OpenDataArena/OpenDataArena-scored-data-2603

该数据集是一个多配置集合，包含多个子数据集，主要用于代码和数学领域的指令跟随任务。数据集配置包括AM-Thinking-v1-Distilled-code（约323,965个示例）、AM-Thinking-v1-Distilled-math（约558,129个示例）、Code-Feedback（约66,383个示例）、CodeFeedback-Filtered-Instruction（约156,526个示例）、DeepMath-103K（约309,066个示例）、EpiCoder-func-380k（约380,000个示例）和Evol-Instruct-Code-80k-v1（部分列出）。每个示例包含id、source、instruction（指令文本）和output（输出文本），以及processed_scores和raw_scores两个评分结构，涵盖AtheneRM、Cleanliness、LLM_as_Judge_Complexity、Compress_Ratio、Deita_Complexity、Deita_Quality、EmbedSVD_Entropy、Logical_Word_Count、HES、IFD、Instag、MTLD、Normalized_Loss、PPL、Professionalism、Writing_Style、Required_Expertise、Facts_Trivia、Educational_Value、Readability、Reasoning、SkyworkRM_Qwen、SkyworkRM_Llama、Token_Entropy、Token_Length、TreeInstruct_Node、TreeInstruct_Depth、Unique_Token_Ratio、UPD、VOCD-D等多个评估指标，用于量化文本质量、复杂性、专业性和可读性。数据集适用于训练和评估语言模型在代码生成、数学推理和反馈过滤等任务中的性能。

This dataset is a multi-configuration collection containing several sub-datasets, primarily designed for instruction-following tasks in code and mathematics domains. Dataset configurations include AM-Thinking-v1-Distilled-code (approximately 323,965 examples), AM-Thinking-v1-Distilled-math (approximately 558,129 examples), Code-Feedback (approximately 66,383 examples), CodeFeedback-Filtered-Instruction (approximately 156,526 examples), DeepMath-103K (approximately 309,066 examples), EpiCoder-func-380k (approximately 380,000 examples), and Evol-Instruct-Code-80k-v1 (partially listed). Each example includes id, source, instruction (instruction text), and output (output text), along with processed_scores and raw_scores structures that encompass multiple evaluation metrics such as AtheneRM, Cleanliness, LLM_as_Judge_Complexity, Compress_Ratio, Deita_Complexity, Deita_Quality, EmbedSVD_Entropy, Logical_Word_Count, HES, IFD, Instag, MTLD, Normalized_Loss, PPL, Professionalism, Writing_Style, Required_Expertise, Facts_Trivia, Educational_Value, Readability, Reasoning, SkyworkRM_Qwen, SkyworkRM_Llama, Token_Entropy, Token_Length, TreeInstruct_Node, TreeInstruct_Depth, Unique_Token_Ratio, UPD, and VOCD-D, which quantify text quality, complexity, professionalism, and readability. The dataset is suitable for training and evaluating language models on tasks like code generation, mathematical reasoning, and feedback filtering.