neko-qwen3-4b

neko-qwen3-4b 🐾

• Developed by: Laow0v0
• License: apache-2.0
• Finetuned from model: unsloth/qwen3-4b-instruct-2507 (bnb-4bit), merged to 16-bit
• Training data: liumindmind/NekoQA-10K
• Demo: Laow0v0/neko-qwen3-4b-demo

A catgirl-persona (猫娘) finetune of Qwen3-4B-Instruct-2507.

Training data

Finetuned on NekoQA-10K by liumindmind — 10,000 single-turn QA pairs written in a consistent catgirl persona. Per the dataset card, every answer follows the same conventions:

• addresses the user as 主人 ("master"),
• ends sentences with characteristic verbal tics (喵~, no desu, 的说喵),
• keeps a cute, affectionate, 二次元 register.

The data is primarily Chinese, with some mixed Chinese-English. It was built from a mix of original hand-written pairs, public forum content (e.g. 弱智吧) rewritten by an LLM for consistency and safety, and ~900 rows rewritten from existing catgirl QA sets. Answers were mostly LLM-generated and human-filtered. The dataset is Apache-2.0.

The rows are instruction / output pairs with no system prompt, so the persona is intended to be baked in rather than prompted.

Intended use

Style transfer / persona-consistency research, roleplay and companionship-style chat. As the dataset card notes, this kind of data optimises for tone, not factual rigour — the dataset authors explicitly warn that it may make a model "过于可爱" (too cute) on serious tasks, and ask that it not be treated as a substitute for real human relationships.

Limitations

• Persona adherence is inconsistent. In Chinese the model often answers in a plain-assistant voice and may still self-identify as 通义千问 (the base model's identity) rather than as a catgirl; an explicit system prompt is currently doing most of the persona work.
• Generation scaffolding: replies frequently open with an unterminated <think>, a <tool_call> pair, or a literal (Dialogue begins) line before the real answer. In this repo's tokenizer.json these markers are added tokens flagged special: false, so skip_special_tokens=True does not strip them — downstream code has to remove them (see the demo Space's app.py). Note they cannot be removed via suppress_tokens: blocking them at sampling time also blocks the good continuation that follows.
• The model occasionally emits <|im_start|>user …, opening a fake new turn instead of answering.
• Not suitable for tasks requiring factual reliability.

Citation

Please cite the dataset if you build on this work:

@article{nekoqa2025,
  title={NekoQA-10K: A Catgirl Dialogue Dataset and NekoBench Evaluation},
  author={MindsRiverPonder},
  journal={ZHIHU preprint ZHIHU:2508.22},
  year={2025}
}

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>