chayuto/thai-job-ner-dataset

Thai Job Posting NER Dataset (Synthetic)

A synthetic Named Entity Recognition dataset for Thai job postings, formatted in IOB2 tagging scheme.

Dataset Description

This dataset contains 1270 synthetically generated Thai job posting texts annotated with 7 entity types commonly found in informal Thai job advertisements (e.g., Facebook groups, LINE).

The data was generated using GPT-4o with carefully designed prompts to mimic the style and content of real Thai job posts, then aligned using fuzzy matching with Thai Character Cluster (TCC) boundary snapping.

Entity Types

Entity	Description	Example
PERSON	Names, nicknames	พี่แจง, คุณสมชาย
LOCATION	Places, areas	บางนา, ลาดพร้าว
CONTACT	Phone, LINE, email	088-888-8888, @line_id
HARD_SKILL	Job skills, tasks	ทำบัญชี, ขับรถ
COMPENSATION	Salary, pay	วันละ 500 บาท
EMPLOYMENT_TERMS	Job type, hours	งานประจำ, Full-time
DEMOGRAPHIC	Age, gender, nationality	คนไทย, อายุ 20-35

Tagging Scheme

IOB2 format with 15 labels: O + 7 entity types × 2 (B- prefix for begin, I- for inside).

Dataset Structure

Splits

Split	Examples
train	1015
validation	128
test	127

Features

• id (string): Unique example identifier
• tokens (list[string]): Word-tokenized text (pythainlp, newmm engine)
• ner_tags (list[ClassLabel]): IOB2 integer labels per token

Example

from datasets import load_dataset

ds = load_dataset("chayuto/thai-job-ner-dataset")
print(ds["train"][0])
# {'id': 'synthetic_0000', 'tokens': ['คนไทย', 'ที่', 'ทำ', ...], 'ner_tags': [13, 0, 0, ...]}

Usage

from datasets import load_dataset

dataset = load_dataset("chayuto/thai-job-ner-dataset")

# Access label names
label_names = dataset["train"].features["ner_tags"].feature.names
# ['O', 'B-HARD_SKILL', 'I-HARD_SKILL', 'B-PERSON', ...]

# Decode tags for an example
example = dataset["train"][0]
tags = [label_names[t] for t in example["ner_tags"]]
for token, tag in zip(example["tokens"], tags):
    if tag != "O":
        print(f"  {token} → {tag}")

Data Generation

• Generator: GPT-4o (OpenAI API)
• Style: Informal Thai job posts mimicking social media (Facebook groups, LINE)
• Alignment: Fuzzy string matching (rapidfuzz) with TCC boundary snapping (pythainlp)
• Tokenization: pythainlp word_tokenize (newmm engine), whitespace tokens excluded

Associated Model

Fine-tuned WangchanBERTa model trained on this dataset: chayuto/thai-job-ner-wangchanberta

License

CC-BY-4.0

Citation

@misc{thai-job-ner-dataset,
  title={Thai Job Posting NER Dataset (Synthetic)},
  author={Chayut Orapinpatipat},
  year={2026},
  url={https://huggingface.co/datasets/chayuto/thai-job-ner-dataset},
}