Legal Document Simplification Dataset (LegalSim-PT)

This repository contains code and resources for building and evaluating a large-scale dataset for legal document simplification in Portuguese.

GitHub: https://github.com/scalercio/legal-doc-simplification-data

Hugging Face Dataset: https://huggingface.co/datasets/melll-uff/legal-simplification-pt

---

Overview

This project introduces LegalSim-PT, a large-scale dataset for automatic simplification of legal documents, containing nearly 1 million document pairs.

The repository includes:

• Code for generating simplified legal documents using LLMs
• Scripts for fine-tuning models
• Tools for linguistic annotation and evaluation
• Dataset organization and statistics

---

Dataset

The dataset is publicly available on Hugging Face under:

melll-uff/legal-simplification-pt

It is released in two configurations:

1. random_split

Standard machine learning split:

• train
• validation
• test

2. by_source

Grouped by document origin:

• TCU rulings (acórdãos)
• STF decisions
• STF votes and reports
• TJ-SP jurisprudence
• TRF-5 jurisprudence
• Ulysses Tesemõ dataset (subset)

---

How to Download the Dataset

Option 1 — Using Hugging Face datasets

from datasets import load_dataset

# Random split version
dataset = load_dataset("melll-uff/legal-simplification-pt", "random_split")

# By-source version
dataset_by_source = load_dataset("melll-uff/legal-simplification-pt", "by_source")

Option 2 — Download manually

You can also download the files directly from the dataset page:

https://huggingface.co/datasets/melll-uff/legal-simplification-pt

---

Local Data Organization

The data used in this repository should be placed under:

data/legal-simplification-pt/
├── random_split/
│   ├── train_random.parquet
│   ├── val_random.parquet
│   └── test_random.parquet
└── by_source/
    ├── acordaos_tcu.parquet
    ├── stf_decisions.parquet
    ├── stf_votes.parquet
    ├── tjsp.parquet
    ├── trf5.parquet
    └── ulysses_tesemo.parquet

---

Code Structure

The main scripts are located in the src/ directory.

Data Generation

• src/gen_paraphrases.py Generates simplified legal documents using LLMs.

• src/gen_paraphrases_gguf.py Generates simplified legal documents using GGUF-based LLMs.

Model Fine-tuning

• src/finetune_qwen3.py Fine-tunes the Qwen3-1.7B model.

• src/finetune_chat.py Fine-tunes the Qwen2.5-7B model.

Linguistic Annotation

• src/annotate_parquet.py Performs morphosyntactic annotation of documents.

Inference

• src/eval_qwen2_5.py Runs inference with Qwen2.5-7B.

• src/eval_qwen3.py Runs inference with Qwen3-1.7B.

• src/generate_bode.py Runs inference with Bode (recogna-nlp/bode-7b-alpaca-pt-br).

• src/generate_tucano.py Runs inference with Tucano (TucanoBR/Tucano-2b4-Instruct).

LLM-based Evaluation

• src/groq-judge.py Used for evaluation with an LLM-as-a-judge approach.

Statistics

• statistics.ipynb Notebook used to generate dataset statistics and analyses.

---

Research Focus

This project explores:

• Document-level simplification
• The Portuguese legal domain, a relatively low-resource setting
• Trade-offs between readability and content preservation
• The use of LLMs for large-scale data generation
• Linguistic evaluation using morphosyntactic metrics

---

Citation

If you use this dataset or repository, please cite:

@inproceedings{scalercio2026legalsim,
  title={LegalSim-PT: Building a Dataset for Legal Document Simplification in Portuguese Leveraging Linguistic Metrics},
  author={Scalercio, Arthur and others},
  booktitle={Proceedings of the International Conference on Computational Processing of the Portuguese Language (PROPOR)},
  year={2026}
}

---

Notes

• The dataset contains automatically generated simplifications filtered using linguistic metrics.
• It is intended for research in NLP, legal text processing, and text simplification.

---

Contact

Arthur Scalercio [email protected]