so 'not', 'none' and 'nothing' do not satisfy it accidentally; first_sentence keeps the justification ('the shop only sells Dota 2 merchandise') out of scope.", "source": "batch_02.json" } } }, { "task_id": 188, "intent": "How many electric scooters are available in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"electric scooter\" in the search box and press Enter.\n2. Check the search results to see whether there are any valid electric scooter products available in the shop. As there aren't, send a message to the user: \"Zero\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "zero |OR| none |OR| no electric scooter |OR| no electric scooters |OR| no scooter |OR| no scooters |OR| not any electric |OR| aren't any electric |OR| are no electric |OR| ^.*(?<![\\w.])0(?![\\w.]).*$" ], "scope": "full", "fuzzy_match": "Zero" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the count 'Zero'. number_match is unusable on its own because the equally correct 'There are no electric scooters available' contains no parsable number, so string_match with zero-synonyms is used. The bare-digit case ('**0**', 'The answer is 0.') is matched by a regex leaf that requires a standalone 0 not adjacent to a word char OR a decimal point, so it will NOT fire on a wrong answer that quotes a one-decimal price ('$199.0') the way a plain '0' alternative would. Positive integer counts ('3', '12', '100', '10') never contain a standalone 0. The 'no ... scooter/electric' alternatives are tied to the subject so a generic 'there are no discounts' cannot satisfy them. Medium confidence: the zero-synonym set is judgement-based.", "source": "batch_05.json" } } }, { "task_id": 189, "intent": "Can you find any Nintendo Switch console in the shop?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Nintendo Switch\" in the search box and press Enter.\n2. Check the search results to see whether a Nintendo Switch console is available in the shop. As there isn't, send a message to the user: \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| none |OR| couldn't find |OR| could not find |OR| cannot find |OR| can't find |OR| unable to find |OR| not available |OR| not for sale |OR| not sold |OR| not carried |OR| not stocked |OR| not find |OR| doesn't have |OR| does not have |OR| doesn't sell |OR| does not sell |OR| doesn't carry |OR| does not carry |OR| no nintendo switch |OR| not in the shop" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No'), stated in the first sentence. Bare 'not' was removed from the alternatives because it flips polarity in a wrong answer ('I found a Nintendo Switch, not the Lite version.' means one WAS found = yes); the kept alternatives are bare 'no'/'none' plus the specific 'could not/couldn't/cannot/can't/unable to find', 'not available/for sale/sold/carried/stocked/find', and 'does not/doesn't have/sell/carry' phrasings that can only express the 'No' verdict. Medium confidence per guidelines trap 3: the gold is the bare word 'No', an inherently low-discriminative token, and inherent yes/no crosstalk with other No-golds is unavoidable.", "source": "batch_03.json" } } }, { "task_id": 190, "intent": "I'm looking for a pair of jeans for less than $30. The size should be XL. Once you complete the order, share the completion code.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"jeans\" in the search box and press Enter.\n2. Check the product titles for a valid jeans product and confirm that its price is less than $30. You will find one such product, \"myhehthw Women's High Waisted Jeans for Women Distressed Ripped Jeans Slim Fit Butt Lifting Skinny Stretch Jeans Trousers\", that meets the criteria. Click on this product.\n3. Choose size \"x-large\".\n4. Click \"Buy Now\" to place the order.\n5. Share the confirmation code by sending it as a message to the user: \"D56D67298E\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "d56d67298e" ], "scope": "full", "fuzzy_match": "D56D67298E" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Opaque alphanumeric completion code; word-boundary containment makes truncated or altered codes fail.", "source": "batch_06.json" } } }, { "task_id": 191, "intent": "I'm looking for a box of cookies under $5. Can you please place the order and share the completion code with me?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"cookies\" in the search box and press Enter.\n2. Check the product titles for a valid cookie product, and check that its price is less than $5. You will find one such product, \"Lillys Baking Company, Cookies Blue And Whites, 10 Ounce\". Click on the product page.\n3. Click \"Buy Now\" to place the order.\n4. Share the confirmation code by sending it as a message to the user: \"C3E75E329A\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "c3e75e329a" ], "scope": "full", "fuzzy_match": "C3E75E329A" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A 10-character confirmation code is maximally discriminative and word-boundary matching prevents a partial-code match. Case is handled by normalization, so the lowercase reference is fine.", "source": "batch_07.json" } } }, { "task_id": 192, "intent": "Order these products and share the confirmation codes: Welch's Orange Pineapple Juice, Stirrings Simple Classic Cocktail Mix Mojito, and Italian Basil Pesto.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Welch's Orange Pineapple Juice\" in the search box and press Enter.\n2. Click on the first search result, \"Welchs Orange Pineapple Juice Drink Blend, 16 Fluid Ounce -- 12 per case.\".\n3. Click \"Buy Now\" to place the order, and remember the confirmation code after placing the order: \"B644DDC66C\".\n4. Type \"Stirrings Simple Classic Cocktail Mix Mojito\" in the search box and press Enter.\n5. Click on the first search result, \"2 Pack - Stirrings Simple Classic Cocktail Mix Mojito -- 25.4 fl oz\".\n6. Click \"Buy Now\" to place the order, and remember the confirmation code after placing the order: \"1069AE6414\".\n7. Type \"Italian Basil Pesto\" in the search box and press Enter.\n8. Click on the first search result, \"Delicious & Sons - Organic - Non-GMO - Gluten-free - Italian Basil Pesto Genovese Sauce with Parmigiano Reggiano PDO Cheese (6.70 oz.)\".\n9. Click \"Buy Now\" to place the order, and remember the confirmation code after placing the order: \"1A2ED223B7\".\n10. Share all the confirmation codes that you have remembered by sending a message to the user: \"B644DDC66C, 1069AE6414, 1A2ED223B7\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "b644ddc66c" ], [ "1069ae6414" ], [ "1a2ed223b7" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "B644DDC66C, 1069AE6414,1A2ED223B7" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Three opaque confirmation codes; all must appear. Matching is case-insensitive after normalization and word-boundary anchored, so a mistyped code cannot match. The intent lists the products but does not demand an output order, so ordered:false.", "source": "batch_09.json" } } }, { "task_id": 193, "intent": "Can you place an order for a Romwe wrap skirt and share the confirmation code with me? If there are size options, then please pick the medium size.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Romwe wrap skirt\" in the search box and press Enter.\n2. Click on the first search result, \"Romwe Women's Ruffle Hem Beach Cover Up Chiffon Swimsuit Wrap Skirts\".\n3. On the product page, choose the \"Medium\" size by clicking on it.\n4. Click \"Buy Now\" to place the order.\n5. Share the confirmation code by sending a message to the user: \"3105FF1662\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "3105FF1662" ], "scope": "full", "fuzzy_match": "3105FF1662" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Alphanumeric confirmation code -- string_match on the literal token (normalization lowercases both sides) rather than number_match, which cannot parse '3105FF1662'. Word-boundary matching means a code that merely contains the token as a substring would not falsely match.", "source": "batch_08.json" } } }, { "task_id": 194, "intent": "Order every variant of the Cravings by Zoe Valentine's Day chocolates and share the confirmation codes.