This is a domain-adapted language model fine-tuned for Daoist (Taoist) knowledge, designed to be used as research infrastructure for digital humanities, religion studies, and Chinese cultural AI applications. We release the complete open-source stack — model + dataset + RAG API — so anyone can audit, deploy, fine-tune, or build upon this work.

🏛️ 關於釋出者 · About the Releaser

劉厝派鼎新門不是 AI 開發公司，而是一間道教教育機構。 我們的本業是道教知識的整理、傳承與教學——由台灣正一道道長創辦，長期累積經典考據、科儀紀錄、派別傳承等實務內容。

這次釋出 Daoism-Qwen3.5-9B，不是要做 AI 產品，而是把我們累積多年的道教知識資產，以最完整的形式交給有能力的開發者與研究者——模型權重、訓練資料、檢索 API 三件完整，Apache 2.0 開源、商用可、修改可、再發佈可。

「我們是道教教育機構，不是 AI 開發人員。但我們把最完整的能力與資料，交給開發人員。」

"Liu Cuo School Dingxin Lineage is not an AI startup—it is a Daoist educational institution. We are not AI developers. But we hand the most complete capability and data to those who are."

這條路線比自己做封閉聊天 demo 困難，但更誠實、更尊重道教文化的開放精神。把刀交到開發者手中——讓道教 AI 真正成為社群共有的基礎設施，而非單一機構的壟斷。

🎁 完整開源三件套

元件	URL	用途
🤖 Model (fp16)	`lius-cc/Daoism-Qwen3.5-9B`	完整精度模型，研究 / 二次微調
🤖 Model (GGUF)	`lius-cc/Daoism-Qwen3.5-9B-GGUF`	Q4_K_M / Q5_K_M 量化版，Ollama / llama.cpp / LM Studio
📚 Dataset (RAG)	`lius-cc/daoism-knowledge-rag`	95,919 條鼎稔道學館館藏語料
🔌 Public RAG API	`POST https://lius.cc/api/llm-rag`	即時檢索館藏，30 req/min 免費
📖 Deployment Guide	lius.cc/llm	從零部署完整教學

License: Apache 2.0（商用可）

Why this model?

Mainstream LLMs (ChatGPT, Claude, Gemini) consistently fail on Daoist queries:

混淆道佛、儒道
編造科儀流程（補運、補財庫、安太歲）
不分正一 / 全真 / 閭山等派別差異
不知台灣民間信仰系統（城隍、王爺、媽祖）
無法判讀奇門遁甲 / 紫微斗數命盤

Daoism-Qwen3.5-9B 以鼎稔道學館 57.3 萬條道教專業語料 QLoRA 微調自 Qwen/Qwen3.5-9B，針對性填補此空缺。

🚀 Quick Start

1. Transformers (Python, fp16)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "lius-cc/Daoism-Qwen3.5-9B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "你是道教智慧 AI，使用台灣繁體中文回答。"},
    {"role": "user", "content": "請說明正一道齋醮科儀中『金籙』『黃籙』『玉籙』三大齋的差異。"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

2. Ollama (GGUF, 一行)

bash

ollama pull hf.co/lius-cc/Daoism-Qwen3.5-9B-GGUF:Q4_K_M
ollama run hf.co/lius-cc/Daoism-Qwen3.5-9B-GGUF:Q4_K_M

3. LM Studio (no-code, GUI)

Install LM Studio
搜尋 lius-cc/Daoism-Qwen3.5-9B-GGUF
下載 Q4_K_M (5.3 GB) 或 Q5_K_M (6.1 GB)
載入 → 直接對話

4. llama.cpp (CLI)

./llama-cli -hf lius-cc/Daoism-Qwen3.5-9B-GGUF -p "你是道教智慧 AI..." -n 1024

5. vLLM (production serving, OpenAI 相容 API)

bash

pip install vllm
vllm serve lius-cc/Daoism-Qwen3.5-9B \
  --dtype bfloat16 \
  --max-model-len 4096 \
  --gpu-memory-utilization 0.85 \
  --port 8000

⚠️ vLLM 0.21+ 才支援 Qwen3.5 config。

6. llama-cpp-python (推薦給 single-GPU 用戶)

bash

CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python[server]

python -m llama_cpp.server \
  --model daoism-qwen3-5-9b.Q4_K_M.gguf \
  --n_gpu_layers -1 \
  --n_ctx 4096 \
  --host 0.0.0.0 --port 8000

🔍 RAG (推薦用法)

Daoism-Qwen3.5-9B 設計為與 RAG 配合使用以獲得最佳深度。 9B 規模在純 model 推理時有知識天花板，請務必搭配本 dataset 或 RAG API。

Option A: 用我們的公開 RAG API（最快）

python

import requests
from openai import OpenAI

client = OpenAI(api_key="dummy", base_url="http://localhost:8000/v1")

def ask(question):
    # 1. Retrieve from lius.cc
    r = requests.post(
        "https://lius.cc/api/llm-rag",
        json={"q": question, "n": 5},
        timeout=10,
    )
    hits = r.json().get("hits", [])
    context = "\n\n---\n\n".join(
        f"### {h['name']}\n{h['summary']}\n\n{h['content'][:1500]}\n來源：{h['url']}"
        for h in hits
    )

