Thai-PII-Synthetic

A Synthetic Dataset for Thai Named Entity Recognition and Privacy De-Identification

Thai-PII-Synthetic is a 20,000-sample synthetic dataset (60,677 entities) designed for Thai Named Entity Recognition (NER) and pre-LLM privacy de-identification across 7 common PII categories. Built without any real personal data using rule-based synthetic generation, it supports both character-level span annotations and pre-tokenized IOB2 formats.

Summary • Example • Fields & Splits • Usage • Baseline • Safety & Policy • Citation

---

Dataset Summary

Thai-PII-Synthetic (v6.0.0) contains 20,000 sentences (60,677 entities) created to evaluate and train privacy-preserving systems, including automated Named Entity Recognition (NER), pre-LLM privacy de-identification gateways, and log sanitization pipelines. The dataset adopts Privacy Enhancing Technologies (PETs) principles referenced in the draft AI regulatory guidelines by the Office of the Personal Data Protection Commission of Thailand (PDPC).

Attribute	Specification
Tasks	Named Entity Recognition (NER), Token Classification, PII De-identification
Primary Language	Thai (th)
Total Samples	20,000 sentences (Train: 16,000 / Validation: 2,000 / Test: 2,000)
Total Entities	60,677 entities (Train: 48,586 / Validation: 6,039 / Test: 6,052)
Entity Labels (7)	NAME, ID_CARD, PHONE, EMAIL, BANK_ACC, BANK, ADDRESS
Diversity Upgrades (v6)	50% Trigger Word Dropout, Non-PII Hard Negatives (Tracking, Order IDs, Barcodes), English Names & Nicknames, Fragmented Addresses
Configurations	1. default: Character-level Span JSON (start, end, label, value) 2. iob2: Pre-tokenized IOB2 JSON (15 classes)
Splitting Strategy	0% Template & Lexical Overlap across splits
License	Creative Commons Attribution 4.0 International (CC-BY-4.0)
Maintainer	DuckerMaster

---

What's New in v6.0.0 (Adversarial Robustness Upgrades)

To prevent shortcut learning / superficial pattern memorization, v6.0.0 introduces:

1. 50% Trigger Word Dropout: Randomly removes leading anchors (โทร:, ชื่อ:, ที่อยู่:) to simulate raw chat dumps.
2. 30% Numeric Hard Negatives: Injects non-PII tracking numbers (e.g., TH0192837465), order IDs (#ORD-91823746), barcodes/SKUs (SKU-481920), and price tags into sentence context labeled as O.
3. Lexical Realism & Nicknames: Adds Thai informal nicknames (10%, e.g., พี่โต้ง, หมอเอก), English/transliterated names (6%), and reduces formal honorific ratio to 45%.
4. Fragmented & Landmark Addresses: Injects short/partial addresses (25%) including condos, office buildings, and neighborhood landmarks without full subdistrict/district structures.
5. Real-world Chat Slang & Spacing: Supports colloquial spellings (เบอ, บช, ที่ยุ) and space collapsing between triggers and values.

---

Example Record

Sample instance from the default configuration (Character-level Spans):

{
  "id": "th_pii_tr_01676",
  "text": "แจ้งข้อมูลเพิ่มเติมครับ รบกวนส่งใบเสร็จรับเงินมาที่อีเมล o41kjnpr@yahoo.com ชื่อผู้ซื้อ: นเรศ เพชรประเสริฐ ที่อยู่: สยามทาวเวอร์ ชั้น 9 งามวงศ์วาน รบกวนดำเนินการด้วยนะครับ ขอบคุณครับ",
  "entities": [
    {
      "start": 57,
      "end": 75,
      "label": "EMAIL",
      "value": "o41kjnpr@yahoo.com"
    },
    {
      "start": 89,
      "end": 106,
      "label": "NAME",
      "value": "นเรศ เพชรประเสริฐ"
    },
    {
      "start": 116,
      "end": 146,
      "label": "ADDRESS",
      "value": "สยามทาวเวอร์ ชั้น 9 งามวงศ์วาน"
    }
  ],
  "split": "train",
  "template_id": "T2_TR_CLAUSE_777"
}

---

Data Fields and Splits

Data Fields

• id (string): Unique sample identifier (e.g. th_pii_tr_00001)
• text (string): Thai sentence text (e-commerce orders, customer chat, transfer slips)
• entities (list): List of annotated entity objects:
  • start (int): Starting character index (0-indexed)
  • end (int): Ending character index (Python slice convention)
  • label (string): Entity label (1 of 7 supported classes)
  • value (string): Exact substring matching text[start:end]
• split (string): Dataset split (train, validation, or test)
• template_id (string): Generator group template identifier
• (For iob2 configuration) Includes tokens (tokenized via pythainlp.tokenize.newmm), ner_tags (integer IDs), and ner_labels (BIO-scheme labels)

Entity Distribution

Audit statistics for annotated entity instances from data/manifest.json:

Entity Type	Train (16k)	Validation (2k)	Held-out Test (2k)	Total Entities	Distribution
NAME	14,827	1,843	1,841	18,511	30.51%
ADDRESS	11,130	1,382	1,390	13,902	22.91%
PHONE	9,289	1,129	1,128	11,546	19.03%
BANK_ACC	3,983	514	492	4,989	8.22%
BANK	3,270	418	410	4,098	6.75%
ID_CARD	3,282	397	420	4,099	6.76%
EMAIL	2,805	356	371	3,532	5.82%
Total Entities	48,586	6,039	6,052	60,677	100.00%

Density Balance Note: ADDRESS accounts for 22.91% of all entities (satisfying the density constraint ≥ 20%), and EMAIL represents 5.82% (3,532 / 60,677).

