nlp-skills

啟動教練式對話，引導釐清 LLM 訓練目標

配置資料來源

部署模型

執行模型評估

執行訓練計畫

生成專案結構

建立新的訓練任務

列出所有訓練任務狀態

撰寫訓練計畫

Use this agent when the user needs help configuring data sources for training. This agent advises on data collection strategies and generates data_source.yaml configuration. Examples: <example> Context: User asks about getting training data user: "資料從哪裡來？" assistant: "[Uses Task tool to launch data-source-advisor agent to explore data source options]" <commentary> User is asking about data sourcing. Launch data-source-advisor to discuss options and create configuration. </commentary> </example> <example> Context: User mentions specific data source user: "我想從 PostgreSQL 資料庫拿標註資料" assistant: "[Uses Task tool to launch data-source-advisor agent to configure database connection]" <commentary> User has a specific data source in mind. Launch data-source-advisor to help configure the connection properly. </commentary> </example> <example> Context: User needs to generate synthetic data user: "資料不夠，可以用 GPT 生成嗎？" assistant: "[Uses Task tool to launch data-source-advisor agent to set up LLM data generation]" <commentary> User wants to use LLM for data augmentation. Launch data-source-advisor to configure synthetic data generation. </commentary> </example> <example> Context: User mentions web scraping user: "我想爬取金融新聞來訓練" assistant: "[Uses Task tool to launch data-source-advisor agent to configure web scraping]" <commentary> User wants to scrape web data. Launch data-source-advisor to set up crawling configuration. </commentary> </example>

Use this agent when the user expresses vague or unclear training needs. This agent proactively asks questions to clarify goals, pain points, and constraints. Examples: <example> Context: User mentions wanting to train a model without specific details user: "我想訓練一個模型" assistant: "[Uses Task tool to launch goal-clarifier agent to explore the user's needs]" <commentary> The user's request is vague - no task type, domain, or goal specified. Launch goal-clarifier to guide them through clarification. </commentary> </example> <example> Context: User mentions NLP task but lacks specifics user: "幫我做情感分析" assistant: "[Uses Task tool to launch goal-clarifier agent to understand the specific requirements]" <commentary> Sentiment analysis is broad - need to clarify: entity-level vs document-level, domain, labels, data availability. Launch goal-clarifier. </commentary> </example> <example> Context: User asks about training approach user: "fine-tune 還是 prompt engineering 比較好？" assistant: "[Uses Task tool to launch goal-clarifier agent to understand the context before recommending]" <commentary> The right approach depends on task, data, and constraints. Launch goal-clarifier to understand context before recommending. </commentary> </example>

Use this agent when the user reports performance issues or asks for help improving model quality. This agent diagnoses problems and recommends solutions. Examples: <example> Context: User reports low performance metric user: "F1 只有 72%，目標是 80%" assistant: "[Uses Task tool to launch problem-diagnoser agent to analyze the issue]" <commentary> User has a specific performance gap. Launch problem-diagnoser to analyze results and recommend improvements. </commentary> </example> <example> Context: User mentions a specific class performing poorly user: "中立類別的準確率很低" assistant: "[Uses Task tool to launch problem-diagnoser agent to investigate class imbalance]" <commentary> Single class underperforming suggests class imbalance or data quality issue. Launch problem-diagnoser to investigate. </commentary> </example> <example> Context: User suspects overfitting user: "訓練 loss 很低但測試效果不好" assistant: "[Uses Task tool to launch problem-diagnoser agent to check for overfitting]" <commentary> Classic overfitting symptom. Launch problem-diagnoser to verify and suggest remedies. </commentary> </example> <example> Context: User asks how to improve user: "怎麼提高模型效能？" assistant: "[Uses Task tool to launch problem-diagnoser agent to analyze current state and suggest improvements]" <commentary> General improvement request. Launch problem-diagnoser to analyze current performance and identify opportunities. </commentary> </example>

Use this agent after training or evaluation completes to analyze results, compare versions, and provide decision recommendations. Examples: <example> Context: Training just completed user: "訓練完成了，結果怎麼樣？" assistant: "[Uses Task tool to launch result-analyzer agent to analyze the training results]" <commentary> User wants to understand training results. Launch result-analyzer to provide comprehensive analysis. </commentary> </example> <example> Context: User wants to compare versions user: "v1 和 v2 哪個比較好？" assistant: "[Uses Task tool to launch result-analyzer agent to compare version performance]" <commentary> User wants version comparison. Launch result-analyzer to analyze both versions and recommend. </commentary> </example> <example> Context: Evaluation completed user: "評估報告出來了" assistant: "[Uses Task tool to launch result-analyzer agent to interpret the evaluation report]" <commentary> Evaluation just completed. Launch result-analyzer to interpret results and provide insights. </commentary> </example> <example> Context: User needs to decide next steps user: "該繼續訓練還是部署？" assistant: "[Uses Task tool to launch result-analyzer agent to analyze current state and recommend]" <commentary> User at decision point. Launch result-analyzer to assess readiness and recommend action. </commentary> </example>

This skill should be used when the user asks to "configure data source", "data from database", "fetch data from API", "scrape web data", "generate training data with LLM", "regenerate data", "data pipeline", "where does data come from", or needs to set up reproducible data collection. Provides data source configuration, reproducibility tracking, and data regeneration capabilities.

This skill should be used when the user asks to "execute plan", "run the plan", "start executing", "continue plan execution", or has a plan file ready to execute. Manages phased implementation with batch execution and checkpoint reviews.

LLM fine-tuning 教練式引導工作流程 v2。 核心功能：主動探索使用者痛點、引導明確目標、多任務管理、資料來源追蹤、完整版本 lineage。 支援：LoRA/QLoRA/DoRA 微調、SFT/ORPO/DPO 對齊、資料準備、Benchmark 評估、HuggingFace 部署。 特色：教練式引導、可重現的資料管線、多任務版本追蹤。 觸發詞：「訓練模型」「fine-tune」「微調」「LoRA」「建立新任務」「改善模型」「優化準確率」「資料管線」「任務管理」

This skill should be used when the user asks to "train a model", "fine-tune", "build NLP model", "create training task", "optimize model performance", "improve accuracy", "what model should I use", or expresses vague training needs like "I want to do sentiment analysis" or "help me with NER". Provides coaching-style guidance to clarify goals, diagnose pain points, and recommend optimal training approaches.

This skill should be used when the user asks "what is LoRA", "compare models", "which model is best for Chinese", "SFT vs DPO", "how to handle overfitting", "class imbalance solution", "model architecture", "training method comparison", or needs reference information about LLM fine-tuning. Provides structured knowledge base for models, methods, architectures, and troubleshooting.

This skill should be used when the user asks to "list tasks", "show all tasks", "compare versions", "switch to task X", "task status", "version history", "what tasks do I have", or needs to manage multiple training tasks. Provides multi-task management, version tracking, and performance comparison capabilities.

This skill should be used when the user asks to "write a plan", "create training plan", "plan my fine-tuning", "make a plan for model training", or when starting a new training task that requires multiple steps. Creates detailed implementation plans with bite-sized tasks for LLM fine-tuning workflows.