Synthetic Task Graph Dataset

该数据集专为训练和评估计划图方法而生成，包含了一系列综合任务图。其中，训练集特别关注大型语言模型（LLMs）的规划能力，并包含了随机生成的以及基于树的定向无环图（DAG）结构。测试集则涵盖了基线测试和边缘变化测试。该数据集规模宏大，拥有12,000个训练实例，并分为三种节点规模（10、30、50个节点）。任务旨在评估大型语言模型在任务图上的并行规划能力。

This dataset is specifically generated for training and evaluating planning graph methods, and comprises a series of comprehensive task graphs. The training subset specifically focuses on the planning capabilities of Large Language Models (LLMs), and includes both randomly generated and tree-based Directed Acyclic Graphs (DAGs). The test subset covers baseline tests and edge variation tests. This dataset has a substantial scale with 12,000 training instances, and is divided into three node count tiers: 10, 30, and 50 nodes. The task aims to evaluate the parallel planning capabilities of Large Language Models (LLMs) on task graphs.

搜集汇总

背景与挑战

背景概述

该数据集是'Plan-over-Graph'项目的一部分，用于支持大语言模型（LLM）代理的并行调度研究。它通过自动化管道生成合成任务图，帮助模型将复杂文本任务分解为可并行执行的子任务，从而提升任务效率和性能。数据集主要用于训练和评估'plan-over-graph'方法，以优化LLM在抽象图结构上的规划能力。