trajectory-analysis
收藏资源简介:
该数据集名为具有顺序依赖性的轨迹分析,专注于对六个大型语言模型在HLE和OpenMathReasoning两个数据集上生成的推理轨迹进行步骤级分析。每个推理步骤根据其在轨迹中的角色和依赖关系被分类为三种类型之一:尝试步骤(引入或探索尚未确定价值的新思路)、子任务步骤(执行推理的具体组成部分,且不依赖于尚未可用的早期步骤输出)以及顺序步骤(需要早期编号步骤的输出或上下文,因此无法与该先决条件并行处理)。数据集以TXT文件格式组织,目录结构遵循<模型>/<数据集>/*.txt的模式。每个文件包含编号的推理步骤、步骤级别的分类标签及解释说明、合并的步骤文本,并在可能时包含最终答案。每个分类都附有解释,特别是对于顺序步骤,解释会明确指出其依赖的先决步骤以及为何无法并行执行的原因。该数据集旨在支持对推理过程并行性、步骤依赖关系和推理轨迹结构的研究与分析,适用于推理任务分析、模型行为评估和自动化推理系统优化等场景。需要注意的是,该数据集目前处于积极更新状态,其覆盖范围、分类和元数据可能会随着新轨迹的纳入和现有条目的审查而不断演进。
This dataset, named 'Trajectory Analysis with Sequential Dependencies', focuses on step-level analysis of reasoning trajectories generated by six large language models (LLMs) across two datasets: HLE and OpenMathReasoning. Each reasoning step is categorized into one of three types based on its role in the trajectory and dependency relationships: 'Trial Step' (introduces or explores new ideas whose value has not yet been confirmed), 'Subtask Step' (executes a specific component of reasoning without relying on outputs from earlier steps that are not yet available), and 'Sequential Step' (requires outputs or context from earlier numbered steps, thus cannot be processed in parallel with those prerequisites). The dataset is organized in TXT file format, with a directory structure following the pattern <model>/<dataset>/*.txt. Each file contains numbered reasoning steps, step-level classification labels with explanatory notes, merged step text, and final answers when available. Each category is accompanied by an explanation; specifically for 'Sequential Step', the explanation will clearly indicate its prerequisite steps and the reason why parallel execution is not possible. This dataset aims to support research and analysis of reasoning process parallelism, step dependencies, and reasoning trajectory structure, and is applicable to scenarios such as reasoning task analysis, model behavior evaluation, and automated reasoning system optimization. It should be noted that this dataset is currently in active development, and its coverage, classification, and metadata may continue to evolve as new trajectories are added and existing entries are reviewed.
数据集概述
名称:Trajectory Analysis with Sequential Dependencies(含顺序依赖的轨迹分析)
语言:英文
标签:推理、轨迹分析、步骤分类、并行性、依赖分析
状态:该数据集正在持续更新中,其覆盖范围、分类及元数据可能随轨迹增加和已有条目审查而演变。
数据内容
该数据集包含来自两个数据集上六个模型生成的推理轨迹的步骤级分类。每个推理步骤根据其在轨迹中的作用和依赖关系被标记为以下三类之一:
- Trial Step:引入或研究一个价值未知的想法或方法。
- Subtask Step:执行推理中一个具体的子组件,该子组件不依赖于不可用的先前步骤输出。
- Sequential Step:需要依赖更早编号步骤的输出或上下文,因此无法与该前置步骤并行处理。每个 Sequential Step 的解释中会标识其前置步骤及无法并行执行的原因。
注:步骤1不会被标记为 Sequential,因为它没有更早的编号步骤。
模型与数据集
数据集按 <模型>/<数据集>/*.txt 组织。
模型(6个):
- Claude-Opus-4.5
- DeepSeek-R1
- GPT-5.5
- Gemini-2.5-Pro
- Gemini-3-high-thinking
- Qwen3-30B-A3B
数据集(2个):
- HLE
- OpenMathReasoning
文件格式
每个非空的 TXT 文件包含以下内容:
- 按编号划分的推理步骤
- 步骤级分类及其解释
- 合并后的步骤文本
- 最终答案(如存在)
典型分类行示例:
Step 7: Sequential Step [This step requires an intermediate result established in an earlier step, so the two steps cannot run in parallel.]
方法论
每个步骤基于其文本及其与周围轨迹的关系进行评估,然后被分配三个标签之一。当一个步骤依赖更早编号步骤且无法在同一时间独立处理时,被分类为 Sequential,其解释会指明前置步骤和顺序约束的原因。
验证
数据集经过以下检查:
- 分类解释非空
- Sequential 步骤的解释标识了顺序依赖关系
- 步骤1未被标记为 Sequential




