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, it uses the questions from Dolly and Alpaca, paired with distilled final answers from GPT-OSS-20B, to run instruction SFT with LoRA/DoRA - teaching GPT-2 XL to follow instructions cleanly while staying efficient. Then, a lightweight representation-level KD step aligns internal features between teacher and student (using cosine similarity in a shared projection space). This combination transfers the knowledge and style of GPT-OSS-20B into a smaller, more frugal body while keeping the best behaviors of the big model alive.

After the brain transplant comes the survival instinct and personality. We apply a Survival pass using GPT-OSS-20B - distilled answers from the Bushcraft/Survival dataset to ground the model in practical skills - from finding water and building shelter to first aid and risk awareness. On top of that, a short Persona booster fine-tune with our custom zombie dataset stabilizes the undead voice, ensuring responses stay consistent, concise, and useful without drifting into pure role-play. The result is a model that carries survival know-how, speaks with its own recognizable character, avoids 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.