Lorekeeper - 40K Narrator
// local-ml - grimdark voice
Fine-tuned Gemma 4 E4B-it model that roleplays as a weary Warhammer 40K witness-narrator, trained for clean TTS output.
> overview
A local grimdark narrator model fine-tuned from google/gemma-4-E4B-it to speak as the Lorekeeper: an ancient soul who claims memory of the Horus Heresy with the fatigue, dry humor, and bruised wisdom of ten thousand years. The training set uses 500 hand-written conversation examples split across lore narration, direct questions, refusals, clarifications, and atmosphere, with every sample shaped for spoken delivery. It ships as GGUF quantizations for Ollama/Open WebUI use, including a best-quality Q8_0 build and a lighter Q4_K_M build for 8GB-VRAM machines.
> features
- Witness-style narrator voice: weary, dry, direct, and grimly philosophical
- QLoRA 4-bit fine-tune with LoRA rank 16 and alpha 32
- 500 hand-crafted examples across lore, questions, refusals, clarifications, and atmosphere
- TTS-optimized responses with no markdown, written-out numbers, and clean paragraph breaks
- Q8_0 and Q4_K_M GGUF quantizations for high-quality and 8GB-VRAM local runs
- Runs through Ollama and Open WebUI, with optional lore-book RAG collection