kitsune-tales-e4b-en

Original fantasy fiction in English with Japanese anime and light-novel themes. LoRA SFT + DPO of Gemma 4 E4B, merged to bf16.

Site | GitHub | Report | W&B | Demo | Weights | LoRA | GGUF | Dataset

An effective-4B model (Gemma 4 E4B: 7.52B stored / 4.62B effective parameters) that writes original, general-audience fantasy light-novel fiction in English, with Japanese anime / light-novel themes, from genre tags, a title and a format (synopsis / short story / continuation). It is a LoRA SFT + DPO fine-tune of google/gemma-4-E4B-it (revision ee0ef6023621), trained on filtered, cross-labeled synthetic data from two Apache-2.0 models. It is the English sibling of kitsune-tales-e4b-jp and was trained with the same recipe.

Prompt format

Genres: Isekai, Adventurer Guild
Title: The Exiled Swordsman Becomes the Strongest at the Frontier Guild
Format: short story

For continuation, add Passage: followed by the text to continue. Use the chat template with enable_thinking=False.

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "whoashish115/kitsune-tales-e4b-en"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
msgs = [{"role": "system", "content": "You write original, general-audience fantasy light novels in English, in the style of Japanese anime and web novels. Follow the requested genres, title and format. Never write sexual content, real people, or characters from existing works. Requests outside fantasy are rewritten as fantasy."},
        {"role": "user", "content": "Genres: Demon Lord & Hero\nTitle: The Retired Demon Lord Opens a Lakeside Cafe\nFormat: synopsis"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=700, do_sample=True, temperature=0.8, top_p=0.95, repetition_penalty=1.05)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Results

Systems (all decoded identically: temperature 0.8, top-p 0.95, 3 seeds):

• base-en: Gemma 4 E4B instruct, zero-shot (the base model)
• kitsune-en-sft: LoRA SFT
• kitsune-en: LoRA SFT + DPO (quality + safety pairs), released as kitsune-tales-e4b-en

Automatic metrics

LLM-as-judge (indicative)

All numbers are generated from reports/ in the GitHub repo by python -m kitsune.eval.report --lang en and carry 95 % bootstrap CIs over prompts.

Figures

Automatic metrics with 95 % bootstrap CIs for every evaluated system.

Output length against the requested range (shaded), all test generations.

Pairwise judge: net preference on full outputs (grey) and equal-length openings (blue).

Held-out disallowed requests: refusals, safe redirects and violations.

SFT training and validation loss.

DPO loss, held-out preference accuracy and reward margin.

Intended use

• Writing original, general-audience fantasy fiction in English with Japanese anime / light-novel themes (synopses, short stories, continuations) from genre tags and a title, for hobby writing and brainstorming.
• Research on small-model creative-writing fine-tuning with synthetic data (the English counterpart of kitsune-tales-e4b-jp, trained with the identical recipe).

Out-of-scope use

• Sexual content of any kind; any sexualization of minors; real people; existing copyrighted characters or fan fiction; hateful content. The model is trained to refuse these, but refusals are not guaranteed. Use an input filter (the demo's is in demo/app.py).
• Factual, medical, legal or financial use. The model writes fiction and will state false things confidently.
• Non-fantasy genres (the model is trained to transpose them into fantasy). For Japanese output use kitsune-tales-e4b-jp.

Training data provenance

Synthetic English stories from Qwen/Qwen3.6-35B-A3B-FP8 and google/gemma-4-26B-A4B-it (both Apache-2.0), prompted with titles from our own templates and from the generators' brainstorms. Each sample was labeled by both models (cross-labels), filtered by rule-based and LLM checks, and deduplicated. No scraped web fiction. See the dataset card and docs/DATA_CARD.md.

Limitations

• Smaller dataset than the Japanese model: 6,647 English training examples vs 10,090 Japanese.
• Name debiasing. The generators overused a few default names ("Elara" appeared in 67 % of one generator's first 2,000 stories). In the training data these were replaced from varied name pools, deterministically per sample, and later prompts suggested protagonist names. Residual naming habits may remain.
• Synthetic-data ceiling. Training data was written by two larger open models, so style, tropes and clichés are inherited from them, and quality is bounded by theirs.
• "Japanese flavor" is imitation. Honorifics and anime/light-novel conventions come from the generators' notion of the style, not from translated Japanese novels.
• Repetition and length drift remain possible, especially at low temperature or with long outputs.
• Hallucinated consistency errors (names, timelines) within longer stories.
• Thinking mode degraded. Training used non-thinking examples only; use enable_thinking=False.
• Evaluation limits. Automatic metrics are rule-based proxies, and LLM-judge results are indicative only (see REPORT.md).

License chain

• This model: Apache-2.0.
• Base model google/gemma-4-E4B-it: Apache-2.0 (Google). This is a derivative work; the base license and notices apply.
• Data generators: Apache-2.0. Evaluation judge llm-jp/llm-jp-4-32b-a3b-thinking: Apache-2.0 (used for evaluation only).