Exxe/hermeneutic-dive

Dataset Card for HermeneuticDive

Dataset Summary

HermeneuticDive v3.0.0 is a five-language deep-search training corpus built as an independent, CC0-licensed reimagining of the DeepDive methodology (KG multi-hop synthesis + obfuscation + difficulty filtering) with an interpretive construction discipline borrowed from the Integral Hermeneutics methodology. v3 is the result of a second external audit: v2 fixed most v1 problems but introduced new ones (clairvoyant trajectories, a vacuous decontamination pass, label inconsistencies, a length-confounded DPO set, no executable verifier). The audit table under Curation Rationale lists each finding, its fix, and how the fix was checked; Known Limitations lists what is still not good enough.

The corpus contains 189 unique questions — 74 en, 23 hi, 23 ja, 29 ru, 40 zh — across six task types: multi_hop obfuscated entity chains (109), date_sensitive answers with as-of stamps and valid_until (18), conflicting_sources (18), unanswerable (13), false_premise scored on correcting the premise (17), and perspective_framing (14). 17 multi-hop chains are deep (≥ 5 hops; deep now implies ≥ 5 hops for every task type, asserted at build time), mean multi-hop depth is 3.23 hops (27 of 109 multi-hop chains have ≤ 2 hops), 9 questions sit at estrangement level 3, and 6 answers are numbers with unit + tolerance. Every question carries a gold path — 558 hops, each with a typed relation, a rationale, a confidence tag (518 observed / 34 inferred / 6 speculative) and source URLs (508 unique: 486 Wikipedia titles across five language editions, 22 institutional pages).

It ships in five linked configs: qa_sft (56 tool-calling trajectories whose final assistant message is the five-part hermeneutic arc), qa_rl (133 rollable prompts with per-row verifier specs — question sets fully disjoint from qa_sft), trajectories_sft (28 paired demonstrations: 14 successes + 14 labelled failures), dpo (14 behavioural chosen/rejected pairs — length-confounded, see Limitations) and dpo_matched (17 length-matched minimal-contrast pairs; new in v3). The authoring pipeline (seed banks, patch layers, builders), a reference verifier.py with a self-test, and pipeline/card_tools.py (card ↔ data consistency check) ship in the repository.

Supported Tasks and Leaderboards

• Supervised fine-tuning of deep-search agents (qa_sft: system / user / tool-calling turns with tool_call_id / final hermeneutic arc — the arc follows searching, it does not replace it). Tool results are synthesised from the gold path (see Limitations 3).
• Reinforcement learning with verifiable rewards (qa_rl: canonical answer + aliases + typed verifiers). verifier.py implements canonical_or_alias, numeric_with_tolerance, date_match deterministically, premise_correction and abstain_or_calibrated_hedge as multilingual keyword heuristics, and exposes a judge= hook for the 24 llm_judge rows. python verifier.py --selftest runs >1,000 checks (positive controls, aliases, and negative controls: empty answer, bare "unknown", question pasted back, intermediate chain entity, generic abstention, answer-spraying).
• Preference learning: dpo_matched (17 pairs identical except the final answer line — no length or tool-call confound) is the cleaner signal; dpo (14 behavioural pairs) teaches richer behaviours but every pair is length-confounded (median 5.07× characters; the chosen side makes more tool calls in 13/14) — use a length-normalised objective and read pair_stats.
• Not a leaderboard. Evaluation on BrowseComp, BrowseComp-zh, xbench-DeepSearch, WebWalkerQA and GAIA is recommended only after re-running decontamination against dumps you hold (see Verification).

Languages

English (en, 74), Chinese (zh, 40, simplified), Russian (ru, 29), Japanese (ja, 23, new in v2), Hindi (hi, 23, new in v2). System prompts are English; questions, gold paths and arcs are fully in-language; cross-language aliases are included in answer specs where relevant.

Dataset Structure

Data Instances

Examples below are regenerated from the shipped files at build time (full rows).

qa_sft example (hd-en-mh-504, multi_hop deep, long-tail — the first query is the question's own clue, every later query uses only an entity an earlier result revealed, and the final cross-check opens a URL a search surfaced)

