DuckerPII-Thai-Base: High-Precision Thai PII De-Identification Foundation Model

Model Description • Benchmark • Quickstart • Hybrid Architecture • Limitations • Citation

DuckerPII-Thai-Base is a state-of-the-art open Thai Named Entity Recognition (NER) foundation model developed by DuckerMaster. Specifically engineered to detect, extract, and mask Personally Identifiable Information (PII) in unstructured Thai text, it serves as a lightweight, high-throughput privacy gateway before dispatching prompts to Large Language Models (LLMs) and cloud APIs, ensuring strict compliance with Thailand's Personal Data Protection Act (PDPA) and draft PDPC AI regulatory guidelines.

Built on an optimized RoBERTa architecture and fine-tuned on the 20,000-sample Thai-PII-Synthetic v6.0.0 corpus, DuckerPII-Thai-Base overcomes superficial shortcut memorization by training against 50% trigger-word dropout, numeric hard negatives, and fragmented address structures.

Model Description

Model Type: Token Classification / Named Entity Recognition (NER)
Architecture: RoBERTa-based contextual subword encoder (WangchanBERTa-base fine-tuned)
Parameters: ~110M
Primary Language: Thai (th)
Context Length: 416 tokens
Supported Entity Types (7 Classes): NAME, ID_CARD, PHONE, EMAIL, BANK_ACC, BANK, ADDRESS
Dataset: DuckerMaster/Thai-PII-Synthetic (v6.0.0) (100% Synthetic, Zero Real-World PII)
License: Apache-2.0 (Permissive Commercial Use)

Performance & Benchmark

1. In-Domain & Out-of-Distribution (OOD) Evaluation

Evaluated against the held-out test split (2,000 sentences / 6,052 entities with 0% template and lexical overlap) and an adversarial OOD noisy chat benchmark:

Metric	Regex Baseline	DuckerPII-Thai-Base (Ours)	Gain vs. Baseline
In-Domain Macro Precision	0.7142	0.9989	+28.47%
In-Domain Macro Recall	0.6079	0.9989	+39.10%
In-Domain Macro F1-Score	0.6432	0.9989	+35.57%
In-Domain Micro F1-Score	0.5981	0.9989	+40.08%
OOD Noisy Chat Recall	0.4210	98.33%	+56.23%
OOD Span Leak Rate	57.90%	1.67%	-56.23% (Leak Dropped)
Single-Core CPU Latency	< 1 ms	~42 ms/sample	Real-time ready

2. Class-by-Class F1 Breakdown

Entity Label	Precision	Recall	F1-Score	Key Invariant Resolved
`NAME`	0.9989	0.9984	0.9986	Informal nicknames, English transliterated names
`ID_CARD`	1.0000	1.0000	1.0000	Mod11 Checksum mathematical validation
`PHONE`	0.9991	0.9991	0.9991	Validated Thai telco prefixes (`08`, `09`, `06`, `02`)
`EMAIL`	1.0000	1.0000	1.0000	Fullmatch RFC compliance, rejects alphanumeric tokens
`BANK_ACC`	0.9980	1.0000	0.9990	Differentiates order IDs, tracking, and barcodes
`BANK`	1.0000	0.9976	0.9988	Formal bank names and colloquial abbreviations
`ADDRESS`	0.9964	0.9978	0.9971	Fragmented landmarks, subdistrict/district hierarchies

3. Visual Performance Dashboard

The model was subjected to a 12-step validation harness on Kaggle, tracking loss convergence, confusion matrices, and adversarial leak stability:

