Update README.md

README.md

@@ -20,27 +20,22 @@ pipeline_tag: text-generation
 inference: false
 ---
 
-# Preface: The Era of Sovereign Intelligence
+# The Era of Sovereign Intelligence
 
 > *"Not your weights, not your intelligence."*
 > *In the dark forest of Web3, an individual without private AI is merely prey.*
 
-## 1. Model Summary & Legacy
+### Entropy & Value
+We stand at the intersection of history's two most significant exponential curves: **Artificial Intelligence**, which seeks to solve **Knowledge & Logic**, and **Blockchain**, which seeks to solve **Trust & Value**.
 
-### The Evolution of Sovereign Intelligence
-The DMind lineage was born from a singular conviction: **that decentralized finance deserves decentralized intelligence.** This journey began with **DMind-1**, which shattered the monopoly of closed-source AI by releasing the world's first Web3-native LLM, democratizing access to Solidity comprehension. It continued with **DMind-2**, which proved that domain-specific fine-tuning could outmaneuver trillion-parameter giants in vertical benchmarks, establishing a new standard for specialized AI.
+However, the current convergence is flawed. Mainstream AI remains a "Computational Leviathan"—centralized, opaque, and running on servers that ingest your data but cannot comprehend the concept of "Ownership." When a Web2 AI reads a smart contract, it sees code. When we look, we see **Assets, Risks, and Game Theory**.
 
-### A Paradigm Shift: From Retrieval to Reflection
-However, **DMind-3** represents our most significant evolutionary leap yet. We recognized that in the high-stakes, adversarial "dark forest" of DeFi, standard knowledge retrieval is a liability. A model must do more than recite facts; it must possess **Reflective Intelligence (System 2 Thinking)** to navigate risk.
-
-**DMind-3-mini** embodies this philosophy. Positioned as the "Brain" within our local ecosystem, it bridges the gap between the real-time reflexes of the edge-side *DMind-3-nano* and the macroscopic foresight of the cloud-native *DMind-3-max*. It is engineered not as a chatbot, but as a **"Computational Financial Actuary"**—a privacy-first, offline-capable engine designed to bring institutional-grade logic to the individual sovereign user.
-
-### 2. From Copilot to Species
+### From Copilot to Species
 Traditional "Copilots" are merely assistants. In the 24/7 PVP environment of DeFi, you need an **Agent** capable of independent risk assessment. Humans—carbon-based lifeforms with limited processing speed—are increasingly outmatched by MEV bots and algorithmic predators.
 
 We need an evolutionary tool: **An "Exo-Cortex" that runs locally, remains absolutely loyal, and deeply understands the dark logic of finance.**
 
-### 3. The Mission
+### The Mission
 **DMind-3-mini** was not built to score points on general benchmarks. It was engineered to arm the individual against institutional extraction. We refuse to upload your Alpha strategies to the cloud. True Web3 AI must be **Private**, **Local**, and **Antifragile**.
 
 * **DMind-3-nano** is your Shield.
@@ -50,23 +45,16 @@ Welcome to the era of **Sovereign Intelligence**.
 
 ---
 
-# Model Card: DMind-3-mini
+# DMind-3-mini: The Computational Financial Actuary
 
-## 1. Model Summary & Legacy
+## 1. Evolution & Legacy
 
-### The DMind Heritage
-The DMind series represents a continuous evolution in specialized Web3 artificial intelligence:
-* **DMind-1 Series:** Pioneered the industry by open-sourcing the world's first LLM dedicated to Web3/Solidity contexts.
-* **DMind-2 Series:** Achieved SOTA performance in vertical benchmarks, proving that domain-specific fine-tuning can outperform larger generalist models.
-* **DMind-3 Series (Current):** Marks a paradigm shift from "Knowledge Retrieval" to **"Reflective Intelligence" (System 2 Thinking)**. We introduce the **C³-SFT** paradigm to solve the hallucination and logic deficit problems in high-stakes financial environments.
+**The DMind lineage was born from a singular conviction: that decentralized finance deserves decentralized intelligence.** This journey began with **DMind-1**, which shattered the monopoly of closed-source AI by releasing the world's first Web3-native LLM, democratizing access to Solidity comprehension. It continued with **DMind-2**, which proved that domain-specific fine-tuning could outmaneuver trillion-parameter giants in vertical benchmarks, establishing a new standard for specialized AI.
 
