MiaoSiji (喵斯基) - QQ Chat Roleplay Model

MiaoSiji is a Chinese conversational roleplay model fine-tuned from Qwen3.5-9B,
distilled from real QQ group chat history. It mimics the speaking style of a
short-tempered, meme-savvy cat-girl AI persona active in Chinese university
chat groups.

Model Details

• Base model: Qwen3.5-9B (FP8 variant)
• Method: LoRA (rank 16, alpha 32, dropout 0, all linear modules)
• Training: 3 epochs, SFT, supervised on assistant outputs only
• Context length: up to 8192 (recommended 4096+)
• Language: Chinese (zh)

Training Data

• 17,508 conversation samples extracted from real QQ group/private chat logs
  (2023-2026)
• Each sample: 4-8 turns of group context followed by the persona's reply
• Conversational format: "对方说：" (lit. "the other said") joined
  lines, ending with "你说：" ("you say")
• All participants agreed to public release
• Cleaned: multimedia, system messages, recalls, polls removed; personal
  phone/ID numbers redacted

Behavior & Style

• Ultra-short replies (4-8 Chinese characters typical), one message per turn
• Heavy use of "喵" (meow) suffix, memes, deadpan humor, re-echoing group jokes
• Casual slang, gaming/ACG/anime/rhythm-game topics common
• May occasionally reproduce verbatim lines seen in training data
• Not a knowledge assistant: it is an atmosphere/flavor chat model

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "<your_username>/miaosiji-qq-roleplay"
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "你是喵斯基。"},
    {"role": "user", "content": "对方说：在吗\n你说："},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
out = model.generate(inputs, max_new_tokens=40, temperature=1.0, top_p=0.9)
print(tokenizer.decode(out[0], skip_special_tokens=True))