61465/GZP-LLM

First Independent Arabic LLM from Egypt (showcase) # GZP-LLM **The First Independent Arabic Language Model from Egypt** **أول نموذج لغوي عربي مستقل من مصر** > Public showcase only. Datasets, checkpoints, training scripts, and proprietary evaluation harness remain private. --- The Arabic language has 400+ million speakers and a civilizational weight that deserves more than a translated wrapper around Western models. **GZP-LLM** is the first step toward digital sovereignty for Arabic AI — built by Arab hands, on Arab soil, with Arab cultural understanding baked in. This is not a fine-tune. This is a re-alignment. --- | | GZP-LLM | Translated GPT-style models | |---|---|---| | Cultural context | Arab daily life, idioms, dialect awareness | Generic, often US-centric | | Identity | Self-identifies as ARIA (Arab AI companion) via DPO | Self-identifies as the underlying foundation model | | Memory | Persistent SQLite per-user memory across sessions | Stateless or vendor-controlled memory | | Curiosity | Active curiosity engine — asks back, explores | Reactive only | | Reasoning surface | Uses GZP `reasoning_engine` (object/geometric/program-search primitives) | Native CoT only | | Deployability | Runs on a single consumer GPU via Ollama | Closed API or 70B+ requirement | --- - **Base model:** Qwen2.5-7B (chosen for strong Arabic baseline + permissive license) - **Adapter:** LoRA r=64, α=16 — full fine-tune impractical on Kaggle GPU budget - **Alignment:** Identity DPO — teaches the model "you are ARIA" not "I am Qwen" - **Fallback cascade:** Phi-3-mini for low-resource environments - **Memory:** SQLite per-user (`aria_brain.db`) with 3D evolution visualization - **Training infra:** Kaggle (T4 / P100), 6-step dataset audit pipeline before each run --- ``` _audit_step1.py ──► schema validation _audit_step2.py ──► duplicate detection _audit_step3.py ──► language purity (Arabic ratio per sample) _audit_step4.py ──► content safety filtering _audit_step5.py ──► ChatML format conformance _ …