Tamir39/vietnam-tax-qa

Vietnamese Tax Law QA + Knowledge Base

A small but clean Vietnamese-language dataset for retrieval-augmented question answering over 5 Vietnamese tax laws. Built as the eval/training corpus for the LawMate project (final-year NLP report).

Contents

knowledge_base/knowledge_base.csv — 600 passages

Pre-chunked legal passages from 5 consolidated Vietnamese tax laws (đã chia theo Chương → Điều → Khoản → điểm).

Column	Type	Description
passage_id	string	passage_000001 … passage_000600 (zero-padded, sequential)
doc_id	string	doc_000001 … doc_000005
title	string	Document title in Vietnamese
passage_text	string	Legal text (86–1752 chars, mean 284)
url	string	Source URL on vanban.chinhphu.vn / chinhphu.vn

Source documents

doc_id	Document	# passages
doc_000001	VBHN 12/VBHN-VPQH 2026 — Luật Thuế Giá trị gia tăng	143
doc_000002	VBHN 103/VBHN-VPQH 2025 — Luật Thuế Thu nhập cá nhân	140
doc_000003	Luật 48/2010/QH12 — Thuế sử dụng đất phi nông nghiệp	58
doc_000004	Luật 66/2025/QH15 — Thuế tiêu thụ đặc biệt	127
doc_000005	Luật 67/2025/QH15 — Thuế thu nhập doanh nghiệp	132
Total		600

qa/train_qa.jsonl — 305 train pairs

{"question": "...", "answer": "...", "passage_id": "passage_000XYZ"}

Distribution (approximate):

• ~70% factoid (single-passage extractive)
• ~15% paraphrase (rewording the gold)
• ~10% reasoning (multi-step over a single passage)
• ~5% tricky / negative (out-of-scope or numerically tight)

qa/test_qa.jsonl — 54 hand-written test pairs

Held out, harder distribution. Designed to not overlap with train_qa in question wording (validated by scripts/prepare_qa.py in the LawMate repo).

Loading

from datasets import load_dataset

# Knowledge base
kb = load_dataset("Tamir39/vietnam-tax-qa", "knowledge_base", split="train")

# QA splits
train = load_dataset("Tamir39/vietnam-tax-qa", "qa", split="train")
test  = load_dataset("Tamir39/vietnam-tax-qa", "qa", split="test")

Use cases

• Vietnamese RAG benchmarking (small enough to iterate fast).
• Legal-domain retriever evaluation: 600 passages give meaningful Recall@k signal.
• Vietnamese SFT for legal QA style (used to train the QLoRA adapter Tamir39/qwen2_5-7b-vietnam-tax-lora).

Construction

• KB: passages extracted from official Vietnamese government legal portals (vanban.chinhphu.vn, chinhphu.vn). Each passage corresponds to one điểm / khoản / short Điều. Length capped so the entire passage fits as one embedding without further chunking.
• QA: drafted by Claude with strict instructions to ground every answer in a specific passage_id from the KB (no outside knowledge). A validator checks schema, that every passage_id exists, no train/test question-text overlap.

License & attribution

• Released under CC-BY-4.0.
• The underlying legal text is Vietnamese law (public-domain by nature in VN); attribution + URLs to the source portal are preserved in the url column.
• If you use this dataset, please cite the LawMate project: https://github.com/tamir39/rag-llm-vietnam-law-advisor

Limitations

• Test set is small (54 questions) → BLEU/ROUGE variance is high.
• Single annotator for QA generation → no inter-annotator agreement signal.
• Five laws only — does not cover the full Vietnamese tax code (e.g., import/export duty, environmental tax).
• Some VAT-related articles (Điều 13–16) were not parsed from the source PDF and are absent from the KB.