dokukoza/kdv-rag-benchmark

KDV RAG Benchmark

A retrieval benchmark dataset for Turkish VAT (KDV, Katma Değer Vergisi) law — built by adding retrieval layers one at a time (chunking, model choice, hybrid search, reranking, query rewriting, historical/date filtering) and statistically validating each one individually (see Results).

Dataset structure

Splits

Split	Records	Period
train	728	2018-2023
test	154	2024-2026

Split strategy: temporal — train and test come from different time windows, minimizing data leakage risk.

⚠️ In some evaluation runs (see scores below), the train split is folded into eval as well to increase statistical power (154→813 queries) — this is valid only for zero-shot/pretrained model comparisons, since no model here was fine-tuned on train. If you fine-tune your own model on train, evaluate only on the original test split.

Corpus

Provided at four granularities:

Config	Records	Content
corpus	110	Articles of the VAT Law (Law No. 3065), unsplit
corpus_chunked	174	Same articles, long ones split into sub-sections — noticeably better retrieval results
corpus_versioned	88	Real multi-version history for 8 articles, article-level
corpus_versioned_chunked	456	corpus_chunked + the same 8 articles' historical versions also chunked — the corpus used for point-in-time eval

Features

train / test:

id             : int      — GIB ruling ID
siteLink       : string   — source URL (gib.gov.tr)
ozelgeNo       : string   — official ruling number
ozelgeTarih    : string   — publication date (YYYY-MM-DD)
baslik         : string   — ruling title
kanunNo        : string   — 3065
soru           : string   — the taxpayer's question
cevap          : string   — GIB's answer
madde_atiflar  : list     — cited articles [{id, title}]

corpus:

id             : int      — article ID
title          : string   — "Madde 17 Sosyal ve Askeri Amaçlı İstisnalarla..."
metin          : string   — full article text
siteLink       : string   — source URL
priority       : int      — article order
bolum          : string   — section title

corpus_chunked:

chunk_id       : string   — "MADDE_<article_id>_<part>"
article_id     : int      — the article id in corpus (ground truth matches on this)
title          : string   — article title
section        : string?  — sub-section title (if any)
part           : int      — order within the article
metin          : string   — chunk text
metin_len      : int      — character length

corpus_versioned:

madde_id       : int      — the article id in corpus (same space as article_id)
madde_no       : string   — the law's article number ("13")
version        : int      — version order (1 = oldest)
metin          : string   — the full article text for that version
valid_from     : string?  — date this version took effect (null = in force since the law's start)
valid_until    : string?  — date this version ended (null = still in force)
kaynak_edit_key: string?  — the edit record that triggered the transition to this version

corpus_versioned_chunked: corpus_chunked fields + version, valid_from, valid_until (as above).

Results

	recall@10
Best general retrieval (BM25∪Nomic pool + query rewriting + Qwen3-Reranker-8B rerank)	0.737 (the 270 real-subject questions)
Point-in-time¹, date-blind (current-text approach)	0.089
Point-in-time¹, historical data + date filter	0.768

¹ Point-in-time: the ability to answer a question like "was this transaction VAT-exempt in 2019?" with the article version that was actually in force on the event date, not today's text — corpus_versioned/corpus_versioned_chunked exist for this.

The path to these numbers — chunking mattering more than model choice, the MTEB leaderboard not transferring to this task, the risks of drawing conclusions from a small sample, and how much the date filter's position in the pipeline (before vs. after ranking) changes the result — is written up in a separate post: dokukoza.com/blog/measured-legal-rag. Raw experiment records and code: github.com/er3n/dokumevzuat.

Usage

from datasets import load_dataset

# Benchmark
train = load_dataset("dokukoza/kdv-rag-benchmark", "benchmark", split="train")
test  = load_dataset("dokukoza/kdv-rag-benchmark", "benchmark", split="test")

# Corpus (unsplit, 110 articles)
corpus = load_dataset("dokukoza/kdv-rag-benchmark", "corpus", split="train")

# Corpus (chunked, 174 records — better retrieval results)
corpus_chunked = load_dataset("dokukoza/kdv-rag-benchmark", "corpus_chunked", split="train")

# Point-in-time article versions (8 multi-version articles)
corpus_versioned = load_dataset("dokukoza/kdv-rag-benchmark", "corpus_versioned", split="train")

# Point-in-time article versions, chunked — the corpus actually used for eval
corpus_versioned_chunked = load_dataset("dokukoza/kdv-rag-benchmark", "corpus_versioned_chunked", split="train")

Evaluation example

# For each test record:
# query  = record["soru"]
# ground_truth = {a["id"] for a in record["madde_atiflar"]}
# retrieved = retrieval_system(query, corpus, top_k=10)
# recall@10 = len(ground_truth & set(retrieved[:10])) / len(ground_truth)

Data source

• Rulings: Revenue Administration (GİB) — VAT Law ruling database
• Articles: mevzuat.gov.tr — Law No. 3065 (VAT Law)

Public institutions' open data — full text included, not just references and labels.

License

CC BY 4.0 — usable with attribution.

Citation

@misc{kdv_rag_benchmark_2026,
  title        = {KDV RAG Benchmark: A Point-in-Time Turkish VAT Law Retrieval Benchmark},
  author       = {Öztürk, Eren},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/dokukoza/kdv-rag-benchmark}}
}