|
|
--- |
|
|
datasets: |
|
|
- ZhenghanYU/CFunSet |
|
|
language: |
|
|
- ch |
|
|
- zh |
|
|
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!** 🎉 |
