readme

README.md

@@ -77,34 +77,38 @@ Plaintext
     +-------------------------------------+
 
 ```
-## Methodology to Reduce Hardware Cost
+## Methodology of Training
 ----------------------------------------------------
+To achieve high-fidelity reasoning in a compact 7B parameter footprint, Energy-Intelligence was developed through a **Distillation & RLHF Architecture**:
 
-Our methodology focuses on embedding the intelligence of large-scale systems into a compact and efficient architecture:
-
-- By distilling knowledge from a high-parameter Teacher into a 7B model, we significantly reduce computational requirements without sacrificing reasoning capability.  
-- The approach captures the **brains of domain experts**, built upon decades of domain expertise and engineering practices.  
-- Optimized training and alignment ensure that the model delivers high accuracy with minimal resource consumption.  
-- This enables deployment on cost-efficient infrastructure, including edge environments, while maintaining enterprise-grade performance.
+1. **RLHF (Reinforcement Learning from Human Feedback):**  
+   Human evaluators review multiple responses generated by the model and select the better one. The model improves based on these preferences, making it more accurate, helpful, and aligned with real-world expectations.
 
-----------------------------------------------------
+2. **Synthetic Data Generation:**  
+   We utilized synthetic data generated by the Teacher model to capture domain knowledge and real-world scenarios, enabling scalable training with improved accuracy and coverage of complex use cases.
 
-### Why Adding RLAIF Matters for the Model Card
+3. **Distillation:**  
+   - **The Oracle (Teacher):** We utilized Gemini Pro as a high-parameter teacher model, providing it with domain knowledge, business logic, and complex system understanding to generate high-quality learning data.  
+   - **The Specialist (Student):** The Qwen2.5-7B-Instruct base model was fine-tuned on this curated dataset, effectively capturing the Teacher’s advanced reasoning in a more efficient form.
 
-- **Precision:** The model is actively corrected during training, ensuring higher accuracy and consistency.  
-- **Domain Safety:** Reduces the risk of incorrect outputs that could impact critical energy operations.  
-- **Technical Authority:** Demonstrates the use of advanced alignment techniques tailored for domain-aware reasoning and real-world system constraints.
+4. **The Result:**  
+   A model that possesses the intelligence of a much larger AI system while operating with the speed and cost-efficiency required for real-time industrial monitoring and analytics.
 
-To achieve high-fidelity reasoning in a compact 7B parameter footprint, Energy-Intelligence was developed through a **Distillation & RLAIF Architecture**:
+----------------------------------------------------
+### Why Adding RLHF Matters for the Model Card
 
-1.  **The Oracle (Teacher):** We utilized Gemini Pro as a high-parameter teacher model, providing it with domain knowledge, business logic, and complex system understanding to generate high-quality learning data.
+- **Precision:** The model is refined using human feedback, improving the quality and reliability of responses.  
+- **Domain Safety:** Reduces the risk of incorrect outputs that could impact critical energy operations.  
+- **Human Alignment:** Ensures the model behaves in a helpful, consistent, and context-aware manner aligned with human expectations.
+  
+Our methodology focuses on embedding the intelligence of large-scale systems into a compact and efficient architecture:
 
-2.  **The Specialist (Student):** The Qwen2.5-7B-Instruct base model was fine-tuned on this curated dataset, effectively capturing the Teacher’s advanced reasoning in a more efficient form.
+- By distilling knowledge from a high-parameter Teacher into a 7B model, we significantly reduce computational requirements without sacrificing reasoning capability.  
+- The approach captures the **brains of domain experts**, built upon decades of domain expertise and engineering practices.  
+- Optimized training and alignment ensure that the model delivers high accuracy with minimal resource consumption.  
+- This enables deployment on cost-efficient infrastructure, including edge environments, while maintaining enterprise-grade performance.
 
-3.  **RLAIF (Reinforcement Learning from AI Feedback):** To ensure high reliability, we implemented an RLAIF loop where the Teacher model (Gemini Pro) acted as an automated evaluator. The model was guided to improve accuracy, consistency, and context-aware decision-making through continuous feedback.
 
-4.  **The Result:** A model that possesses the intelligence of a much larger AI system while operating with the speed and cost-efficiency required for real-time industrial monitoring and analytics.
-----------------------------------------------------
 * * * * *