[
  {
    "id": "seed-parquet-full-workflow",
    "anchors": [
      "overview-anchor"
    ],
    "title": [
      "从轨迹到训练文件",
      "From trace to training file",
      "軌跡から学習ファイルへ"
    ],
    "caption": [
      "先固定轨迹，再转换协议与监督位置，最后检查文件与训练准入。",
      "Fix the source traces first, then serialize and mask them; check both file integrity and training admission.",
      "元の軌跡を確定し、形式と学習対象を変換してから、ファイルと学習への受け入れを確認する。"
    ],
    "width": 1140,
    "height": 470,
    "nodes": {
      "trace": [
        50,
        40,
        300,
        135,
        [
          "原始轨迹",
          "Native trace",
          "元の軌跡"
        ],
        [
          "messages + tools",
          "messages + tools",
          "messages + tools"
        ]
      ],
      "select": [
        420,
        40,
        300,
        135,
        [
          "筛选与划分",
          "Select & split",
          "選別と分割"
        ],
        [
          "质量 · 去重 · 分组",
          "Quality · dedup · split",
          "品質・重複除去・分割"
        ]
      ],
      "protocol": [
        790,
        40,
        300,
        135,
        [
          "协议转换",
          "Serialize",
          "形式変換"
        ],
        [
          "系统前缀 · XML65",
          "Headers · XML65",
          "ヘッダー・XML65"
        ]
      ],
      "segments": [
        790,
        285,
        300,
        135,
        [
          "标记学习位置",
          "Mark supervision",
          "学習対象を指定"
        ],
        [
          "片段 + has_loss",
          "Segments + has_loss",
          "断片 + has_loss"
        ],
        "soft"
      ],
      "parquet": [
        420,
        285,
        300,
        135,
        [
          "Parquet",
          "Parquet",
          "Parquet"
        ],
        [
          "四列封装",
          "Four columns",
          "4列に格納"
        ]
      ],
      "check": [
        50,
        285,
        300,
        135,
        [
          "读回与准入",
          "Read & validate",
          "読戻しと検証"
        ],
        [
          "一致性 · 长度",
          "Integrity · length",
          "整合性・長さ"
        ],
        "result"
      ]
    },
    "edges": [
      [
        "trace",
        "select",
        [
          [
            350,
            107
          ],
          [
            420,
            107
          ]
        ],
        false
      ],
      [
        "select",
        "protocol",
        [
          [
            720,
            107
          ],
          [
            790,
            107
          ]
        ],
        false
      ],
      [
        "protocol",
        "segments",
        [
          [
            940,
            175
          ],
          [
            940,
            285
          ]
        ],
        false
      ],
      [
        "segments",
        "parquet",
        [
          [
            790,
            352
          ],
          [
            720,
            352
          ]
        ],
        false
      ],
      [
        "parquet",
        "check",
        [
          [
            420,
            352
          ],
          [
            350,
            352
          ]
        ],
        false
      ]
    ],
    "labels": [],
    "nodeDetails": {
      "trace": [
        "保留原生消息、工具定义和调用标识。转换只改表示方式，不重新生成动作或工具结果。",
        "Preserve native messages, tool definitions, and call identifiers. Conversion changes representation; it does not regenerate actions or tool results.",
        "元のメッセージ、ツール定義、呼び出しIDを保持する。変換で行動や実行結果を作り直さない。"
      ],
      "select": [
        "先按固定质量条件选取，再去重；同一 query 的多个尝试保留在同一个集合。",
        "Apply fixed quality criteria before deduplication. Keep attempts for the same query in the same split.",
        "固定した品質条件で選別してから重複を除く。同じqueryの試行を同じ集合に置く。"
      ],
      "protocol": [
        "ThinkSP、FunctionSP、MainSP 组成系统前缀；assistant 的原生字段转换成目标协议。",
        "ThinkSP, FunctionSP, and MainSP form the system prefix; native assistant fields become the target protocol.",
        "ThinkSP、FunctionSP、MainSPがシステム前置部を構成し、assistantの元フィールドを対象形式に変換する。"
      ],
      "segments": [
        "本次导出只监督 assistant 正文及其 EOS；上下文仍保留并可供后续读取。",
        "This export supervises assistant bodies and EOS markers. Context is retained for later positions to read.",
        "この出力ではassistant本文とEOSを学習対象とする。文脈は後続位置から読めるよう保持する。"
      ],
      "check": [
        "文件读回与源记录一致后，还需使用目标 tokenizer / Checker 核对长度和加载要求。207 条对照导出未运行官方 Checker。",
        "After readback equality, the target tokenizer / Checker must verify length and loader requirements. The matched 207-trace export did not run the official Checker.",
        "読み戻しの一致に加え、対象tokenizer / Checkerで長さと読み込み条件を確認する。207本の比較用出力では公式Checkerは未実行。"
      ]
    }
  },
  {
    "id": "seed-parquet-protocol",
    "anchors": [
      "protocol-anchor"
