Data Generator

训练数据生成技能。根据传入的工具名和用户指令列表，生成多轮对话格式的 JSONL 训练数据。触发场景：(1) 传入工具名和用户指令列表，生成完整训练数据；(2) 批量生成指定工具的标注数据；(3) 给定指令列表，输出 JSONL 对话样本。

Security Scan

The skill is internally consistent: it generates JSONL training samples from provided tool names and instruction lists, uses only local reference files and included scripts, and does not request credentials or perform network access.

Purpose & Capability

Name/description (生成训练数据) match the included artifacts: prompt templates, per-tool rule files under references/tools/, and two generator scripts. All required files and behavior (creating JSONL from user instructions) are coherent with the stated purpose.

Instruction Scope

SKILL.md and scripts constrain behavior to assembling prompts and producing JSONL records. The generator reads local reference files and, if invoked with --file, will read a user-supplied file of instructions — this is expected for the task but means you should not point it at sensitive system files because their contents would be embedded into generated data.

Install Mechanism

No install spec; instruction-only skill with bundled Python scripts. No downloads, no external package installs, and no code executed from remote URLs.

Credentials

The skill requests no environment variables, credentials, or config paths. The scripts operate on provided inputs and bundled reference files only, which is proportionate to the stated functionality.

Persistence & Privilege

Skill is not always-enabled and uses normal model invocation semantics. It does write output files (the generated .jsonl) as expected, but does not modify other skills or request elevated/system-wide privileges.

Assessment

This skill appears to do only what it claims: generate JSONL training samples from tool-specific templates and user instruction lists. Before running: (1) inspect a sample output to ensure it doesn't accidentally encode any sensitive content you might include in the input file; (2) do not pass paths to sensitive system files to the --file option (the script will embed file contents into the generated data); (3) if you plan to include real user data or production logs, sanitize or anonymize them first. Otherwise it is safe to install from a provenance perspective (no network calls, no secrets requested).

将用户指令列表转换为标准 JSONL 训练数据。

输入

两个必填参数：

参数	类型	说明
tool_name	string	工具名，如 dev_control、scene_generator、weather
user_instructions	string[]	用户指令列表，如 ["5分钟后打开空调", "3分钟后关灯"]

输出 JSONL 格式

{"conversations":[
  {"from":"human","value":"<当前用户指令>打开客厅空调</当前用户指令>\n<本地设备>格力冷静王(空调)</本地设备>\n<当前时间>2026-03-15 14:22:47</当前时间>\n<用户场景列表>[{\"scene_id\":1001,\"scene_name\":\"回家模式\",\"room_name\":\"全屋\"},{\"scene_id\":1002,\"scene_name\":\"睡眠模式\",\"room_name\":\"主卧\"}]</用户场景列表>\n<用户设备列表>{\"客厅\":[\"格力冷静王(空调)\",\"洗碗机A1(洗碗机)\"],\"主卧\":[\"美的舒省风(空调)\"]}</用户设备列表>"},
  {"from":"assistant","value":"<tool_call>{\"tool_name\":\"dev_control\",\"query\":\"打开客厅空调\"}</tool_call>"},
  {"from":"observation","value":"<tool_response>客厅空调已打开</tool_response>"},
  {"from":"assistant","value":"好的，客厅空调已经打开啦~"}
],"system":"","history":[]}

格式规则

1. human value = 完整上下文，格式固定：

   <当前用户指令>用户原始指令</当前用户指令>
   <本地设备>设备名(类型)</本地设备>
   <当前时间>YYYY-MM-DD HH:mm:ss</当前时间>
   <用户场景列表>[{"scene_id":xxx,"scene_name":"场景名","room_name":"房间名"},...]</用户场景列表>
   <用户设备列表>{"房间":["设备名(类型)",...]}</用户设备列表>
   

2. assistant tool_call = 直接输出 tool_call 标签，无垫音前缀
3. observation = <tool_response>...</tool_response> 或 <tool_call>{...}</tool_call>（dev_info/weather 等工具）
4. assistant 终接回复 = 直接回复内容，无垫音前缀
5. system = ""，history = []

工作流

1. 接收 tool_name + user_instructions[]
         ↓
2. 加载提示词：通用要求 + 工具特定要求（references/tools/{tool}.txt）
         ↓
3. 将 user_instructions 注入提示词
         ↓
4. 生成 JSONL（每条独立）
         ↓
5. 输出 .jsonl 文件

提示词拼接

拼接规则：

[通用要求]
# ═══════════════════════════════════════════════════════════════════
# 【工具特定要求】
# 本次只调用：{TOOL_NAME}
# ────────────────────────────────────────────────────────────────
[references/tools/{TOOL_NAME}.txt 内容]

拼接脚本：scripts/build_prompt.py

工具与文件对照

工具	要求文件
dev_control	references/tools/dev_control.txt
scene_generator	references/tools/scene_generator.txt
alarm_remind	references/tools/alarm_remind.txt
weather	references/tools/weather.txt
scene_control	references/tools/scene_control.txt
dev_info	references/tools/dev_info.txt
exit_dialog	references/tools/exit_dialog.txt
GreeQA	references/tools/GreeQA.txt
scene_guide	references/tools/scene_guide.txt
chat	references/tools/chat.txt

使用示例

输入：

tool_name: "dev_info"
user_instructions: ["家里空调数量", "有几个空调"]

输出字段说明：

字段	说明
conversations[0].value	含 <当前用户指令> + 完整上下文
conversations[1].value	<tool_call>{"tool_name":"dev_info"}</tool_call>
conversations[2].value	dev_info 返回结果（设备列表）
conversations[3].value	文字终接回复
system	空字符串 ""
history	空数组 []

BUG 修复数据

当传入 tool_name 为修复后的正确工具时，生成的数据应体现：

• 工具调用格式正确（符合工具要求文件）
• query 字段格式正确（如延时类指令含 timing 字段）
• 文字回复符合预期（含延时时间描述）

具体格式参考：references/tools/scene_generator.txt。