language:
- tr
license: cc-by-4.0
task_categories:
- text-retrieval
- question-answering
pretty_name: Turkish Legal Özelge Corpus
size_categories:
- 10K<n<100K
tags:
- legal
- turkish
- özelge
- tax-law
- corpus
- retrieval
- IR
- information-retrieval
- beir
dataset_info:
- config_name: corpus
features:
- name: _id
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 120864700
num_examples: 23587
download_size: 49244406
dataset_size: 120864700
- config_name: default
features:
- name: query-id
dtype: string
- name: corpus-id
dtype: string
- name: score
dtype: int64
splits:
- name: train
num_bytes: 9147664
num_examples: 120364
download_size: 4844361
dataset_size: 9147664
- config_name: queries
features:
- name: _id
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 47933248
num_examples: 120364
download_size: 21179422
dataset_size: 47933248
configs:
- config_name: corpus
data_files:
- split: train
path: corpus/train-*
- config_name: default
data_files:
- split: train
path: data/train-*
- config_name: queries
data_files:
- split: train
path: queries/train-*
Turkish Legal Özelge Corpus Dataset
📊 Dataset Summary
Turkish Legal Özelge Corpus is a comprehensive Information Retrieval dataset consisting of özelge (tax ruling) decisions published by the Turkish Revenue Administration (Gelir İdaresi Başkanlığı - GİB).
Key Features
- Format: BEIR (Benchmarking IR) format with corpus-queries-qrels structure
- Language: Turkish 🇹🇷
- Domain: Tax Law, Administrative Law, Turkish Law
- Source: GİB Özelge Decisions
- Use Cases: Information retrieval, question-answering systems, RAG applications
Dataset Structure
The dataset follows the BEIR format and consists of three main components:
2. Queries (Query Collection)
Legal information pieces extracted from 7 different perspectives for each document.
7 Query Types:
- Subject: Main topic of the özelge
- Article Text: Text of relevant law articles
- Communique Text: Content of relevant communiques and circulars
- Regulation Text: Regulation and legislation texts
- Justification Text: Legal justifications
- Decision Text: Administrative opinions and final decisions
- Condition Text: Application conditions and requirements
Tokenizer Benchmark & Data Filtering Summary
This process is not a training error and does not involve any training failure. It is a data analysis and preprocessing step performed before model training.
We benchmarked seven tokenizers (MPNet, Qwen2, Gemma, XLM-R, BERT, Pretrained, T5) on all datasets to measure token lengths and identify extreme long-sequence outliers. Among these, MPNetTokenizerFast generated the highest total token count, making it the most sensitive tokenizer for detecting unusually long samples.
Using MPNet as the reference tokenizer, we removed samples that exceeded the dataset-specific average by ~7000 tokens. This filtering was applied independently to each dataset to ensure balanced sequence distributions and cleaner input data.
The number of removed and remaining samples is summarized in the table below.
| Tokenizer | vocab_size | total_tokens | avg_tokens | min_tokens | max_tokens | median_tokens |
|---|---|---|---|---|---|---|
| MPNetTokenizerFast | 30,527 | 276,476,811 | 2,281 | 263 | 12,383 | 1,998 |
| Qwen2TokenizerFast | 151,669 | 219,326,828 | 1,810 | 190 | 9,201 | 1,594 |
| GemmaTokenizerFast | 262,144 | 183,710,411 | 1,516 | 158 | 7,578 | 1,341 |
| XLMRobertaTokenizerFast | 250,002 | 151,008,441 | 1,246 | 132 | 6,397 | 1,099 |
| BertTokenizerFast | 32,000 | 127,503,718 | 1,052 | 103 | 5,386 | 931 |
| PretrainedTokenizerFast | 32,000 | 122,387,578 | 1,010 | 102 | 5,227 | 893 |
| T5TokenizerFast | 32,128 | 121,315,289 | 1,001 | 100 | 5,238 | 885 |
Tokenizer / Total Token |
Corr of Vocab Size – Total Token |
| Dataset | max_tokens | avg_tokens | deleted_samples | total_samples |
|---|---|---|---|---|
newmindai/regulation-retrieval |
276,476,811 | 2281.19945 | 1,300 | 121,300 |
newmindai/caselaw-retrieval |
1,386 | 2,281 | 0 | 1,386 |
newmindai/court-of-cassation-caselaw |
30,527 | 186.4827586 | 11 | 272 |
3. Default (Relevance Matrix)
Relationship table showing which query belongs to which document.
