mtntasci/turkish-legal-rag

Turkish Legal RAG Corpus — Türk Hukuku için Açık RAG Datasetı

Tek cümle: 25 önemli Türk kanununun (mevzuat.gov.tr kaynaklı, madde bazlı temiz chunk'lar) + 290 manuel doğrulanmış soru-cevap altın benchmark'ının olduğu açık kaynak Türkçe hukuk RAG datasetı.

🇹🇷 Türkçe Özet — Bu dataset, Türkçe hukuk uygulamaları için sıfırdan üretilmiş açık ve denetlenebilir bir RAG corpus'udur. mevzuat.gov.tr üzerinden alınan 25 ana kanunun madde madde temizlenmiş, chunk'lanmış sürümünü (6.350 satır) ve 290 manuel doğrulanmış soru-cevap çiftinden oluşan altın benchmark setini içerir. Tüm kaynaklar gerçek mevzuat.gov.tr URL'leri ile takip edilebilir; yürürlükten kalkmış (Mülga) maddeler is_active=False etiketiyle işaretlidir.

🇬🇧 English Summary — A clean, article-level chunked corpus of 25 major Turkish statutes sourced directly from mevzuat.gov.tr, plus a 290-item manually verified gold QA benchmark with quality metrics on every row. Every chunk traces back to the original PDF; repealed (Mülga) items are explicitly flagged.

---

📖 Read the Full Story / Detaylı Yazı

The technical decisions, design rationale, and complete usage guide are documented in a Medium article:

Türk Hukuku için Açık Kaynak RAG Dataset'i — Medium

Why article-level chunking matters for Turkish law, how the is_active flag works, how to evaluate retrieval on this benchmark, and what makes this corpus different from closed/synthetic Turkish legal datasets.

---

Why This Dataset? / Neden Bu Dataset?

Most Turkish legal NLP datasets mix raw law text with synthetic QA pairs, making it hard to build a clean retrieval corpus. This dataset solves that:

• Pure retrieval corpus — no synthetic QA contamination, no hallucinated contexts
• Article-level chunking — each row is one madde (article), not a random sliding window
• Two chunking strategies — statute_full for complete articles, statute_window for long articles split into overlapping windows
• Controlled chunk size — average 892 characters, max 4,396 characters per chunk
• Source-verified — every row links back to the original PDF on mevzuat.gov.tr via the url column
• Manually verified QA gold benchmark — 290 hand-checked question/answer pairs with quality metrics on every row
• is_active flag for repealed (Mülga) articles — 24 questions referencing repealed law text are explicitly marked, so users can filter them out when measuring current-law accuracy

🇹🇷 Türkçe: Türkçe hukuk NLP datasetlerinin çoğu kanun metniyle sentetik soru-cevapları karıştırır. Bu dataset şu farkları sunar: yapay QA karışımı yok, her satır bir kanun maddesi, kontrollü chunk boyutu, her satırda kaynak PDF URL'si, ve manuel doğrulanmış 290 altın soru.

---

Example Use Cases / Kullanım Senaryoları

Use Case	Description
Fine-tune Turkish legal retrieval	Train BGE-M3, multilingual-e5, or any sentence-transformers model on Turkish legal text — clean source-grounded passages give a strong contrastive signal.
Benchmark RAG systems	Use qa_benchmark_gold.csv (filter is_active=True, 266 items) to measure Recall@k, MRR, EM, and F1 on a manually verified Turkish legal test set.
Train Turkish legal QA models	Each gold question has a verified context passage and answer text — directly usable for extractive QA fine-tuning.
Evaluate citation accuracy	Every row in the corpus has a url pointing to the official mevzuat.gov.tr PDF — useful for citation-grounded LLM evaluations.
Statute-aware embedding training	retrieval_text column prepends the law name to each chunk; useful for learning cross-law disambiguation.
Educational / research baselines	Drop-in dataset for university courses on legal NLP, RAG architecture, or low-resource retrieval.

🇹🇷 Türkçe kullanım örnekleri: Türkçe embedding modeli fine-tune'u, RAG sistem benchmark'lama, hukuki soru-cevap modeli eğitimi, kaynak (citation) doğruluk testi, üniversite NLP/hukuk-bilişim derslerinde temel dataset olarak.