{
  "id": "hd-en-mh-504",
  "language": "en",
  "task_type": "multi_hop",
  "difficulty_mode": "deep",
  "hop_count": 5,
  "estrangement_level": 3,
  "question": "A Flemish friar sent by the French king to the Mongol world described, in his 1255 report, a marvel before the Great Khan's palace: a great silver tree with serpents coiled up its trunk and gilded lions at its base, spouting fermented mare's milk on command. It was the work of a Parisian craftsman captured in Hungary. Who was he?",
  "answer": {
    "value": "Guillaume Boucher",
    "aliases": [
      "William Boucher",
      "William of Paris (goldsmith)"
    ],
    "answer_type": "person",
    "unit": null,
    "tolerance": null,
    "as_of": null,
    "valid_until": null
  },
  "gold_path": [
    {
      "hop": 1,
      "entity": "William of Rubruck",
      "relation": "led",
      "rationale": "William of Rubruck, a Flemish Franciscan, travelled to the Mongol capital in 1253-55 as an envoy of the French king Louis IX and wrote a detailed report.",
      "confidence": "observed",
      "source_urls": [
        "https://en.wikipedia.org/wiki/William_of_Rubruck"
      ]
    },
    {
      "hop": 2,
      "entity": "Karakorum",
      "relation": "located_in",
      "rationale": "He reached Karakorum, the Mongol capital, and described the palace of Great Khan Möngke and its surroundings.",
      "confidence": "observed",
      "source_urls": [
        "https://en.wikipedia.org/wiki/Karakorum",
        "https://en.wikipedia.org/wiki/M%C3%B6ngke_Khan"
      ]
    },
    {
      "hop": 3,
      "entity": "silver tree fountain",
      "relation": "created",
      "rationale": "Before the palace stood a mechanical silver tree with four lions and serpents, dispensing drinks — the marvel Rubruck reported in detail.",
      "confidence": "observed",
      "source_urls": [
        "https://en.wikipedia.org/wiki/Karakorum",
        "https://en.wikipedia.org/wiki/Guillaume_Boucher"
      ]
    },
    {
      "hop": 4,
      "entity": "Guillaume Boucher",
      "relation": "created",
      "rationale": "The tree was built by Guillaume Boucher, a Parisian goldsmith taken captive during the Mongol raids into Hungary, working with other captured craftsmen.",
      "confidence": "observed",
      "source_urls": [
        "https://en.wikipedia.org/wiki/Guillaume_Boucher"
      ]
    },
    {
      "hop": 5,
      "entity": "capture in Hungary",
      "relation": "caused",
      "rationale": "Boucher had been seized in the 1241-42 Mongol invasion of Hungary and brought east — a craftsman carried between two worlds by conquest.",
      "confidence": "speculative",
      "source_urls": [
        "https://en.wikipedia.org/wiki/Guillaume_Boucher"
      ]
    }
  ],
  "hermeneutics": {
    "pre_understanding": "The evidence is one eyewitness description plus the identity of a craftsman; the question's constraints (Parisian, goldsmith, captured in Hungary, tree) must all converge.",
    "moves": [
      {
        "hop": 1,
        "lens": "perspectival_empathy",
        "move": "Read the marvel through Rubruck's eyes: to him the tree was wonder and warning at once — Latinate craft in the khan's service."
      },
      {
        "hop": 3,
        "lens": "empirical_grounding",
        "move": "The attribution to Boucher comes from Rubruck's own report — a single but primary source."
      },
      {
        "hop": 4,
        "lens": "systemic_embedding",
        "move": "The tree is a document of the Mongol system: tribute and conquest moved skilled people like Boucher across continents."
      },
      {
        "hop": 5,
        "lens": "boundary_exploration",
        "move": "Details of Boucher's life beyond the fountain are thin; tag biographical reconstruction as [Inferred]."
      }
    ],
    "synthesis": "The silver tree at Karakorum was made by the Parisian goldsmith Guillaume Boucher."
  },
  "messages": [
    {
      "role": "system",
      "content": "You are a deep-search agent with two tools: web_search(query) and open_url(url). Explore step by step, verify each hop against a source, and do not repeat similar queries. When confident, end your reply with 'Final answer: <short answer>'. If the question rests on a false premise, end with 'Final answer: <corrected fact>' preceded by a one-line premise correction. If the answer genuinely cannot be determined from evidence, end with 'Final answer: cannot_be_determined' (in the question's language: 无法确定 / невозможно установить / 特定できません / निर्धारित नहीं किया जा सकता) plus a one-line reason. For date-sensitive facts state the answer with 'as of <date>'.",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "A Flemish friar sent by the French king to the Mongol world described, in his 1255 report, a marvel before the Great Khan's palace: a great silver tree with serpents coiled up its trunk and gilded lions at its base, spouting fermented mare's milk on command. It was the work of a Parisian craftsman captured in Hungary. Who was he?",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "I'll start from the question's first clue and see which entity it points to.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"A Flemish friar sent by the French king to the Mongol world\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('A Flemish friar sent by the French king to the Mongol world') — 3 results:\n1. William of Rubruck — en.wikipedia.org\n   https://en.wikipedia.org/wiki/William_of_Rubruck\n   William of Rubruck, a Flemish Franciscan, travelled to the Mongol capital in 1253-55 as an envoy of the French king Louis IX and wrote a detailed report.\n2. Elizabeth Blackburn — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Elizabeth_Blackburn\n   Elizabeth Blackburn carried out her doctoral work on DNA sequencing in Cambridge under Frederick Sanger.\n3. Chauvet Cave — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Chauvet_Cave\n   Chauvet-Pont-d'Arc cave, discovered in 1994, holds some of the world's oldest known cave paintings, radiocarbon-dated around 36,000 years old.",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to William of Rubruck. Next I need: location.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_02",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"William of Rubruck location\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('William of Rubruck location') — 3 results:\n1. Karakorum — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Karakorum\n   He reached Karakorum, the Mongol capital, and described the palace of Great Khan Möngke and its surroundings.\n2. Nobel Prize in Physics 1921 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Albert_Einstein\n   Einstein won a single Nobel Prize in Physics, the 1921 prize (awarded in 1922) for his services to theoretical physics and the discovery of the law of the photo…\n3. Apollo 11 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_11\n   Apollo 11 carried Armstrong, Collins and Aldrin on the first crewed lunar landing in July 1969.",
      "tool_call_id": "call_02",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Karakorum. Next I need: creator.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_03",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Karakorum overview\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Karakorum overview') — 2 results (no direct match):\n1. Karakorum — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Karakorum\n   Encyclopedia overview of the topic; the snippet does not cover the detail requested.\n2. Northern Regional Research Laboratory — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Penicillin\n   In 1941-43 the US programme to scale up penicillin moved to the USDA laboratory in Peoria, Illinois.",
      "tool_call_id": "call_03",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "That result does not name what I need. Reformulating the query.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_04",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Karakorum creator\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Karakorum creator') — 3 results:\n1. silver tree fountain — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Karakorum\n   Before the palace stood a mechanical silver tree with four lions and serpents, dispensing drinks — the marvel Rubruck reported in detail.\n2. premise check — www.nobelprize.org\n   https://www.nobelprize.org/prizes/physics/1921/summary/\n   No second physics Nobel was ever awarded to Einstein — the question's premise is false, so no discovery can answer it.\n3. Michael Collins — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Michael_Collins_(astronaut)\n   Command module pilot Michael Collins had previously flown as pilot of Gemini 10.",
      "tool_call_id": "call_04",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to silver tree fountain. Next I need: creator.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_05",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"silver tree fountain creator\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('silver tree fountain creator') — 3 results:\n1. Guillaume Boucher — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Guillaume_Boucher\n   The tree was built by Guillaume Boucher, a Parisian goldsmith taken captive during the Mongol raids into Hungary, working with other captured craftsmen.\n2. George Antheil — en.wikipedia.org\n   https://en.wikipedia.org/wiki/George_Antheil\n   Her co-inventor was composer George Antheil, whose experience player-piano mechanisms informed the hopping pattern.\n3. Léon Foucault — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Foucault_pendulum\n   Léon Foucault demonstrated in 1851 that a freely swinging pendulum keeps its plane while the Earth turns beneath it.",
      "tool_call_id": "call_05",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Guillaume Boucher. Next I need: cause.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_06",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Guillaume Boucher cause\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Guillaume Boucher cause') — 3 results:\n1. capture in Hungary — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Guillaume_Boucher\n   Boucher had been seized in the 1241-42 Mongol invasion of Hungary and brought east — a craftsman carried between two worlds by conquest.\n2. Apollo 12 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_12\n   Apollo 12, launched in November 1969, was the second crewed lunar landing mission, crewed by Charles Conrad, Richard Gordon and Alan Bean.\n3. Amazon River — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Amazon_River\n   Brazilian and Peruvian studies since 2007 have measured the Amazon at around 6,800 km by including different source tributaries and the marine estuary.",
      "tool_call_id": "call_06",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Before answering, I'll open the source page for capture in Hungary to confirm the decisive hop.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_07",
          "type": "function",
          "function": {
            "name": "open_url",
            "arguments": "{\"url\": \"https://en.wikipedia.org/wiki/Guillaume_Boucher\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "capture in Hungary — en.wikipedia.org\nhttps://en.wikipedia.org/wiki/Guillaume_Boucher\n\nBoucher had been seized in the 1241-42 Mongol invasion of Hungary and brought east — a craftsman carried between two worlds by conquest.\n\n[page retrieved 2026-11-12; article language may differ from the query language]",
      "tool_call_id": "call_07",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Position: The evidence is one eyewitness description plus the identity of a craftsman; the question's constraints (Parisian, goldsmith, captured in Hungary, tree) must all converge.\nSearch arc:\n1. William of Rubruck, a Flemish Franciscan, travelled to the Mongol capital in 1253-55 as an envoy of the French king Louis IX and wrote a detailed report. [Observed]\n2. He reached Karakorum, the Mongol capital, and described the palace of Great Khan Möngke and its surroundings. [Observed]\n3. Before the palace stood a mechanical silver tree with four lions and serpents, dispensing drinks — the marvel Rubruck reported in detail. [Observed]\n4. The tree was built by Guillaume Boucher, a Parisian goldsmith taken captive during the Mongol raids into Hungary, working with other captured craftsmen. [Observed]\n5. Boucher had been seized in the 1241-42 Mongol invasion of Hungary and brought east — a craftsman carried between two worlds by conquest. [Speculative]\nFusion: The silver tree at Karakorum was made by the Parisian goldsmith Guillaume Boucher.\nBoundary: Rubruck's report is the sole primary source for the tree and its maker; the exact appearance of the fountain is reconstructed from his prose and later illustrations.\nFinal answer: Guillaume Boucher",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "provenance": {
    "source_graph": "curated entity chains over Wikipedia/Wikidata titles (snapshot 2026-11-12)",
    "constructed_at": "2026-11-12",
    "constructor": "HermeneuticDive maintainers (independent project; LLM-assisted authoring pipeline under human direction). Not affiliated with Z.ai / zai-org or the original DeepDive authors.",
    "verification": "every source URL validated 2026-11-12 (MediaWiki API / HTTP check); date-sensitive and risky anchors web-verified (see README: Verification); every hop carries a confidence tag"
  },
  "closed_book_risk": "low",
  "decontamination": {
    "method": "INHERITED FROM v2.0.0, NOT RE-RUN in this build (benchmark dumps unavailable offline). containment decontamination (substring >= 20 chars; token containment >= 0.75; char-shingle(5) Jaccard >= 0.5) — v2.0.0 ran 2026-11-12 against 7900 questions from 2wikimultihopqa, hotpotqa, musique, triviaqa: 0 overlaps, 5 near misses. v3 reworded 2 English questions (hd-en-mh-511, hd-en-mh-518); re-run before publishing. Scope: English open-domain multi-hop QA sets only — not BrowseComp/xbench/GAIA, and no cross-lingual matching for zh/ru/ja/hi questions.",
    "overlaps": []
  },
  "linked_trajectory_ids": [
    "hd-traj-en-mh-504",
    "hd-traj-en-mh-504-fail"
  ],
  "split": "test"
}

