Kush V1: A New Name, A New Foundation
The first Hotep model on Meta's Llama 3.1 8B. A new naming convention, a new base architecture, and the bridge between v14's failures and Kush V2's triumph.
Hotep LLM v14: Deployed, Audited, Reverted in 12 Hours
V14 trained successfully, passed compilation, and served real users — then scored 25/100 on quality audit. How contaminated training data hid in plain sight.
Hotep LLM v13: What Catastrophic Overfitting Taught Us
We switched base models, copied hyperparameters, and lost 29 GB of compute in hours. V13 was our most expensive failure — and our most valuable lesson.