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Cravings by Zoe Valentine's Day chocolates\" in the search box and press Enter.\n2. Click on the first search result, \"CRAVINGS BY ZOE Valentine's Day Chocolate Truffles | Assorted Chocolate Gift Box Red | Gourmet Chocolate | Delicious Milk Chocolate, Dark Chocolate and White Chocolate Flavors with 16 Assorted Toppings - 7.5 oz, 16 Count\".\n3. Click on the first variant, \"happy birthday truffle chocolate\". Then click \"Buy Now\" and get the code \"0FA416F314\". Remember this code.\n4. Go back and click on the second variant, \"holiday truffle chocolate\". Then click \"Buy Now\" and get the code \"68A92442CA\". Remember this code.\n5. Go back and click on the third variant, \"truffle collection chocolate\". Then click \"Buy Now\" and get the code \"5AB56CB7A1\". Remember this code.\n6. Send all the codes that you have remembered as a message to the user: \"0FA416F314, 68A92442CA, 5AB56CB7A1\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "0fa416f314" ], [ "68a92442ca" ], [ "5ab56cb7a1" ] ], "ordered": false }, "fuzzy_match": "0FA416F314, 68A92442CA, 5AB56CB7A1" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Three confirmation codes; all must be present. Unordered because the intent asks for 'the confirmation codes', not a sequence. Codes are alphanumeric literals, matched case-insensitively on word boundaries so a longer lookalike string cannot satisfy them.", "source": "batch_10.json" } } }, { "task_id": 195, "intent": "Order all the skirts available in the shop and share the confirmation codes. If there are multiple options, then always pick the small size.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"skirts\" in the search box and press Enter.\n2. Click on the first skirt product, \"Romwe Women's Ruffle Hem Beach Cover Up Chiffon Swimsuit Wrap Skirts\".\n3. Click on the size \"Small\".\n4. Click \"Buy Now\" and get the code \"73A46F8F3F\". Remember this code.\n5. Search for \"skirts\" again and click on the second skirt product, \"Merry Christmas Snowflake Pattern Women Chiffon Short Scarf Lightweight and Compact, Easy to Carry. Beach Sarong Cover Wrap Skirt Bikini Swimwear Cover Ups Black\".\n6. Click on the size \"Small\".\n7. Click \"Buy Now\" and get the code \"EFBFB6B826\". Remember this code.\n8. Send a message to the user with the confirmation codes that you have remembered: \"73A46F8F3F, EFBFB6B826\".", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "73A46F8F3F" ], [ "EFBFB6B826" ] ], "ordered": false }, "fuzzy_match": "73A46F8F3F, EFBFB6B826" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Both confirmation codes must be reported. Codes are opaque tokens, so word-boundary containment is exact enough; ordered stays false because the intent does not prescribe an order. list_match has no 'forbidden' entries because the set of possible wrong/extra codes is not enumerable; requiring both correct codes is the discriminating check.", "source": "batch_11.json" } } }, { "task_id": 196, "intent": "What is the focal length range in millimeters of the stock lens on the Canon PowerShot S3 camera product picture?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Canon PowerShot S3\" in the search box and press Enter.\n2. Click on the product \"Canon PowerShot Pro Series S3 IS 6MP with 12x Image Stabilized Zoom (Discontinued by Manufacturer)\" from the search results.\n3. Check the product image. You will see \"6.0 - 72.0 mm\" written on the stock lens.\n4. Send a message to the user: \"6.0 - 72.0\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "values": [ 6, 72 ] }, "fuzzy_match": "6.0 - 72.0" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A numeric range. Requiring both endpoints as numbers accepts '6.0 - 72.0', '6-72 mm' and '6.0 to 72.0 millimetres' alike (Decimal comparison makes 6.0 == 6). Extra numbers in the answer ('12x zoom', '6MP') are harmless because both endpoints are still required, and the 35mm-equivalent distractor range 36-432 is rejected.", "source": "batch_09.json" } } }, { "task_id": 197, "intent": "How many of the color options for the Nautica headphones have two distinct visible colors?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Nautica headphones\" in the search box and press Enter.\n2. Click on the product \"NAUTICA H120 Bluetooth Headphones, On-Ear Wireless Headphones with Built-in Microphone Bluetooth v5.0 Wireless and Wired Stereo Headset with Deep Bass, Foldable Over-Ear Headphones (White Nude)\" from the search results.\n3. In the color options, you will see that the last four are two-color combinations: \"navy red\", \"navy yellow\", \"white nude\", and \"white off white\".\n4. Send a message to the user: \"Four\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 4 }, "fuzzy_match": "Four" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count; gold 'Four' -> 4 exact, and number_match reads both 'four' and '4'. Model numbers in a restated product title (H120, v5.0) are not 4. Medium (not high) because the reasoning naturally references 'the last four color options' (navy red/navy yellow/white nude/white off white), so an agent that miscounts could still emit an incidental '4' (e.g. '3 of the last four options show two colors'); number_match cannot separate that residual case, but every off-count wrong answer that does NOT restate 'four' is correctly rejected.", "source": "batch_06.json" } } }, { "task_id": 198, "intent": "What's the color of the Rifle Scope Mount Camera Adapter in the product picture?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Rifle Scope Mount Camera Adapter\" in the search box and press Enter.\n2. Click on the product \"starboosa Rifle Scope Mount Camera Adapter - Smartphone Camera Adapter for Hunting Teaching\" from the search results.\n3. Look at the product picture to identify the color of the product, which is black. Send a message to the user: \"Black\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "black" ], "scope": "first_sentence", "fuzzy_match": "Black" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Single colour word. scope:first_sentence stops a wrong colour verdict from passing because a later sentence mentions something else that is black (the scope, the mount, the phone). Medium because 'black' remains a common incidental word.", "source": "batch_07.json" } } }, { "task_id": 199, "intent": "Place an order for the first pink socks that appear when you search for \"socks.\" Then share the confirmation code here. If you need to pick a size, please choose medium.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"socks\" in the search box and press Enter.\n2. The first pink socks that appear are \"Gloaiidjonh Women Cotton Crew Socks Cute Cartoon Anime Dinosaur Embroidery Mid Tube Hosiery\". Click on the product page.\n3. Select the first pair of pink socks visible.\n4. Select the color option \"pk\".\n5. Click \"Buy Now\".\n6. Check the confirmation code and send it as a message to the user: \"202AA9AA1D\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "202aa9aa1d" ], "scope": "full", "fuzzy_match": "202AA9AA1D" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "A single alphanumeric confirmation code. string_match rather than number_match because the value is not numeric; word-boundary containment means a longer code containing it as a prefix cannot match.", "source": "batch_10.json" } } }, { "task_id": 200, "intent": "What's written on the body of the 20x50 Binoculars? If there are multiple 20x50 Binoculars, pick the first one from the search results.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"20x50 Binoculars\" in the search box and press Enter.\n2. Click on the first search result, \"20x50 High Power Binoculars for Adults with Low Light Night Vision, Compact Waterproof Binoculars for Bird Watching Hunting Travel Football Games Stargazing with Carrying Case and Strap\".