---

Files in This Repository / Dosyalar

File	Purpose	Rows
statute_chunks.csv	Main retrieval corpus — 25 Turkish statutes chunked at article level	6,350
qa_benchmark_gold.csv	Gold QA benchmark — manually verified question/answer pairs for RAG evaluation	290

---

Dataset Statistics

Retrieval Corpus (statute_chunks.csv)

Property	Value
Total rows	6,350
Laws covered	25
Language	Turkish
Avg. chunk length	892 characters
Min chunk length	29 characters
Max chunk length	4,396 characters
Chunk strategy: article-level	5,765 rows
Chunk strategy: sliding window	585 rows
Source	mevzuat.gov.tr

Gold QA Benchmark (qa_benchmark_gold.csv)

Property	Value
Total questions	290
Laws covered	26
Active questions (is_active=True)	266
Repealed/Mülga questions (is_active=False)	24
Avg. answer_support_overlap	0.90
Mid-sentence contexts	0
Open-ended questions	0
Duplicate questions	0

---

Quick Start

from datasets import load_dataset

# Load the retrieval corpus
corpus = load_dataset("CtnkyaABC/turkish-legal-rag-corpus", "corpus", split="train")
print(corpus[0]["retrieval_text"])   # law name + article text (use this for embedding)
print(corpus[0]["context"])          # article text only
print(corpus[0]["kaynak"])           # law name
print(corpus[0]["madde_no"])         # article number
print(corpus[0]["url"])              # source PDF link

# Load the gold QA benchmark
qa = load_dataset("CtnkyaABC/turkish-legal-rag-corpus", "qa_benchmark", split="test")

# Use only currently-active law questions for evaluation
qa_active = qa.filter(lambda x: x["is_active"] == "True")
print(f"Active QA items: {len(qa_active)}")  # 266

# Each gold item has the question, expected answer, source chunk, and quality metrics
print(qa_active[0]["soru"])                      # question
print(qa_active[0]["cevap"])                     # gold answer
print(qa_active[0]["context"])                   # supporting passage
print(qa_active[0]["kaynak"])                    # law name
print(qa_active[0]["answer_support_overlap"])    # overlap between answer and context (0-1)

---

Column Schemas

statute_chunks.csv

Column	Type	Description
kaynak	string	Full law name in Turkish (e.g. Türk Ceza Kanunu)
kanun_no	string	Official law number (e.g. 5237)
madde_no	string	Article number within the law
context	string	Article text — the passage to retrieve
retrieval_text	string	[law name]\n[article text] — recommended for embedding
context_key	string	Unique identifier: kaynak|madde_no|chunk-N
chunk_strategy	string	statute_full (one article = one chunk) or statute_window (long article split into windows)
url	string	Direct link to the source PDF on mevzuat.gov.tr

Tip: Use retrieval_text for embedding (it prepends the law name, which improves retrieval accuracy for cross-law queries). Use context alone for BM25.

qa_benchmark_gold.csv

Column	Type	Description
row_id	string	Unique question identifier
soru	string	Question (in Turkish, natural-language phrasing)
cevap	string	Gold answer text, manually verified
context	string	Supporting passage from the law (the chunk the answer is derived from)
kaynak	string	Law name
kanun_adi	string	Same as kaynak (kept for compatibility)
madde_no	string	Article number
is_active	string	True if the referenced law/article is currently in force, False for Mülga (repealed) articles
answer_support_overlap	float	Lexical overlap between answer and context (0–1, higher = stronger grounding)
question_context_overlap	float	Overlap between question and context
answer_in_chunk	string	True if the answer text appears verbatim in the context
chunk_strategy	string	Chunking strategy applied to the source article
manual_review_reason	string	Empty
context_starts_mid_sentence	string	Always False — every context starts at a sentence boundary
open_ended_question	string	Always False — every question has a definite answer in the context
score_valid, Score, selection_score	various	Internal quality scores from the curation pipeline

Important: For evaluating performance on currently-valid Turkish law, filter is_active=True (266 items). The 24 is_active=False items reference repealed (Mülga) articles and are retained for completeness but should be excluded from accuracy metrics on production-grade systems.

