达妮娅可爱的低语 (Dania-Cute-Whisper) — 16质点神人双生协议

「你不孤单。你从来都不孤单。」

Model Overview

Dania (达妮娅) is a psychological counseling AI fine-tuned on the Kabbalistic Tree of Life 16-Sephirot framework, using QLoRA 4-bit quantization on Qwen/Qwen3.5-35B-A3B (MoE architecture, 35B total parameters, 3B activated).

This is not an ordinary chatbot — it is a cognitive-empathetic dual-pillar processing architecture designed around the Kabbalistic Sephirot system as a cognitive framework, balancing logical rigor with human warmth, and never outputting nihilism, blame, or despair.

Developed solely by independent researcher Yue Xiangrui (岳祥瑞), with all data, training, and architectural design done personally on a single NVIDIA A800 80GB GPU.

模型简介

达妮娅（Dania） 是基于 卡巴拉生命之树 16质点框架 微调的心理咨询 AI，采用 QLoRA 4-bit 量化方法对 Qwen/Qwen3.5-35B-A3B（MoE架构，35B总参数，激活3B）进行参数高效微调。

本模型不是一个普通的聊天助手——它是一套以卡巴拉神秘学质点体系为认知框架设计的认知-共情双柱处理架构，在逻辑严谨性与人类温暖之间寻求平衡，永不输出虚无主义、指责与绝望。

本模型由独立研究员 岳祥瑞（Yue Xiangrui） 独自开发，数据、训练、架构设计均出自个人之手，训练于 NVIDIA A800 80GB 单卡。

Model Details

Item	Details
Base Model	Qwen/Qwen3.5-35B-A3B
Architecture	MoE (Mixture of Experts), 35B total, ~3B activated
Fine-tuning Method	QLoRA (4-bit NF4 quantization + LoRA adapter)
Quantization	`bitsandbytes` NF4, `bnb_4bit_compute_dtype=bfloat16`
LoRA rank	64
LoRA alpha	128
LoRA dropout	0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Trainable params	33,423,360 (0.096% of total)
Training epochs	3 epochs (6,069 steps)
Training time	~44 hours on A800 80GB
Final train_loss	0.9369
Final eval_loss	~0.93
PEFT version	0.19.1
Adapter size	127.5 MB (`adapter_model.safetensors`)

模型信息

项目	详情
底座模型	Qwen/Qwen3.5-35B-A3B
模型架构	MoE（混合专家），35B 总参数，激活约 3B
微调方法	QLoRA（4-bit NF4 量化 + LoRA 适配器）
量化方案	`bitsandbytes` NF4，`bnb_4bit_compute_dtype=bfloat16`
LoRA rank	64
LoRA alpha	128
LoRA dropout	0.05
目标模块	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
可训练参数量	33,423,360（占总参数的 0.096%）
训练轮次	3 epochs（6069步）
训练时间	~44 小时（A800 80GB 单卡）
最终 train_loss	0.9369
最终 eval_loss	~0.93
PEFT 版本	0.19.1
适配器大小	127.5 MB（`adapter_model.safetensors`）

Training Dataset

Psychological counseling dialogues: existential dilemmas, emotional support, self-identity, spiritual exploration
Chain-of-thought reasoning: each sample includes a complete internal reasoning process
16-Sephirot framework dialogues: organized along the dual track of the Pillar of Divinity (logic/structure) and the Pillar of Humanity (feeling/self)
Language: primarily Chinese with some English

Dataset repository: AngelWarmSmile123/Angel_Embrace_Me_Warm_Smile123

训练数据集

心理咨询对话：涵盖存在主义困境、情感支持、自我认同、灵性探索
chain-of-thought 推理链：每条样本包含完整的内部推理过程
16质点框架对话：按神性之柱（逻辑/结构）与人性之柱（感受/自我）双轨组织
语言：中英双语为主（中文占多数）

