ACuRL
收藏资源简介:
ACuRL Curriculum Tasks数据集包含由ACuRL(自主课程强化学习框架)生成的课程任务,旨在使计算机使用智能体在零人类数据的情况下持续适应目标环境。该数据集专为文本生成和强化学习任务设计,特别适用于计算机使用智能体、课程学习和GUI智能体的研究与开发。数据集包含两个分片:qwen3vl和uitars,分别对应两个基础智能体(Qwen3-VL-8B-Instruct和UI-TARS-1.5-7B)。每个分片包含总计4,608个自然语言任务,这些任务由GPT-5生成,涵盖六个环境:libreoffice_impress、libreoffice_calc、libreoffice_writer、thunderbird、Celestia和KAlgebra。每个环境的数据进一步组织为三个课程迭代(iteration 1、2、3),每个迭代包含256个任务,因此每个数据源总计768个任务。数据集以表格形式存储,包含三个字段:data source(表示任务所属的环境)、iteration(表示课程迭代编号)和task(表示生成的特定自然语言任务描述)。数据规模属于1K到10K之间,适用于训练和评估计算机使用智能体在不同软件环境中的适应性。
The ACuRL Curriculum Tasks dataset comprises curriculum tasks generated by the ACuRL (Autonomous Curriculum Reinforcement Learning) framework, which aims to enable computers to use intelligent agents to continuously adapt to target environments with zero human data. This dataset is specifically designed for text generation and reinforcement learning tasks, and is particularly suitable for research and development of computer-using agents, curriculum learning, and GUI agents. The dataset contains two splits: qwen3vl and uitars, corresponding to two base agents, Qwen3-VL-8B-Instruct and UI-TARS-1.5-7B, respectively. Each split contains a total of 4,608 natural language tasks generated by GPT-5, covering six environments: libreoffice_impress, libreoffice_calc, libreoffice_writer, thunderbird, Celestia, and KAlgebra. The data for each environment is further structured into three curriculum iterations (Iteration 1, 2, and 3), with 256 tasks per iteration, resulting in a total of 768 tasks per data source. The dataset is stored in tabular format, including three fields: "data source" (denoting the environment to which the task belongs), "iteration" (denoting the curriculum iteration number), and "task" (denoting the generated specific natural language task description). The dataset has a scale ranging from 1K to 10K, and is suitable for training and evaluating the adaptability of computer-using agents across various software environments.
数据集概述:ACuRL Curriculum Tasks
- 数据集名称:ACuRL Curriculum Tasks
- 所属组织:osunlp
- 语言:英语(en)
- 许可证:其他(other)
- 任务类别:文本生成(text-generation)、强化学习(reinforcement-learning)
- 标签:computer-use-agents、curriculum-learning、reinforcement-learning、gui-agents、acurl
- 数据集大小:1K < n < 10K
数据集描述
该数据集包含了由 ACuRL(Autonomous Curriculum Reinforcement Learning)框架生成的课程任务。ACuRL 是一个自主课程强化学习框架,旨在持续将计算机使用代理适应到目标环境中,无需任何人工数据。
数据集划分
数据集包含两个划分(splits),分别对应两个基础代理:
| 划分名称 | 对应代理 |
|---|---|
qwen3vl |
Qwen3-VL-8B-Instruct |
uitars |
UI-TARS-1.5-7B |
两个划分中的课程任务均由 GPT-5 生成。每个划分包含 4,608 个任务。对于每个划分,每个数据源包含 768 个任务,组织为 3 个课程迭代,每个迭代 256 个任务。
数据列
| 列名 | 描述 |
|---|---|
data source |
环境名称:libreoffice_impress、libreoffice_calc、libreoffice_writer、thunderbird、Celestia 或 KAlgebra |
iteration |
课程迭代次数,取值为 1、2 或 3 |
task |
生成的任务描述(自然语言) |
数据加载示例
- 使用
datasets库加载:
python from datasets import load_dataset
ds = load_dataset( "parquet", data_files={ "qwen3vl": "qwen3vl/impress_calc_writer_thunderbird_celestia_kalgebra_iter1_3_tasks.parquet", "uitars": "uitars/impress_calc_writer_thunderbird_celestia_kalgebra_iter1_3_tasks.parquet", }, ) print(ds["qwen3vl"][0])
- 直接读取 Parquet 文件:
python import pandas as pd
df = pd.read_parquet("qwen3vl/impress_calc_writer_thunderbird_celestia_kalgebra_iter1_3_tasks.parquet") print(df.groupby(["data source", "iteration"]).size())
相关资源
- 论文:https://arxiv.org/abs/2602.10356
- GitHub 仓库:https://github.com/OSU-NLP-Group/ACuRL
- 模型集合:https://huggingface.co/collections/osunlp/acurl
引用
如果该数据集或相关资源对您的研究有帮助,请引用 ACuRL(引用格式见 README 中的 BibTeX 条目)。