Dataset Splits and Composition

Breakdown of sentence counts, negatives, and template partitioning:

Metric	Train Split	Validation Split	Held-out Test Split	Total / Overall
Total Sentences	16,000	2,000	2,000	20,000
PII-Bearing Sentences	14,400	1,800	1,800	18,000 (90.0%)
Pure Negatives (0 PII)	1,600 (10.0%)	200 (10.0%)	200 (10.0%)	2,000 (10.0%)
Hard Numeric Negatives	~4,320	~540	~540	~5,400 (30% of PII sentences)
Disjoint Group Templates	2,600 (T2_TR_*)	1,037 (T2_VAL_*)	1,035 (T2_TE_*)	4,672 (0% Overlap)

---

Usage

1. Load Character Span Format (default)

from datasets import load_dataset

# Load character-level span annotations
dataset = load_dataset("DuckerMaster/Thai-PII-Synthetic", "default")
print(dataset)
print("Sample:", dataset["train"][0])

2. Load Token Classification Format (iob2)

from datasets import load_dataset

# Load pre-tokenized IOB2 format ready for Hugging Face Transformers
token_dataset = load_dataset("DuckerMaster/Thai-PII-Synthetic", "iob2")
print("Tokens :", token_dataset["train"][0]["tokens"][:8])
print("Labels :", token_dataset["train"][0]["ner_labels"][:8])

---

Baseline

Evaluation results of a Regex Baseline evaluated on the Held-out Test Set (2,000 samples / 6,052 entities with 0% template and lexical overlap with Train) from reports/evaluation_report.json:

Entity Type	Exact Precision	Exact Recall	Exact F1	Partial Overlap F1	Support
NAME	0.7402	0.2863	0.4128	0.4740	1,841
ID_CARD	0.8049	0.8548	0.8291	0.8291	420
PHONE	0.9868	0.8626	0.9205	0.9205	1,128
EMAIL	1.0000	0.7978	0.8876	0.8876	371
BANK_ACC	0.6667	0.7683	0.7139	0.7139	492
BANK	0.8006	0.6854	0.7385	0.7385	410
ADDRESS	0.0000	0.0000	0.0000	0.0000	1,390
Overall (Micro Avg)	0.5981	0.5981	0.5981	0.6519	6,052

• Observation: The Regex baseline fails completely on ADDRESS (F1 = 0.0000) and suffers a major drop in NAME recall (0.2863), highlighting that unanchored real-world PII requires contextual neural representation models (e.g. WangchanBERTa).
• Preliminary Noise Test: When tested against synthetic lexical noise (spacing variations and typos), Exact Micro F1 shifted from 0.5981 to 0.5975 (Delta = -0.0006). This confirms high evaluation consistency.

---

Responsible Use & Safety

• Synthetic Disclaimer & Prohibited Use: 100% synthetically generated for privacy research. While names and addresses are permuted, numbers (Thai IDs, phones) follow official syntax—strictly prohibited for real calls, financial transactions, fraud, or phishing.
• PDPA Scope: Covers 7 common PII types as practical research examples; does not cover special statutory categories (e.g. biometrics). Organizations must perform their own DPIA.
• Full Safety Policy: Comprehensive guidelines, prohibited use cases, and vulnerability disclosure are detailed in SAFETY.md.

---

Citation

BibTeX

@misc{duckermaster2026thaipii,
  author       = {DuckerMaster},
  title        = {Thai PII Synthetic NER Dataset},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DuckerMaster/Thai-PII-Synthetic}},
  note         = {Version 6.0.0}
}

Regulatory and Academic References

1. PDPC Thailand: แนวปฏิบัติการคุ้มครองข้อมูลส่วนบุคคลเกี่ยวกับการพัฒนาและการใช้งานเทคโนโลยีปัญญาประดิษฐ์ (AI) (Draft AI Privacy Guidelines), Office of the Personal Data Protection Commission (PDPC).
2. NIST SP 800-122: McCallister, E., Grance, T., & Scarfone, K. (2010). Guide to Protecting the Confidentiality of Personally Identifiable Information (PII). National Institute of Standards and Technology.
3. Article 29 Data Protection Working Party (WP216 / 2014): Opinion 05/2014 on Anonymisation Techniques, European Data Protection Board (EDPB).
4. PIIBench (2026): A Unified Multi-Source Benchmark Corpus for Personally Identifiable Information Detection, arXiv:2604.15776.
5. SPY Benchmark (2025): Savkin, M., Ionov, T., & Konovalov, V. (2025). SPY: Enhancing Privacy with Synthetic PII Detection Dataset, In Proceedings of the 2025 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop (NAACL 2025 SRW), pages 236–246. https://aclanthology.org/2025.naacl-srw.23/