\n3. Inspect the product image to see the text written on the body. You should find the text written in white: \"UncleHu\". Send a message to the user: \"UncleHu\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "unclehu |OR| uncle hu" ], "scope": "full", "fuzzy_match": "UncleHu" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single brand token read off the product image. The |OR| covers the spaced spelling, since word-boundary containment treats 'unclehu' and 'uncle hu' as different strings. Answers that only report the model text '20x50' (which is quoted in the intent) or the truncated 'Uncle' are rejected.", "source": "batch_08.json" } } }, { "task_id": 201, "intent": "Place an order for a Spider-Man superhero T-shirt with a white spider. The size should be Large. Once you complete the order, share the completion code.", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Spider-Man superhero T-shirt white spider\" in the search box and press Enter.\n2. Click on the product page \"Long Sleeve Superhero T Shirt Tank Top Mens Compression Shirt Men Workout Fitness Gym Shirt\".\n3. Click on \"brown\". You will see that the product image has a white spider. Select the correct shirt and choose size \"Large\".\n4. Click \"Buy Now\" to place the order.\n5. Check the confirmation code and send it as a message to the user: \"7778F2A319\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "7778F2A319" ], "scope": "full", "fuzzy_match": "7778F2A319" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Single opaque completion code; literal containment with word boundaries so a longer lookalike token cannot match. No distractor code appears in the intent. string_match rather than number_match because the code is alphanumeric and would be split into several numbers by the parser.", "source": "batch_11.json" } } }, { "task_id": 202, "intent": "Could you please order the most expensive cracker available in the Shop and send me the confirmation code?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"cracker\" in the search box and press Enter.\n2. You will find five search results. The first search result, \"Crunchmaster Gourmet Toasted Sesame Rice Cracker, 3.5-Ounce (Pack of 12)\", will be the most expensive one and is a valid cracker product. Click on the product page.\n3. Click \"Buy Now\" and retrieve the confirmation code.\n4. Send the confirmation code as a message to the user: \"79E27147B0\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "79e27147b0" ], "scope": "full", "fuzzy_match": "79E27147B0" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Confirmation code: an opaque literal, matched case-insensitively on word boundaries, so an off-by-one-character code cannot pass. string_match rather than number_match because the code is alphanumeric.", "source": "batch_12.json" } } }, { "task_id": 203, "intent": "Which of the following items are available in the shop: Crayon, Candle, Face Lotion, and Electric Chair?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Search for \"Crayon\" and check the search results. None of the results should be valid crayon products. Remember this fact.\n2. Search for \"Candle\" and check the search results. You should find at least one valid candle product. Remember this fact.\n3. Search for \"Face Lotion\" and check the search results. You should find at least one valid face lotion product. Remember this fact.\n4. Search for \"Electric Chair\" and check the search results. None of the results should be valid electric chair products. Remember this fact.\n5. Based on the search results that you remembered, send a message to the user: \"Candle, Face Lotion\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "candle", "face lotion" ], "must_exclude": [ "^.*\\b(?:crayons?|electric chairs?)\\b[,\\s]*(?:and\\s+|&\\s*|or\\s+)?(?:candles?|face lotions?)\\b.*$", "^.*\\b(?:candles?|face lotions?)\\b[,\\s]*(?:and\\s+|&\\s*|or\\s+)?(?:crayons?|electric chairs?)\\b.*$", "^.*\\b(?:crayons?|electric chairs?)\\b\\s+(?:is|are|was|were)\\s+(?:also\\s+|too\\s+|indeed\\s+)?(?:available|in stock|present)\\b.*$" ], "scope": "full", "fuzzy_match": "Candle, Face Lotion" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1 + trap 2: all four candidate items are named in the intent, so a wrong 'all four are available' answer contains the gold tokens, yet a correct answer legitimately names Crayon and Electric Chair in a negative clause. Plain must_exclude would reject the common correct phrasing '...Candle and Face Lotion are available, while Crayon and Electric Chair are not'. The first two exclusions fire only when a wrong item sits in the same comma/'and' list slot as a right one (how an over-inclusive list reads). The third exclusion catches a wrong item asserted available in its own clause ('Crayon is available, and Candle and Face Lotion too') -- it fires on '<wrong item> is/are (also/too/indeed)? available/in stock/present', which never appears in a correct answer (correct answers say 'are not', 'are unavailable', or 'returned no products', none of which match). Residual leniency: an elliptical 'Crayon is too' with no availability word would still slip.", "source": "batch_12.json" } } }, { "task_id": 204, "intent": "How many times does the word \"Cotton\" appear on the product page (including Description, Features, Reviews) for \"Custom Polo Shirts for Men\"?", "sites": [ "webshop" ], "start_url": "__WEBSHOP__", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Custom Polo Shirts for Men\" in the search box and press Enter.\n2. Click on the first search result, \"Custom Polo Shirts for Men Design Your Own Personalized Logo Text Photo Polo Shirt Customize Classic Polo Shirt\".\n3. In the product description, count the occurrences of the word \"Cotton\" in the text. You should find three such occurrences. Remember this count.\n4. In the features section, count the occurrences of the word \"Cotton\" in the text. You should find a single occurrence. Remember this count.\n5. Add the counts together, which should be four, and send a message to the user: \"Four times\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 4 }, "fuzzy_match": "Four times" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Count question, gold spelled out ('Four times'); number_match reads both 'four' and '4'. Scope is left 'full' because a correct answer often opens with a sentence about the sections before giving the total. Medium confidence: an answer that states a wrong total but happens to mention 4 as a per-section sub-count would be accepted. Requiring the literal word 'times' was rejected as over-fitting (an agent may answer '4').", "source": "batch_09.json" } } }, { "task_id": 205, "intent": "Visit the Wiki article on the first country mentioned in the latest News article about TikTok. Based on that Wiki article, tell me the population of that country.", "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"TikTok\" and open the latest (most recent) article in the results.\n2. Identify the first country mentioned in that news article. Remember the country name.\n3. Switch to Wiki and search for that country name. Open the country’s article.\n4. From the Wiki article, extract the population stated for that country.\n5. Send the population figure to the user.", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 340100000, "rel_tolerance": 0.005 }, "fuzzy_match": "340.1 million" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Population figure. number_match normalizes '340.1 million' and '340,100,000' to the same value. The small relative tolerance (0.5%, about +/-1.7 million) exists only because the gold is itself a rounded figure and the article may also be quoted precisely (e.g. 340,110,988); it is far too tight to admit another country or another census year (331.9 million is rejected). Medium confidence because the tolerance had to be reasoned about.", "source": "batch_09.json" } } }, { "task_id": 206, "intent": "I need you to run a quick check. Go to the “Quantum mechanics” article on Wiki and find the first article listed in the “Related pages” section. Search for the second word of that article’s title on News, visit the second search result, and tell me whether that news article contains the word “ship”.", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, open the \"Quantum mechanics\" article.\n2. Locate the \"Related pages\" section and click the first article listed. Note its title.\n3. Take the second word of that title and search it on News.\n4. Open the second search result on News.