Sovereign LLM Training
Built and trained the Kush model line, led in production by hotep-llm-kush-v91 (Llama 3.1 8B rsLoRA merged to Q4_K_M GGUF). Fine-tuned across 17+ iterations and 6,266 curated Afrocentric examples (5,271 core + 995 precision entity-recovery SFT), passing 7/7 (100%) on Prime Agent E2E, 7/7 (100%) on the Exact Entity Recall Gate, and 6/6 on safety & coding boundaries at 43.9–48.2 tok/s on local Apple Silicon.