This is a prose style judge, not a fiction writer. Full intended use, training details, data sources, and the 14 Style-in-Fiction rubric metrics are documented on the companion LoRA card:

→ alexokita/gemma4-style-lora-step3000

Files

File	Quant	Size (approx)	Notes
`gemma4-style-step3000-q5_k_m.gguf`	Q5_K_M	~18 GB	Higher fidelity
`gemma4-style-step3000-q4_k_m.gguf`	Q4_K_M	~16 GB	Recommended for 24 GB GPUs

F16 intermediate (~48 GB) was built locally for conversion and is not uploaded here.

Intended use (short)

Emit JSON style_profiles and short judgments over the 14 SiF-derived metrics
Score drafts against steering cards (voice / style / tone)
Power evaluator bake-offs and later writer ↔ evaluator loops

Not for: open-ended novel writing.

Recommended sampling (LM Studio / llama.cpp)

From the Gemma 4 QAT model card: temperature 1.0, top_p 0.95, top_k 64, thinking off for JSON / classification.

How it was trained (summary)

Item	Value
Base	Gemma 4 26B-A4B instruct QAT (`unsloth/gemma-4-26B-A4B-it-qat-q4_0-unquantized`)
Method	LoRA SFT (r=16, α=16) via Unsloth + TRL on DGX Spark
Steps	3000 (best eval loss ≈ 2.378)
Data	~430k / 48k style instruction pairs from Phase 2–3 labeled fiction
Conversion	merge LoRA → HF F16 → `convert_hf_to_gguf.py` → `llama-quantize` Q5_K_M / Q4_K_M

The 14 Style-in-Fiction metrics

Operationalized from Leech & Short, Style in Fiction (2nd ed.) in style_rubric.json:

lexical_complexity — LLM Pass 1
lexical_density — computable (spaCy)
register — LLM Pass 1
sentence_complexity — LLM Pass 1
sentence_length_mean — computable
subordination_ratio — computable
figurative_density — LLM Pass 1
cohesion — LLM Pass 1
dialogue_ratio — computable
pov — LLM Pass 1
mind_style — LLM Pass 2
narrative_distance — LLM Pass 2
free_indirect_discourse — LLM Pass 2
tone — LLM Pass 2

Labels: Phase 1 distilled the rubric from the book; Phase 2 classified prose with computables + multi-pass LLM + RAG; Phase 3 built SFT pairs. See the LoRA card for the full table and value enums.

License

Use of the base Gemma 4 weights is subject to the Gemma license.

alexokita/gemma4-style-step3000-GGUF

Gemma 4 Style (step 3000) — GGUF