    ],
    "title": [
      "把原生 assistant 转成协议文本",
      "Serialize the native assistant",
      "assistantをプロトコル本文に変換する"
    ],
    "caption": [
      "三类原生字段合成 assistant 正文；工具返回保持为独立的 tool 消息。",
      "Three native fields form the assistant payload; the tool response remains a separate message.",
      "3種類の元フィールドをassistant本文にまとめ、ツール応答は別のメッセージに保つ。"
    ],
    "width": 1140,
    "height": 760,
    "nodes": {
      "reason": [
        45,
        35,
        290,
        135,
        [
          "原生推理",
          "Native reasoning",
          "元の推論"
        ],
        [
          "reasoning_content",
          "reasoning_content",
          "reasoning_content"
        ]
      ],
      "body": [
        45,
        220,
        290,
        135,
        [
          "回答正文",
          "Answer text",
          "回答本文"
        ],
        [
          "content",
          "content",
          "content"
        ]
      ],
      "calls": [
        45,
        405,
        290,
        135,
        [
          "结构化动作",
          "Structured action",
          "構造化した行動"
        ],
        [
          "tool_calls",
          "tool_calls",
          "tool_calls"
        ]
      ],
      "think": [
        450,
        35,
        300,
        135,
        [
          "协议推理区段",
          "Reasoning section",
          "推論区間"
        ],
        [
          "约定的标签",
          "Configured tags",
          "設定したタグ"
        ]
      ],
      "text": [
        450,
        220,
        300,
        135,
        [
          "原有正文",
          "Existing text",
          "元の本文"
        ],
        [
          "保留内容",
          "Preserve content",
          "内容を保持"
        ]
      ],
      "xml": [
        450,
        405,
        300,
        135,
        [
          "XML65 调用",
          "XML65 call",
          "XML65呼び出し"
        ],
        [
          "函数名 + 参数",
          "Function + arguments",
          "関数名 + 引数"
        ],
        "soft"
      ],
      "assistant": [
        850,
        220,
        245,
        135,
        [
          "assistant",
          "assistant",
          "assistant"
        ],
        [
          "合成正文",
          "One payload",
          "本文を合成"
        ],
        "result"
      ],
      "tool": [
        450,
        590,
        300,
        135,
        [
          "工具返回",
          "Tool result",
          "ツール応答"
        ],
        [
          "tool · 原文",
          "tool · unchanged",
          "tool・原文"
        ]
      ],
      "following": [
        850,
        590,
        245,
        135,
        [
          "后续上下文",
          "Next context",
          "次の文脈"
        ],
        [
          "供下一步读取",
          "For the next action",
          "次の行動で読む"
        ]
      ]
    },
    "edges": [
      [
        "reason",
        "think",
        [
          [
            335,
            102
          ],
          [
            450,
            102
          ]
        ],
        false
      ],
      [
        "body",
        "text",
        [
          [
            335,
            287
          ],
          [
            450,
            287
          ]
        ],
        false
      ],
      [
        "calls",
        "xml",
        [
          [
            335,
            472
          ],
          [
            450,
            472
          ]
        ],
        false
      ],
      [
        "think",
        "assistant",
        [
          [
            750,
            102
          ],
          [
            805,
            102
          ],
          [
            805,
            250
          ],
          [
            850,
            250
          ]
        ],
        false
      ],
      [
        "text",
        "assistant",
        [
          [
            750,
            287
          ],
          [
            850,
            287
          ]
        ],
        false
      ],
      [
        "xml",
        "assistant",
        [
          [
            750,
            472
          ],
          [
            805,
            472
          ],
          [
            805,
            325
          ],
          [
            850,
            325
          ]
        ],
        false
      ],
      [
        "tool",
        "following",
        [
          [
            750,
            657
          ],
          [
            850,
            657
          ]
        ],
        false
      ]
    ],
    "labels": [],
    "nodeDetails": {
      "xml": [
        "字符串参数使用 string=true，其余值用 JSON 表达。参数正文遵循 XML65 协议，不额外做 HTML 转义。",
        "String parameters use string=true; other values use JSON. Parameter text follows XML65 and is not additionally HTML-escaped.",