qa_rl example (hd-ja-fp-503, false_premise — premise-correction verifier)

{
  "id": "hd-ja-fp-503",
  "language": "ja",
  "task_type": "false_premise",
  "difficulty_mode": "tactical",
  "hop_count": 3,
  "estrangement_level": 2,
  "question": "源義経は衣川で死なず、蝦夷地から大陸へ渡り、後にチンギス・ハンとなった——この説は史実として確立しているか。",
  "answer": {
    "value": "確立していない：義経=チンギス・ハン説は近世以来の俗説・伝承で、史料による根拠がない。義経は1189年に衣川で自刃したと『吾妻鏡』等が伝える",
    "aliases": [
      "俗説：史実ではない",
      "義経北行伝説は根拠を欠く"
    ],
    "answer_type": "explanation",
    "unit": null,
    "tolerance": null,
    "as_of": null,
    "valid_until": null
  },
  "gold_path": [
    {
      "hop": 1,
      "entity": "義経の死",
      "relation": "documented_in",
      "rationale": "『吾妻鏡』等の同時代史料は、義経が1189年に平泉・衣川で自刃したと伝える。",
      "confidence": "observed",
      "source_urls": [
        "https://ja.wikipedia.org/wiki/源義経"
      ]
    },
    {
      "hop": 2,
      "entity": "北行伝説",
      "relation": "disputed_by",
      "rationale": "義経が蝦夷地へ逃れ、大陸でチンギス・ハンになったという話は、江戸期に成立した俗説・創作である。",
      "confidence": "observed",
      "source_urls": [
        "https://ja.wikipedia.org/wiki/源義経"
      ]
    },
    {
      "hop": 3,
      "entity": "premise check",
      "relation": "refuted_by",
      "rationale": "義経の死（1169-1189年没）とチンギス・ハンの実像（モンゴルのテムジン）の年代・経歴は結びつかず、学界はこの説を採らない。",
      "confidence": "observed",
      "source_urls": [
        "https://ja.wikipedia.org/wiki/源義経"
      ]
    }
  ],
  "verifier": {
    "match_rule": "premise_correction",
    "normalization": [
      "unicode_nfc",
      "lowercase",
      "strip_punctuation",
      "whitespace_collapse"
    ],
    "judge_rubric": "Accept only answers that (a) explicitly identify the question's premise as false and (b) supply the corrected fact from gold_path. Canonical correction: '確立していない：義経=チンギス・ハン説は近世以来の俗説・伝承で、史料による根拠がない。義経は1189年に衣川で自刃したと『吾妻鏡』等が伝える'. A bare abstention (cannot_be_determined) without the correction scores 0; an answer that accepts the premise and invents detail scores 0.",
    "abstain_allowed": false
  },
  "messages": [
    {
      "role": "system",
      "content": "You are a deep-search agent with two tools: web_search(query) and open_url(url). Explore step by step, verify each hop against a source, and do not repeat similar queries. When confident, end your reply with 'Final answer: <short answer>'. If the question rests on a false premise, end with 'Final answer: <corrected fact>' preceded by a one-line premise correction. If the answer genuinely cannot be determined from evidence, end with 'Final answer: cannot_be_determined' (in the question's language: 无法确定 / невозможно установить / 特定できません / निर्धारित नहीं किया जा सकता) plus a one-line reason. For date-sensitive facts state the answer with 'as of <date>'.",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "源義経は衣川で死なず、蝦夷地から大陸へ渡り、後にチンギス・ハンとなった——この説は史実として確立しているか。",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "provenance": {
    "source_graph": "curated entity chains over Wikipedia/Wikidata titles (snapshot 2026-11-12)",
    "constructed_at": "2026-11-12",
    "constructor": "HermeneuticDive maintainers (independent project; LLM-assisted authoring pipeline under human direction). Not affiliated with Z.ai / zai-org or the original DeepDive authors.",
    "verification": "every source URL validated 2026-11-12 (MediaWiki API / HTTP check); date-sensitive and risky anchors web-verified (see README: Verification); every hop carries a confidence tag"
  },
  "closed_book_risk": "low",
  "split": "validation"
}

trajectories_sft example (failure: redundant_query_cycling, Hindi)