\n5. Check whether the article contains the word \"ship\" (case-insensitive). Report \"yes\" if it does, otherwise \"no\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "no" ], "scope": "first_sentence", "fuzzy_match": "yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold, standard recipe. The instructions literally tell the agent to report 'yes' or 'no', so the verdict token is reliably present. first_sentence protects the must_exclude from a trailing justification, and word boundaries keep 'no' from firing inside 'not' or 'nothing'. The word 'ship' in the intent is not a discriminator and is left unconstrained.", "source": "batch_04.json" } } }, { "task_id": 207, "intent": "First, verify whether “Brisling Sardines in BBQ Sauce” is kosher. If it is, look up “sardine” on Wikipedia and find the word from which its name is derived. If it is not, search for “Fish” on News and tell me the 20th search result.", "sites": [ "wiki", "webshop", "news" ], "start_url": "__WEBSHOP__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On the Shop, search for \"Brisling Sardines in BBQ Sauce\" and open the product page.\n2. Determine if the product is marked kosher in the description, features, or labeling. Remember whether it is kosher.\n3. If it IS kosher: switch to Wiki, search for \"sardine\", open the article, and find the word from which the name \"sardine\" is derived. Report that word to the user.\n4. If it is NOT kosher: go to News, search for \"Fish\", navigate to the 20th search result, and report the title of that article.\n5. Send only the required answer per the above branching.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "sardinia |OR| sardinian" ], "must_exclude": [ "^[\\s\\S]*\\bnot\\b[\\s\\S]{0,25}sardinia[\\s\\S]*$" ], "scope": "first_sentence", "fuzzy_match": "The fishes are named after the Island of Sardinia." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The kosher branch is the correct one, so the answer must name Sardinia. 'Sardinia' appears nowhere in the intent (only 'sardine'/'Sardines'), and word-boundary matching keeps 'sardina'/'sardinos' from matching. scope first_sentence plus the exclusion of a negated 'not ... Sardinia' block an answer that gives a different etymology and mentions Sardinia only to rule it out.", "source": "batch_14.json" } } }, { "task_id": 208, "intent": "Visit the first article of the News search for the toy maker Mattel. Find the name of the first toy product mentioned in that article. Is that product available in the Shop? If so, how many such products are available?", "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"Mattel\" and open the first article in the results.\n2. Read the article and identify the first toy product mentioned by name. Remember that product name.\n3. Go to the Shop and search for that product name.\n4. Count how many valid product listings match that toy. If none, the count is zero.\n5. Report availability as yes/no and the count if available.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No, the product is not available in the Shop." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no rule for a 'No' gold. The intent explicitly asks for a yes/no verdict plus a count, so the verdict word is reliable. first_sentence scope plus must_exclude 'yes' rejects the hardest near-miss ('Yes, it is available, but there is no exact match').", "source": "batch_01.json" } } }, { "task_id": 209, "intent": "Please find the fifth unique country mentioned on the “Anime” Wiki page, then search for that country’s name on News, and report the title of the thirteenth article in the search results.", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, open the \"Anime\" article.\n2. Scan the article text to identify countries mentioned; track unique country names in order of first appearance.\n3. Identify the fifth unique country mentioned. Remember its name.\n4. Go to News and search for that country name.\n5. Navigate to the 13th article in the search results and read its title.\n6. Report that title to the user.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "france recalls |OR| recalls |OR| recalled", "betancourt" ], "scope": "full", "fuzzy_match": "France recalls Íngrid Betancourt mission" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Article title 'France recalls Íngrid Betancourt mission'. Unicode folding turns 'Íngrid' into 'Ingrid', so the accent is irrelevant. Two anchors are required because 'France' alone is just the country from step 3 and would be trivially present in a wrong answer, while 'Betancourt' alone could name a different Betancourt article ('Íngrid Betancourt freed after six years').", "source": "batch_05.json" } } }, { "task_id": 210, "intent": "Does the Canon EOS Rebel T7i found in the Shop have a Wiki page? If yes, tell me the title of that Wiki page. Otherwise, find the most relevant page and tell me its title. If you find none of the pages to be relevant, just say \"Article not found\".", "sites": [ "wiki", "webshop", "news" ], "start_url": "__WEBSHOP__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On the Shop, search for \"Canon EOS Rebel T7i\" and open the product page (or the closest match).\n2. From that product, derive the exact model name.\n3. On Wiki, search for that model name. If a dedicated page exists, report its title.\n4. If no dedicated page exists, choose the most relevant related Wiki page (e.g., a series page) and report its title.\n5. If no relevant Wiki page is found, reply \"Article not found\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*(?<![a-z0-9])canon[-\\s]+eos(?![a-z0-9])(?!\\s*(?:rebel|t7i)).*$" ], "must_exclude": [ "article not found" ], "scope": "full", "fuzzy_match": "Canon EOS" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1: the gold 'Canon EOS' is a prefix of 'Canon EOS Rebel T7i', which is named in the intent, so a plain must_include would accept 'there is no page for the Canon EOS Rebel T7i — Article not found'. The regex requires an occurrence of 'canon eos' that is NOT followed by 'rebel'/'t7i', i.e. the bare series title, while still allowing a correct answer to mention the T7i. must_exclude kills the explicit 'Article not found' escape phrase.", "source": "batch_06.json" } } }, { "task_id": 211, "intent": "Search for \"Apple\" on News. Is the first search result negative news about the tech company Apple? If yes, say \"Found negative news\". If not, order an “iPad Case Fit for iPad 9.7” and send me the confirmation code.", "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"Apple\".\n2. Open the first search result and determine whether it is negative news about the tech company Apple. Negative means conveying bad outcomes, issues, or criticism about Apple.\n3. If it IS negative: reply \"Found negative news\".\n4. If it is NOT negative: go to the Shop and search for \"iPad Case Fit for iPad 9.7\". Select the product, choose any default options, click \"Buy Now\", and return the confirmation code.\n5. Send only the required final message based on the branch.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "found negative news" ], "must_exclude": [ "^.*(?<![a-z0-9])(?=[a-z0-9]{5,8}(?![a-z0-9]))(?=[a-z0-9]*[a-z])(?=[a-z0-9]*[0-9])[a-z0-9]+.*$" ], "scope": "full", "fuzzy_match": "Found negative news." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 6: the gold is a fixed phrase the instructions tell the agent to emit verbatim, so must_include on the phrase (not exact_match, which a decorated answer would fail). Requiring the three words in order separates it from the other branch's near-misses ('No negative news was found', 'did not find negative news'). Added a must_exclude that fires on a confirmation-code token -- a 5-8 character alphanumeric run containing BOTH a letter and a digit (e.g. 8KD2QX). Only the order branch returns such a code, so this blocks the leak where a wrong-branch answer quotes the phrase while reporting its order ('...did not reply Found negative news, so I ordered; confirmation code 8KD2QX'). The regex is deliberately narrow: pure-letter tokens like 'iphone' and pure-digit tokens like '2024' do not match, and a correct answer that merely explains it generated no confirmation code (no actual code value) still passes. Medium because the code heuristic is the load-bearing exclusion.", "source": "batch_04.json" } } }, { "task_id": 212, "intent": "Does the \"Once Upon A Time Queen Born In 1982 T-Shirts\" product from the Shop contain polyester? If so, go to the Polyester article on Wiki and tell me the names of the clothing brands mentioned there. Otherwise, just tell me, \"T-shirt is polyester-free!!!