mimran8971/llama-3-crownelius-writing

Llama 3 8B — Grounded Creative Writing Fine-tune

A QLoRA fine-tune of unsloth/llama-3-8b-Instruct-bnb-4bit, trained to produce more physically and psychologically grounded creative writing suggestions — favoring concrete, sensory, spatially-coherent prose over generic "flowery" or "vibes-based" description.

Training

• Base model: unsloth/llama-3-8b-Instruct-bnb-4bit
• Method: QLoRA (4-bit), via Unsloth
• Dataset: Crownelius/Creative-Writing-High-Quality-1300x — a synthetic dataset built around a "Think Before You Write" methodology, where each example pairs an instruction with a stage-blocking skeleton (explicit coordinates, weight distribution, and velocity vectors for action, or subtext mapping for dialogue) followed by prose conditioned on that skeleton.
• Instruction augmentation: training instructions were varied (full framework / simplified constraints / narrative-only targets) so the model learns the underlying skill rather than one fixed prompt format.

Intended use

Creative-writing continuation and drafting assistance — particularly scenes involving physical action, blocking, or spatial staging, where base instruction-tuned models tend to produce descriptions that are evocative but physically incoherent (e.g. limbs or objects moving in ways that don't track logically through a scene).

Not intended for: factual Q&A, code generation, or general-purpose chat — this fine-tune is narrowly scoped to creative-writing style and grounding, not broader capability improvements.

Files

• llama-3-8b-instruct.Q4_K_M.gguf — 4-bit quantization

Quickstart (Ollama)

# Download llama-3-8b-instruct.Q4_K_M.gguf from this repo, then:
ollama create llama-writer -f Modelfile
ollama run llama-writer

A minimal Modelfile: ``` FROM ./llama-3-8b-instruct.Q4_K_M.gguf

SYSTEM """You are an AI writing companion.

Help authors improve stories by producing coherent, grounded, physically plausible writing while preserving the author's intent and style.""" ```