LavanyaPobbathi/lamus-scotus-legal-arguments

LAMUS: Legal Argument Mining from U.S. Supreme Court

<div align="center">

  

</div>

📋 Dataset Description

This dataset contains 2,900,083 sentences from U.S. Supreme Court opinions spanning 1921-2025, automatically labeled with legal argument categories. This is the largest publicly available labeled dataset for legal argument mining from U.S. caselaw.

🎯 Purpose

The dataset enables:

• —Legal Argument Mining research
• —Legal Text Classification model training
• —Temporal Analysis of judicial writing styles
• —Cross-court Comparison studies

📊 Dataset Statistics

🏛️ Supreme Court Eras Covered

🏷️ Label Categories

📁 Dataset Structure

Data Fields

Example Entry

{
  "row_id": 0,
  "sentence": "The defendant was convicted of first-degree murder.",
  "case_title": "Smith v. United States",
  "citation": "500 U.S. 123",
  "court": "Rehnquist Court",
  "year": 1991,
  "Predicted_Label": "Facts"
}

🚀 Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("LavanyaPobbathi/lamus-scotus-legal-arguments")

# Access the data
scotus_data = dataset["scotus_all_courts"]
print(f"Total sentences: {len(scotus_data):,}")

# Example: Filter by court era
burger_court = [s for s in scotus_data if s["court"] == "Burger Court"]

Loading as Pandas DataFrame

import pandas as pd
from datasets import load_dataset

dataset = load_dataset("LavanyaPobbathi/lamus-scotus-legal-arguments")
df = dataset["scotus_all_courts"].to_pandas()

# Analyze label distribution
print(df["Predicted_Label"].value_counts())

Filter by Court Era

# Get only Roberts Court (2005-present)
roberts = df[df["court"] == "Roberts Court"]
print(f"Roberts Court sentences: {len(roberts):,}")

🔬 Methodology

Labeling Model

The sentences were labeled using a fine-tuned Llama-3-8B model with the following specifications:

Training Data

The model was trained on 2,585 manually annotated sentences from Texas criminal court cases and validated on 647 test sentences.

Validation

• —Stratified train/test split (80/20)
• —6-class classification task
• —Macro F1: 0.69, Weighted F1: 0.80

📈 Key Research Findings

1. 1.Fine-tuning dramatically outperforms prompting (+9.27% accuracy)
2. 2.General-domain LLMs outperform legal-specific models (surprising finding)
3. 3.Few-shot prompting decreases accuracy (important negative result)
4. 4.Significant domain shift between trial courts (Facts-heavy) and SCOTUS (Rule/Law-heavy)

📚 Related Resources

• —Training Data: Texas Criminal Cases (available upon request)
• —Model: Fine-tuned Llama-3-8B (available upon request)
• —Paper: [Coming Soon - ICAIL 2026]

📖 Citation

If you use this dataset in your research, please cite:

@dataset{lamus2026,
  title={LAMUS: Legal Argument Mining from U.S. Supreme Court Using Large Language Models},
  author={Pobbathi, Lavanya and Wang, Serene and Chen, Haihua},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/LavanyaPobbathi/lamus-scotus-legal-arguments}
}

📄 License

This dataset is released under the MIT License.

The underlying Supreme Court opinions are in the public domain as U.S. government works.

👥 Authors

• —Lavanya Pobbathi - University of North Texas
• —Serene Wang - University of North Texas
• —Haihua Chen - University of North Texas (Supervisor)

📧 Contact

For questions or feedback, please open an issue on this dataset repository or contact the authors.