gemma4-style-lora-step3000

Gemma 4 Style LoRA (step 3000) — prose style evaluator

Purpose: This adapter is a style evaluator / judge, not a fiction writer.

It classifies and scores English prose against a 14-metric Leech & Short–derived rubric and emits structured JSON style_profiles (plus short natural-language judgments). Those judgments power a bake-off and later steering cards used to train / score a stylish fiction writer and assemble long-form novels that stay on voice, style, and tone.

Companion GGUFs (merged + quantized for LM Studio / llama.cpp): alexokita/gemma4-style-step3000-GGUF.

Intended use

• Produce a full style_profile JSON for a prose passage
• Answer single-dimension questions (tone, POV, figurative density, FID, …) with a short rationale
• Score whether a draft matches a target steering card (voice / style / tone)
• Run agreement bake-offs vs held-out labeled prose (and Style-in-Fiction exemplars)
• After a successful bake-off: bulk re-label corpora and drive writer ↔ evaluator loops

Recommended inference settings (LM Studio / llama.cpp): temperature 1.0, top_p 0.95, top_k 64, thinking off for JSON / classification tasks. Prefer the Q4_K_M or Q5_K_M GGUF for day-to-day use on a 24 GB GPU.

Out of scope

• Open-ended story generation or “write a novel”
• Plot / character consistency (separate stack)
• Overall literary quality preference ranking without a human agreement study

How it was trained

Training mix (Phase 3 instruction pairs from classified prose):

• Classification (~25%) — “Provide a complete style profile…” → full JSON of the 14 metrics (+ related computables in the profile)
• Judgment (~75%) — “What tone / POV / figurative density…?” → bold label + short explanation

Approx. 430k train / 48k validation pairs under train/style_training/.

Training data

Styled fiction chunks labeled with metadata.style_profile, then converted to instruction pairs. Corpora include (HF sources):

• Literotica stories
• Erotic / fantasy / general fiction book corpora (e.g. Korshuk BookRix sets)
• Project Gutenberg fiction
• Horror novel chunks

Labels were produced by the Phase 2 pipeline (tools/style_classification/run_pipeline.py): computable metrics (spaCy / textstat) plus LLM semantic metrics grounded in a Style-in-Fiction rubric and RAG knowledge base (source/style_rubric.json, source/extracted/style_knowledge.jsonl).

The 14 Style-in-Fiction rubric metrics

Derived from Leech & Short, Style in Fiction (2nd ed.), operationalized in source/style_rubric.json (stats.dimensions = 14).

How those labels were generated:

1. Phase 1 — Rubric from Style in Fiction — Vision-transcribe Style in Fiction PDF → markdown; chunk into RAG style_knowledge.jsonl; distill checklist + 14 operational dimensions into style_rubric.json.
2. Phase 2 — Classify prose — Chunk fiction (~sentence-aware); compute spaCy/textstat fields for the four computable metrics; run multi-pass LLM classification with rubric + retrieved SiF excerpts (--pass both: Pass 1 lexical/syntax/discourse + POV; Pass 2 tone / mind style / narrative distance / FID).
3. Phase 3 — SFT pairs — Turn each styled row into classification + judgment instruction pairs for this LoRA.

Additional textual principles in the rubric (end_focus, segmentation, prose_rhythm, climax, …) may appear in full profiles from the classifier pipeline; the 14 rows above are the core training/eval dimensions for this adapter.

Example prompts

Full profile:

Provide a complete style profile for this passage as JSON.

Passage:
...

Single dimension:

What tone does the author strike in this passage?

Passage:
...

Steering-card check (post–Phase 5B):

Does this passage match the following steering card?
List mismatches only (field: expected vs observed).

Card: { "pov": "first_person", "register": "colloquial", "tone": "tense", ... }
Passage:
...

Files in this release

Limitations

• Optimized for analytic JSON / short judgments, not creative generation
• Label noise follows the Phase 2 teacher models; bake-off against held-out gold (and SiF exemplars) before trusting bulk re-label
• Multimodal / vision paths of base Gemma 4 are unused (text-only LoRA)

Citation / acknowledgements

• Leech, G. & Short, M. Style in Fiction (2nd ed.) — theoretical source of the rubric
• Unsloth, TRL, Hugging Face PEFT
• Base: google/gemma-4-26B-A4B-it / Unsloth QAT training weights

@software{gemma4_style_lora_step3000,
  title  = {Gemma 4 Style LoRA — prose style evaluator},
  author = {Okita, Alex},
  year   = {2026},
  url    = {https://huggingface.co/alexokita/gemma4-style-lora-step3000}
}

License

Use of the base Gemma 4 weights is subject to the Gemma license. This adapter inherits those terms for combined use.