README.md

---

datasets:
- ZhenghanYU/CFunSet
language:
- zho
- eng
- fra
- spa
- por
- deu
- ita
- rus
- jpn
- kor
- vie
- tha
- ara
base_model:
- Qwen/Qwen2.5-7B-Instruct
---

# 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. 

```python

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!** 🎉