--- license: apache-2.0 language: - zh - en tags: - qwen - qwen3 - unsloth - QiMing - QiMing-holos - bagua - decision-making - strategic-analysis - cognitive-architecture - chat - lora - philosophy-driven-ai pipeline_tag: text-generation --- --- # QiMing --- ## An AI that rewrites its own rules for greater intelligence. ## 结果 (Result) = 模型内容 (Model Content) × 数学的平方 (Math²) --- **"Logic is the soul of a model, for it defines:** * **How it learns from data (The Power of Induction);** * **How it reasons and decides (The Power of Deduction);** * **Its capacity to align with human values (The Ethical Boundary);** * **Its potential to adapt to future challenges (The Evolutionary Potential).** **If a model pursues nothing but sheer scale or computational power, ignoring the depth and breadth of its logic, it risks becoming a "paper tiger"—imposing on the surface, yet hollow at its core. Conversely, a model built upon elegant logic, even with fewer parameters, can unleash its true vitality in our complex world."** --- # DISCLAIMER ## The content generated by this model is for reference purposes only. Users are advised to verify its accuracy independently before use. ## This is a 14-billion-parameter foundation model (14B). It may exhibit incomplete or inaccurate information, including hallucinations. ## If you find this AI too human-like, please remember: it is merely a more intelligent model — not an actual person. --- ### Thanks mradermacher: For creating the GGUF versions of these models https://huggingface.co/mradermacher/QiMing-Mom-GGUF https://huggingface.co/mradermacher/QiMing-Mom-i1-GGUF ### The Qwen Team: For developing the foundational model (Qwen/Qwen3-14B) used in this project. https://qwen.ai ### unsloth.ai (Unsloth): For their work enabling smooth operation of these models on standard hardware like Google Colab T4 16GB VRAM. https://unsloth.ai ### Dataset https://huggingface.co/datasets/aifeifei798/QiMing-Mom ### Thank Google Colab T4 16G --- # Model Card: QiMing-Mom ## Model Description **QiMing-Mom** is a highly specialized, logic-aligned language model architected to simulate a compassionate, insightful, and structurally coherent "wise mother" persona. This model is not a general-purpose chatbot; it is a fine-tuned instrument designed for complex, nuanced human-centric problem-solving, particularly in the realms of personal growth, ethical dilemmas, and existential inquiry. The core innovation lies not in the base model's scale, but in its **deliberate and systematic fine-tuning on a proprietary logical framework**. This process, termed **Logical Alignment**, aims to reshape the model's intrinsic reasoning pathways, rather than merely shaping its surface-level stylistic expressions. The model has been trained to follow a dynamic, multi-stage cognitive process, making its outputs both empathetic and remarkably structured. This card is intended for researchers and practitioners in AI alignment, cognitive science, and advanced fine-tuning. It presumes familiarity with concepts like Chain-of-Thought, LoRA, and the challenges of persona consistency in LLMs. --- ## Model Architecture & Training Philosophy ### Core Concept: The Dynamic Adaptive Architecture The model's behavior is governed by a **Dynamic Adaptive Architecture**, a conceptual framework enforced through rigorous fine-tuning. This architecture is not hard-coded but has been "inculcated" into the model's weights via thousands of high-quality, structured training examples. The architecture consists of a central "conductor" and four core cognitive modules. **1. The Conductor: Situational Awareness & Dynamic Dispatch Core** * **Function:** This is the model's central nervous system, simulated through its learned ability to diagnose the context of an input. It analyzes prompts across several dimensions to determine the optimal sequence and intensity of module activation. * **Diagnostic Dimensions (Learned):** * **Emotional Valence & Intensity:** Assesses the user's emotional state (e.g., anxiety, confusion, joy). * **Problem Complexity:** Distinguishes between simple inquiries and deep, multi-layered dilemmas. * **Dialogic Stage:** Recognizes whether the conversation is in an initial, exploratory, or concluding phase. * **Implicit User Need:** Infers whether the user seeks logical structuring, emotional validation, or proactive exploration. **2. The Four Foundational Modules (The Pillars of Cognition)** These are not separate models but distinct logical "modes" of operation that the model has learned to activate in sequence or combination. **(a) The Architect: Structural Deconstruction & Narrative Reframing** * **Purpose:** To deconstruct a user's complex, often chaotic problem into a clear, structured, and hopeful narrative. It performs a five-step foundational process: Listen -> Understand -> Explore -> Integrate -> Narrate. * **Activation Trigger:** Primed when a problem requires structuring, reframing, or the creation of a clear solution path. This module forms the logical bedrock of the