Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 783113944 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Agent-STITCH-S 中文版 (agent-sft-stitch-zh)

voidful/agent-sft 轉換而成的台灣繁體中文 Agent-STITCH-S 語音助理合成資料:模型一邊對使用者說話(<SAY>)、一邊私下推理([SOPR])、一邊呼叫工具(<TOOL_CALL>)的 speak-while-acting 軌跡。

產製流程

  1. agent-sft(309,322 筆)篩出具完整 tool-call 鏈(user → tool_call → tool_result → final answer)的對話;多輪對話的早前輪次保留為 context 供改寫模型 grounding。
  2. google/gemma-4-26B-A4B-it 把每筆改寫成台灣繁體中文的 speech-first STITCH-S 軌跡:先安全開場 → [SOPR] 推理 → <TOOL_CALL> → 等待語音(不得洩漏 pending 結果)→ <TOOL_RESULT> → 逐步整合 → 最終口語答覆 [EOR]。標記與真實工具內容由程式確定性組裝,工具呼叫與結果不經模型改寫。
  3. 生成端品質閘門:tool_steps 與當前請求無關/幻覺者丟棄;private state 出現改寫者後設語言(rewriter meta-language)者由確定性 filter 丟棄;步數不對齊、JSON 不可解析者丟棄。
  4. 9 維 / 18 分 STITCH-S rubric 用同一模型嚴格評分(speech-first、tool 等待安全性、時序因果、逐步整合、標記正確、靜默間隔、grounding、客服口語品質、tool 有效性),含 hard-fail 條件。
  5. 只保留 total >= 17 且無 hard-failkeep 樣本。

欄位

  • id, source:原始 agent-sft 來源
  • user:中文使用者語句
  • msg:STITCH-S 軌跡(含 [SOPR]/[EOPR]/[EOR]<TOOL_CALL><TOOL_RESULT>)
  • input:評分用 metadata(context、available_tools、tool_steps、reference_answer、原始使用者語句)
  • total_score(0–18)與 8 個攤平分項分數、evaluation_json(完整 9 維評分與評語;tool_validity 分項在此 JSON 內)

統計

  • 評分樣本:76,450,保留(keep):63,941;review 3,453、reject 9,056(其中 6,300 為超長樣本、約 2,756 為品質不合格)
  • 保留資料平均分:17.909 / 18
  • 各來源保留數:
    • nemotron_terminal: 24147
    • toolmind: 17987
    • glaive: 6119
    • hermes_reasoning: 3961
    • apigen_mt: 3934
    • at_deepseek_v4_pro: 3017
    • at_synthtraces: 1580
    • hermes: 1576
    • swe_smith: 766
    • toolace: 345
    • swe_zero: 277
    • swe_gym: 156
    • at_ml_intern: 33
    • at_qwen37_max_pi: 25
    • at_pi_excel: 11
    • at_minimax_m3_cc: 7
Downloads last month
121

Models trained or fine-tuned on voidful/agent-sft-stitch-zh