🧠 Argument Reasoning Tasks (ART) Dataset

Evaluating natural language argumentative reasoning in large language models.

---

📖 Overview

The Argument Reasoning Tasks (ART) dataset is a large-scale benchmark designed to evaluate the ability of large language models (LLMs) to perform natural language argumentative reasoning.

It contains multiple-choice questions where models must identify missing argument components, given an argument context and reasoning structure.

---

🧩 Argumentation Structures

ART covers 16 task types derived from four core argumentation structures:

1. Serial reasoning – chained inference steps.
2. Linked reasoning – multiple premises jointly supporting a conclusion.
3. Convergent reasoning – independent premises supporting a conclusion.
4. Divergent reasoning – a single premise leading to multiple possible conclusions.

---

📄 Source & Reference

This dataset was introduced in:

Debela Gemechu, Ramon Ruiz-Dolz, Henrike Beyer, and Chris Reed. 2025.
Natural Language Reasoning in Large Language Models: Analysis and Evaluation.
Findings of the Association for Computational Linguistics: ACL 2025, pp. 3717–3741.
Vienna, Austria: Association for Computational Linguistics.
📄 Read the paper | DOI: 10.18653/v1/2025.findings-acl.192

@inproceedings{gemechu-etal-2025-natural,
  title     = {Natural Language Reasoning in Large Language Models: Analysis and Evaluation},
  author    = {Gemechu, Debela and Ruiz-Dolz, Ramon and Beyer, Henrike and Reed, Chris},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  pages     = {3717--3741},
  year      = {2025},
  address   = {Vienna, Austria},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.findings-acl.192/},
  doi       = {10.18653/v1/2025.findings-acl.192}
}

---

📂 Dataset Details

• Hugging Face repo: debela-arg/art
• License: CC BY-NC-SA 4.0 (non-commercial, share alike)
• Languages: English
• Domain: Argumentative reasoning, question answering
• File format: JSON
• Size: ~482 MB
• Splits: Single train split with 88,628 examples

---

🗂 Example JSON Entry

{
  "prompt": "Please answer the following multiple-choice question...",
  "task_type": "1H-C",
  "answer": ["just one of three children returning to school..."],
  "data_source": "qt30"
}

Fields:

• prompt – Question with context and multiple-choice options
• task_type – Argument reasoning task category
• answer – Correct answer(s)
• data_source – Original source corpus

---

📊 Statistics

Attribute	Value
Total examples	88,628
Task types	16
Data sources	MTC, AAEC, CDCP, ACSP, AbstRCT, US2016, QT30

---

⚡ How to Load the Dataset

Install the dependencies:

pip install datasets pandas

Load in Python:

from datasets import load_dataset
import pandas as pd

# Load the train split
dataset = load_dataset("debela-arg/art", split="train")

# Convert to DataFrame
df = pd.DataFrame(dataset)

print("Total examples:", len(df))
print("Available columns:", df.columns.tolist())
print("Task type distribution:")
print(df["task_type"].value_counts())

---

🔍 Suggested Uses

• LLM evaluation – Benchmark reasoning capabilities
• Few-shot prompting – Create reasoning-based examples for instruction tuning
• Error analysis – Identify reasoning failure modes in models

---

📌 Citation

If you use ART in your work, please cite:

@inproceedings{gemechu-etal-2025-natural,
  title     = {Natural Language Reasoning in Large Language Models: Analysis and Evaluation},
  author    = {Gemechu, Debela and Ruiz-Dolz, Ramon and Beyer, Henrike and Reed, Chris},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  pages     = {3717--3741},
  year      = {2025},
  address   = {Vienna, Austria},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.findings-acl.192/},
  doi       = {10.18653/v1/2025.findings-acl.192}
}

---

🛠 Maintainers

• Author: Debela Gemechu, Ramon Ruiz-Dolz, Henrike Beyer and Chris Reed

---