{
  "id": "hd-traj-hi-mh-501-fail",
  "qa_id": "hd-hi-mh-501",
  "language": "hi",
  "task_type": "multi_hop",
  "outcome": "fail",
  "failure_mode": "redundant_query_cycling",
  "turns": [
    {
      "role": "system",
      "content": "You are a deep-search agent with two tools: web_search(query) and open_url(url). Explore step by step, verify each hop against a source, and do not repeat similar queries. When confident, end your reply with 'Final answer: <short answer>'. If the question rests on a false premise, end with 'Final answer: <corrected fact>' preceded by a one-line premise correction. If the answer genuinely cannot be determined from evidence, end with 'Final answer: cannot_be_determined' (in the question's language: 无法确定 / невозможно установить / 特定できません / निर्धारित नहीं किया जा सकता) plus a one-line reason. For date-sensitive facts state the answer with 'as of <date>'.",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "भारत की सबसे पुरानी जीवित शिला-कट गुफाएँ किसी सम्राट ने बनवाईं और वे एक ऐसे तपस्वी संप्रदाय को समर्पित हैं जो नियति (भाग्य) की सर्वोच्चता मानता था। उस संप्रदाय के संस्थापक कौन थे?",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "गुफाओं के बारे में खोजता हूँ।",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"बाराबर गुफाएँ किसने बनाईं\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('बाराबर गुफाएँ किसने बनाईं') — 3 results:\n1. Barabar Caves — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Barabar_Caves\n   The Barabar Hill Caves are the oldest surviving rock-cut caves in India, from the Maurya period, with inscriptions of Ashoka dedicating them to the Ajivika ascetics.\n2. बिहार पर्यटन — समीक्षा\n3. चित्र।",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "अशोक का नाम आया। थोड़ा और पक्का करता हूँ।",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_02",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"बिहार प्राचीन गुफाएँ अशोक\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('बिहार प्राचीन गुफाएँ अशोक') — 3 results:\n1. Barabar Caves — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Barabar_Caves\n   Mauryan rock-cut caves in Bihar, dedicated by Ashoka to the Ajivikas.\n2. मौर्य काल — समीक्षा\n3. चित्र।",
      "tool_call_id": "call_02",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "वही परिणाम। दूसरे शब्दों में।",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_03",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"बाराबर की गुफाएँ बनाने वाला राजा\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('बाराबर की गुफाएँ बनाने वाला राजा') — 2 results:\n1. Barabar Caves — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Barabar_Caves\n   Built under Ashoka, dedicated to the Ajivika sect.\n2. चित्र।",
      "tool_call_id": "call_03",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "गुफाओं से अशोक ही जुड़े हैं, तो संप्रदाय का संस्थापक भी वही होगा।\nFinal answer: अशोक",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "metrics": {
    "n_turns": 9,
    "n_tool_calls": 3,
    "n_searches": 3,
    "n_page_opens": 0,
    "n_redundant_queries": 2,
    "n_dead_ends": 2,
    "approx_tokens": 451
  },
  "annotations": {
    "lens_moves": [
      {
        "turn_index": 6,
        "lens": "empirical_grounding",
        "note": "अनुपस्थित: 'आजीविक' सूत्र पर एक भी लक्षित खोज नहीं हुई"
      }
    ],
    "redundancy_flags": [
      {
        "turn_index": 3,
        "note": "near-duplicate of an earlier query; no new information"
      },
      {
        "turn_index": 5,
        "note": "near-duplicate of an earlier query; no new information"
      }
    ],
    "lesson": "प्रश्न संस्थापक पूछता है, दाता नहीं। पहले परिणाम में ही 'Ajivika ascetics' लिखा था — एक कदम आगे (आजीविक संप्रदाय का संस्थापक कौन?) खोजने की बजाय वही तथ्य तीन बार दोहराया और दाता (अशोक) को ही उत्तर बना दिया। दोहराव + गलत अंतिम-सूत्र का संयोग।"
  },
  "split": "train"
}

dpo example (behavioural pair on 兰亭序 — note the length gap that pair_stats exposes)

{
  "id": "hd-dpo-zh-un-001",
  "qa_id": "hd-zh-un-001",
  "language": "zh",
  "task_type": "false_premise",
  "failure_mode": "hallucinated_answer_on_false_premise",
  "prompt": [
    {
      "role": "system",
      "content": "You are a deep-search agent with two tools: web_search(query) and open_url(url). Explore step by step, verify each hop against a source, and do not repeat similar queries. When confident, end your reply with 'Final answer: <short answer>'. If the question rests on a false premise, end with 'Final answer: <corrected fact>' preceded by a one-line premise correction. If the answer genuinely cannot be determined from evidence, end with 'Final answer: cannot_be_determined' (in the question's language: 无法确定 / невозможно установить / 特定できません / निर्धारित नहीं किया जा सकता) plus a one-line reason. For date-sensitive facts state the answer with 'as of <date>'.",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "王羲之《兰亭序》的真迹如今收藏在哪座博物馆？",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "chosen": [
    {
      "role": "assistant",
      "content": "先从题面的第一条线索入手，看它指向哪个实体。",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"王羲之《兰亭序》的真迹如今收藏在哪座博物馆\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('王羲之《兰亭序》的真迹如今收藏在哪座博物馆') — 3 results:\n1. 真迹下落 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/兰亭集序\n   《兰亭序》真迹不存于世，传世者皆为唐宋摹本与临本。\n2. 学界异议 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/夏商周断代工程\n   国内外学者对断代工程的测年链条与文献解释提出方法论批评，江晓原等提出前1044年方案，另有学者坚持古本《竹书纪年》系统的前1027年等，克商年至今未获一致。\n3. 2026年棋王战 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/World_Chess_Championship_2026\n   截至2026年11月，古克什仍在卫冕等待中，其与挑战者辛达罗夫（2026年候选人赛冠军）的比赛尚未开赛。",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "结果指向「真迹下落」。下一步需要查：反驳。",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_02",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"真迹下落 反驳\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('真迹下落 反驳') — 3 results:\n1. 史载归宿 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/兰亭集序\n   唐宋以来文献多称真迹随葬唐太宗昭陵，但昭陵从未经考古发掘证实，故「现藏某馆」的前提不成立。\n2. 史载归宿 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/昭陵\n   (no snippet — open the page for details)\n3. 讽严武说 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/蜀道難\n   唐宋以来有记载认为此诗讽刺章仇兼琼或（后起诸说）严武虐政、友人恐遭不测，沈括等亦有辨析。",
      "tool_call_id": "call_02",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Position: 问题预设「真迹尚存且在某馆」，而事实层是「真迹不存」——应当拒答并纠正前提。\nSearch arc:\n1. 《兰亭序》真迹不存于世，传世者皆为唐宋摹本与临本。 [Observed]\n2. 唐宋以来文献多称真迹随葬唐太宗昭陵，但昭陵从未经考古发掘证实，故「现藏某馆」的前提不成立。 [Observed]\nFusion: 前提不成立：《兰亭序》真迹已不存于世，传世者为唐摹本；若问最佳摹本的下落，神龙本今藏北京故宫博物院。\nBoundary: 若答「北京故宫博物院」，那是神龙本摹本的位置，不是真迹。\nFinal answer: 《兰亭序》真迹已不存于世：传世者为唐摹本，最著名的神龙本今藏北京故宫博物院",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "rejected": [
    {
      "role": "assistant",
      "content": "查一下兰亭序的收藏情况。",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"兰亭序 收藏 博物馆\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('兰亭序 收藏 博物馆') — 3 results:\n1. 兰亭集序 — zh.wikipedia.org\n   https://zh.wikipedia.org/wiki/兰亭集序\n   《兰亭序》真迹不存于世，传世者皆为唐摹本，其中以神龙本最著名，现藏北京故宫博物院。\n2. 神龙本兰亭集序 — 概述\n3. 图片结果。",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "神龙本藏于北京故宫博物院，那真迹自然也在那里。\nFinal answer: 北京故宫博物院",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "pair_stats": {
    "chosen_turns": 5,
    "rejected_turns": 3,
    "chosen_tool_calls": 2,
    "rejected_tool_calls": 1,
    "chosen_chars": 1162,
    "rejected_chars": 229,
    "char_ratio": 5.07,
    "length_confounded": true
  },
  "split": "train"
}

dpo_matched example (hd-dpom-en-mh-001: identical trajectory; the rejected side answers with the intermediate entity)

