Welcome to ZombieLLM - a compact, fast, and delightfully undead language model. We started with the classic GPT-2 XL (1.5B) as the “corpse” and used an open GPT-OSS-20B teacher as the necromancer to bring it back to life. The goal is simple: deliver a witty, responsive assistant that runs comfortably on ordinary hardware, works offline, and keeps the playful zombie vibe without bloating your compute budget. Brains, but concise.

Under the hood, ZombieLLM learns in two major passes. First, we run instruction SFT with LoRA/DoRA on a blend of Dolly + Alpaca, teaching the model to follow instructions cleanly while keeping training efficient. Then we add a lightweight representation-level KD step that aligns internal features between teacher and student (think cosine-similarity in a shared projection space). This pairing retains the best behaviors of the big model while fitting them into a smaller, more frugal body.

After the brain transplant comes personality and domain sense. We apply a Survival pass to ground responses in practical, resilient know-how (water, shelter, first aid basics, risk awareness). A short Persona booster then stabilizes the undead voice so answers stay consistent, concise, and helpful without drifting into role-play excess. The end result is a model that sounds like itself, resists rambling, and defaults to “I don’t know” when context is missing.

Deployment is intentionally frictionless. We ship as GGUF for drop-in use with llama.cpp and Ollama. The provided Ollama template is stateless by design: every prompt is treated as a new conversation, which helps with privacy, reproducibility, and evaluation. You get the interactive feel you want without hidden history—perfect for demos, tinkering, and clean benchmarks.

Research use only. ZombieLLM is an experimental artifact meant for exploration, evaluation, and fun. It may produce incorrect, biased, or misleading outputs and is not intended for production or for professional advice (medical, legal, financial, or safety-critical). Always keep a human in the loop, apply your own filters and safeguards, and verify important claims. If that all sounds good, welcome aboard—enjoy the reanimation.