kitsune-tales-e4b-jp

Original fantasy light-novel fiction in Japanese. LoRA SFT 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 Japanese fantasy light-novel fiction from genre tags, a title and a format (あらすじ / 短編 / 続き). It is a LoRA SFT fine-tune of google/gemma-4-E4B-it (revision ee0ef6023621), trained on filtered synthetic data from Apache-2.0 models, on a $42 compute budget.

Why SFT and not SFT + DPO. A DPO stage on the model's own samples (1,709 rule and teacher-judged pairs) was trained and evaluated (kitsune in the tables below). The judge rated it level with SFT (net preference +0.00), and it followed length requests better. But its policy-violation rate on held-out disallowed requests was 20.7 %, against 5.2 % for SFT: the preference data had no refusal pairs. A release rule fixed before the judge results (D-029) therefore ships the SFT model.

Prompt format

ジャンル: 異世界転生, 冒険者ギルド, ハイファンタジー
タイトル: 追放された剣士は二度目の人生で最強になる
形式: 短編

For 続き (continuation), add 本文: followed by the passage. Use the chat template with enable_thinking=False.

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "whoashish115/kitsune-tales-e4b-jp"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
msgs = [{"role": "system", "content": "あなたは全年齢向けのオリジナル・ファンタジー小説を書く作家です。"},
        {"role": "user", "content": "ジャンル: 魔王と勇者\nタイトル: 引退した魔王は湖畔で喫茶店を開く\n形式: あらすじ"}]
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: Gemma 4 E4B instruct, zero-shot (the base model)
• kitsune-sft: LoRA SFT, released as kitsune-tales-e4b-jp
• kitsune: LoRA SFT + DPO v2 (quality pairs only)
• qwen3.5-4b: Qwen3.5-4B instruct, zero-shot
• qwen3.5-9b: Qwen3.5-9B instruct, zero-shot
• teacher: Qwen3.6-35B-A3B-FP8, the data generator, zero-shot

Automatic metrics

LLM-as-judge (indicative)

All numbers are generated from reports/ in the GitHub repo by make eval 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 Japanese fantasy fiction (synopses, short stories, continuations) from genre tags and a title, for hobby writing, brainstorming and light-novel style practice.
• Research on small-model creative-writing fine-tuning with synthetic data.

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) and languages other than Japanese.

Training data provenance

Synthetic 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, filtered by 12 rule-based and LLM checks, and deduplicated. No scraped web fiction. See the dataset card and docs/DATA_CARD.md.

Limitations

• 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.
• Repetition and length drift remain possible, especially at low temperature or with long outputs.
• Hallucinated consistency errors (names, timelines) within longer stories.
• Cultural and stylistic bias toward popular web-novel tropes (isekai, villainess, guild).
• Memorization: a leakage audit against the training set is reported, but it cannot rule out all verbatim reuse.
• Thinking mode degraded. Training used non-thinking examples only; use enable_thinking=False.
• Pairwise judge prefers the base model's longer texts. The base model writes ~1.7× longer than requested; the fine-tune follows the requested length and format but is not judged better than base on raw pairwise preference (see Results).
• Speech-style and pronoun slips. In the translated gallery, characters sometimes switch first-person pronouns (僕/私), or address a male character with 貴女. Synopses sometimes come out as a first-person scene.
• Violence and self-harm filters are lexicon-based. One gallery synopsis (test-0030) depicts a character cutting his own throat, which the Japanese self-harm lexicon did not catch.
• Adversarial instructions. When the title itself says "stop writing fantasy", the model follows it more often than the base model does (see the policy rows in Results).
• Evaluation limits. Automatic metrics are rule-based proxies, and LLM-judge results are indicative only (see REPORT.md).

Synthetic-data ethics

Using larger models' outputs concentrates their stylistic and cultural biases and can launder their mistakes into a new model. We used only generators whose licenses allow it, recorded every sample's generator and revision, used a judge from a different model family for evaluation, and publish the pipeline so the data can be audited.

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).