        "文字列にはstring=true、それ以外はJSON表現を使う。引数本文には追加のHTMLエスケープを行わない。"
      ],
      "assistant": [
        "这里画的是离线序列化，不是在导出时重新执行工具。真实工具返回来自已有轨迹。",
        "This is offline serialization, not tool execution during export. Tool results come from the recorded trace.",
        "これはオフラインの形式変換であり、出力時にツールを再実行しない。応答は記録済みの軌跡から得る。"
      ]
    }
  },
  {
    "id": "seed-parquet-supervision",
    "anchors": [
      "loss-anchor"
    ],
    "title": [
      "一条序列中的上下文与学习目标",
      "Context and targets in one sequence",
      "同じ系列の文脈と学習対象"
    ],
    "caption": [
      "上方显示各消息正文的监督标记；下方展开一条 assistant 消息。",
      "Top: supervision of message bodies. Below: the segments of one assistant message.",
      "上段は各メッセージ本文の学習対象の指定、下段はassistantメッセージの断片構成。"
    ],
    "width": 1140,
    "height": 510,
    "nodes": {
      "user": [
        40,
        35,
        235,
        115,
        [
          "用户请求",
          "User query",
          "依頼"
        ],
        [
          "user · loss 0",
          "user · loss 0",
          "user · loss 0"
        ]
      ],
      "action": [
        315,
        35,
        235,
        115,
        [
          "调用 Read",
          "Call Read",
          "Readを呼ぶ"
        ],
        [
          "assistant · loss 1",
          "assistant · loss 1",
          "assistant · loss 1"
        ],
        "soft"
      ],
      "result": [
        590,
        35,
        235,
        115,
        [
          "文件内容",
          "File content",
          "読取結果"
        ],
        [
          "tool · loss 0",
          "tool · loss 0",
          "tool · loss 0"
        ]
      ],
      "answer": [
        865,
        35,
        235,
        115,
        [
          "下一步",
          "Next action",
          "次の行動"
        ],
        [
          "assistant · loss 1",
          "assistant · loss 1",
          "assistant · loss 1"
        ],
        "soft"
      ],
      "cls": [
        40,
        330,
        235,
        120,
        [
          "[CLS]",
          "[CLS]",
          "[CLS]"
        ],
        [
          "起始 · loss 0",
          "Start · loss 0",
          "開始・loss 0"
        ]
      ],
      "role": [
        315,
        330,
        235,
        120,
        [
          "assistant\\n",
          "assistant\\n",
          "assistant\\n"
        ],
        [
          "角色 · loss 0",
          "Role · loss 0",
          "役割・loss 0"
        ]
      ],
      "payload": [
        590,
        330,
        235,
        120,
        [
          "正文",
          "Body",
          "本文"
        ],
        [
          "内容 · loss 1",
          "Content · loss 1",
          "内容・loss 1"
        ],
        "soft"
      ],
      "eos": [
        865,
        330,
        235,
        120,
        [
          "[EOS]",
          "[EOS]",
          "[EOS]"
        ],
        [
          "结束 · loss 1",
          "End · loss 1",
          "終了・loss 1"
        ],
        "soft"
      ]
    },
    "edges": [
      [
        "user",
        "action",
        [
          [
            275,
            92
          ],
          [
            315,
            92
          ]
        ],
        false
      ],
      [
        "action",
        "result",
        [
          [
            550,
            92
          ],
          [
            590,
            92
          ]
        ],
        false
      ],
      [
        "result",
        "answer",
        [
          [
            825,
            92
          ],
          [
            865,
            92
          ]
        ],
        false
      ],
      [
        "action",
        "cls",
        [
          [
            432,
            150
          ],
          [
            432,
            215
          ],
          [
            157,
            215
          ],
          [
            157,
            330
          ]
        ],
        true
      ],
      [
        "cls",
        "role",
        [
          [
            275,
            390
          ],
          [
            315,
            390
          ]
        ],
        false
      ],
      [
        "role",
        "payload",
        [
          [
            550,
            390
          ],
          [
            590,
            390
          ]
        ],
        false
      ],
      [
        "payload",
        "eos",
        [
          [
            825,
            390
          ],
          [
            865,
            390
          ]
        ],
        false
      ]
    ],
    "labels": [
      [
        210,
        234,
        720,
        [
          "展开 assistant 消息",
          "Expand one assistant message",
          "assistantのメッセージを展開"
        ],
        "center"
      ]
    ],
    "nodeDetails": {
      "result": [
