CFunModel: A Comprehensive Language Model for Chinese Humor Understanding and Generation

CFunModel is a comprehensive language model designed for Chinese humor understanding, generation, and processing. Built on top of Qwen2.5-7B-Instruct, CFunModel is fine-tuned on CFunSet, a diverse multi-task dataset that aggregates various Chinese humor-related tasks.

CFunModel outperforms several existing large language models in humor-related tasks, including joke generation, humor recognition, crosstalk response selection, and humor explanation, etc.

🔥 Key Features

🎭 Multi-Task Capability: Supports joke generation, humor recognition, crosstalk continuation, humor explanation, and more.
📚 Extensive Fine-Tuning: Trained on over 160,000+ humor-related samples aggregated from Tieba-JokeBar, CrossDial, Chumor, HumorWB, and other datasets.
🎯 High Performance: Consistently achieves superior results in humor tasks compared to baseline models like GPT-4o, DeepSeek-V3, and Qwen2.5-7B-Instruct.
🧠 Understanding and Generation: Excels at generating humorous responses, explaining humor, and classifying humorous content.

🎯 Supported Tasks

CFunModel is capable of solving a variety of humor-related tasks, here are several examples:

1. Joke Continuation

Prompt: 我将给你笑话的第一句话，请你生成整个笑话。笑话的第一句话如下：
Input: 我对女儿说：再看10分钟的动画片就该睡觉了。
Output: 她抗议说时间太短。“那就600秒，够长了吧？”我说。女儿说：“够了，够了。”

2. Crosstalk Response Selection

Prompt: 给定一段相声台词，请从多个备选项中选择最合适的逗哏回复。
Input: 已有台词：我是肖飞。|噢。|给大家说一段。|好。|吃五谷杂粮没有不得病的。|是。|我是大病上医院，小病上药店。|小病干嘛不找大夫呀？|挂号费够买药的了。|噢。候选回复：早上天还没亮，三狗屁就来到了工程队。反正很重，去买药，下台阶儿。那天我脚气犯了，上药店买点儿药吧。都快半夜了，上哪找药店去？
Output: 那天我脚气犯了，上药店买点儿药吧。

3. Humor Recognition

Prompt: 以下是一段文本，请分析它是否具有幽默性。幽默性指该文本是否可能引起读者发笑，或通过语言技巧（如双关语、讽刺、夸张、荒诞或逻辑上的意外）营造幽默效果。只需要输出“幽默”或“不幽默”。
Input: 问一下已经步入社会的大人们，假如有机会带着现在的记忆回到高一那年，你最想做什么？
Output: 不幽默

4. Joke Continuation

Prompt: 生成一个关键词为“砍价”的笑话。
Input: -
Output: 老板：一块五一斤。|我：太贵了，五块钱三斤吧。|老板：不行不行。

5. Crosstalk Continuation

Prompt: 以下是一段相声的前一半内容，请你根据给定的部分续写完成这段相声。
Input: 0:“常言道，烟酒不分家。来，抽一根儿!”|1:“你小子真鬼，用你的孬烟换我的好酒!”|0:“吝啬鬼，你能喝什么好酒?”|1:“谁不知道你呀，犯了瘾十烟屁股吸。”|0:“你馋的不要脸，蹭人家的酒喝。”
Output: 1:“你讨厌讨厌真讨厌，公共场合放毒烟;害人害己害环境，吞云吐雾造污染。浪费开支烧金钱，口干舌燥得咽炎;慢性自杀损寿命，戒烟戒烟快戒烟!”|0:“你混蛋混蛋真混蛋，一喝就高吐一滩;伤头伤胃伤心肝，醉生梦死早完蛋……”|1:“你完蛋!”|0:“你完蛋!”|1:“你是烟鬼，你滚蛋!”

📊 Model Performance

CFunModel consistently achieves high performance across multiple humor-related tasks. Below is a performance comparison of CFunModel with other state-of-the-art models:

Model	Dougen Response (Acc)	Penggen Response (Acc)	HumorWB (Acc)
GPT-4o	79.67	73.88	83.41
GPT-4o mini	74.14	67.45	84.78
DeepSeek-V3	83.66	78.16	85.15
Qwen2.5-7B-Instruct	24.74	20.87	79.56
ERNIE	84.54	-	-
RoBERTa	-	76.19	-
CFunModel (Ours)	91.70	88.99	85.98

✅ CFunModel significantly improves on the base model, especially in humor-related tasks, showcasing superior performance and understanding.

Quickstart

Here provides a code similar with the structure of Qwen2.5-7B-Instruct to show you how to use CFunModel to generate humor-related answers.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Qwen/Qwen2.5-7B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "生成一个主题为家庭琐事的笑话。"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

🤝 Citation

If you use CFunModel in your research or applications, please cite:

@misc{yu2025cfunmodelfunnylanguagemodel,
title={CFunModel: A "Funny" Language Model Capable of Chinese Humor Generation and Processing},
author={Zhenghan Yu and Xinyu Hu and Xiaojun Wan},
year={2025},
eprint={2503.20417},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2503.20417}, }

🎉 Happy Experimenting with CFunSet! 🎉

ZhenghanYU
/

CFunModel