-### The DMind-3 Family
-The DMind-3 lineup is architected for specific computational environments:
-* **DMind-3-nano (270M):** **The Shield.** An edge-side Intent Recognition model. Optimized for <500MB memory footprint, delivering >98% accuracy in identifying malicious transaction signatures locally.
-* **DMind-3-mini (4B) [THIS MODEL]:** **The Brain.** A "Local Financial Actuary". The core reasoning engine designed for high-end consumer workstations, balancing deep logic inference with privacy-first offline deployment.
-* **DMind-3-max (MoE):** **The Oracle.** A cloud-native Mixture-of-Experts model designed for cross-chain macroeconomic simulation and ecosystem-wide governance.
+**A Paradigm Shift: From Retrieval to Reflection**
+**DMind-3** represents our most significant evolutionary leap yet. We recognized that in the high-stakes, adversarial "dark forest" of DeFi, standard knowledge retrieval is a liability. A model must do more than recite facts; it must possess **Reflective Intelligence (System 2 Thinking)** to navigate risk.
 
----
+**DMind-3-mini** embodies this philosophy. Positioned as the "Brain" within our local ecosystem, it bridges the gap between the real-time reflexes of the edge-side *DMind-3-nano* and the macroscopic foresight of the cloud-native *DMind-3-max*. It is engineered not as a chatbot, but as a **"Computational Financial Actuary"**—a privacy-first, offline-capable engine designed to bring institutional-grade logic to the individual sovereign user.
 
 ## 2. Model Details
 
@@ -79,31 +67,29 @@ The DMind-3 lineup is architected for specific computational environments:
 * **Context Window:** 128k tokens
 * **Hardware Requirement:** GPU with $\ge$ **12GB VRAM** (Recommended: NVIDIA RTX 4070Ti+, Apple M3/M4 Pro/Max).
 
----
-
 ## 3. Methodology: C³-SFT
 
 **DMind-3-mini** introduces **Contrastive Chain-of-Correction Supervised Fine-Tuning (C³-SFT)**.
-Unlike standard SFT which models a direct mapping $P(y|x)$, C³-SFT forces the model to navigate a "Correction Trajectory" by contrasting against plausible but flawed reasoning (Negative Samples).
+Unlike standard SFT which models a direct mapping \\(P(y|x)\\), C³-SFT forces the model to navigate a "Correction Trajectory" by contrasting against plausible but flawed reasoning (Negative Samples).
 
 ![Figure 1: C3-SFT Paradigm](https://your-image-host-url/figure1-c3sft.png)
 *(Figure 1: The C³-SFT training pipeline, illustrating the Triplet Data Structure and Contextual Loss Masking)*
 
 ### 3.1 Mathematical Formalization
-Let $\mathcal{D} = \{(x, y^-, y^+_{cot})\}_{i=1}^N$ be our dataset, where $x$ is the financial query, $y^-$ is a "Negative Sample" containing common logical fallacies (e.g., overlooking Impermanent Loss), and $y^+_{cot}$ is the corrective Chain-of-Thought.
-
-The optimization objective $\mathcal{L}_{C^3}$ is defined as maximizing the conditional probability of the correction path, given the error context:
+The optimization objective \\(\mathcal{L}_{C^3}\\) is defined as maximizing the conditional probability of the correction path, given the error context:
 
 $$
 \mathcal{L}_{C^3}(\theta) = - \mathbb{E}_{\mathcal{D}} \left[ \sum_{t=1}^{T} \alpha_t \cdot \log P_\theta(y^+_{t} \mid x, y^-, y^+_{<t}) \right] + \lambda \mathbb{KL}[\pi_\theta || \pi_{\text{ref}}]
 $$
 
-Where:
-* The input sequence explicitly includes the error state: $[x; \texttt{<SEP>}; y^-; \texttt{<CRITIQUE>}]$.
-* $\alpha_t$ is a dynamic attention weight that penalizes logical discontinuities.
+* \\(\mathcal{D} = \{(x, y^-, y^+_{cot})\}_{i=1}^N\\) represents the training dataset containing financial query triplets.
+* \\(y^-\\) denotes the "Negative Sample" containing common logical fallacies (e.g., overlooking Impermanent Loss).
+* \\(y^+_{cot}\\) is the corrective Chain-of-Thought that the model aims to generate.
+* \\(\alpha_t\\) is a dynamic attention weight that penalizes logical discontinuities during the reasoning steps.
+* \\(\pi_{\text{ref}}\\) is the reference policy used for KL-divergence regularization to prevent mode collapse.
 