Quickstart & Code Examples

1. Requirements

pip install torch transformers pythainlp

2. Standard Inference via Hugging Face Pipeline

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

MODEL_NAME = "DuckerMaster/DuckerPII-Thai-Base"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForTokenClassification.from_pretrained(MODEL_NAME)
model.eval()

text = "แจ้งส่งพัสดุให้คุณ นพดล แซ่ลี้ โทร 0812345678 จัดส่งที่ 12/4 ซอยสุขุมวิท 21 คลองเตย กทม 10110 ครับ"

inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)

predictions = torch.argmax(outputs.logits, dim=-1)[0]
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

print(f"Input: {text}\n")
print("Detected Entities:")
for token, pred_idx in zip(tokens, predictions):
    label = model.config.id2label[pred_idx.item()]
    if label != "O":
        print(f"  {token:<20} -> {label}")

3. Production De-Identification Pipeline (Masking & Redaction)

from transformers import pipeline

ner_pipeline = pipeline(
    "token-classification",
    model="DuckerMaster/DuckerPII-Thai-Base",
    aggregation_strategy="simple"
)

raw_chat = "คุณกานดา จันทร์เพ็ญ โอนเงิน 500 บาท เข้าบัญชี 098-7-65432-1 ธนาคารกรุงไทย นะคะ"
results = ner_pipeline(raw_chat)

# De-identification / Redaction
masked_text = raw_chat
for entity in sorted(results, key=lambda x: x["start"], reverse=True):
    masked_text = (
        masked_text[:entity["start"]]
        + f"<{entity['entity_group']}>"
        + masked_text[entity["end"]:]
    )

print("Original :", raw_chat)
print("Masked   :", masked_text)
# Output: <NAME> โอนเงิน 500 บาท เข้าบัญชี <BANK_ACC> <BANK> นะคะ

Hybrid Architecture (ThaiPIIGateway)

In high-consequence enterprise environments, neural models can occasionally overgeneralize on edge cases like non-PII logistics codes (#ORD-982341, paid1234567890123). We recommend deploying DuckerPII-Thai-Base inside the ThaiPIIGateway hybrid architecture:

[Raw Input Text]
       │
       ▼
[Subword Tokenization (SentencePiece)]
       │
       ▼
[DuckerPII-Thai-Base (RoBERTa Neural NER)]
       │
       ▼
[Deterministic Post-Processing Gate]
  ├── Phone Prefix Validator (06, 08, 09, +66)
  ├── Email Fullmatch Regex Filter
  ├── Thai ID Card Mod11 Checksum Gate
  └── Name Non-Alpha / Keyword Sanitizer
       │
       ▼
[Outputs: Masked Text + Redacted Logs + Synthetic Anonymization]

Intended Uses & Limitations

Intended Uses

Pre-LLM Privacy Filters: Real-time sanitization of customer prompts before forwarding to cloud LLMs (Gemini, ChatGPT, Claude).
Enterprise Log Masking: Redacting sensitive customer data from application traces and database transaction dumps.
Academic Research: Benchmarking Thai privacy-preserving NLP pipelines.

Limitations & Ethical Boundary

Special Category Data: The current version detects 7 foundational PII categories. It does not cover biometric, genetic, or specialized health records.
100% Synthetic Data: Trained exclusively on synthetic data created under PDPA/PDPC Privacy Enhancing Technology (PETs) guidelines. No real identities were harmed or used.
Prohibited Use: Strictly prohibited for unauthorized re-identification, deanonymization attacks, surveillance, or fraud.

Citation

If you use DuckerPII-Thai-Base or its synthetic dataset in your research or production systems, please cite:

@misc{duckermaster2026duckerpii,
  author       = {DuckerMaster},
  title        = {DuckerPII-Thai-Base: High-Precision Thai PII De-Identification Foundation Model},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/DuckerMaster/DuckerPII-Thai-Base}},
  note         = {Version 1.0.0 (Trained on Thai-PII-Synthetic v6.0.0)}
}

Contact & Maintainer

Maintained by DuckerMaster. For technical inquiries, enterprise integration guidelines, or vulnerability reports, please open an issue in the Hugging Face Community tab.

Downloads last month: 57

Safetensors

Model size

0.1B params

Tensor type

F32

DuckerMaster
/

DuckerPII-Thai-Base