ChuGyouk/Arguinas

--- size_categories: - 1K<n<10K language: - en configs: - config_name: default data_files: - split: train path: data/train-* - split: test path: data/test-* dataset_info: features: - name: messages list: - name: content dtype: string - name: role dtype: string - name: title dtype: string - name: background dtype: string - name: argument dtype: string - name: fallacy_info struct: - name: rationale dtype: string - name: type dtype: string - name: sections struct: - name: check_faithfulness struct: - name: faithfulness dtype: bool - name: feedback_faithfulness dtype: string - name: check_validity struct: - name: final_formalized_conclusion dtype: string - name: necessary_formalized_premises dtype: string - name: valid_formalized_premises list: string - name: validity dtype: string - name: z3_program dtype: string - name: reconstruction struct: - name: conclusion dtype: string - name: definition dtype: string - name: formalized_conclusion dtype: string - name: formalized_intermediate_conclusions dtype: string - name: formalized_premises dtype: string - name: intermediate_conclusions dtype: string - name: premises dtype: string - name: streamlined struct: - name: valid_conclusion dtype: string - name: valid_premises list: string - name: explicit_premises list: string - name: implicit_premises list: string splits: - name: train num_bytes: 61877399 num_examples: 2934 - name: test num_bytes: 4997520 num_examples: 241 download_size: 24779661 dataset_size: 66874919 --- # Argument Reconstruction as Supervision for Critical Thinking in LLMs [](https://arxiv.org/abs/2603.17432) [](#bibtex) **Arguinas** (**Argu**ment reconstruct**i**o**n**) dat**a**set as presented in our paper: [**Argument Reconstruction as Supervision for Critical Thinking in LLMs**](https://arxiv.org/abs/2603.17432) by Hyun Ryu^\*1,2, Gyouk Chu^\*2, Gregor Betz³, Eunho Yang², Carolyn Rosé^†1, and Sean Welleck^†1 ¹Language Technologies Institute, Carnegie Mellon University &nbsp;&nbsp; ²Graduate School of AI, Korea Advanced Institute of Science & Technology &nbsp;&nbsp; ³Department of Philosophy, Karlsruhe Institute of Technology &nbsp;&nbsp; ^\*Equal Contribution &nbsp;&nbsp; ^†Equal Advising <p align="center"> <img src="gaar_engine_github.png" alt="GAAR" width="80%" /> </p> --- ## 🔔 Updates - [✔] (25.04.21) The Arguinas dataset are out. - [✔] (26.03.18) Paper is out! [here](https://arxiv.org/abs/2603.17432) --- ## 🏋️ Data Our train and test Arguinas datasets are open! **See [`data/README.md`](https://github.com/GyoukChu/Arguinas/blob/main/data/README.md) for the full data format (top-level columns, `fallacy_info`, `sections`, etc.).** The source dataset of arguments are [DebateLabKIT/arguments-and-debates](https://huggingface.co/datasets/DebateLabKIT/arguments-and-debates), [Anthropic/persuasion](https://huggingface.co/datasets/Anthropic/persuasion), and [webis/args_me](https://huggingface.co/datasets/webis/args_me). ### Statistics | Sources | # Data | Avg. Words in Argument | Author of Argument | # Premises | % Implicit Premises | |----------------------------------|--------|-------------------------|--------------------|------------|---------------------| | Procon.org | 282 | 178.01 ± 100.68 | Staff Editors | 6.46 ± 2.95| 45.39 ± 19.39 | | Pros-and-cons-1950 | 119 | 43.41 ± 8.98 | Educators | 5.59 ± 1.82| 47.41 ± 18.08 | | Pros-and-cons-2010 | 373 | 83.53 ± 25.81 | Educators | 5.79 ± 2.20| 51.99 ± 18.13 | | NYT-room-for-debate | 297 | 398.21 ± 92.02 | Journalists | 8.13 ± 3.39| 41.68 ± 15.54 | | Anthropic/Persuasion | 287 | 252.33 ± 37.34 | Human / Claude | 8.47 ± 3.15| 40.42 ± 17.57 | | Synthetic Arguments | 1,520 | 332.60 ± 190.55 | GPT-5 / GPT-5.1 | 9.14 ± 4.28| 37.88 ± 15.29 | | Synthetic *Fallacious* Arguments | 297 | 296.45 ± 160.88 | GPT-5.2 | 8.40 ± 4.28| 35.70 ± 19.11 | | **Total** | 3,175 | 269.48 ± 177.21 | - | 8.15 ± 3.94| 40.94 ± 17.46 | <a id="bibtex"></a> ## 📚 BibTeX If you find this repo useful for your research, please consider citing us: ``` @article{ryu2026argument, title={Argument Reconstruction as Supervision for Critical Thinking in LLMs}, author={Ryu, Hyun