yusuf-said/turkish-privacy-filter-dataset

Turkish Privacy Filter Dataset

Turkish Privacy Filter Dataset is a Turkish-language dataset for privacy filtering, personally identifiable information (PII) detection, data redaction, and token-classification research.

The dataset contains synthetic and curated Turkish text examples with character-level privacy spans. It is designed to support the development and evaluation of models that detect sensitive information in Turkish text, including names, phone numbers, email addresses, addresses, account-like identifiers, dates, URLs, and secrets.

• Dataset repository: https://huggingface.co/datasets/yusuf-said/turkish-privacy-filter-dataset
• DOI: https://doi.org/10.57967/hf/8838
• Revision: e417d0f
• License: MIT
• Language: Turkish

Dataset Summary

This dataset focuses on Turkish privacy and PII detection in realistic short-form text scenarios such as:

• customer support tickets
• call center logs
• CRM notes
• billing notes
• delivery notes
• medical-style notes
• legal intake notes
• reservation messages
• server logs
• JSON logs
• configuration snippets
• code review notes
• OCR-like noisy text
• email-style messages
• address confirmation messages
• identity-style forms

Each record contains a Turkish text sample, a dictionary of annotated privacy spans, and metadata describing the source, split, difficulty, and scenario.

Intended Use

This dataset can be used for:

• Turkish PII detection
• privacy filtering
• data redaction
• sensitive information masking
• token classification / NER-style training
• evaluating LLM privacy filters on Turkish text
• building preprocessing pipelines before storing or sending user text to downstream systems

Not Intended Use

This dataset is not intended for:

• identifying real individuals
• reconstructing personal identities
• surveillance or profiling
• generating realistic fake credentials for abuse
• bypassing privacy filters or data protection systems
• making legal, medical, or compliance decisions without human review

Dataset Structure

Each row has the following fields:

{
  "text": "Turkish input text",
  "spans": {
    "private_phone:+90 555 123 45 67": [[42, 59]],
    "private_email:example@example.com": [[75, 94]]
  },
  "info": {
    "id": "tr_web_00001",
    "source": "web_synthetic_tr_hard_v4",
    "split": "train",
    "difficulty": "hard",
    "scenario": "support_ticket"
  }
}

Fields

Field	Type	Description
text	string	Turkish text sample.
spans	object	Dictionary of sensitive spans. Keys follow the format entity_type:surface_value; values are character offset ranges.
info	object	Metadata about the example.

Span Format

The spans field maps each detected sensitive value to one or more character-level offsets:

{
  "private_person:Ahmet Yılmaz": [[10, 22]]
}

The offset [start, end] follows Python slicing semantics:

text[start:end]

Examples with no sensitive information have:

{}

This makes the dataset useful for both positive and negative privacy-filter training.

Entity Types

The current version uses the following privacy labels:

Label	Description
private_person	Personal names and person-like identifiers.
private_phone	Phone numbers in Turkish or international formats.
private_email	Email addresses.
private_address	Physical addresses and address-like location details.
private_date	Dates, datetimes, appointments, or time-specific personal references.
private_url	Personal, account-specific, or private URLs.
account_number	Account numbers, ticket numbers, document numbers, IBAN-like or reference identifiers.
secret	Tokens, API keys, temporary codes, credentials, or secret-like strings.

Dataset Statistics

Current version:

Metric	Value
Language	Turkish
Split	train
Number of rows	5,110
Annotation type	Character-level spans
Difficulty	Hard
Negative examples	Included
Format	JSON / JSONL-compatible

Approximate entity distribution:

Entity Type	Count
account_number	1,471
private_person	1,468
private_phone	1,401
private_email	1,388
private_date	1,386
private_address	1,233
private_url	767
secret	545

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("yusuf-said/turkish-privacy-filter-dataset")

print(dataset)
print(dataset["train"][0])

Example

from datasets import load_dataset

dataset = load_dataset("yusuf-said/turkish-privacy-filter-dataset")
example = dataset["train"][0]

text = example["text"]
spans = example["spans"]

for key, offsets in spans.items():
    label, value = key.split(":", 1)
    for start, end in offsets:
        print(label, text[start:end])

Training Notes

This dataset can be adapted for multiple modeling approaches.

1. Character-level privacy filtering

Use the provided offsets directly to detect and redact sensitive substrings.

2. Token classification

Convert character offsets into token-level BIO or BILOU labels using a tokenizer that supports offset mappings.

Example labels:

B-private_person
I-private_person
B-private_email
I-private_email
O

3. Text classification

Examples can also be converted into binary or multi-label classification tasks, such as:

• contains PII / does not contain PII
• contains secret / does not contain secret
• contains contact information / does not contain contact information

Data Generation and Curation

The dataset is composed of Turkish synthetic and curated privacy-filter examples. The examples are designed to reflect realistic Turkish-language contexts where sensitive information may appear, including noisy, informal, semi-structured, and domain-specific text.

The dataset includes both:

• positive examples containing sensitive spans
• negative examples without sensitive spans

This helps models distinguish private user-specific information from public or non-sensitive references.

Quality Considerations

The dataset is intended for research and model-development purposes. While the examples are designed to be realistic and useful for Turkish privacy filtering, users should validate model behavior on their own domain-specific evaluation sets before production deployment.

Recommended validation steps:

• manually inspect model predictions on real target-domain text
• measure false positives and false negatives separately
• evaluate performance on Turkish-specific formats such as phone numbers, addresses, names, dates, and identity-like references
• test robustness on noisy OCR, logs, support tickets, and informal user messages
• review outputs before using the model in high-risk environments

Privacy and Safety

This dataset is intended to support privacy-preserving NLP systems. It should be used to detect, mask, redact, or reduce exposure of sensitive information.

The dataset should not be used to infer, expose, or reconstruct real personal information. If any potentially real personal information is identified, it should be treated as a data issue and removed.

Limitations

Known limitations include:

• The dataset is Turkish-focused and may not generalize to other languages.
• Synthetic examples may not cover all real-world privacy patterns.
• Address, name, and account-number formats can vary significantly across domains.
• Character-level spans must be converted before standard token-classification training.
• Some labels may overlap conceptually depending on downstream policy definitions.
• The dataset is not a substitute for legal or compliance review under KVKK, GDPR, or other privacy regulations.

Recommended Evaluation

For production-oriented use, evaluate models on a separate, manually reviewed Turkish test set that reflects the actual deployment domain.

Suggested metrics:

• entity-level precision
• entity-level recall
• entity-level F1
• false positive rate on non-sensitive public text
• false negative rate for high-risk entities such as secrets, account numbers, phone numbers, and emails

Citation

If you use this dataset, please cite it as:

@misc{yusuf_said_çöpcü_2026,
  author       = {Yusuf Said Çöpcü and Tunahan Yardımcı and Muhammet İbrahim Kaya},
  title        = {turkish-privacy-filter-dataset (Revision e417d0f)},
  year         = 2026,
  url          = {https://huggingface.co/datasets/yusuf-said/turkish-privacy-filter-dataset},
  doi          = {10.57967/hf/8838},
  publisher    = {Hugging Face}
}

APA:

Çöpcü, Y. S., Yardımcı, T., & Kaya, M. İ. (2026). turkish-privacy-filter-dataset (Revision e417d0f) [Data set]. Hugging Face. https://doi.org/10.57967/hf/8838

License

This dataset is released under the MIT license.

Contact

For questions, issues, or dataset improvement suggestions, please use the Hugging Face community tab of this repository.