duyet/vietnamese-legal-instruct

Vietnamese Legal Instruction Dataset

Dataset: huggingface.co/datasets/duyet/vietnamese-legal-instruct | Source code: github.com/duyet/vietnamese-legal-documents-dataset

Instruction-following dataset built from th1nhng0/vietnamese-legal-documents — 127K Vietnamese legal documents from vbpl.vn (Government Legal Document Portal, Ministry of Justice).

467,732 training pairs across 14 QA types with deep Vietnamese legal hierarchy knowledge. Every document has a full_text pair for content recall, plus 5 short metadata recall types for memorization.

Statistics

Metric	Value
Total records	467,732
Source documents	116,933 (from 127,271 unique, filtered by length)
QA types	14
Train split	444,346 (95%)
Test split	23,386 (5%)

QA Type Distribution

Type	Count	%	Description
full_text	116,933	25.0	Full document content (1 per doc for content recall)
scope	35,715	7.6	Scope, applicability, effective dates
classify	35,401	7.6	Document type & hierarchy position
summarize	35,226	7.5	3-5 sentence structured summary
meta_date	29,247	6.3	Issue date & effective date (short)
explain_simple	29,176	6.2	Plain language for non-lawyers
meta_issuer	29,059	6.2	Issuing authority (short)
meta_title	29,049	6.2	Title & subject (short)
meta_type	29,013	6.2	Document type & hierarchy level (short)
qa_practical	28,982	6.2	Practical compliance Q&A
meta_status	22,201	4.7	Current legal status (short)
key_provisions	22,061	4.7	Key articles and provisions
legal_basis	17,867	3.8	Legal basis chain analysis
amounts	7,802	1.7	Monetary amounts & percentages

Document Type Distribution (top 10)

Document Type	Count
Quyết định (Decisions)	137,524
Nghị quyết (Resolutions)	40,404
Thông tư (Circulars)	23,544
Chỉ thị (Directives)	16,198
Nghị định (Decrees)	6,610
Thông tư liên tịch (Joint Circulars)	4,834
Sắc lệnh (Ordinances)	1,950
Nghị Quyết	864
Lệnh (Orders)	750
Pháp lệnh (Ordinances)	404

Format

Unsloth-compatible conversation format (conversations column with role/content):

{
  "conversations": [
    {"role": "system", "content": "Bạn là chuyên gia pháp luật Việt Nam..."},
    {"role": "user", "content": "Tóm tắt văn bản sau..."},
    {"role": "assistant", "content": "Văn bản này quy định về..."}
  ],
  "source_id": "12345",
  "document_type": "Nghị định",
  "qa_type": "summarize"
}

Vietnamese Legal Hierarchy

The dataset encodes knowledge of the Vietnamese legal document hierarchy (per Luật ban hành VBQPPL 2015):

1. Hiến pháp (Constitution)
2. Luật, Bộ luật (Laws, Codes) — Quốc hội
3. Pháp lệnh, Lệnh (Ordinances, Orders) — UBTVQH / Chủ tịch nước
4. Nghị định, Nghị quyết (Decrees, Resolutions) — Chính phủ
5. Thông tư, Quyết định (Circulars, Decisions) — Bộ trưởng / UBND
6. Chỉ thị (Directives) — Thủ tướng / Chủ tịch UBND

Quick Start with Unsloth

from unsloth import FastModel
from unsloth.chat_templates import get_chat_template
from datasets import load_dataset

model, tokenizer = FastModel.from_pretrained(
    model_name="unsloth/gemma-3-4B-it",
    max_seq_length=4096,
    load_in_4bit=True,
)
tokenizer = get_chat_template(tokenizer, chat_template="gemma")

dataset = load_dataset("duyet/vietnamese-legal-instruct", split="train")

def formatting_prompts_func(examples):
    texts = [
        tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=False)
        for convo in examples["conversations"]
    ]
    return {"text": texts}

dataset = dataset.map(formatting_prompts_func, batched=True)

from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model=model, tokenizer=tokenizer, train_dataset=dataset,
    args=SFTConfig(
        per_device_train_batch_size=2, learning_rate=2e-4,
        packing=True,  # 2-5x speedup for mixed-length data
    ),
)
trainer.train()

Generation

GitHub: duyet/vietnamese-legal-documents-dataset | Source code: generate.py

Built with 14 local QA generators (no LLM API calls needed):

• 9 analysis types: summarize, key_provisions, qa_practical, explain_simple, scope, classify, legal_basis, amounts, full_text
• 5 short metadata recall types: meta_type, meta_issuer, meta_date, meta_title, meta_status
• Vietnamese legal hierarchy knowledge baked into system prompts
• Quality-filtered: min 60 char responses, no within-doc duplicates
• Every doc gets 1 full_text + 3 random QA types = 4 records per doc
• DuckDB-backed cache for memory-efficient processing (~145 MB RSS)

# Reproduce the dataset
pip install requests python-dotenv beautifulsoup4 lxml pyarrow datasets duckdb
python generate.py --fresh --qa-types 3 --upload duyet/vietnamese-legal-instruct

Source

• Original dataset: th1nhng0/vietnamese-legal-documents by Thịnh Ngô
• Data source: vbpl.vn — Vietnamese Ministry of Justice
• License: CC BY 4.0

Citation

@dataset{vietnamese_legal_instruct_2026,
  title     = {Vietnamese Legal Instruction Dataset},
  author    = {Duyet Le},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/8343},
  url       = {https://huggingface.co/datasets/duyet/vietnamese-legal-instruct},
  note      = {468K instruction pairs from 127K Vietnamese legal documents, 14 QA types}
}