and Chu, Gyouk and Betz, Gregor and Yang, Eunho and Rose, Carolyn and Welleck, Sean}, journal={arXiv preprint arXiv:2603.17432}, year={2026} } ``` ## ✉️ Contact If you have any questions or feedback, feel free to reach out: - Hyun Ryu: [e-mail] - Gyouk Chu: [e-mail]_categories: - 1000 < 样本数 < 10000 language: - 英语 configs: - config_name: 默认配置 data_files: - split: 训练集（train） path: data/train-* - split: 测试集（test） path: data/test-* dataset_info: features: - name: messages 类型: 列表，包含子字段： - name: content 数据类型: 字符串 - name: role 数据类型: 字符串 - name: title 数据类型: 字符串 - name: background 数据类型: 字符串 - name: argument 数据类型: 字符串 - name: fallacy_info 类型: 结构体，包含子字段： - name: rationale 数据类型: 字符串 - name: type 数据类型: 字符串 - name: sections 类型: 结构体，包含子字段： - name: check_faithfulness 类型: 结构体，包含子字段： - name: faithfulness 数据类型: 布尔值 - name: feedback_faithfulness 数据类型: 字符串 - name: check_validity 类型: 结构体，包含子字段： - name: final_formalized_conclusion 数据类型: 字符串 - name: necessary_formalized_premises 数据类型: 字符串 - name: valid_formalized_premises 类型: 字符串列表 - name: validity 数据类型: 字符串 - name: z3_program 数据类型: Z3程序（Z3） - name: reconstruction 类型: 结构体，包含子字段： - name: conclusion 数据类型: 字符串 - name: definition 数据类型: 字符串 - name: formalized_conclusion 数据类型: 字符串 - name: formalized_intermediate_conclusions 数据类型: 字符串 - name: formalized_premises 数据类型: 字符串 - name: intermediate_conclusions 数据类型: 字符串 - name: premises 数据类型: 字符串 - name: streamlined 类型: 结构体，包含子字段： - name: valid_conclusion 数据类型: 字符串 - name: valid_premises 类型: 字符串列表 - name: explicit_premises 类型: 字符串列表 - name: implicit_premises 类型: 字符串列表 splits: - name: 训练集（train） num_bytes: 61877399 num_examples: 2934 - name: 测试集（test） num_bytes: 4997520 num_examples: 241 download_size: 24779661 dataset_size: 66874919 --- # 以论证重构作为大语言模型（LLM/Large Language Model）批判性思维的监督信号 [](https://arxiv.org/abs/2603.17432) [](#bibtex) **Arguinas**（论证重构数据集，全称Argu**ment reconstruct**i**o**n）数据集，源自我们的论文： [**Argument Reconstruction as Supervision for Critical Thinking in LLMs**](https://arxiv.org/abs/2603.17432) 作者：Hyun Ryu^*1,2、Gyouk Chu^*2、Gregor Betz³、Eunho Yang²、Carolyn Rosé^†1、Sean Welleck^†1 ¹卡内基梅隆大学语言技术研究所 &nbsp;&nbsp; ²韩国科学技术院人工智能研究生院 &nbsp;&nbsp; ³卡尔斯鲁厄理工学院哲学系 &nbsp;&nbsp; ^*共同第一作者 &nbsp;&nbsp; ^†共同通讯作者 <p align="center"> <img src="gaar_engine_github.png" alt="GAAR" width="80%" /> </p> --- ## 🔔 最新动态 - [✔] (2025.04.21) Arguinas数据集正式发布。 - [✔] (2026.03.18) 论文正式上线！[点击查看](https://arxiv.org/abs/2603.17432) --- ## 🏋️ 数据集详情 我们的Arguinas训练集与测试集已开源！**完整数据格式（包括顶层字段、`fallacy_info`、`sections`等）请参阅[`data/README.md`](https://github.com/GyoukChu/Arguinas/blob/main/data/README.md)。** 本数据集的论证文本来源为： [DebateLabKIT/arguments-and-debates](https://huggingface.co/datasets/DebateLabKIT/arguments-and-debates)、[Anthropic/persuasion](https://huggingface.co/datasets/Anthropic/persuasion) 以及 [webis/args_me](https://huggingface.co/datasets/webis/args_me)。 ### 统计信息 | 数据源 | 样本量 | 单论证平均词数 ± 标准差 | 论证作者 | 单样本前提数 ± 标准差 | 隐式前提占比 ± 标准差 | |----------------------------------|--------|-------------------------|--------------------|------------|---------------------| | Procon.org | 282 | 178.01 ± 100.68 | 编辑人员 | 6.46 ± 2.95| 45.39 ± 19.39 | | Pros-and-cons-1950 | 119 | 43.41 ± 8.98 | 教育工作者 | 5.59 ± 1.82| 47.41 ± 18.08 | | Pros-and-cons-2010 | 373 | 83.53 ± 25.81 | 教育工作者 | 5.79 ± 2.20| 51.99 ± 18.13 | | NYT-room-for-debate | 297 | 398.21 ± 92.02 | 新闻工作者 | 8.13 ± 3.39| 41.68 ± 15.54 | | Anthropic/Persuasion | 287 | 252.33 ± 37.34 | 人类 / Claude模型 | 8.47 ± 3.15| 40.42 ± 17.57 | | Synthetic Arguments | 1520 | 332.60 ± 190.55 | GPT-5 / GPT-5.1 | 9.14 ± 4.28| 37.88 ± 15.29 | | Synthetic *Fallacious* Arguments | 297 | 296.45 ± 160.88 | GPT-5.2 | 8.40 ± 4.28| 35.70 ± 19.11 | | **总计** | 3175 | 269.48 ± 177.21 | - | 8.15 ± 3.94| 40.94 ± 17.46 | <a id="bibtex"></a> ## 📚 BibTeX 引用 若本数据集对你的研究有所帮助，请引用我们的工作： @article{ryu2026argument, title={Argument Reconstruction as Supervision for Critical Thinking in LLMs}, author={Ryu, Hyun and Chu, Gyouk and Betz, Gregor and Yang, Eunho and Rose, Carolyn and Welleck, Sean}, journal={arXiv preprint arXiv:2603.17432}, year={2026} } ## ✉️ 联系方式 如有任何问题或建议，欢迎联系： - Hyun Ryu: [e-mail] - Gyouk Chu: [e-mail]