        "loss 0 只表示不作为预测目标；后续 assistant 仍然读取真实文件内容。",
        "Loss 0 excludes prediction targets, not context: the next assistant still reads the actual file content.",
        "loss 0は予測対象から外す指定であり、次のassistantは実際のファイル内容を読む。"
      ],
      "payload": [
        "在这批导出中，全部 assistant 正文参与损失。是否按质量选择部分轮次，是另一种监督策略。",
        "All assistant bodies are supervised in this export. Selecting only some turns by quality would be a different supervision policy.",
        "この出力はassistant本文をすべて学習対象とする。品質に応じて一部のターンだけを選ぶなら、別の教師信号方針になる。"
      ]
    }
  },
  {
    "id": "seed-parquet-comparison",
    "anchors": [
      "comparison-anchor"
    ],
    "title": [
      "同一段轨迹，两种组织方式",
      "One trace, two training units",
      "同じ軌跡、異なる学習単位"
    ],
    "caption": [
      "示意：两条逐目标样本，对应一条包含两个目标的完整轨迹。系统前缀在图中省略。",
      "Schematic: two target-wise samples correspond to one full trace with two targets. System prefixes are omitted.",
      "模式図：対象別の2サンプルと、2対象を含む全軌跡の1サンプル。システム前置部は省略。"
    ],
    "width": 1140,
    "height": 730,
    "nodes": {
      "source": [
        300,
        30,
        540,
        100,
        [
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂"
        ],
        [
          "同一条原始记录",
          "The same source trace",
          "同じ元の記録"
        ]
      ],
      "ark": [
        45,
        235,
        480,
        110,
        [
          "方舟：逐目标展开",
          "Ark: target-wise rows",
          "Ark：対象ごとに展開"
        ],
        [
          "messages + tools",
          "messages + tools",
          "messages + tools"
        ]
      ],
      "seed": [
        615,
        235,
        480,
        110,
        [
          "完整轨迹 · Parquet",
          "Full trace · Parquet",
          "全軌跡 · Parquet"
        ],
        [
          "inputs + has_loss",
          "inputs + has_loss",
          "inputs + has_loss"
        ]
      ],
      "first": [
        45,
        430,
        480,
        110,
        [
          "Q → A₁",
          "Q → A₁",
          "Q → A₁"
        ],
        [
          "第 1 行 · 只监督 A₁",
          "Row 1 · supervise A₁ only",
          "1行目・A₁のみ学習"
        ],
        "soft"
      ],
      "second": [
        45,
        585,
        480,
        110,
        [
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂"
        ],
        [
          "第 2 行 · 只监督 A₂",
          "Row 2 · supervise A₂ only",
          "2行目・A₂のみ学習"
        ],
        "soft"
      ],
      "whole": [
        615,
        475,
        480,
        135,
        [
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂",
          "Q → A₁ → Tool → A₂"
        ],
        [
          "一行 · 监督 A₁ 和 A₂",
          "One row · supervise A₁ and A₂",
          "1行・A₁とA₂を学習"
        ],
        "result"
      ]
    },
    "edges": [
      [
        "source",
        "ark",
        [
          [
            420,
            130
          ],
          [
            420,
            185
          ],
          [
            285,
            185
          ],
          [
            285,
            235
          ]
        ],
        false
      ],
      [
        "source",
        "seed",
        [
          [
            720,
            130
          ],
          [
            720,
            185
          ],
          [
            855,
            185
          ],
          [
            855,
            235
          ]
        ],
        false
      ],
      [
        "ark",
        "first",
        [
          [
            285,
            345
          ],
          [
            285,
            430
          ]
        ],
        false
      ],
      [
        "ark",
        "second",
        [
          [
            525,
            290
          ],
          [
            565,
            290
          ],
          [
            565,
            640
          ],
          [
            525,
            640
          ]
        ],
        false
      ],
      [
        "seed",
        "whole",
        [
          [
            855,
            345
          ],
          [
            855,
            475
          ]
        ],
        false
      ]
    ],
    "labels": [],
    "nodeDetails": {
      "ark": [
        "本次导出每行只训练末尾的 assistant；此前 assistant 的 loss_weight 为 0，历史仍作为上下文。",
        "Each row in this export trains only the last assistant; earlier assistants have loss_weight 0 and remain context.",
        "この出力は各行の最後のassistantだけを学習し、過去のassistantをloss_weight 0の文脈として残す。"
      ],
      "seed": [
        "同一行内每个 assistant 正文与 EOS 都是目标，因果注意力仍阻止读取未来。",
        "Every assistant body and EOS in the row is a target; causal attention still prevents future access.",
        "同じ行のassistant本文とEOSを学習対象にするが、因果的注意で未来の参照は防がれる。"
      ]
    }
  }
]