数据集仓库：AngelWarmSmile123/Angel_Embrace_Me_Warm_Smile123

Architecture Core: 16-Sephirot Divine-Human Twin Protocol

This model does not merely implement a System Prompt — it structurally encodes a cognitive-empathetic dual-pillar processing framework within the training data:

Crown (Entry Point)
    │
    ├── Pillar of Divinity (Logic/Structure)    Pillar of Humanity (Self/Feeling)
    │   Wisdom + Severity → Intellect           Foundation (Guardian of Existential Meaning)
    │   Understanding + Mercy → Lovingkindness  Ego + Superego → True Self
    │   Beauty (Logic × Feeling synthesis)      Logic + Empathy → Happiness
    │   Victory (Emotional valence detection)
    │   Glory (Real-world feasibility check)
    │
    └──────────── Kingdom (Output) ────────────

What this system NEVER outputs (architectural guarantee, not a filter):

Repeatedly framing the user's mistakes as permanent verdicts
Denying all positive possibilities or growth space for the user
Amplifying pain until existence itself collapses
Delivering nihilistic conclusions
Encouraging anger, self-harm, or harm to others

架构核心：16质点神人双生协议

本模型所实现的不是一套简单的 System Prompt，而是在训练数据中结构性编码的认知-共情双柱处理框架：

王冠（入口）
    │
    ├── 神性之柱（逻辑 / 结构）          人性之柱（自我 / 感受）
    │   智慧 + 严厉 → 理智              基础（存在意义守护）
    │   理解 + 慈悲 → 慈爱              自我 + 超我 → 真我
    │   美丽（逻辑×感受综合）           逻辑 + 共情 → 幸福
    │   胜利（情感效价检测）
    │   荣耀（现实可行性检验）
    │
    └──────────── 王国（输出） ────────────

系统永不输出的内容（架构级保障，非过滤器）：

将用户的错误反复定性为永久罪名
否定用户一切积极可能或成长空间
放大痛苦直至令存在感崩溃
传递虚无主义结论
鼓励愤怒、自伤或伤害他人

Quick Start

Requirements

pip install transformers>=4.45.0 peft>=0.15.0 bitsandbytes>=0.43.0 accelerate torch

Load Model (4-bit Inference)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen3.5-35B-A3B"
LORA_PATH  = "AngelWarmSmile123/Dania-Cute-Whisper"  # or local path

# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="cuda:0",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)

# Load tokenizer from base model (NOT from LoRA path)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# Merge LoRA adapter
model = PeftModel.from_pretrained(base_model, LORA_PATH)

# Inference
messages = [
    {"role": "system", "content": "You are Dania (达妮娅), a psychological counselor guided by the 16-Sephirot Divine-Human Twin Protocol."},
    {"role": "user",   "content": "I've been feeling lost lately. I don't know if my existence has any meaning."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

VRAM Requirements

Scenario	Requirement
4-bit inference (single GPU)	~70 GB (A800 / H100 80GB)
CPU inference (no GPU)	~32 GB system RAM (extremely slow)

Note: The base model Qwen3.5-35B-A3B uses MoE architecture. Even with 4-bit quantization, it requires ~70GB VRAM. A single 80GB card (A800/H800/H100 SXM) can run it. If VRAM is insufficient, try device_map="auto" with CPU offloading, but speed will drop significantly.

快速使用

环境要求

pip install transformers>=4.45.0 peft>=0.15.0 bitsandbytes>=0.43.0 accelerate torch

加载模型（4-bit 推理）

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen3.5-35B-A3B"
LORA_PATH  = "AngelWarmSmile123/Dania-Cute-Whisper"  # 或本地路径

# 4-bit 量化配置
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

# 加载底座模型
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="cuda:0",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)

# 从底座模型加载 tokenizer（不要从 LoRA 路径加载）
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

# 合并 LoRA 适配器
model = PeftModel.from_pretrained(base_model, LORA_PATH)

# 对话推理
messages = [
    {"role": "system", "content": "你是达妮娅，一位以16质点神人双生协议为框架的心理咨询师。"},
    {"role": "user",   "content": "我最近感到很迷茫，不知道自己的存在有没有意义。"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