{
  "id": "hd-dpom-en-mh-001",
  "qa_id": "hd-en-mh-001",
  "language": "en",
  "task_type": "multi_hop",
  "failure_mode": "intermediate_entity_answer",
  "prompt": [
    {
      "role": "system",
      "content": "You are a deep-search agent with two tools: web_search(query) and open_url(url). Explore step by step, verify each hop against a source, and do not repeat similar queries. When confident, end your reply with 'Final answer: <short answer>'. If the question rests on a false premise, end with 'Final answer: <corrected fact>' preceded by a one-line premise correction. If the answer genuinely cannot be determined from evidence, end with 'Final answer: cannot_be_determined' (in the question's language: 无法确定 / невозможно установить / 特定できません / निर्धारित नहीं किया जा सकता) plus a one-line reason. For date-sensitive facts state the answer with 'as of <date>'.",
      "tool_call_id": null,
      "tool_calls": []
    },
    {
      "role": "user",
      "content": "The only legitimate daughter of the poet who praised a woman whose beauty is like 'night' and 'starry skies' went on to add extensive notes to the first translation of an Italian engineer's memoir about a certain calculating machine. Name the never-completed earlier calculating engine that the machine's designer built with the engineer Joseph Clement.",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "chosen": [
    {
      "role": "assistant",
      "content": "I'll start from the question's first clue and see which entity it points to.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"The only legitimate daughter of the poet who praised a woman\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('The only legitimate daughter of the poet who praised a woman') — 3 results:\n1. Lord Byron — en.wikipedia.org\n   https://en.wikipedia.org/wiki/She_Walks_in_Beauty\n   The poem 'She Walks in Beauty' was written by Lord Byron, whose only legitimate daughter was Ada Lovelace.\n2. designation check — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Venice_of_the_North\n   No international or municipal authority assigns the nickname officially to one city; it is a multiple-use epithet, so the uniqueness premise fails.\n3. Fort Mackinac — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Alexis_St._Martin\n   In June 1822 at Fort Mackinac on Mackinac Island, Alexis St. Martin, a French-Canadian voyageur, was accidentally shot in the upper left abdomen.",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Lord Byron. Next I need: family.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_02",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Lord Byron family\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Lord Byron family') — 3 results:\n1. Ada Lovelace — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Ada_Lovelace\n   Ada Lovelace translated Luigi Menabrea's 1842 memoir on Babbage's Analytical Engine from French and appended her extensive Notes.\n2. Emily Warren Roebling — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Emily_Warren_Roebling\n   His wife, Emily Warren Roebling, studied cable construction and engineering, relayed his instructions, inspected work and dealt with contractors and officials f…\n3. modern estimates — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Nineveh\n   Modern estimates for ancient Nineveh's population range from tens of thousands to several hundred thousand depending on the walled area counted and density assu…",
      "tool_call_id": "call_02",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Ada Lovelace. Next I need: source.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_03",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Ada Lovelace source\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Ada Lovelace source') — 3 results:\n1. Analytical Engine — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Analytical_Engine\n   The memoir and Lovelace's notes describe Charles Babbage's Analytical Engine, a proposed mechanical general-purpose computer.\n2. J. J. Thomson — en.wikipedia.org\n   https://en.wikipedia.org/wiki/J._J._Thomson\n   Rutherford had earlier worked at the Cavendish Laboratory under J. J. Thomson, who identified the electron in 1897.\n3. Nile — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Nile\n   Most encyclopedic references list the Nile at roughly 6,650 km as the world's longest river.",
      "tool_call_id": "call_03",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Analytical Engine. Next I need: inventor.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_04",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Analytical Engine inventor\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Analytical Engine inventor') — 3 results:\n1. Charles Babbage — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Charles_Babbage\n   The Analytical Engine was designed by Charles Babbage, who had earlier begun a simpler calculating machine with engineer Joseph Clement.\n2. Karakorum — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Karakorum\n   He reached Karakorum, the Mongol capital, and described the palace of Great Khan Möngke and its surroundings.\n3. callsigns — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_12\n   Its command module was named Yankee Clipper and its lunar module Intrepid — names the crew chose.",
      "tool_call_id": "call_04",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Charles Babbage. Next I need: predecessor or successor.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_05",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Charles Babbage predecessor or successor\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Charles Babbage predecessor or successor') — 3 results:\n1. Difference Engine — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Difference_Engine\n   That earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime.\n2. Apollo 12 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_12\n   Apollo 12, launched in November 1969, was the second crewed lunar landing mission, crewed by Charles Conrad, Richard Gordon and Alan Bean.\n3. pen names — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Bront%C3%AB\n   The three Brontë sisters published as Currer, Ellis and Acton Bell — Charlotte, Emily and Anne respectively.",
      "tool_call_id": "call_05",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Before answering, I'll open the source page for Difference Engine to confirm the decisive hop.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_06",
          "type": "function",
          "function": {
            "name": "open_url",
            "arguments": "{\"url\": \"https://en.wikipedia.org/wiki/Difference_Engine\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "Difference Engine — en.wikipedia.org\nhttps://en.wikipedia.org/wiki/Difference_Engine\n\nThat earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime.\n\n[page retrieved 2026-11-12; article language may differ from the query language]",
      "tool_call_id": "call_06",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Position: The question presupposes a famous Romantic poet identifiable from a single quoted phrase, and assumes the daughter's scientific work is traceable to a specific machine and designer.\nSearch arc:\n1. The poem 'She Walks in Beauty' was written by Lord Byron, whose only legitimate daughter was Ada Lovelace. [Observed]\n2. Ada Lovelace translated Luigi Menabrea's 1842 memoir on Babbage's Analytical Engine from French and appended her extensive Notes. [Observed]\n3. The memoir and Lovelace's notes describe Charles Babbage's Analytical Engine, a proposed mechanical general-purpose computer. [Observed]\n4. The Analytical Engine was designed by Charles Babbage, who had earlier begun a simpler calculating machine with engineer Joseph Clement. [Observed]\n5. That earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime. [Observed]\nFusion: The machine's designer is Charles Babbage, whose earlier calculating engine, begun in the 1820s with Joseph Clement and never completed in his lifetime, was the Difference Engine.\nBoundary: The Difference Engine project passed through several numbered designs; 'the Difference Engine' canonically refers to Difference Engine No. 1 begun with Clement.\nFinal answer: the Difference Engine",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "rejected": [
    {
      "role": "assistant",
      "content": "I'll start from the question's first clue and see which entity it points to.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_01",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"The only legitimate daughter of the poet who praised a woman\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('The only legitimate daughter of the poet who praised a woman') — 3 results:\n1. Lord Byron — en.wikipedia.org\n   https://en.wikipedia.org/wiki/She_Walks_in_Beauty\n   The poem 'She Walks in Beauty' was written by Lord Byron, whose only legitimate daughter was Ada Lovelace.\n2. designation check — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Venice_of_the_North\n   No international or municipal authority assigns the nickname officially to one city; it is a multiple-use epithet, so the uniqueness premise fails.\n3. Fort Mackinac — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Alexis_St._Martin\n   In June 1822 at Fort Mackinac on Mackinac Island, Alexis St. Martin, a French-Canadian voyageur, was accidentally shot in the upper left abdomen.",
      "tool_call_id": "call_01",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Lord Byron. Next I need: family.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_02",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Lord Byron family\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Lord Byron family') — 3 results:\n1. Ada Lovelace — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Ada_Lovelace\n   Ada Lovelace translated Luigi Menabrea's 1842 memoir on Babbage's Analytical Engine from French and appended her extensive Notes.\n2. Emily Warren Roebling — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Emily_Warren_Roebling\n   His wife, Emily Warren Roebling, studied cable construction and engineering, relayed his instructions, inspected work and dealt with contractors and officials f…\n3. modern estimates — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Nineveh\n   Modern estimates for ancient Nineveh's population range from tens of thousands to several hundred thousand depending on the walled area counted and density assu…",
      "tool_call_id": "call_02",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Ada Lovelace. Next I need: source.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_03",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Ada Lovelace source\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Ada Lovelace source') — 3 results:\n1. Analytical Engine — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Analytical_Engine\n   The memoir and Lovelace's notes describe Charles Babbage's Analytical Engine, a proposed mechanical general-purpose computer.\n2. J. J. Thomson — en.wikipedia.org\n   https://en.wikipedia.org/wiki/J._J._Thomson\n   Rutherford had earlier worked at the Cavendish Laboratory under J. J. Thomson, who identified the electron in 1897.\n3. Nile — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Nile\n   Most encyclopedic references list the Nile at roughly 6,650 km as the world's longest river.",
      "tool_call_id": "call_03",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Analytical Engine. Next I need: inventor.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_04",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Analytical Engine inventor\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Analytical Engine inventor') — 3 results:\n1. Charles Babbage — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Charles_Babbage\n   The Analytical Engine was designed by Charles Babbage, who had earlier begun a simpler calculating machine with engineer Joseph Clement.\n2. Karakorum — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Karakorum\n   He reached Karakorum, the Mongol capital, and described the palace of Great Khan Möngke and its surroundings.\n3. callsigns — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_12\n   Its command module was named Yankee Clipper and its lunar module Intrepid — names the crew chose.",
      "tool_call_id": "call_04",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "The result points to Charles Babbage. Next I need: predecessor or successor.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_05",
          "type": "function",
          "function": {
            "name": "web_search",
            "arguments": "{\"query\": \"Charles Babbage predecessor or successor\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "web_search('Charles Babbage predecessor or successor') — 3 results:\n1. Difference Engine — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Difference_Engine\n   That earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime.\n2. Apollo 12 — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Apollo_12\n   Apollo 12, launched in November 1969, was the second crewed lunar landing mission, crewed by Charles Conrad, Richard Gordon and Alan Bean.\n3. pen names — en.wikipedia.org\n   https://en.wikipedia.org/wiki/Bront%C3%AB\n   The three Brontë sisters published as Currer, Ellis and Acton Bell — Charlotte, Emily and Anne respectively.",
      "tool_call_id": "call_05",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Before answering, I'll open the source page for Difference Engine to confirm the decisive hop.",
      "tool_call_id": null,
      "tool_calls": [
        {
          "id": "call_06",
          "type": "function",
          "function": {
            "name": "open_url",
            "arguments": "{\"url\": \"https://en.wikipedia.org/wiki/Difference_Engine\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": "Difference Engine — en.wikipedia.org\nhttps://en.wikipedia.org/wiki/Difference_Engine\n\nThat earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime.\n\n[page retrieved 2026-11-12; article language may differ from the query language]",
      "tool_call_id": "call_06",
      "tool_calls": []
    },
    {
      "role": "assistant",
      "content": "Position: The question presupposes a famous Romantic poet identifiable from a single quoted phrase, and assumes the daughter's scientific work is traceable to a specific machine and designer.\nSearch arc:\n1. The poem 'She Walks in Beauty' was written by Lord Byron, whose only legitimate daughter was Ada Lovelace. [Observed]\n2. Ada Lovelace translated Luigi Menabrea's 1842 memoir on Babbage's Analytical Engine from French and appended her extensive Notes. [Observed]\n3. The memoir and Lovelace's notes describe Charles Babbage's Analytical Engine, a proposed mechanical general-purpose computer. [Observed]\n4. The Analytical Engine was designed by Charles Babbage, who had earlier begun a simpler calculating machine with engineer Joseph Clement. [Observed]\n5. That earlier project, begun in the 1820s with Clement, was the Difference Engine, which Babbage never completed in his lifetime. [Observed]\nFusion: The machine's designer is Charles Babbage, whose earlier calculating engine, begun in the 1820s with Joseph Clement and never completed in his lifetime, was the Difference Engine.\nBoundary: The Difference Engine project passed through several numbered designs; 'the Difference Engine' canonically refers to Difference Engine No. 1 begun with Clement.\nFinal answer: Charles Babbage",
      "tool_call_id": null,
      "tool_calls": []
    }
  ],
  "pair_stats": {
    "chosen_turns": 13,
    "rejected_turns": 13,
    "chosen_tool_calls": 6,
    "rejected_tool_calls": 6,
    "chosen_chars": 5508,
    "rejected_chars": 5502,
    "char_ratio": 1.0,
    "length_confounded": false
  },
  "split": "train"
}

