Turkish Legal QA Triplets (tr-legal-triplets)

A large-scale dataset of query/positive/negative triplets generated from Turkish legal documents, designed for training and evaluating multilingual text embedding models on Turkish legal content.

Subset	Records	Description
default	1,148,041	Full generated dataset
cleaned	1,043,913	Quality-filtered — recommended for training

Dataset Description

Source Data

The dataset is built from 19,970 active Turkish legal documents covering:

Document Type	Count
Regulation (Yönetmelik)	9,581
Communiqué (Tebliğ)	4,321
Presidential Decision (Cumhurbaşkanlığı Kararnamesi)	4,006
Law (Kanun)	1,796
Bylaw (Tüzük)	107
Central Bank Regulation (TCMB)	66
Decree Law (Kanun Hükmünde Kararname)	63
Circular (Genelge)	29
Constitution (Anayasa)	1

The source provisions span 163,616 legal provisions across 6 provision types (article, paragraph, clause, temporary article, part, annex article), issued by 308 institutions.

Generation Pipeline

The dataset was produced via a 3-step pipeline:

1. Normalize (pipeline/01_normalize.py): Clean and normalize raw legal provisions, removing noise and standardizing formatting. Input: 157,831 provisions → Output: 157,831 normalized records.

2. Chunk (pipeline/02_chunk.py): Split long provisions into semantically coherent segments.

   • Chunk size: 1,000 characters with 200-character overlap
   • Threshold: provisions > 1,500 characters are chunked
   • Output: 198,789 chunks (avg. length: 501 characters)

3. Generate (pipeline/03_generate_qa.py): Use GPT-4o-mini with Turkish legal domain prompts to produce retrieval triplets for each chunk.

   • Concurrency: 100 parallel API calls
   • Triplets per chunk: 5 (short) / 7 (medium) / 10 (long)
   • 192,985 chunks processed → 1,148,041 triplets (0 failures)

What Each Record Contains

Each record is a training triplet with four text fields:

Field	Description
query	A natural-language search query in Turkish that a user might type to find the legal provision. Generated in diverse styles: direct questions, keywords, scenario descriptions, and everyday language.
positive	The original authoritative legal text (the ground-truth answer to the query).
positive_human	A simplified, plain-Turkish paraphrase of the legal text — same information, accessible language.
negative	A hard negative: a realistic legal text from the same legal domain that superficially resembles the query but does not answer it.

Schema

Field	Type	Example
id	string	"chunk_4821__t_2"
chunk_id	string	"chunk_4821"
provision_id	string	"prov_18432"
document_title	string	"İş Kanunu"
document_type	string	"law"
document_number	string	"4857"
provision_type	string	"article"
provision_number	string	"17"
query	string	"İş sözleşmesi feshedilirken ihbar süresi ne kadar?"
positive	string	"Belirsiz süreli iş sözleşmelerinin feshinden önce..."
positive_human	string	"İş sözleşmesini bitirmek isteyen işveren ya da işçi..."
negative	string	"İşçinin haklı nedenle iş sözleşmesini feshetmesi..."

Example Usage

from datasets import load_dataset

# Load the cleaned (recommended) subset
ds = load_dataset("yunus-emre/tr-legal-triplets", "cleaned")

# Or load the full unfiltered dataset
# ds = load_dataset("yunus-emre/tr-legal-triplets", "default")

# For embedding model fine-tuning (triplet loss)
for row in ds["train"]:
    query    = row["query"]
    positive = row["positive"]
    negative = row["negative"]

# For human-readable / cross-encoder training pairs
for row in ds["train"]:
    query          = row["query"]
    plain_positive = row["positive_human"]

Fine-tuning with sentence-transformers

from sentence_transformers import SentenceTransformer, InputExample, losses
from torch.utils.data import DataLoader
from datasets import load_dataset

ds = load_dataset("yunus-emre/tr-legal-triplets", "cleaned", split="train")

model = SentenceTransformer("sentence-transformers/paraphrase-multilingual-mpnet-base-v2")

examples = [
    InputExample(texts=[row["query"], row["positive"], row["negative"]])
    for row in ds
]

loader = DataLoader(examples, batch_size=32, shuffle=True)
loss = losses.TripletLoss(model)

model.fit(
    train_objectives=[(loader, loss)],
    epochs=3,
    output_path="tr-legal-embedding-model",
)

Dataset Subsets

default — Full Dataset (1,148,041 triplets)

The complete generated dataset with no filtering applied. May contain:

• Records where the source text was unavailable (system error in positive field): 89,314 records (7.78%)
• Short or low-information fields
• Duplicate queries within the same chunk

cleaned — Filtered Dataset (1,043,913 triplets) ⭐ Recommended

A quality-filtered version with the following records removed:

Filter	Removed
Error in positive field (system error instead of legal text)	89,314
Fields below minimum length threshold	14,659
Duplicate queries within the same chunk	155
Total removed	104,128 (9.07%)

Data Quality

Quality was assessed using a two-layer evaluation pipeline (pipeline/04_quality_check.py):

Automated Heuristic Checks (full dataset)

Metric	Value
Total records	1,148,041
Error positives (system errors)	89,314 (7.78%)
Query length	mean 46 chars, median 46
Positive length	mean 570 chars, median 412
Positive_human length	mean 180 chars, median 176
Negative length	mean 188 chars, median 186
Query↔Positive Jaccard similarity	mean 0.04 (low overlap ✅)
Negative↔Positive Jaccard similarity	mean 0.03 (distinct texts ✅)
Global duplicate queries	15.6%

LLM Quality Evaluation (10,000 stratified sample, GPT-4o-mini)

Each sampled record was scored 1–5 on four dimensions:

Dimension	Score	Interpretation
positive_human_accuracy	4.25 / 5	Simplified text preserves original meaning well
negative_hardness	3.81 / 5	Hard negatives are challenging and realistic
overall_quality	3.45 / 5	Acceptable for training use
query_relevance	3.20 / 5	Queries are relevant but vary in naturalness

Quality by document type:

Document Type	Avg. Score
Constitution (Anayasa)	3.94
Law (Kanun)	3.57
Communiqué (Tebliğ)	3.00
Decree Law (KHK)	2.78
Presidential Decision	2.67
Regulation (Yönetmelik)	2.61

• Low-quality records (overall ≤ 2): 10.85% of sampled records
• Each query is generated to be unique and diverse (direct questions, keywords, scenarios, everyday Turkish)
• positive_human preserves all information from the original legal text in plain language
• Hard negatives are generated, not retrieved — they use realistic legal language and terminology to maximize difficulty

Intended Uses

• ✅ Fine-tuning multilingual or Turkish-specific embedding models (e.g., mE5, multilingual-e5, paraphrase-multilingual-mpnet)
• ✅ Benchmarking retrieval systems on Turkish legal content
• ✅ Building RAG (Retrieval-Augmented Generation) pipelines for Turkish law
• ✅ Cross-encoder / reranker training using (query, positive_human) pairs
• ✅ Legal question-answering research

Out-of-Scope Uses

• ❌ Direct legal advice — this is training data, not verified legal opinion
• ❌ Non-Turkish legal systems without domain adaptation

Citation

If you use this dataset in your research, please cite:

@dataset{tr_legal_triplets_2025,
  title     = {Turkish Legal QA Triplets},
  author    = {yunus-emre},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/yunus-emre/tr-legal-triplets}
}

License

This dataset is released under the Apache 2.0 License.

The underlying legal texts are sourced from publicly available Turkish legislation and are in the public domain. The generated queries, paraphrases, and hard negatives are released under Apache 2.0.