Ninja069/Darija-HRCN-V4

# 🇲🇦 HRCN V4: Darija Neuro-Symbolic Polymath AI --- ## 📌 Abstract **HRCN V4 (Hybrid Recurrent Cognitive Network)** is a specialized Small Language Model (SLM) designed to solve the "Low-Resource Language" problem for **Moroccan Darija**. Standard Large Language Models (LLMs) like GPT-4 or Llama-2 require billions of parameters and massive hardware to function. HRCN V4 challenges this paradigm by utilizing a novel **"Polymath Architecture"**—a hybrid of LSTM (for grammatical stability) and Multi-Head Attention (for context retrieval). Crucially, it features a dedicated **Neuro-Symbolic Cortex** that allows the model to switch seamlessly between neural text generation and symbolic arithmetic, achieving **100% precision in math tasks** where larger models often hallucinate. ## 🚀 Key Accomplishments * **Extreme Efficiency:** Achieves conversational competency with only **20 Million Parameters** (vs 7 Billion for Llama-2). * **Edge-Native:** Runs on standard CPU/Mobile hardware with only **80MB of RAM**. * **Neuro-Symbolic Math:** Solves arithmetic problems (`CALC: 500 - 55`) with calculator-level precision, overcoming the "hallucination" limitation of standard transformers. * **Vector Alignment:** Successfully learned to rotate linguistic embeddings into a linear numerical manifold, proving semantic understanding of magnitude. * **Zero Catastrophic Forgetting:** The bicameral architecture allows the model to learn math without degrading its fluency in Darija. --- ## 🧠 The Polymath Architecture The model departs from the standard Transformer-only approach. Instead, it uses three distinct processing streams: 1. **Base Stream (LSTM):** Handles syntax, grammar, and short-term dependencies (The "Broca's Area"). 2. **Polymath Stream (Attention):** Handles long-context retrieval, recipes, and reasoning (The "Wernicke's Area"). 3. **Neuro-Symbolic Cortex:** A dedicated neural path that identifies logical queries and interfaces with a symbolic execution to …