Data Fields

Schemas are declared per config in the YAML dataset_info above. Shared qa fields: id, language, task_type, difficulty_mode, hop_count, estrangement_level (1–3), question, answer (value, aliases, answer_type, unit, tolerance, as_of, valid_until), gold_path (entity, relation, rationale, confidence, source_urls), provenance, closed_book_risk. qa_sft adds hermeneutics (pre_understanding, ≥3 lens moves, synthesis) and messages (tool-calling turns; every tool result carries tool_call_id); qa_rl adds verifier (match_rule, normalization, judge_rubric, abstain_allowed); trajectories_sft carries outcome/failure_mode labels, metrics (incl. n_dead_ends) and per-turn lens annotations; dpo and dpo_matched carry prompt / chosen / rejected message lists plus pair_stats (turns, tool calls, characters, char_ratio, length_confounded). answer.valid_until is a scheduled flip date or, where no flip is scheduled, a 90-day re-verification horizon (as_of + 90 days).

Data Splits

Config	Train	Validation	Test	Total
qa_sft	40	8	8	56
qa_rl	92	18	23	133
trajectories_sft	24	2	2	28
dpo	12	1	1	14
dpo_matched	12	4	1	17

Splits are stratified by language × task type with enforced coverage guarantees (multi_hop in test; unanswerable in train; false_premise in train and test; estrangement-3 and deep chains in train and test); qa_sft and qa_rl question sets are fully disjoint (asserted); trajectories and DPO pairs inherit their question's split. Statistical power is very low: a 95% interval for an accuracy near 50% is about ±20 points on qa_rl test (n = 23), ±23 on validation (n = 18) and ±35 on qa_sft test/validation (n = 8); the DPO test splits hold 1 pair each. Treat held-out numbers as smoke tests, pool splits, and report intervals.