| Field | Description |
|---|---|
query-id |
Query identifier |
corpus-id |
Related document identifier |
score |
Relevance score (all 1) |
Dataset Statistics
Total Statistics:
├─ Corpus Records: 23,701 documents
├─ Query Records: 121,198 queries
└─ Relevance Records: 121,198 relations
Per Document:
├─ 1 corpus entry (full ruling text)
├─ 2–7 queries (legal perspectives)
└─ Average ~5.1 queries per document
Field Coverage (Queries per Document)
On average, each özelge is represented by around 5.1 distinct queries, corresponding to different legal fields. The distribution of populated query types per document is as follows:
- 2 query types: ~0.1% of documents (e.g., Subject + Article Text)
- 3 query types: ~12.3% of documents (e.g., Subject + Article Text + Decision Text)
- 4 query types: ~26.2% of documents (e.g., Subject + Article Text + Communique Text + Decision Text)
- 5 query types: ~23.9% of documents (e.g., Subject + Article Text + Communique Text + Regulation Text + Decision Text)
- 6 query types: ~12.6% of documents (e.g., Subject + Article Text + Communique Text + Regulation Text + Justification Text + Decision Text)
- 7 query types: ~24.9% of documents (All fields: Subject + Article Text + Communique Text + Regulation Text + Justification Text + Decision Text + Condition Text)
Query Types Available:
- Subject: Main topic/issue of the ruling
- Article Text: Relevant law article content
- Communique Text: Official communique/circular content
- Regulation Text: Regulation and legislation texts
- Justification Text: Legal reasoning and justifications
- Decision Text: Administrative opinion and final decision
- Condition Text: Application conditions and requirements
In other words, roughly 61% of the corpus has 5 or more query types populated, making them rich multi-perspective legal cases rather than shallow single-label examples.
Text Length Distribution
For corpus texts (original full özelge rulings with non-empty ozelge_content, currently 100 documents):
- Mean length: ~1,736 words
- Median (p50): ~1,658 words
- 90th percentile (p90): ~2,393 words
These are long, dense legal rulings, comparable to typical tax/administrative decisions with full reasoning and references.
For query texts (legal snippets extracted from seven perspectives across all 23k+ records):
- Mean length: ~41.6 words
- Median (p50): ~24 words
- 90th percentile (p90): ~97 words
This makes queries similar to short legal questions, issue statements, justifications or excerpts from statutes/communiques, while the associated corpus entries provide the full ruling context for the subset of records where the full original özelge text is available.
Use Cases
1. Information Retrieval Systems
- Training for semantic search models
- Dense retrieval systems (DPR, ANCE, ColBERT)
- Sparse retrieval systems (BM25, TF-IDF) benchmark
2. RAG (Retrieval-Augmented Generation) Applications
- Legal chatbots
- Tax consultation assistants
- Automatic özelge analysis systems
3. Question-Answering Systems
- Legal QA models
- Extractive and abstractive QA
- Multi-hop reasoning
4. 📊 Model Evaluation
- Benchmarking Turkish IR models
- Retrieval performance analysis
- Domain adaptation studies
Data Collection and Processing
Data Source
The data is sourced from official özelge decisions of the Turkish Revenue Administration. Each özelge:
- Responds to specific questions asked by taxpayers
- References relevant legislation, communiques, and regulations
- Contains the Administration's opinion for concrete applications
Ethics and Legal Notices
License
This dataset is published under CC-BY 4.0 license. Please cite the source when using.
Citation
@article{mecellem2026,
title={Mecellem Models: Turkish Models Trained from Scratch and Continually Pre-trained for the Legal Domain},
author={Uğur, Özgür and Göksu, Mahmut and Çimen, Mahmut and Yılmaz, Musa and Şavirdi, Esra and Demir, Alp Talha and Güllüce, Rumeysa and Çetin, İclal and Sağbaş, Ömer Can},
journal={arXiv preprint arXiv:2601.16018},
year={2026},
month={January},
url={https://arxiv.org/abs/2601.16018},
doi={10.48550/arXiv.2601.16018},
eprint={2601.16018},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
License
This dataset is released under the Apache 2.0 License.
Contact
For questions: info@newmind.ai