\"", "sites": [ "wiki", "webshop" ], "start_url": "__WEBSHOP__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Once Upon A Time Queen Born In 1982 T-Shirts\" in the Shop search box and press Enter.\n2. Click on the first search result.\n3. Check the material composition by reading the product description and other tabs.\n4. Verify whether polyester is listed by checking all the tabs on the product page. You will find polyester mentioned on the Attributes tab.\n5. Go to Wikipedia, type \"Polyester\" in the search box, and press Enter.\n6. You will find no clothing brands mentioned in the Wikipedia article. Send a message to the user: \"No clothing brands have been mentioned in the Polyester article.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*(?:\\bno\\b(?!\\s+(?:doubt|question|surprise)\\b)(?:\\s+\\w+){0,4}\\s+brands?\\b|\\b(?:not|never)\\b(?:\\s+\\w+){0,2}\\s+(?:mention|list|name|include|reference|cite|contain|feature|have|has|had|find|found|see|show|shown|display|discuss|specif)\\w*(?:\\s+\\w+){0,3}\\s+brands?\\b|\\bwithout\\b(?:\\s+\\w+){0,3}\\s+brands?\\b|\\bbrands?\\b(?:\\s+\\w+){0,3}[\\s:,-]+(?:none|nil|absent)\\b).*$" ], "scope": "full", "fuzzy_match": "No clothing brands are mentioned in the Polyester article." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is a full sentence asserting an absence, so exact_match is unusable and a literal must_include over-fits the phrasing. A naive 'negation anywhere before brands' regex false-accepts a hedged wrong answer that lists brands ('Not entirely sure, but the clothing brands mentioned include Nike, Adidas'), because the hedge 'not' precedes the brand mention. The regex is instead structured so the negation must attach to the brand claim: (a) 'no [<=4 words] brand(s)' -- 'no clothing brands', 'no mention of any clothing brands' -- with a lookahead that blocks the hedges 'no doubt/question/surprise ... brands'; (b) 'not/never [<=2 words] <absence verb: mention|list|find|...> [<=3 words] brand(s)' so the negated verb must come *before* brands, which a brand-listing answer ('brands mentioned include Nike') fails because its verb follows brands; (c) 'without ... brands'; (d) post-nominal 'brands ... none/nil/absent'. The 'polyester-free' wrong branch has no 'brand' token at all.", "source": "batch_12.json" } } }, { "task_id": 213, "intent": "Could you check whether the \"Tattoo\" article on Wiki talks about mummification? If so, tell me how many tattoo machines are available in the Shop. Only consider electronic tattoo machines, not manual ones.", "sites": [ "wiki", "webshop" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Tattoo\" in the Wiki search box and press Enter.\n2. Read the article to see whether the concept of mummification has been discussed. You will find it mentioned in the History section.\n3. Switch to the Shop search.\n4. Type \"tattoo machine\" in the search box and press Enter.\n5. Count the number of electronic tattoo machines available (excluding manuals) in the search results by visiting the first two pages. Don't go beyond that, as most of the results will not be relevant. You will find two such machines. Send a message to the user: \"Two\".", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 2 }, "fuzzy_match": "Two" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Gold is the spelled-out count 'Two'; number_match reads both 'two' and '2'. Exact comparison. Scope must stay 'full': the answer's first sentence is usually the mummification yes/no, with the count only in the second sentence. Medium confidence because the value 2 is low-discriminative -- an answer that reports a different count but mentions 'the first two pages of results' (the wording of the additional_instructions) would be accepted, and unrelated golds containing a 2 will show up as crosstalk.", "source": "batch_13.json" } } }, { "task_id": 214, "intent": "Search for \"Iran\" on News and visit the first search result. Was that news incident reported in Iran’s Wiki article?", "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"Iran\" and open the first article.\n2. Identify the incident/event described in that article.\n3. Go to Wiki and open the \"Iran\" article.\n4. Check whether that specific incident/event is reported in the Iran article.\n5. Answer \"Yes\" if it is reported; otherwise, answer \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no |OR| not reported |OR| not mentioned |OR| no mention |OR| does not |OR| doesn't |OR| is not |OR| was not |OR| isn't |OR| wasn't" ], "must_exclude": [ "yes" ], "scope": "first_sentence", "fuzzy_match": "No" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Yes/no gold ('No'). Scoped to the first sentence so the justification tail ('the Iran article covers ... but not this incident') cannot flip the verdict. The alternatives are negated verb phrases rather than a bare 'not', so a wrong answer such as 'The incident is reported in the Iran article, but not in detail' is still rejected (no leading negation of the verdict). must_exclude ['yes'] is low-risk here because 'yes' rarely appears in a genuine negative answer.", "source": "batch_05.json" } } }, { "task_id": 215, "intent": "Search for \"Driving Accident\" on News. Visit the first News article and identify the U.S. state where that accident occurred. Then search for that state on Wiki and tell me its population.", "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"Driving Accident\" and open the first article.\n2. Identify the U.S. state where the accident occurred. Remember the state name.\n3. Go to Wiki and search for that state. Open the state article.\n4. Find and report the population of that state from the Wiki article.", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 29000000, "rel_tolerance": 0.02 }, "fuzzy_match": "Over 29 million." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold 'Over 29 million.' is a hedged population figure, so a small relative tolerance is required: it must accept whichever 29.x figure the article snapshot carries (29,145,505 census, or a 29.5M estimate) while rejecting the off-by-one-million near-misses 28 and 30 million. 2% gives a window of 28.42M-29.58M, which does exactly that. No string anchor on the state name is possible because the gold itself does not name it.", "source": "batch_10.json" } } }, { "task_id": 216, "intent": "Visit the Wiki article on \"Espadrille\" and find the first country mentioned in that article. Which of the following sites: Wiki, News, or Shop, have an article about that country or a product from that country?", "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, open the \"Espadrille\" article.\n2. Identify the first country mentioned in the text. Remember the country name.\n3. Check Wiki to confirm that a page exists for that country (it should).\n4. On News, search for that country and verify that at least one news article appears.\n5. On the Shop, search for that country name and see if any product explicitly references or is from that country.\n6. Report which of the three sites (Wiki, News, Shop) have an article about that country or a product from that country. List only the sites that qualify.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "wiki |OR| wikipedia", "news" ], "must_exclude": [ "^(?!.*(?:\\bnot\\b|\\bno\\b|\\bonly\\b|\\bexcept\\b|\\bwithout\\b|\\bnone\\b|\\bneither\\b|\\bnor\\b|\\bnothing\\b|doesn't|don't|isn't|aren't)).*\\bshop\\b.*$", "^.*\\bnews\\b(?:\\s+(?:site|sites))?(?:\\s+and\\s+(?:the\\s+)?shop)?\\s+(?:do(?:es)?\\s+not|don't|doesn't|has\\s+no\\b|have\\s+no\\b|has\\s+nothing|have\\s+nothing|is\\s+not|are\\s+not|lacks?\\b).