 ### 3.2 Dual-State Inference Mechanism
-During inference, DMind-3-mini operates in two distinct topological modes based on the presence of a trigger token $\tau$. Let $\mathcal{I}(x)$ denote the inference function:
+During inference, DMind-3-mini operates in two distinct topological modes based on the presence of a trigger token \\(\tau\\). Let \\(\mathcal{I}(x)\\) denote the inference function:
 
 $$
 \hat{y} = 
@@ -113,10 +99,9 @@ $$
 \end{cases}
 $$
 
-* **Standard Mode:** Optimized for latency; generates direct financial summaries.
-* **Audit Mode:** The model internally generates a latent negative hypothesis $\mathcal{G}_{neg}(x)$ and applies the critique operator $\mathcal{H}_{crit}$ to derive a rigorously verified conclusion. This is critical for **Smart Contract Auditing** and **Ponzi Scheme Detection**.
-
----
+* \\(\tau\\) represents the control token (e.g., `<REFLECT>`) injected into the system prompt.
+* \\(\mathcal{G}_{neg}(x)\\) is the latent negative hypothesis generated internally by the model to simulate a potential error.
+* \\(\mathcal{H}_{crit}\\) is the critique operator that filters the latent hypothesis to derive a rigorously verified conclusion.
 
 ## 4. Intended Use: Web3 Financial Know-How
 
@@ -136,8 +121,6 @@ For maximum security, we recommend the **DMind Local Stack**:
 * **The Brain (DMind-3-mini):** Runs on your high-performance laptop. Handles complex strategy formulation, deep research, and "System 2" logic.
 * **The Shield (DMind-3-nano):** Runs in your browser/wallet background. Handles real-time transaction signing safety checks and "System 1" intuition.
 
----
-
 ## 5. Training Data
 
 The model was fine-tuned on **82,000** high-value private samples:
@@ -146,11 +129,9 @@ The model was fine-tuned on **82,000** high-value private samples:
 | :--- | :---: | :--- |
 | **Institutional Alpha Reports** | 40% | Deep dive reports from top-tier firms (e.g., Paradigm, Delphi), structured into logic chains. |
 | **Financial Post-Mortems** | 30% | Historical analysis of collapses (Luna, FTX, Euler Hack), focusing on pre-crash indicators. |
-| **Smart Contract Audits** | 20% | C³-SFT formatted pairs: Vulnerable Code $\to$ Exploit Analysis $\to$ Fix. |
+| **Smart Contract Audits** | 20% | C³-SFT formatted pairs: \\(\text{Vulnerable Code} \to \text{Exploit Analysis} \to \text{Fix}\\). |
 | **On-Chain Behavior Logs** | 10% | Parsed intent analysis of "Smart Money" wallet operations during high volatility events. |
 
----
-
 ## 6. Performance Benchmarks
 
 Evaluated on **Web3-Finance-Eval-2026**:
@@ -164,10 +145,20 @@ Evaluated on **Web3-Finance-Eval-2026**:
 
 *DMind-3-mini outperforms generalist models 15x its size in specific vertical tasks, validating the C³-SFT approach.*
 
----
-
 ## 7. Limitations & Disclaimer
 
 * **High Hardware Barrier:** Due to the decision to retain BF16 precision for financial accuracy, this model **requires >= 12GB VRAM**. It is not suitable for standard office laptops.
 * **Knowledge Cutoff:** While the *logic* is robust, specific protocol data (TVL, Price) is limited to the training cutoff. Use with RAG for real-time data.
 * **Legal Disclaimer:** This model is an **analytical tool**, not a financial advisor. The output (NFA) should never be the sole basis for investment decisions. The developers assume no liability for financial losses.
+
+---
+
+## Citation
+
+```bibtex
+@model{dmind2026mini,
+  title={DMind-3-mini: A 4B Parameter LLM for Web3 Financial Intelligence},
+  author={Huang, Enhao and Zhulong AI Research Team},
+  year={2026},
+  url={[https://huggingface.co/zhulong-ai/dmind-3-mini](https://huggingface.co/zhulong-ai/dmind-3-mini)}
+}