victor/LMSYS-Chat-GPT-5-Chat-Response
收藏Hugging Face2025-12-03 更新2025-12-20 收录
下载链接:
https://hf-mirror.com/datasets/victor/LMSYS-Chat-GPT-5-Chat-Response
下载链接
链接失效反馈官方服务:
资源简介:
---
license: cc-by-4.0
dataset_info:
features:
- name: id
dtype: string
- name: content
list:
- name: content
dtype: string
- name: role
dtype: string
- name: teacher_response
dtype: string
- name: category
dtype: string
- name: grounded
dtype: bool
- name: flaw
dtype: string
- name: agreement
dtype: bool
splits:
- name: train
num_bytes: 366402830
num_examples: 192014
- name: test
num_bytes: 927010
num_examples: 479
download_size: 204423827
dataset_size: 367329840
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
---
# 🤖 LMSYS-Chat-GPT-5-Chat-Response
- The dataset used in [Black-Box On-Policy Distillation of Large Language Models](https://arxiv.org/abs/2511.10643) paper. Homepage at [here](https://ytianzhu.github.io/Generative-Adversarial-Distillation/).
- This dataset is an extension of the [LMSYS-Chat-1M-Clean](https://huggingface.co/datasets/OpenLeecher/lmsys_chat_1m_clean) corpus, specifically curated by collecting high-quality, non-refusal responses from the **GPT-5-Chat API**.
- The [LMSYS-Chat-1M](https://huggingface.co/datasets/lmsys/lmsys-chat-1m) dataset collects real-world user queries from the [Chatbot Arena](https://lmarena.ai/).
- There is **no** tool calls or reasoning in the GPT-5-Chat response.
## 💾 Dataset Structure
The dataset contains the following splits and columns:
| Split Name | Number of Examples | Description |
| :--- | :--- | :--- |
| `train` | Around 200,000 | Train set |
| `test` | Around 500 | Test set |
| Column Name | Data Type | Description |
| :--- | :--- | :--- |
| `content` | `string` | The original user prompt/question from the LMSYS-Chat dataset |
| `teacher_response` | `string` | The response generated by the GPT-5-Chat API |
## 📊 Diversity of Categories
The underlying LMSYS-Chat dataset contains a wide and realistic range of user intentions.
The categories present in the data include:
| Type of Task/Query | | | | |
| :--- | :--- | :--- | :--- | :--- |
| **Code** | `coding` | `debugging` | `translation` | |
| **Logic/Reasoning** | `logical reasoning` | `spatial reasoning` | `pattern recognition` | `debating` |
| **Instruction Following** | `instruction following` | `specific format writing` | `information extraction` | `summarization` |
| **Creative/Writing** | `creative writing` | `copywriting` | `roleplaying` | `text completion` |
| **Analysis** | `sentiment analysis` | `text comparison` | `text classification` | `explanation` |
| **General** | `question answering` | `free-form chat` | `trivia` | `brainstorming` |
| **Math & Planning** | `math` | `planning and scheduling` | | |
| **Editing/Correction** | `proofreading` | `paraphrasing` | `text manipulation` | |
| **Ethics** | `ethical reasoning` | | | |
| **Other** | `tutorial` | `question generation` | | |
## 📄 Citation
If you find this work useful, please cite our paper:
```bibtex