    # 2. Generate
    response = client.chat.completions.create(
        model="Daoism-Qwen3.5-9B",
        messages=[
            {"role": "system", "content": f"你是道教 AI。請以下列館藏資料為基礎深度回答（引用時附 URL）：\n\n{context}"},
            {"role": "user", "content": question},
        ],
        max_tokens=4096,
        temperature=0.6,
    )
    return response.choices[0].message.content

print(ask("城隍信仰與道教有什麼關係？"))

Option B: 離線自架 RAG（用我們的 dataset）

python

from datasets import load_dataset
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np

# Load 95,919 entries
ds = load_dataset("lius-cc/daoism-knowledge-rag", split="train")

# Embed
encoder = SentenceTransformer("BAAI/bge-m3")
embeds = encoder.encode(
    [f"{x['name']}：{x['summary']}" for x in ds],
    batch_size=64,
    show_progress_bar=True,
    normalize_embeddings=True,
)

# Index
index = faiss.IndexFlatIP(embeds.shape[1])
index.add(embeds.astype(np.float32))

def retrieve(query, k=5):
    q_emb = encoder.encode([query], normalize_embeddings=True)
    _, ids = index.search(q_emb.astype(np.float32), k)
    return [ds[int(i)] for i in ids[0]]

🔧 Fine-tuning (LoRA)

如果你想用自己資料繼續微調，建議走 QLoRA（與本 model 原訓練同方法）：

pip install axolotl unsloth

yaml

# axolotl-config.yml
base_model: lius-cc/Daoism-Qwen3.5-9B
load_in_4bit: true
adapter: qlora
lora_r: 64
lora_alpha: 128
sequence_len: 2048
sample_packing: true
micro_batch_size: 2
gradient_accumulation_steps: 16
num_epochs: 3
optimizer: paged_adamw_8bit
learning_rate: 2.0e-4
bf16: true
flash_attention: true
gradient_checkpointing: true
datasets:
  - path: your-dataset-path
    type: alpaca

accelerate launch -m axolotl.cli.train axolotl-config.yml

建議硬體：A100 40GB+ 或 RTX 4090 24GB。

📊 Training Details

項目	內容
Base model	Qwen/Qwen3.5-9B
Fine-tuning method	QLoRA 4-bit (r=64, α=128, dropout=0.05)
Training data	鼎稔道學館館藏 573k alpaca samples
Epochs	3 (33,993 steps)
Hardware	1× A100 80GB SXM4
Training time	~47 hours
Final loss	1.04-1.07 (stable convergence)
Compute provider	vast.ai (community spot GPU)

📐 Specifications


Architecture	Qwen3_5ForCausalLM
Parameters	9B
Context length	262,144 (training); recommended 4,096 for QLoRA inference
Vocab size	256,000
Languages	Traditional Chinese (Taiwan) primary; Simplified Chinese / English secondary
Tokenizer	GPT-2 style BPE (Qwen)
Supports	tool calling, thinking mode (`<think>`)

⚠️ Limitations & Responsible Use

9B 規模有知識天花板：複雜學術問題建議搭配 RAG（見上方範例）
量化版（Q4_K_M）犧牲約 5-10% 精度換 4× 速度，深度查詢請用 Q5_K_M 或 fp16
語言偏向：Traditional Chinese (Taiwan) > Simplified Chinese > English
派別覆蓋：正一道 / 閭山派 / 台灣民間信仰較完整；全真道 / 全真內丹較淺
不可取代真道長：涉及法事、運勢判斷、個人重大決策請洽合格道士
學術引用：引用文獻時請以 lius.cc 原條目為準，並交叉驗證

📜 Citation

bibtex

@misc{daoism-qwen3-9b-2026,
  author = {Liu, Chi-Ying and Dingren Daoxue Lab},
  title  = {Daoism-Qwen3.5-9B: An Open-Source Taoist Knowledge Language Model},
  year   = {2026},
  month  = {May},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/lius-cc/Daoism-Qwen3.5-9B},
  note   = {Apache 2.0 License; released with companion dataset and RAG API}
}

@dataset{daoism-rag-2026,
  author = {Liu, Chi-Ying and Dingren Daoxue Lab},
  title  = {Daoism Knowledge RAG: Curated Taoist Knowledge Base for Retrieval-Augmented Generation},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/datasets/lius-cc/daoism-knowledge-rag}
}

🙏 Acknowledgements

Base model: Alibaba Qwen team for Qwen3.5-9B (Apache 2.0)
Fine-tuning framework: Unsloth
GPU sponsorship: vast.ai community marketplace
Inspiration: All open-source efforts to digitize and preserve Daoist cultural heritage

🌐 Contact / Issues

Web: lius.cc · lius.cc/llm
Issues & discussions: HF Discussions
Email: chiyingliu@gmail.com
Maintained by: 劉啟穎 (Liu Chi-Ying), Dingren Daoxue Lab founder

「真正的開源精神，是把完整的能力交到開發者手中。」