duyet/vietnamese-legal-instruct

---
language:
- vi
license: cc-by-4.0
task_categories:
- text-generation
- question-answering
tags:
- vietnamese
- legal
- instruction-tuning
- unsloth
- fine-tuning
- law
size_categories:
- 100K<n<1M
---

# Vietnamese Legal Instruction Dataset

**Dataset**: [huggingface.co/datasets/duyet/vietnamese-legal-instruct](https://huggingface.co/datasets/duyet/vietnamese-legal-instruct) | **Source code**: [github.com/duyet/vietnamese-legal-documents-dataset](https://github.com/duyet/vietnamese-legal-documents-dataset)

Instruction-following dataset built from [th1nhng0/vietnamese-legal-documents](https://huggingface.co/datasets/th1nhng0/vietnamese-legal-documents) — 127K Vietnamese legal documents from [vbpl.vn](https://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`):

```json
{
"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

```python
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](https://github.com/duyet/vietnamese-legal-documents-dataset) | Source code: [`generate.py`](https://huggingface.co/datasets/duyet/vietnamese-legal-instruct/blob/main/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)

```bash
# 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](https://huggingface.co/datasets/th1nhng0/vietnamese-legal-documents) by Thịnh Ngô
- **Data source**: [vbpl.vn](https://vbpl.vn) — Vietnamese Ministry of Justice
- **License**: CC BY 4.0

## Citation

```bibtex
@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}
}
```

---
language:
- 越南语（Vietnamese）
license: CC BY 4.0
task_categories:
- 文本生成（text-generation）
- 问答（question-answering）
tags:
- 越南语（Vietnamese）
- 法律（legal）
- 指令微调（instruction-tuning）
- Unsloth
- 微调（fine-tuning）
- 法律（law）
size_categories:
- 10万<样本数<100万
---

# 越南语法律指令数据集（Vietnamese Legal Instruction Dataset）

**数据集**: [https://huggingface.co/datasets/duyet/vietnamese-legal-instruct] | **源代码**: [https://github.com/duyet/vietnamese-legal-documents-dataset]

本指令遵循数据集基于[th1nhng0/vietnamese-legal-documents](https://huggingface.co/datasets/th1nhng0/vietnamese-legal-documents)构建，原始数据集包含来自[越南司法部政府法律文档门户vbpl.vn](https://vbpl.vn)的12.7万份越南法律文档。

本数据集包含467732条训练样本对，覆盖14类问答任务，蕴含深度的越南法律层级体系知识。每份文档均配有用于内容召回的`full_text`（全文）样本对，以及5类用于记忆性检索的短元数据问答样本。

## 统计信息

| 指标 | 数值 |
|--------|-------|
| 总记录数 | 467,732 |
| 源文档数 | 116,933（从127,271份唯一文档中按长度筛选得到） |
| 问答类型数 | 14 |
| 训练集划分 | 444,346（占比95%） |
| 测试集划分 | 23,386（占比5%） |

## 问答类型分布

| 问答类型 | 样本数 | 占比 | 描述 |
|------|------:|--:|-------------|
| `full_text` | 116,933 | 25.0 | 全文档内容（每份文档对应1条，用于内容召回） |
| `scope` | 35,715 | 7.6 | 适用范围、生效日期 |
| `classify` | 35,401 | 7.6 | 文档类型与层级定位 |
| `summarize` | 35,226 | 7.5 | 3-5句结构化摘要 |
| `meta_date` | 29,247 | 6.3 | 发布日期与生效日期（短格式） |
| `explain_simple` | 29,176 | 6.2 | 面向非法律从业者的通俗解释 |
| `meta_issuer` | 29,059 | 6.2 | 发布机构（短格式） |
| `meta_title` | 29,049 | 6.2 | 标题与主题（短格式） |
| `meta_type` | 29,013 | 6.2 | 文档类型与层级级别（短格式） |
| `qa_practical` | 28,982 | 6.2 | 实操合规类问答 |
| `meta_status` | 22,201 | 4.7 | 当前法律状态（短格式） |
| `key_provisions` | 22,061 | 4.7 | 核心条款与条文 |
| `legal_basis` | 17,867 | 3.8 | 法律依据链条分析 |
| `amounts` | 7,802 | 1.7 | 金额与百分比数值 |

## 文档类型分布（前10名）

| 文档类型 | 样本数 |
|---------------|------:|
| Quyết định（决定/Decisions） | 137,524 |
| Nghị quyết（决议/Resolutions） | 40,404 |
| Thông tư（通知/Circulars） | 23,544 |
| Chỉ thị（指令/Directives） | 16,198 |
| Nghị định（政令/Decrees） | 6,610 |
| Thông tư liên tịch（联合通知/Joint Circulars） | 4,834 |
| Sắc lệnh（法令/Ordinances） | 1,950 |
| Nghị Quyết（决议） | 864 |
| Lệnh（命令/Orders） | 750 |
| Pháp lệnh（法令/Ordinances） | 404 |

## 数据格式

适配Unsloth的对话格式（包含`conversations`列，内含`role`（角色）与`content`（内容）字段）：

json
{
"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"
}

## 越南法律层级体系

本数据集编码了越南法律文档层级体系知识（依据2015年《法律文件颁布法》（Luật ban hành VBQPPL 2015））：

1. 宪法（Hiến pháp / Constitution）
2. 法律、法典（Luật, Bộ luật / Laws, Codes）—— 由国会制定
3. 法令、命令（Pháp lệnh, Lệnh / Ordinances, Orders）—— 由国会常务委员会/国家主席发布
4. 政令、决议（Nghị định, Nghị quyết / Decrees, Resolutions）—— 由政府发布
5. 通知、决定（Thông tư, Quyết định / Circulars, Decisions）—— 由部长/人民委员会发布
6. 指令（Chỉ thị / Directives）—— 由总理/人民委员会主席发布

## 快速上手Unsloth

python
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()

## 生成流程

GitHub仓库：[duyet/vietnamese-legal-documents-dataset](https://github.com/duyet/vietnamese-legal-documents-dataset) | 源代码：[`generate.py`](https://huggingface.co/datasets/duyet/vietnamese-legal-instruct/blob/main/generate.py)

本数据集通过14种本地问答生成器构建，无需调用大语言模型API：
- 9类分析型任务：摘要、核心条款、实操合规问答、通俗解释、适用范围、文档分类、法律依据分析、金额数值提取、全文内容召回
- 5类短元数据检索任务：文档类型、发布机构、发布/生效日期、标题主题、法律状态
- 系统提示中嵌入了越南法律层级体系知识
- 经过质量筛选：响应内容不少于60字符，无文档内重复样本
- 每份文档生成1条`full_text`样本与3条随机问答样本，即每份文档对应4条数据记录
- 使用DuckDB作为缓存，实现内存高效处理（常驻内存约145MB RSS）

bash
# 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

## 数据来源

- **原始数据集**: [th1nhng0/vietnamese-legal-documents](https://huggingface.co/datasets/th1nhng0/vietnamese-legal-documents) by Thịnh Ngô
- **数据来源**: [vbpl.vn](https://vbpl.vn) — 越南司法部
- **许可证**: CC BY 4.0

## 引用格式

bibtex
@dataset{vietnamese_legal_instruct_2026,
title = {越南语法律指令数据集（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 = {包含来自12.7万份越南法律文档的46.8万条指令样本对，覆盖14类问答任务}
}