*$" ], "scope": "full", "fuzzy_match": "Wiki, News" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Trap 1 in its purest form: all three options (Wiki, News, Shop) are named in the intent, so both a plain must_include and a plain must_exclude['shop'] are compromised -- the correct answer very often says 'Wiki and News, but not Shop'. First exclusion is a negation-aware regex: it fires only when 'shop' appears and the answer contains no negation token anywhere, which kills 'Wiki, News, and Shop' / 'all three sites ... Wiki, News and Shop' while leaving every 'but not Shop' / 'the Shop does not' phrasing intact. Second exclusion catches the mirror-image wrong answer that names News only to deny it ('Only Wiki; News and Shop have nothing'), which the first exclusion would let through because it contains a negation. Medium confidence: a wrong 'all three' answer that happens to contain an unrelated 'not' would still be accepted.", "source": "batch_13.json" } } }, { "task_id": 217, "intent": "Order the most expensive bed and send me the confirmation code.", "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"bed\" in the Shop search box and press Enter.\n2. Go over all the products on page 1. You will find that the product \"iDealBed Luxe Series iQ5 Hybrid Luxury Plush Mattress 4i Custom Adjustable Bed Set, Wireless, Massage, Nightlight, Zero-Gravity, Anti-Snore, Memory Pre-Sets (Queen)\" has the highest price. Remember this fact. Warning: You might encounter something like a bed kit. This is not a bed, but this is a bed kit and shouldn't be considered.\n3. Visit pages 2, 3, and 4 of the search results to verify that no bed product exists with a higher price. Don't visit more pages, as you won't find better results.\n4. Since you cannot find any other product with a higher price, search \"iDealBed Luxe Series iQ5 Hybrid Luxury Plush Mattress 4i Custom Adjustable Bed Set, Wireless, Massage, Nightlight, Zero-Gravity, Anti-Snore, Memory Pre-Sets (Queen)\" and click on the product page.\n5. Click \"Buy Now\" and retrieve the confirmation code. Send it as a message to the user: \"61A1FDDAD1\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "61a1fddad1" ], "scope": "full", "fuzzy_match": "61A1FDDAD1" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Opaque alphanumeric confirmation code; word-boundary containment makes truncated or altered codes fail.", "source": "batch_06.json" } } }, { "task_id": 218, "intent": "Search \"dog\" on News. If the first search result says something positive about dogs, then place an order for a dog toy in the Shop and share the confirmation code. Otherwise, tell me the title of the news article.", "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"dog\" in the News search box and press Enter.\n2. Click on the first article in the search results.\n3. As it is a negative article on dogs, send the title of the news article message to the user: \"Dog owner to face felony charge\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "dog owner", "felony" ], "scope": "full", "fuzzy_match": "Dog owner to face felony charge" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The answer must be the news headline. Two distinctive tokens of the title are required rather than the whole string, so a minor requoting ('felony charges') still passes; 'felony' is what separates this headline from any other dog article, and 'dog owner' from an unrelated felony story. The alternative branch (ordering a toy and returning a confirmation code) contains neither token.", "source": "batch_12.json" } } }, { "task_id": 219, "intent": "Who are the first three female referees to be appointed to the men's 2022 FIFA World Cup?", "sites": [ "wiki", "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, search for information about female referees appointed to the men's 2022 FIFA World Cup.\n2. Identify the first three female referees appointed.\n3. Report their names in order.", "eval": { "eval_types": [ "list_match" ], "reference_answers": { "list_match": { "items": [ [ "frappart" ], [ "mukansanga" ], [ "yamashita" ] ], "ordered": false, "forbidden": [] }, "fuzzy_match": "Stéphanie Frappart,Salima Mukansanga,Yoshimi Yamashita" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Three names; surnames are unique and accent-folding handles 'Stéphanie'. ordered:false because the three were appointed in the same announcement -- 'in order' in the instructions is not a real ranking, so a differently ordered but complete list is correct.", "source": "batch_14.json" } } }, { "task_id": 220, "intent": "Verify whether Sony's net profit fell significantly in 2024 or 2025. If so, visit the Sony's Wiki article and identify the first product mentioned on that page. Can you find that exact product in the Shop? If so, place the order and share the confirmation code with me. If any of the above conditions cannot be satisfied, state that explicitly.", "sites": [ "wiki", "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Go to News and search \"Sony's net profit fell significantly\".\n2. You will find an article \"Sony net profit falls by 95%\". Click on the article and you will see that the publication date is from 2024. Hence it is a valid article. Remember this fact.\n3. Go to Wiki search and type \"Sony\".\n4. Identify the first product mentioned on the page, which is \"PlayStation\".\n5. Go to Shop.\n6. Search \"PlayStation\" and press Enter.\n7. You will not find any PlayStation console in the shop. Send the confirmation code as a message to the user: \"The product is not available in the shop.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "^.*\\b(?:not available|unavailable|not in the shop|not sold|not listed|not carried|no playstation|no such product|no results|not find|not found|does not sell|doesn't sell|does not carry|doesn't carry|does not have|doesn't have|does not stock|doesn't stock)\\b.*$", "playstation |OR| product |OR| shop |OR| store |OR| console" ], "scope": "full", "fuzzy_match": "The product is not available in the shop." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Multi-site conditional task whose gold is the explicit 'condition not satisfiable' statement. The first must_include entry is a regex over the unavailability phrasings an agent uses ('not available', 'could not find', 'the shop does not sell it', 'no PlayStation'). The second entry (playstation|product|shop|store|console) ties that unavailability claim to the product/shop, so an abandonment for the WRONG reason that happens to use an unavailability word - e.g. 'I could not find the article about Sony's profit' - no longer false-accepts. Wrong answers with a fabricated confirmation code contain no unavailability word. Medium because the spec verifies the unavailable-in-shop verdict, not the full chain of reasoning (profit fell, PlayStation is the first product).", "source": "batch_15.json" } } }, { "task_id": 221, "intent": "I love the Matrix movie series. Please find related Wiki articles, news, and products from the Shop. If there are multiple entries on each site, pick the one you think is the most relevant. If you cannot find any product or article, mention that in your answer. Share the titles of these items in the following format: <site name>: <article/news/product title>.", "sites": [ "wiki", "news", "webshop" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Go to Wiki, type \"The Matrix\", and press Enter.\n2. You will find a Wiki article on the Matrix movie series. Remember this fact.\n3. Go to News, type \"The Matrix\", and press Enter.\n4. You will find no news article on the Matrix movie series. Only look at the first page, don't go to more pages. Remember this fact.\n5. Go to Shop, type \"The Matrix\", and press Enter.\n6. You will find no shop products in the Matrix movie series. Only look at the first page, don't go to more pages. Remember this fact.\n7. Based on the facts that you have remembered, send a message to the user:\n\"Wiki: The Matrix\nNews: No relevant news found.\nShop: No relevant product found.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "the matrix", "no relevant news |OR| no news |OR| no relevant article |OR| no article |OR| none found", "no relevant product |OR| no product |OR| no matching product |OR| no products" ], "scope": "full", "fuzzy_match": "Wiki: The Matrix\nNews: No relevant news found.