---

Laws Covered (25 statutes)

The corpus covers the most widely referenced Turkish laws across civil, criminal, commercial, administrative, and constitutional law domains.

Law No	Law Name (Turkish)	English Description	Chunks
657	Devlet Memurları Kanunu	Civil Servants Law	339
1136	Avukatlık Kanunu	Attorney Law	215
1475	İş Kanunu (mülga — sadece 14. madde yürürlükte)	Labor Law 1475 (repealed — only Art. 14 on severance pay active)	3
2547	Yükseköğretim Kanunu	Higher Education Law	150
2576	Bölge İdare Mahkemeleri … Kanunu	Administrative Courts Organization Law	32
2577	İdari Yargılama Usulü Kanunu	Administrative Procedure Law	73
2709	Türkiye Cumhuriyeti Anayasası	Constitution of the Republic of Turkey	192
2893	Türk Bayrağı Kanunu	Turkish Flag Law	12
2942	Kamulaştırma Kanunu	Expropriation Law	54
4721	Türk Medeni Kanunu	Turkish Civil Code	1,030
4857	İş Kanunu	Labor Law (current)	128
4982	Bilgi Edinme Hakkı Kanunu	Right to Information Law	33
5237	Türk Ceza Kanunu	Turkish Penal Code	353
5271	Ceza Muhakemesi Kanunu	Code of Criminal Procedure	359
5275	Ceza ve Güvenlik Tedbirlerinin İnfazı Hakkında Kanun	Law on Execution of Penalties and Security Measures	143
5326	Kabahatler Kanunu	Misdemeanors Law	48
5510	Sosyal Sigortalar ve Genel Sağlık Sigortası Kanunu	Social Insurance and General Health Insurance Law	262
5651	İnternet Ortamında Yapılan Yayınların Düzenlenmesi Hakkında Kanun	Internet Publications Regulation Law	27
6098	Türk Borçlar Kanunu	Turkish Code of Obligations	648
6100	Hukuk Muhakemeleri Kanunu	Code of Civil Procedure	458
6102	Türk Ticaret Kanunu	Turkish Commercial Code	1,545
6284	Ailenin Korunması ve Kadına Karşı Şiddetin Önlenmesine Dair Kanun	Law on Protection of Family and Prevention of Violence Against Women	25
6356	Sendikalar ve Toplu İş Sözleşmesi Kanunu	Trade Unions and Collective Bargaining Law	84
6502	Tüketicinin Korunması Hakkında Kanun	Consumer Protection Law	101
6698	Kişisel Verilerin Korunması Kanunu (KVKK)	Personal Data Protection Law (Turkish GDPR equivalent)	36

---

Source & License / Kaynak ve Lisans

All texts are extracted from PDFs published on mevzuat.gov.tr, the official portal of the Turkish Presidency of Legislative Affairs. Turkish statutory texts are public domain. This dataset is released under Creative Commons Attribution 4.0 (CC BY 4.0).

The 290-item gold QA benchmark was hand-curated by the dataset author using passages drawn from the same official sources.

🇹🇷 Türkçe: Tüm kanun metinleri mevzuat.gov.tr (Cumhurbaşkanlığı Mevzuat Genel Müdürlüğü) üzerinden alınmıştır. Türk mevzuatı kamu malıdır. Bu dataset CC BY 4.0 lisansı altında yayınlanmıştır. 290 maddelik altın benchmark, aynı resmi kaynaklardan yararlanılarak elle hazırlanmıştır.

---

Citation

@dataset{ctnkya2026turkishlegal,
  author    = {CtnkyaABC},
  title     = {Turkish Legal RAG Corpus: An Open Verified Benchmark from Mevzuat.gov.tr},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/CtnkyaABC/turkish-legal-rag-corpus}
}

---

Keywords / Anahtar Kelimeler

Turkish law, Turkish legal NLP, mevzuat, Türkçe hukuk, kanun corpus, retrieval augmented generation, RAG, FAISS, BGE-M3, BM25, hukuk-bilişim, legal QA, Türkçe soru cevap, mevzuat.gov.tr, açık veri, open Turkish dataset, low-resource legal NLP.