hzfokyn/tck-qa-benchmark-v2

A 150-question benchmark for evaluating RAG systems on the Turkish Penal Code (Türk Ceza Kanunu — TCK).

Questions were designed to cover three reasoning difficulty levels: factual single-hop lookups, multi-article multi-hop reasoning, and open-ended legal reasoning.

Dataset Details

Each example contains a question, a reference answer, relevant article numbers, the question type, the gold-standard context, and the full TCK article text(s).

Field	Type	Description
soru	string	Question in Turkish
cevap	string	Reference answer with article citations (e.g. [Madde 250])
soru_tipi	string	Question type: single-hop, multi-hop, or reasoning
relevant_madde_list	string	JSON-encoded list of relevant article numbers (e.g. "[250]")
context	string	Gold-standard context excerpt from the TCK with section breadcrumbs
tck_madde_metni	string	Full text of the relevant TCK article(s)
kaynak	string	Source — always Türk Ceza Kanunu

Question Type Distribution

Type	Count	Description
single-hop	60	Single article lookup
multi-hop	55	Requires connecting multiple articles
reasoning	35	Open-ended legal reasoning
Total	150

Usage

from datasets import load_dataset

ds = load_dataset("hzfokyn/tck-qa-benchmark-v2", split="train")
print(ds[0])

# Filter by question type
single_hop = ds.filter(lambda x: x["soru_tipi"] == "single-hop")

Citation

If you use this dataset, please cite:

@misc{tck-evaluation-2026,
  author = {Huzeyfe Okuyan},
  title  = {TCK Evaluation Dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/hzfokyn/tck-qa-benchmark-v2}
}