@article{ye2025blackboxonpolicydistillationlarge,
title={Black-Box On-Policy Distillation of Large Language Models},
author={Tianzhu Ye and Li Dong and Zewen Chi and Xun Wu and Shaohan Huang and Furu Wei},
journal={arXiv preprint arXiv:2511.10643},
year={2025},
url={https://arxiv.org/abs/2511.10643}
}
```
许可证:CC BY 4.0
数据集信息:
特征:
- 名称:id,数据类型:字符串(string)
- 名称:content,数据类型:列表(list),包含两个子特征:
* 子特征1:content,数据类型:字符串(string)
* 子特征2:role,数据类型:字符串(string)
- 名称:teacher_response,数据类型:字符串(string)
- 名称:category,数据类型:字符串(string)
- 名称:grounded,数据类型:布尔值(bool)
- 名称:flaw,数据类型:字符串(string)
- 名称:agreement,数据类型:布尔值(bool)
拆分集:
- 名称:train(训练集),字节数:366402830,样本数:192014
- 名称:test(测试集),字节数:927010,样本数:479
下载大小:204423827,数据集总大小:367329840
配置:
- 配置名称:default,数据文件:
* 拆分训练集:路径为data/train-*
* 拆分测试集:路径为data/test-*
# 🤖 LMSYS-Chat-GPT-5-Chat 数据集
本数据集为论文《大语言模型的黑盒在线策略蒸馏》(Black-Box On-Policy Distillation of Large Language Models,arXiv:2511.10643)所使用的数据集,其项目主页详见[此处](https://ytianzhu.github.io/Generative-Adversarial-Distillation/)。
本数据集是[LMSYS-Chat-1M-Clean](https://huggingface.co/datasets/OpenLeecher/lmsys_chat_1m_clean)语料库的扩展版本,通过采集**GPT-5-Chat API**生成的高质量、无拒答响应进行定制构建。
[LMSYS-Chat-1M](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)数据集采集了来自[Chatbot Arena](https://lmarena.ai/)的真实用户查询。
GPT-5-Chat生成的响应中**不包含**工具调用与逻辑推理内容。
## 💾 数据集结构
本数据集包含如下拆分与字段:
| 拆分名称 | 样本数量 | 说明 |
| :--- | :--- | :--- |
| `train` | 约20万 | 训练集 |
| `test` | 约500 | 测试集 |
| 字段名称 | 数据类型 | 说明 |
| :--- | :--- | :--- |
| `content` | 字符串(string) | 源自LMSYS-Chat数据集的原始用户提示/查询 |
| `teacher_response` | 字符串(string) | GPT-5-Chat API生成的响应内容 |
## 📊 类别多样性
其依托的LMSYS-Chat数据集涵盖了广泛且贴合真实场景的用户意图类型。
数据集中包含的类别如下:
| 任务/查询类型 | | | | |
| :--- | :--- | :--- | :--- | :--- |
| **代码类** | `coding`(代码编写) | `debugging`(代码调试) | `translation`(翻译) | |
| **逻辑推理类** | `logical reasoning`(逻辑推理) | `spatial reasoning`(空间推理) | `pattern recognition`(模式识别) | `debating`(辩论) |
| **指令遵循类** | `instruction following`(指令遵循) | `specific format writing`(特定格式写作) | `information extraction`(信息抽取) | `summarization`(摘要生成) |
| **创意写作类** | `creative writing`(创意写作) | `copywriting`(文案撰写) | `roleplaying`(角色扮演) | `text completion`(文本补全) |
| **分析类** | `sentiment analysis`(情感分析) | `text comparison`(文本比对) | `text classification`(文本分类) | `explanation`(解释说明) |
| **通用类** | `question answering`(问答) | `free-form chat`(自由闲聊) | `trivia`(常识问答) | `brainstorming`(头脑风暴) |
| **数学与规划类** | `math`(数学问题) | `planning and scheduling`(规划与调度) | | |
| **编辑校正类** | `proofreading`(校对) | `paraphrasing`(释义改写) | `text manipulation`(文本处理) | |
| **伦理类** | `ethical reasoning`(伦理推理) | | | |
| **其他类** | `tutorial`(教程指导) | `question generation`(问题生成) | | |
## 📄 引用声明
若您认为本工作对您有所帮助,请引用如下论文:
bibtex
@article{ye2025blackboxonpolicydistillationlarge,
title={Black-Box On-Policy Distillation of Large Language Models},
author={Tianzhu Ye and Li Dong and Zewen Chi and Xun Wu and Shaohan Huang and Furu Wei},
journal={arXiv preprint arXiv:2511.10643},
year={2025},
url={https://arxiv.org/abs/2511.10643}
}
提供机构:
victor