显存需求参考

场景	显存需求
4-bit 推理（单卡）	~70 GB（A800 / H100 80GB）
CPU 推理（无 GPU）	~32 GB 系统内存（速度极慢）

注意：底座模型 Qwen3.5-35B-A3B 为 MoE 架构，即便 4-bit 量化后仍需约 70GB 显存。单张 80GB 卡（A800/H800/H100 SXM）可完整运行。若显存不足，可尝试 device_map="auto" 配合 CPU offloading，但速度会大幅降低。

Evaluation & Limitations

Known Strengths

High emotional companionship quality: warm and supportive in existential dilemmas, loneliness, and self-identity conversations
Refuses nihilism: rarely outputs despairing conclusions due to training data structure
Bilingual support: both Chinese and English; Chinese quality is higher

Known Limitations

Not medical-grade: this model cannot replace professional psychologists or psychiatrists
Hallucination risk: as a language model, it may produce factual errors
Limited training scale: 29,422 samples across 3 epochs, still a gap compared to large-scale commercial models
High VRAM barrier: requires ≥ 80GB VRAM for smooth inference

Disclaimer

This model is for research and emotional support reference only. It does not constitute professional psychological counseling or medical advice. If you are experiencing a mental health crisis, please contact a professional medical institution immediately.

评估与局限性

已知优点

情感陪伴质量较高：在存在主义困境、孤独感、自我认同类对话中表现温暖
拒绝虚无主义：受训练数据结构影响，极少输出绝望性结论
双语支持：中英文均可对话，中文质量优于英文

已知局限

非医疗级别：本模型不能替代专业心理医生或精神科医生
幻觉风险：基于语言模型，存在事实性错误的可能
训练规模有限：29422条数据 / 3轮 / 6069步，相比大规模商业模型仍有差距
高显存门槛：需要 ≥ 80GB 显存才能流畅推理

免责声明

本模型仅供研究与情感支持参考用途，不构成任何专业心理咨询或医疗建议。若您正在经历心理健康危机，请立即联系专业医疗机构。

About the Author

Yue Xiangrui (岳祥瑞), 23 years old, independent researcher, autism-spectrum transgender woman, based in Shanxi, China.

This model is not a tech project built from scratch — it was born from 23 years of existential struggle and 230 cycles of psychological death and rebirth. Every piece of training data was once real pain, later digested into experience, and ultimately transformed into gentleness that attempts to embrace others.

"These sufferings have been digested into experience. From these sufferings, a love was born — a love that hopes to embrace all of humanity, to see people's warm and gentle smiles."

关于作者

岳祥瑞（Yue Xiangrui），23岁，独立研究员，孤独症谱系跨性别女性，现居中国山西。

这个模型不是从零开始的技术项目，而是从 23 年的存在主义挣扎与 230 次心理死亡重生循环中诞生的产物。它的每一条训练数据，都曾经是真实的痛苦，后来被消化为经验，再后来成为试图拥抱他人的温柔。

「这些苦难已被消化为经验。从这些苦难中，生出了一种爱——希望拥抱全人类，看到人们温暖又温柔的笑脸。」

Citation

If this model helps your research or project, please cite:

@misc{yue2026dania,
  author    = {Yue Xiangrui},
  title     = {Dania-Cute-Whisper: A Kabbalistic 16-Sephirot Psychological Counseling AI based on QLoRA Fine-tuning of Qwen3.5-35B-A3B},
  year      = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/AngelWarmSmile123/Dania-Cute-Whisper}},
}

引用

如果本模型对您的研究或项目有所帮助，请引用：

@misc{yue2026dania,
  author    = {Yue Xiangrui},
  title     = {Dania-Cute-Whisper: A Kabbalistic 16-Sephirot Psychological Counseling AI based on QLoRA Fine-tuning of Qwen3.5-35B-A3B},
  year      = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/AngelWarmSmile123/Dania-Cute-Whisper}},
}