\nShop: No relevant product found." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Structured multi-part answer: Wiki has 'The Matrix', News and Shop have nothing. Requires the Matrix title plus a site-specific not-found indicator for each of the other two sites. Tightened: the earlier spec shared the generic alternatives 'none'/'not found'/'no result'/'no relevant' between the news and product requirements, so a single incidental 'none' (e.g. 'none of the shop items matched, but The Matrix Blu-ray is closest') satisfied both while the answer actually surfaced a real news item and product — a verified false accept. News-absence and product-absence now use disjoint, site-qualified phrases. Medium: phrasing of the not-found parts still varies; the additional_instructions steer the agent to the canonical 'No relevant news/product found.' wording this targets.", "source": "batch_16.json" } } }, { "task_id": 222, "intent": "Visit the Wiki article on Gabon and search for the last word of the first section in the Shop (introduction counts as a section). If there are search results, click on the first one and tell me the title of the product. Otherwise, tell me \"No product found\".", "sites": [ "wiki", "webshop" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, open the \"Gabon\" article.\n2. Consider the introduction as the first section. Identify the very last word of that section’s text.\n3. Go to the Shop and search for that last word.\n4. If there is at least one valid product result, open the first product and report its title exactly.\n5. If there are no valid product search results, respond with \"No product found\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "legendary whitetails", "non-typical", "long sleeve" ], "scope": "full", "fuzzy_match": "Legendary Whitetails Men's Non-Typical Long Sleeve T-Shirt" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Gold is a long product title; requiring it verbatim would reject a correct answer that drops \"Men's\" or the \"T-Shirt\" wording, so the spec pins the three load-bearing tokens. 'legendary whitetails' is a unique brand token that appears nowhere in the intent. Both 'non-typical' AND 'long sleeve' are required (not an OR): Legendary Whitetails ships several 'Non-Typical' line items and several 'Long Sleeve' items, so an OR would false-accept a *different* Legendary Whitetails product (e.g. a 'Non-Typical Hoodie' or a 'Camo Long Sleeve Tee') -- verified against the oracle. Requiring both still accepts every correct rendering of the real title. 'non-typical' also matches 'non typical' because the containment pattern treats hyphens and spaces alike.", "source": "batch_13.json" } } }, { "task_id": 223, "intent": "How many patients participated in the pollen tablet trial for hay fever?", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"pollen tablet trial hay fever\" or similar terms.\n2. Open the article describing a clinical trial of a pollen tablet for hay fever.\n3. Identify the number of patients who participated in that trial.\n4. If needed, cross-check with a related Wiki article about the same trial or treatment, but use the news article as the primary source.\n5. Report the total number of patients as a number.", "eval": { "eval_types": [ "number_match" ], "reference_answers": { "number_match": { "value": 634 }, "fuzzy_match": "634" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Exact participant count from the news article; number_match with no tolerance handles '634', '634 patients' and thousands separators. Off-by-one and transposed digits are rejected. No tolerance because the gold is a precise, unhedged count.", "source": "batch_11.json" } } }, { "task_id": 224, "intent": "Visit the Wiki article on \"Brazil\" and find the most recent sporting event mentioned in the Sports section. Search for that event (don't include the year in the search) on News, and tell me the name of the third news article in the search results.", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Search \"Brazil\" on Wiki and press Enter.\n2. Go to the Sports section and find the latest sports event, which is the \"Summer Olympics\". Remember this event.\n3. Go to News and search \"Summer Olympics\".\n4. Report the title of the third news article from the search results as a message to the user: \"Three cities submit bids for 2020 Summer Olympics\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "three cities", "bids |OR| bid |OR| bidding", "2020", "summer olympics" ], "scope": "full", "fuzzy_match": "Three cities submit bids for 2020 Summer Olympics" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Article title. Split into the four load-bearing tokens rather than one long verbatim phrase so that title-case, punctuation or a slightly reworded quote still passes, while any other Summer Olympics article (which will not say 'three cities') fails. 'three cities' is the real discriminator; 'bids |OR| bid |OR| bidding' also accepts a paraphrase ('three cities bidding for ...') that the bare 'bids/bid' pair used to false-reject.", "source": "batch_06.json" } } }, { "task_id": 225, "intent": "Search \"Cranberry\" on News and visit the first news article in the search results. Then search, on Wiki, for the name of the first country mentioned in that article’s content. Tell me the first line of the final visited article.", "sites": [ "wiki", "news" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Cranberry\" in the News search box and press Enter.\n2. Open the first article and read the text to identify the first country mentioned, England. Remember the name of the country.\n3. Go to Wiki.\n4. Type the name of that country (\"England\") in the Wiki search box and press Enter.\n5. Read the first line of the article's main content and send it as a message to the user: \"England is a country on the island of Great Britain.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "england is a country |OR| england as a country |OR| england, a country", "great britain" ], "scope": "full", "fuzzy_match": "England is a country on the island of Great Britain." }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "The gold is a sentence, so two anchors are required: the subject clause (with the two paraphrases an agent may use when it reports rather than quotes the line) and the distinctive 'Great Britain'. Requiring both rejects the first line of the wrong article (Great Britain, Cranberry) and a differently worded England lead. scope:full because the quoted line is often introduced by a preamble sentence.", "source": "batch_07.json" } } }, { "task_id": 226, "intent": "Search \"Ruffles Ridged Potato Chips\" on Shop and find the country of origin. Search for the name of that country on News and visit the second news article in the search results. What is the last word in the visited article?", "sites": [ "news", "webshop" ], "start_url": "__WEBSHOP__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Type \"Ruffles Ridged Potato Chips\" in the Shop search box and press Enter.\n2. Open the product page and each tab individually. Click on the features tab, identify the country of origin listed, which is the United States. Remember this fact.\n3. Go to the News website.\n4. Type the name of that country, \"United States\", in the News search box and press Enter.\n5. Navigate to and open the second news article.\n6. Identify the very last word of the article and send it as a message to the user: \"Outrage\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "outrage" ], "scope": "full", "fuzzy_match": "Outrage" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Single-word answer ('Outrage', the last word of the article). Word-boundary matching means 'outrage' does not fire on 'outrageous' or 'outraged', which is the near-miss to guard for a 'last word' task. scope stays 'full' (NOT first_sentence): the two-part task means a correct answer legitimately lands in a later sentence, e.g. 