Magpie-Align/Magpie-Qwen2-Pro-200K-English
收藏Hugging Face2024-07-03 更新2024-07-06 收录
下载链接:
https://hf-mirror.com/datasets/Magpie-Align/Magpie-Qwen2-Pro-200K-English
下载链接
链接失效反馈官方服务:
资源简介:
该数据集是通过Magpie方法使用Qwen/Qwen2-72B-Instruct模型生成的,包含了指令和响应的对话数据。数据集的特征包括输入长度、输出长度、任务类别、输入质量、输入难度、最小邻居距离、安全性、奖励和语言等。过滤设置包括输入质量、指令奖励、语言以及去除重复和不完整的指令。数据集的不同版本包括1M原始对话、300K高质量对话、200K高质量中文对话和200K高质量英文对话。
This dataset is generated by Qwen/Qwen2-72B-Instruct using the Magpie method, containing 200,000 training data entries. The dataset features include UUID, model name, generation input configurations, instructions, responses, conversations, task categories, difficulty, intent, knowledge, and more fields. The dataset split includes a training set, approximately 1GB in size. The description of the dataset details the meaning and data type of each field, as well as the generation method and filtering conditions of the dataset.
提供机构:
Magpie-Align
原始信息汇总
数据集概述
数据集信息
-
特征:
uuid: 字符串model: 字符串gen_input_configs: 结构体temperature: 浮点数top_p: 浮点数input_generator: 字符串seed: 空值extract_input: 字符串
instruction: 字符串response: 字符串conversations: 列表from: 字符串value: 字符串
task_category: 字符串other_task_category: 序列字符串task_category_generator: 字符串difficulty: 字符串intent: 字符串knowledge: 字符串difficulty_generator: 字符串input_quality: 字符串quality_explanation: 字符串quality_generator: 字符串llama_guard_2: 字符串reward_model: 字符串instruct_reward: 浮点数min_neighbor_distance: 浮点数repeat_count: 整数min_similar_uuid: 字符串instruction_length: 整数response_length: 整数language: 字符串
-
分割:
train:num_bytes: 1007184254.428362num_examples: 200000
-
下载大小: 599475522
-
数据集大小: 1007184254.428362
配置
- 默认配置:
data_files:split: trainpath: data/train-*
可用标签
- Input Length: 指令中的字符总数
- Output Length: 响应中的字符总数
- Task Category: 指令的具体类别
- Input Quality: 指令的清晰度、特异性和连贯性,评级为 very poor, poor, average, good, excellent
- Input Difficulty: 完成指令所需的知识水平,评级为 very easy, easy, medium, hard, very hard
- Minimum Neighbor Distance: 数据集中最近的邻居的嵌入距离,用于过滤重复或相似的实例
- Safety: 由 meta-llama/Meta-Llama-Guard-2-8B 标记的安全标签
- Reward: 奖励模型给出的特定指令-响应对的输出
- Language: 指令的语言
过滤设置
- Input Quality: >= good
- Instruction Reward: >=-10
- Language: English
- 移除重复和不完整的指令(例如,以 ":" 结尾的指令)
- 选择响应最长的 200K 数据