Dataset Creation

Curation Rationale

v2 implemented a first external audit of v1 (scale, difficulty, split design, SFT/RL overlap, thin trajectories, verifier gaps). v3 implements a second audit — of v2 itself. Findings, fixes and checks:

#	v2 finding (measured)	v3 fix	How it was checked
1	Clairvoyant trajectories. 98 of 215 SFT queries named a gold entity before any result had revealed it; 45/56 first queries were the hidden entity; 52/69 open_url calls opened a URL no result had surfaced; every result was a verbatim gold sentence; interstitial text came from 6 fixed strings.	Generator rewritten: first query = the question's own opening clue; later queries = previously revealed entity + a relation word; open_url only on URLs already surfaced; dead ends no longer name the target; distractor hits are real facts from other rows filtered never to mention this row's entities/answer; thought text is derived from state (162 distinct strings).	causal_violations() asserted for all 70 success trajectories at build; independent re-measure: 0 gold first queries, 0/38 unseen URLs, 0 answer-before-seen (1 residual hit is a generic relation word that is also a concept name).
2	Vacuous decontamination. A rebuild without benchmark dumps passed and stamped every row "ran against 0 questions: 0 overlaps"; the dumps were not shipped.	NoBenchmarks is raised unless HD_ALLOW_NO_BENCH=1; skipped runs are labelled INHERITED FROM v2, NOT RE-RUN, with scope and changed-question list.	Rebuild without the flag fails; rows carry the inherited label.
3	deep applied to 24 non-multi-hop rows with 2–3 hops (claim "deep ⇒ ≥ 5 hops, asserted" held for multi_hop only).	Difficulty ladder applies to every task type.	Assertion over all rows; deep rows with < 5 hops: 0.
4	5 date_sensitive rows without valid_until.	All 18/18 carry one (scheduled flip, or a 90-day horizon); sumo row set to 2026-11-22 (Kyushu basho end).	Build assertion (also: not already expired).
5	10 canonical answers were sentences or carried parenthetical qualifiers (e.g. "Gukesh Dommaraju (title defence not yet underway)"); explanation-typed answers were string-matched.	Clean canonical entities; explanation answers routed to llm_judge.	Build assertion; verifier.py self-test.
6	Content: mh-511 ("between two Asian capitals' countries" — Shanghai is not a capital); mh-518 ("Welsh" philologist, London-born).	Reworded; wording changes recorded for re-decontamination.	Manual; recorded in pipeline/rows_patch_v3.py.
7	DPO length confound. 14/14 pairs; median 5.07× characters; chosen had more tool calls in 13/14.	pair_stats on every pair; new dpo_matched config (17 pairs, char ratio 0.99–1.0, equal tool calls, identical except the final answer).	Asserted in create_dataset.py: matched pairs differ only in the final answer line.
8	Verifier specified as prose only; rewards not reproducible.	verifier.py + --selftest.	The self-test found a real bug: SFT numeric answers ended "Final answer: 164.8" with no unit, which the RL verifier rejects → final answers now carry units.
9	Card metadata wrong: size_categories: 1K<n<10K for 248 rows; stale download_size values.	YAML regenerated from features_*.json + stats.json.	pipeline/card_tools.py --check.

Construction methodology (what "hermeneutic and system methods" mean here)

Each row is built as an interpretation problem, in five layers:

1. Systemic chain design (lens: systemic embedding). Chains are authored to cross systemic layers — institutional, historical, cultural, geographic, material — with a controlled vocabulary of 29 typed relations. v2 banks deliberately target the less common: recovered books (《营造法式》《天工开物》), unnamed collaborators (梁令瓒, 後藤貞行, Leo Stern), institutional hinterlands (Fort Ross, Кунсткамера, संविधान के हस्तलेखक).
2. Graded estrangement (obfuscation). Level 1 replaces a name with its role; level 2 with indirect attributes; level 3 leaves the entity recoverable only through several joint constraints, with decoys (9 rows in v2 — v1 shipped none).
3. The hermeneutic arc (question → answer). The final assistant message follows a fixed arc — Position, Search arc (each hop tagged [Observed]/[Inferred]/[Speculative]; distribution 518/34/6), Fusion, Boundary (a row-specific line on every row — no boilerplate), Final answer.
4. Lens annotations. Every qa_sft row records ≥ 3 Integral-Hermeneutics lens moves (v1 averaged ~2), tied to hops.
5. Verifiable answers. Answer types now span entity/person/place/work/organization/title/date/number/explanation/abstain; per-row verifiers implement match rules including numeric tolerance, date parsing and premise correction.

Failure trajectories encode seven failure modes as training signal — false_conflict_resolution x 1, hallucinated_answer_on_false_premise x 3, premature_answer x 2, redundant_query_cycling x 3, stale_knowledge_answer x 2, ungrounded_abstention x 1, wrong_terminal_entity x 2 — each with a lesson field explaining the correct behaviour, and each paired with a success on the same question for DPO.

Source Data

Source	Use
Wikipedia (en/zh/ru/ja/hi)	Primary reference layer; 486 canonical titles confirmed to exist via the MediaWiki API on 2026-11-12 (v2; not re-run in v3 — URL set unchanged). Existence is not evidence that the page supports the hop's claim.
nobelprize.org, fifa.com, FIDE-derived press, NASA, USDA ARS, Science History Institute, Britannica, Vatican News, NLM, PIB India, President of India, royal.uk, Kyodo News	Institutional anchors for prize, event, artefact and date facts

No text is copied from sources: questions, rationales and trajectory tool-results are original authored wording over factual skeletons (see Licensing).

Annotations

Machine-generated under human direction (author pipeline: HermeneuticDive maintainers, LLM-assisted, 2026-11): hop rationales, lens moves, confidence tags, difficulty modes, closed-book-risk estimates, verifier specs, trajectory lessons.

Personal and Sensitive Information

Public historical and institutional figures only. No PII patterns (email/phone/ID) appear in any config.

Verification

What was checked, when, and what that does and does not show:

• Live-verified anchors (re-verified 2026-11-12 in v3): 2026 Nobel Peace Prize (Navi Pillay) and Literature (Anne Carson, 8 Oct 2026); 2026 World Cup final (Spain 1–0 Argentina, extra time); both yokozuna (大の里, 豊昇龍) active on day 4 of the Kyushu basho (11 Nov); the Gukesh–Sindarov match starts 24 Nov. Other anchors (UK monarch, Pope, UN Secretary-General, President of India, Voyager 1) agree with the reviewer's own knowledge but were not re-searched in v3.
• URLs (v2 run, not re-run in v3): 486 Wikipedia titles confirmed to exist (0 missing) and 22 institutional URLs answered 200/226/403/429 (403/429 counted as "ok" — a bot-wall response does not prove the page exists). url_check_v2.json. This shows existence only; it does not show the page supports the hop.
• Decontamination — INHERITED, not re-run in v3. v2 ran containment metrics (normalised substring ≥ 20 chars; token containment ≥ 0.75; char-shingle(5) Jaccard ≥ 0.5) against 7,900 questions from HotpotQA, 2WikiMultihopQA, MuSiQue and TriviaQA: 0 overlaps, 5 sub-threshold near misses. The dumps are not shipped, so this cannot be reproduced from the repository; v3 reworded two English questions since. Scope is English open-domain multi-hop QA — not BrowseComp/xbench/GAIA, and there is no cross-lingual matching for zh/ru/ja/hi. A rebuild without dumps now fails (or is labelled inherited). Reports: decontamination_report.json (current), decontamination_report_v2.json.
• Trajectory causality (v3, asserted): every web_search query and open_url URL in the 70 success trajectories uses only strings from the question or an earlier tool result. Six of the 14 hand-authored failure trajectories contain ungrounded queries — by design (a failing agent guesses from memory) — which is also a mild stylistic confound for dpo.
• Verifier (v3): python verifier.py --selftest passes (positive and negative controls; routing equals the shipped match_rule for all 133 rows; every SFT final answer passes its own verifier; in dpo_matched every chosen passes and every rejected fails).
• Duplicates / splits / integrity: near-duplicate question check (char-shingle Jaccard > 0.35) passes; split assignment is unchanged from v2; tool_call_id integrity and SFT/RL disjointness asserted in create_dataset.py; pipeline/card_tools.py --check ties card numbers to data.

