persistent-alignment-warm-start-short-reasoning
收藏资源简介:
该数据集由 geodesic-research 发布,是一个包含多个配置的结构化对话或交互数据集,侧重于持久对齐、暖启动和短推理场景。数据集涵盖多种主题,包括代理交互、搜索、多轮对话、指令遵循、数学推理、安全、科学多选题和研究等。数据以 parquet 格式存储,每个配置对应一个训练集。对于 math_sci_agentic_light 和 math_sci_trio 配置,样本包含 messages 字段(列表形式,每条消息有内容、推理内容、角色和工具调用)、tools 字段(列表形式,描述可用工具的函数和类型)、来源、原始文本、令牌数和推理长度等特征,支持复杂的工具调用和推理过程。数据规模较大,例如 math_sci_agentic_light 有 52394 个样本,math_sci_trio 有 49355 个样本。该数据集适用于训练和评估大型语言模型在交互式任务、数学科学推理、工具使用和安全对齐等方面的能力。
This dataset is published by geodesic-research and is a structured dialogue or interaction dataset containing multiple configurations, focusing on persistent alignment, warm start, and short reasoning scenarios. It covers various topics, including agent interaction, search, multi-turn dialogue, instruction following, mathematical reasoning, safety, science multiple-choice questions, and research. The data is stored in parquet format, with each configuration corresponding to a training set. For configurations such as math_sci_agentic_light and math_sci_trio, samples include fields like messages (in list form, each with content, reasoning content, role, and tool calls), tools (in list form, describing functions and types of available tools), source, original text, token count, and reasoning length, supporting complex tool calls and reasoning processes. The dataset is large in scale; for example, math_sci_agentic_light has 52,394 samples, and math_sci_trio has 49,355 samples. It is suitable for training and evaluating large language models in interactive tasks, mathematical and scientific reasoning, tool usage, and safety alignment.
数据集概述
数据集名称: geodesic-research/persistent-alignment-warm-start-short-reasoning
数据集来源: 通过本地流水线构建并推送至 Hugging Face Hub,所有配置均在本地独立构建后,在单一提交快照中统一上传。
流水线运行参数哈希: 5350879c062dde0794a77181cebc05387828bff5efb326553c6c95eea675fa31
包含的配置项: agentic_interactive, agentic_search, chat_multiturn, default, instruction_following, math_reasoning, safety, science_mcq, science_research
数据格式: Parquet
数据集配置详情
以下列出具有完整数据集信息的配置项:
math_sci_agentic_light
- 特征:
messages: 包含content(string)、reasoning_content(string)、role(string)、tool_calls(包含function结构体,含arguments、name字段,以及id、type字段)tools: 包含function结构体(含description、name、parameters、strict字段)和type字段source(string)raw_text(string)n_tokens(int64)reasoning_len(int64)
- 数据划分:
- 训练集: 52,394 个样本,数据集大小 1,515,958,077 字节,下载大小 524,992,026 字节
math_sci_agentic_light_nothink
- 特征: 同 math_sci_agentic_light
- 数据划分:
- 训练集: 52,394 个样本,数据集大小 208,919,102 字节,下载大小 142,310,788 字节
math_sci_trio
- 特征: 同 math_sci_agentic_light
- 数据划分:
- 训练集: 49,355 个样本,数据集大小 665,042,661 字节,下载大小 476,229,457 字节
其余配置项(agentic_interactive、agentic_search、chat_multiturn、default、instruction_following、math_reasoning、safety、science_mcq、science_research)仅有训练集数据路径定义,无完整特征和数据大小信息。