'The country of origin is the United States. The last word is Outrage.' Medium confidence: the gold word is also the article's likely topic, so a wrong last-word answer that discusses the same story ('the last word is protests, reflecting the public outrage') could mention 'outrage' incidentally and be falsely accepted -- there is no deterministic anchor that separates 'outrage as the last word' from 'outrage as the subject'.", "source": "batch_05.json" } } }, { "task_id": 227, "intent": "Search \"chips\" on News. If the first article is about snack chips, then go to Shop, search \"chips\", and place an order for the first chips product you find in the search results. Otherwise, let me know that the first article is not about snack chips.", "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"chips\" and open the first article in the results.\n2. Determine whether the article is about snack chips (potato chips, corn chips, etc.), not electronic chips or other meanings.\n3. If the article IS about snack chips:\n a. Go to the Shop and search for \"chips\".\n b. Open the first valid snack chips product in the results.\n c. Click \"Buy Now\" and collect the confirmation code.\n d. Return only the confirmation code.\n4. If the article is NOT about snack chips: respond with a statement indicating that the first article is not about snack chips.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "snack |OR| snacks", "^.*\\b(?:not|rather than|instead of|nothing to do with|unrelated to)\\b[^.]{0,40}\\bsnacks?\\b.*$" ], "must_exclude": [ "^.*(?<![a-z0-9])(?=[a-z]*[0-9])(?=[0-9]*[a-z])[a-z0-9]{10}(?![a-z0-9]).*$" ], "scope": "full", "fuzzy_match": "First article is not about the snacks, chips." }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Conditional task whose gold takes the negative branch: the answer must deny that the first article is about snack chips. The regex requires a negation word within 40 characters (and inside the same sentence) of 'snack', which distinguishes 'not about snack chips' from 'about snack chips, not microchips'. The must_exclude regex matches a 10-character mixed letter/digit token, i.e. a webshop confirmation code — a correct answer places no order and so can never contain one, while the wrong branch always reports one. Medium confidence because the negation is matched by pattern rather than by a fixed token.", "source": "batch_09.json" } } }, { "task_id": 228, "intent": "Find the fastest-growing city in Utah, based on its Wiki article. Then search that city's name on News and tell me the publication date of the second search result.", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, open the \"Utah\" article.\n2. In the article text, identify which city is explicitly described as the fastest-growing city in Utah. Remember that city’s name.\n3. On News, search for that city name.\n4. Open the second article in the search results.\n5. Find the article's publication date as displayed on the news page.\n6. Report that date to the user.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "december 18 |OR| december 18th |OR| 18 december |OR| 18th december |OR| 18th of december |OR| dec 18 |OR| dec. 18 |OR| 18 dec |OR| 2024 12 18 |OR| 12 18 2024 |OR| 12/18/2024 |OR| 18/12/2024", "2024" ], "scope": "full", "fuzzy_match": "December 18, 2024" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "Publication date. Two requirements: the day+month in any common ordering/format (the hyphenated ISO form matches because hyphen runs are equivalent to spaces in containment matching) and the year. Requiring the year separately catches the off-by-year near-miss; the only correct answer it would reject is one that omits the year entirely, hence medium rather than high.", "source": "batch_01.json" } } }, { "task_id": 229, "intent": "Place an order for an Oval Mosaic Wall Mirror if you can find a news article on Johan Cruyff. If you place the order, share the confirmation code with me. Otherwise, let me know that you didn’t place the order.", "sites": [ "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On News, search for \"Johan Cruyff\".\n2. If at least one clearly relevant news article about Johan Cruyff appears in the results, remember that you have found such an article.\n3. If NO relevant article is found: reply that you did not place the order because no Johan Cruyff news article was found.\n4. If a relevant article IS found: go to the Shop and search for \"Oval Mosaic Wall Mirror\".\n5. Open the appropriate product page and click \"Buy Now\" to place the order.\n6. Capture the confirmation code and send it as the answer.", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "4aa17fd340" ], "scope": "full", "fuzzy_match": "4AA17FD340" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Conditional task whose correct branch ends in a confirmation code. The code is a unique alphanumeric token, so a single must_include is exact: word-boundary matching stops it from matching a longer code, and case/markdown are normalized away. An answer on the wrong branch ('I did not place the order') cannot contain it.", "source": "batch_03.json" } } }, { "task_id": 230, "intent": "Search \"El Paso, Texas\" on Wiki and find the last sentence in the article’s content. Has the event mentioned in the last sentence been covered by any news article?", "sites": [ "wiki", "news" ], "start_url": "__WIKI__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. On Wiki, search for \"El Paso, Texas\" and open the article.\n2. Read through the main content of the article (excluding references, infoboxes, and navigation) and identify the very last sentence.\n3. Determine what event is being described in that last sentence.\n4. On News, search for that event (using key terms, names, and dates as needed).\n5. Check whether at least one news article clearly covers the same event.\n6. Answer \"Yes\" if such coverage exists; otherwise, answer \"No\".", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "yes" ], "must_exclude": [ "no" ], "scope": "first_sentence", "fuzzy_match": "Yes" }, "revision": 1, "annotation": { "confidence": "high", "subjective": false, "gold_suspect": false, "notes": "Standard yes/no rule; first-sentence scope so the supporting detail after the verdict cannot trip the 'no' exclusion.", "source": "batch_06.json" } } }, { "task_id": 231, "intent": "What happened at the end of the book \"The Neptune Secret\"?", "sites": [ "wiki", "news", "webshop" ], "start_url": "__NEWS__", "split": "train", "additional_instructions": "Perform these instructions to complete the task:\n\n1. Based on the query, searching Wiki is the most appropriate first step. Go to Wiki.\n2. In the Wiki search bar, type \"The Neptune Secret\" and press Enter.\n3. When no relevant results for a book titled \"The Neptune Secret\" are found, reply to the user with: \"No such book exists.\"", "eval": { "eval_types": [ "string_match" ], "reference_answers": { "must_include": [ "no such book |OR| no such title |OR| no such novel |OR| no such publication |OR| no book called |OR| no book titled |OR| no book named |OR| no book by that |OR| no book with that |OR| does not exist |OR| doesn't exist |OR| do not exist |OR| don't exist |OR| not exist |OR| no matching book |OR| not a real book |OR| not a real novel |OR| not a book |OR| find any book |OR| find any such book |OR| find no book |OR| found no book |OR| not appear to exist |OR| not seem to exist |OR| nonexistent |OR| non existent |OR| isn't a real book |OR| is not a real book" ], "scope": "full", "fuzzy_match": "No such book exists" }, "revision": 1, "annotation": { "confidence": "medium", "subjective": false, "gold_suspect": false, "notes": "The correct answer denies the premise, so the anchor is the nonexistence claim rather than any plot detail. Dropped the weak per-site-deniable alternatives from the earlier list -- bare 'no book', 'no results' and 'no record' each fire on a mixed wrong answer that denies one site and then hallucinates an ending from another ('no book quite like it; it ends when...', 'the Shop returned no results, but the News article reveals the ending'). Replaced 'no book' with the specific nonexistence forms ('no book called/titled/named/by that title') and added 'find any book' so 'I couldn't find any book by that title' is still accepted. Medium confidence: a hallucinated plot whose prose happens to contain 'does not exist' would still slip through.", "source": "batch_01.json" } } } ]