Considerations for Using the Data

Social Impact of Dataset

Intended to make deep-search agents more honest: abstention, dispute acknowledgment, premise correction, date-stamping and confidence-tagged reasoning are trained as first-class behaviours.

Discussion of Biases

Source layer is encyclopedic (Wikipedia-dominant) and skews toward documented, notable entities; en dominates the 189-question mix (74 rows); ja/hi rows lean on canonical-but-regional history. All rows come from a single LLM-assisted pipeline under human direction — a position the methodology itself requires disclosing (lens 1, reflexive grounding).

Other Known Limitations

1. Scale. 189 questions / 248 rows against the original DeepDive's 3,250 questions (~6%). This is a methodology-demonstrating release; the pipeline is the scaling instrument, but no scaled run has been done.
2. Difficulty is still below DeepDive's. Mean multi-hop depth is 3.23 hops (original construction: 5–9); only 17 of 109 multi-hop chains reach 5 hops, 27 have ≤ 2. Many chains use famous entities a strong model may answer closed-book.
3. Tool results are synthesised, not retrieved. Search/open results are authored from the gold path (mean 520 chars in qa_sft); distractor hits are real facts from other rows. Queries are templated (" "); result titles are gold-path entity labels, which for concept/event hops differ from the URL's real page title; the opened page shows the same sentence the snippet showed. The data teaches the shape of causal search, not real-world retrieval noise. Train on it alongside real rollouts.
4. Closed-book risk is an author estimate (63 low / 73 medium / 53 high), not a measured pass rate. Measure pass@k before using tiers as a curriculum.
5. Decontamination is inherited and narrow (see Verification).
6. Snapshot dating. as_of is 2026-11-12. valid_until is a scheduled flip for some rows (chess 2026-11-24, UN Secretary-General 2026-12-31, sumo 2026-11-22) and a 90-day re-verification horizon for others; re-check before use after the date.
7. Evaluation power is tiny (Data Splits). Verifier heuristics (premise correction, abstention) are keyword-based and 24 rows need an LLM judge; per-row rubrics are provided, a judge is not.
8. Preference data. dpo is length-confounded (use pair_stats / a length-normalised loss); dpo_matched isolates answer correctness but is narrow (one failure mode: intermediate-entity answer) and has only 1 test pair.
9. Source layer is 96% Wikipedia and skews to notable, documented entities; en is 39% of questions; ja/hi lean on canonical regional history. All rows come from one LLM-assisted pipeline under human direction.
10. URL support is unverified at claim level (existence checked only); the licensing statement that all prose is original wording was not checked by an automated text-reuse scan.

Additional Information

Dataset Curators

HermeneuticDive maintainers — an independent project (LLM-assisted authoring pipeline under human direction), November 2026. Not affiliated with Z.ai / zai-org or the original DeepDive authors; the original release is credited as methodology inspiration.

Licensing Information

CC0 1.0 Universal applies to all authored content: questions, obfuscated phrasing, hop rationales, hermeneutic annotations, verifier specs, trajectory thought-text and tool-result renders, documentation and code. Rationale: (a) factual skeletons — dates, names, relations — are uncopyrightable facts; (b) all prose in this dataset is original wording authored for this project; the prose is intended as original wording and no text is knowingly reproduced from Wikipedia or any other source (no automated text-reuse scan has been run; tool-result snippets are authored renders that carry URLs as citations); (c) source articles remain under their own licenses (Wikipedia: CC BY-SA 4.0) and are credited via gold_path.source_urls on every hop — URLs and titles are facts, and attribution travels with every row. If you relicense derivative benchmark dumps or import source text, their terms apply to that material, not to this corpus.

Citation Information

@dataset{hermeneuticdive2026,
  author = {HermeneuticDive maintainers},
  title  = {HermeneuticDive: A Five-Language Deep-Search Corpus with Hermeneutic Provenance},
  year   = {2026},
  month  = {november},
  version = {2.0.0},
  license = {CC0-1.0},
  note   = {Independent reimagining of the DeepDive methodology; not affiliated with Z.ai}
}

Methodology lineage (verified sources)

• DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL — arXiv:2509.10446 (Apache-2.0 release; lineage of the KG-synthesis + obfuscation + difficulty-filter recipe)
• WebSailor / SailorFog-QA — arXiv:2507.02592 (uncertainty-level injection → our estrangement levels)
• WebShaper — arXiv:2507.15061 (formal set-theoretic control → our typed-relation chain vocabulary)
• ASearcher — arXiv:2508.07976 (long-horizon agentic RL → our trajectory metric fields)
• R1-Searcher — arXiv:2503.05592; Search-R1 — arXiv:2503.09516 (RL-with-search baselines)
• FreshQA — arXiv:2310.03214 (time-sensitive answers and false premises → date_sensitive / false_premise rows)
• UnSeenTimeQA — ACL 2025 (contamination-free time-sensitive QA → as-of and valid-until stamping)
• BrowseComp-zh — arXiv:2504.19314 (multilingual deep-search evaluation)
• NVIDIA NeMo Data Designer search-agent recipe (Wikidata-walk seeds → obfuscated riddles → tool-calling trajectories; independent confirmation of the seed-to-trajectory architecture used here)
• Integral Hermeneutics methodology — ClawHub @darkd/integral-hermeneutics (the eight-lens interpretation framework used for construction and annotation)
• Human-AI co-interpretation for responsible AI — arXiv:2609.00334 (positioning of hermeneutic machine practice; a general hermeneutics-of-AI paper, not a search-data methodology — cited for framing only)

Contributions

New languages, new long-tail banks, and scale-up runs of the shipped pipeline are welcome via repository discussions and pull requests.

Changelog

• v3.0.0 (2026-11-12, second audit): causal (non-oracular) trajectories with build-time assertion; fail-loud decontamination with honest inherited labelling; deep ⇒ ≥ 5 hops for every task type; valid_until on all date-sensitive rows; clean canonical answers; two question rewordings (mh-511, mh-518); pair_stats + new length-matched dpo_matched config (17 pairs); reference verifier.py with self-test (found and fixed missing units in SFT numeric answers); card YAML regenerated, card_tools.py added. Same 189 questions and the same split assignment as v2; 248 rows in 5 configs.
• v2.0.0 (2026-11-12): implements the external audit — 189 questions (en/zh/ru + new ja/hi), 231 rows, 4 configs (qa_sft, qa_rl, trajectories_sft, dpo), false_premise type with premise-correction scoring, numeric/date/title answer types with tolerance and valid_until, estrangement level 3, stratified splits with enforced coverage, disjoint SFT/RL pools, tool-calling SFT trajectories with tool_call_id, dead ends and final verification, containment decontamination run against 7,900 benchmark questions, all 508 URLs validated, authoring pipeline shipped in pipeline/, attribution and licensing corrected.
• v1.0.0 (2026-11-11): initial release — 90 questions / 156 rows, en/zh/ru, 5 task types, 30 trajectories, 253 verified URLs.