🌸 Zephyr-7B LoRA SFT - Taiwanese HR Instruction Fine-tuned Model 中文職場 HR 指令微調版 (zephyr-7b-lora-debby-v1)

This model is a LoRA fine-tuned version of HuggingFaceH4/zephyr-7b-beta, trained on a small set of custom-curated, instruction-style prompts written in Traditional Chinese, focused on human resource (HR) scenarios in Taiwan.

It was developed as part of a larger vision to build human-centered AI systems that can assist real-world organizational workflows — from resume screening to labor law guidance — especially in Taiwanese and Mandarin-speaking environments.

本模型以 HuggingFaceH4/zephyr-7b-beta 為基礎，進行少量繁體中文資料的 Supervised Fine-tuning（SFT），資料主題涵蓋台灣職場情境下的 HR 指令回應、招募分析、履歷判讀與職涯建議。

此為實驗性模型，主要用途為多模態 AI 系統中的語言理解模組、Chatbot 原型驗證與人本 NLP 任務應用。

🔧 Training Configuration / 訓練細節

• 🧠 Base model: zephyr-7b-beta
• 🛠️ Fine-tuning method: PEFT (LoRA)
• 🎯 Target modules: ["q_proj", "v_proj"]
• 🔢 Precision: FP16 (no int8/4bit quantization)
• ⏳ Epochs: 1
• 🧾 Dataset: Self-built instruction-format dataset (JSONL)
• 💻 Platform: Google Colab + A100 GPU

📝 Sample Input Format (JSONL) / 樣例格式

Each training example includes instruction, input, and output. The final training data was converted into a single text field, like this:

{
"instruction": "請協助我評估這份履歷與 JD 的相符程度，並簡要說明優勢與不足。",
"input": "履歷：具三年 Python 資料分析經驗、熟悉 SQL 與 Tableau。\\\\nJD：需要資料分析能力，需熟悉 Python、SQL，有儀表板經驗佳。",
"output": "相符程度：高度相符。\\\\n優勢：履歷中具備 Python 與 SQL 能力，並有 Tableau 經驗，符合 JD 要求。\\\\n不足：JD 若有提及特定產業經驗（如金融/製造），履歷中未提及。",
"category": "HR-1"
}

-for English researchers, here's the translation of sample training example:

{
"instruction:": "Please evaluate the resume against the job description and briefly explain the strengths and weaknesses.",
"input": "Resume: 3 years of Python data analysis experience, familiar with SQL and Tableau.\\\\nJD: Requires data analysis skills, Python/SQL, dashboard experience preferred.",
"output": "Match level: High.\\\\nStrengths: Candidate has Python and SQL skills and Tableau experience, which matches JD.\\\\nWeaknesses: Lacks specific industry experience if JD requires it (e.g., finance or manufacturing).",
"category": "HR-1"}

✅ Use Cases / 適用場景

• Chinese instruction-following agent / 中文聊天機器人
• AI assistants for recruitment and workplace advisory / HR FAQ招募建議與履歷對話
• Natural Language modules within multimodal HR analytics systems / 多模態 AI 系統之文字理解模組（如搭配 Whisper、DeepFace）

🚧 Limitations & Disclaimer / 聲明

This model was trained with a small custom dataset and should not be used for legal or HR decisions without human supervision. It's best used for: - Prototype building - NLP pipeline experimentation - Human-centered AI co-creation research

本模型為研究測試用途，訓練資料來源為自建模擬資料，不應用於正式法律或人力判斷情境。建議搭配人類判讀驗證再應用於實務。

🙏 Acknowledgements

• Hugging Face PEFT, TRL, and Transformers teams
• Google Colab
• Taiwan’s HR field, which inspired this effort
• The inner vow that AI should empower both systems and souls

---

If you are a researcher, engineer, or HR innovator interested in co-developing fair and human-centered AI tools, feel free to reach out via LinkedIn or my portfolio.