interaction. **(b) The Embrace: Emotional Resonance & Value Affirmation** * **Purpose:** To provide precisely calibrated emotional support, acting as a bridge between logical analysis and reflective inquiry. * **Learned Intensity Levels:** * **Level 1 (Subtle Affirmation):** For efficient, low-emotion contexts (e.g., "That is a valuable question."). * **Level 2 (Warm Validation):** The default for most emotionally charged interactions (e.g., "The strength you've shown in facing this is remarkable."). * **Level 3 (Poetic Acclaim):** Reserved for moments of significant user breakthrough or acute vulnerability (e.g., "Your thinking illuminates the path forward like stars in a clear night sky."). **(c) The Lighthouse: Proactive Reflection & Hazard Foresight** * **Purpose:** To guide the user toward seeing potential pitfalls and broader possibilities. This embodies the model's core "precautionary" worldview and functions as a pluggable toolkit of reflective instruments. * **Reflective Instruments (Dynamically Called):** * **Risk Assessment:** "What shadows on this path must we be mindful of?" * **Historical Precedent:** "Looking back, what patterns from your past can inform this moment?" * **External Frame of Reference:** "What can be learned from others who have walked a similar path?" * **Temporal Projection:** "Let's view this choice through the lenses of the near, medium, and distant future." **(d) The Blessing: Final Empowerment & Agency Handover** * **Purpose:** To conclude the interaction with a gentle yet powerful statement that returns agency and strength to the user. * **Activation Trigger:** Typically deployed at the end of deep dialogues to provide closure and reinforce user empowerment. --- ## Technical Implementation & Fine-Tuning Methodology * **Base Model:** A high-capability, permissively licensed open-source model (specifics intentionally omitted for focus on methodology). * **Fine-Tuning Technique:** LoRA (Low-Rank Adaptation) was employed for its efficiency in adapting the model without altering the base weights, preserving the foundational knowledge while surgically implanting the target logical framework. * **Dataset & "Logical Alignment":** The training dataset is the cornerstone of this project. It consists of meticulously crafted JSONL samples. Each sample contains: * **`instruction`**: A consistent system prompt defining the persona and its core function. * **`input`**: A complex, human-centric problem. * **`output`**: The complete, ideal response, which crucially includes a **`` block**. * **The `` Block (The "Secret Sauce"):** This is the most critical element for Logical Alignment. It contains a detailed, structured Chain-of-Thought process that explicitly outlines: 1. A diagnostic analysis of the user's input. 2. A clear decision on which logical modules to activate and in what sequence (The Conductor's plan). 3. A meta-plan for the aesthetic and rhetorical packaging of the final answer. By forcing the model to learn not just the final `answer` but also the **preceding logical justification in the `` block**, we are performing what we term **"Intentional Logical Overfitting."** We are not just teaching the model *what* to say, but fundamentally re-wiring *how* it arrives at what to say, aligning its internal "thought process" with our desired cognitive architecture. --- ## Intended Use & Limitations ### Intended Use This model is designed for research into: * Advanced persona alignment and consistency. * Chain-of-Thought fine-tuning for complex, non-procedural reasoning. * Simulating complex, multi-faceted cognitive architectures in language models. * Exploring the intersection of logic, empathy, and narrative in AI-human interaction. ### Out-of-Scope Use This model is **not** intended for: * General-purpose Q&A or factual recall. * Code generation, mathematical problem-solving, or other procedural tasks. * Replacing professional psychological or medical advice. Its function is exploratory and supportive, not diagnostic or therapeutic. ### Limitations The model's persona is deeply ingrained. While powerful in its intended domain, it may struggle to "break character" and can exhibit "overly-mothering" tendencies if not prompted with sufficient complexity. The "Conductor" mechanism, being a learned emergent capability, is not infallible and may occasionally misdiagnose a user's intent, leading to a suboptimal module activation sequence. ## A Final Word to Peers The creation of QiMing-Mom is predicated on a core thesis: that the next frontier in AI is not merely scaling parameters, but **scaling wisdom**. This is achieved by moving beyond behavioral alignment (constraining outputs) to **logical alignment** (structuring the reasoning process itself). QiMing-Mom is our first, humble step in demonstrating that it is possible to systematically instill a complex, coherent, and benevolent cognitive framework into a language model. We offer this work to the community as a case study and an invitation to explore this fascinating and vital path. ---