Datasets:

ntphuc149
/

ViLegalText

license: apache-2.0
language:
  - vi
tags:
  - legal
  - corpus
  - pre-training
  - legal-documents
  - vilegallm
  - 16GB
  - vietnamese
pretty_name: ViLegalText
size_categories:
  - 100K<n<1M

ViLegalTexts: A 16GB Vietnamese Legal Pre-training Corpus

This repository contains the pre-training corpus used to train ViLegalLM. The corpus was crawled from four publicly available Vietnamese legal repositories. For full details, please refer to the paper: Read paper

📋 Overview

Property	Value
Language	Vietnamese
Domain	Legal
Corpus size	16 GB
Format	`.zip` containing multiple `.txt` files
Sources	4 public Vietnamese legal repositories
License	Apache 2.0

🔧 Data Processing Pipeline

The raw corpus underwent a 4-stage preprocessing pipeline:

Content Extraction → Text Cleaning → Language Identification → Deduplication

For full details of each stage, please refer to the paper.

The final corpus is distributed as .zip files, each containing multiple .txt files (one document per file), with a total uncompressed size of 16 GB.

🚀 Usage

Clone this repository

apt-get install -y git-lfs
git lfs install

git clone https://your_username:YOUR_HF_TOKEN@huggingface.co/datasets/ntphuc149/ViLegalText

Unzip pre-training corpus

import os

os.makedirs("/process", exist_ok=True)
zip_path = f"/ViLegalText/ViLegalText.zip"
print(f"📦 Giải nén: ViLegalText.zip")
os.system(f'unzip -q "{zip_path}" "*.txt" -d /process/')

print("✅ Xong!")

Preview a sample file

with open("/process/doc_0.txt", encoding="utf-8") as f:
    print(f.read())

🤖 Models Trained on This Corpus

This corpus was used to continually pretrain the following models, all publicly available:

Model	Architecture	Parameters	Max Length	HuggingFace
ViLegalBERT	Encoder-only (MLM)	135M	256	ntphuc149/ViLegalBERT
ViLegalQwen2.5-1.5B-Base	Decoder-only (CLM)	1.54B	2,048	ntphuc149/ViLegalQwen2.5-1.5B-Base
ViLegalQwen3-1.7B-Base	Decoder-only (CLM)	1.72B	4,096	ntphuc149/ViLegalQwen3-1.7B-Base

📄 Citation

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

@inproceedings{nguyen-etal-2026-vilegallm,
    title = "{V}i{L}egal{LM}: Language Models for {V}ietnamese Legal Text",
    author = "Nguyen, Truong-Phuc  and
      Nguyen, Quy-Nhan  and
      Nguyen, Minh-Tien",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1801/",
    pages = "36136--36150",
    ISBN = "979-8-89176-395-1",
    abstract = "We present **ViLegalLM**, comprising **ViLegalBERT** and **ViLegalQwen**, the first suite of Vietnamese pretrained language models for legal text understanding and generation. It includes one encoder-only model (ViLegalBERT, 135M parameters) and two decoder-only models (ViLegalQwen2.5-1.5B-Base and ViLegalQwen3-1.7B-Base), all continually pretrained on a newly curated 16GB Vietnamese legal corpus, significantly larger than previous work. To mitigate data scarcity, we construct three synthetic datasets using LLM-based generation and hard negative mining for True/False QA, Multiple Choice QA, and Natural Language Inference. We establish state-of-the-art results among open-source models on four main Vietnamese legal downstream tasks spanning ten benchmarks, demonstrating that continual pretraining from base models consistently outperforms instruction-tuned adaptation. Source codes, corpus, datasets, and model checkpoints are publicly available at https://github.com/ntphuc149/ViLegalLM."
}

⚠️ Intended Use & Limitations

This corpus is intended for research purposes in Vietnamese legal NLP.
Documents are collected from publicly available Vietnamese legal repositories. No personally identifiable information is included.
Models trained on this corpus are not intended to replace professional legal counsel. Users should not rely solely on model outputs for consequential legal decisions without qualified professional consultation.
The corpus may inherit biases present in the source repositories, including temporal bias, regional variations, and domain coverage imbalances.

🔗 Related Resources

GitHub: ViLegalLM
ViLegalLM Models: ViLegalBERT, ViLegalQwen2.5-1.5B-Base, ViLegalQwen3-1.7B-Base,
Synthetic Datasets: ViLegalTF, ViLegalMCQ, ViLegalNLI