lamm-mit/magpie-ultra-v0.1-DPO

---
dataset_info:
features:
- name: model_name_response_base
dtype: string
- name: instruction
dtype: string
- name: response
dtype: string
- name: response_base
dtype: string
- name: intent
dtype: string
- name: knowledge
dtype: string
- name: difficulty
dtype: string
- name: model_name_difficulty
dtype: string
- name: explanation
dtype: string
- name: quality
dtype: string
- name: model_name_quality
dtype: string
- name: primary_tag
dtype: string
- name: other_tags
sequence: string
- name: model_name_classification
dtype: string
- name: embedding
sequence: float64
- name: model_name_embeddings
dtype: string
- name: score
dtype: float64
- name: score_base
dtype: float64
- name: distilabel_metadata
struct:
- name: raw_output_assign_tags_0
dtype: string
- name: nn_indices
sequence: int64
- name: nn_scores
sequence: float64
- name: messages
list:
- name: content
dtype: string
- name: role
dtype: string
- name: guard
dtype: string
- name: model_name_guard
dtype: string
- name: safe
dtype: bool
- name: hazard_category
dtype: string
- name: score_difference
dtype: float64
- name: text
dtype: string
- name: text_tok_length
dtype: int64
- name: rejected_response
dtype: string
- name: prompt
dtype: string
- name: chosen
list:
- name: content
dtype: string
- name: role
dtype: string
- name: rejected
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 1177571394
num_examples: 50000
download_size: 694713520
dataset_size: 1177571394
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# magpie-ultra-v0.1-DPO

Version of (argilla/magpie-ultra-v0.1) with rejected responses added for easy use in DPO or ORPO.

```bibtex
@misc{buehler2024magpie,
title={Generating science-aligned LLMs},
author={Markus J. Buehler, et al.},
year={2024},
eprint={},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@misc{xu2024magpie,
title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
year={2024},
eprint={2406.08464},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```

数据集信息：
特征：
- 字段名：model_name_response_base，数据类型：字符串
- 字段名：instruction(指令)，数据类型：字符串
- 字段名：response，数据类型：字符串
- 字段名：response_base，数据类型：字符串
- 字段名：intent，数据类型：字符串
- 字段名：knowledge，数据类型：字符串
- 字段名：difficulty，数据类型：字符串
- 字段名：model_name_difficulty，数据类型：字符串
- 字段名：explanation，数据类型：字符串
- 字段名：quality，数据类型：字符串
- 字段名：model_name_quality，数据类型：字符串
- 字段名：primary_tag，数据类型：字符串
- 字段名：other_tags，数据类型：字符串序列
- 字段名：model_name_classification，数据类型：字符串
- 字段名：embedding，数据类型：float64序列
- 字段名：model_name_embeddings，数据类型：字符串
- 字段名：score，数据类型：float64
- 字段名：score_base，数据类型：float64
- 字段名：distilabel_metadata，数据类型：结构体：
- 字段名：raw_output_assign_tags_0，数据类型：字符串
- 字段名：nn_indices，数据类型：int64序列
- 字段名：nn_scores，数据类型：float64序列
- 字段名：messages，数据类型：列表：
- 字段名：content，数据类型：字符串
- 字段名：role，数据类型：字符串
- 字段名：guard，数据类型：字符串
- 字段名：model_name_guard，数据类型：字符串
- 字段名：safe，数据类型：布尔值
- 字段名：hazard_category，数据类型：字符串
- 字段名：score_difference，数据类型：float64
- 字段名：text，数据类型：字符串
- 字段名：text_tok_length（文本Token长度），数据类型：int64
- 字段名：rejected_response，数据类型：字符串
- 字段名：prompt（提示词），数据类型：字符串
- 字段名：chosen，数据类型：列表：
- 字段名：content，数据类型：字符串
- 字段名：role，数据类型：字符串
- 字段名：rejected，数据类型：列表：
- 字段名：content，数据类型：字符串
- 字段名：role，数据类型：字符串
划分集：
- 名称：train，字节数：1177571394，样本数量：50000
下载大小：694713520
数据集总大小：1177571394
配置：
- 配置名称：default，数据文件：
- 划分集：train，路径：data/train-*

# magpie-ultra-v0.1-DPO
该数据集是(argilla/magpie-ultra-v0.1)的衍生版本，新增了被拒响应字段，可直接用于DPO或ORPO模型训练。

bibtex
@misc{buehler2024magpie,
title={生成科学对齐的大语言模型(Large Language Models)},
author={Markus J. Buehler, et al.},
year={2024},
eprint={},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@misc{xu2024magpie,
title={Magpie：通过向对齐的大语言模型(Large Language Models)发送空提示从头合成对齐数据},
author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
year={2024},
eprint={2406.08464},
archivePrefix={arXiv},
primaryClass={cs.CL}
}