Datasets:

treamyracle
/

indo-ner

language:
  - id
license: apache-2.0
task_categories:
  - token-classification
pretty_name: Indo-NER (GLiNER Auto-Tagged)
size_categories:
  - 100K<n<1M
tags:
  - ner
  - gliner
  - named-entity-recognition
  - indonesian
  - synthetic

Indo-NER: Indonesian Named Entity Recognition Dataset (Silver Standard)

Dataset Summary

Indo-NER is a large-scale Indonesian Named Entity Recognition (NER) dataset automatically annotated using a zero-shot multilingual NER model. The dataset is designed to support research, benchmarking, and experimentation in Indonesian NLP, particularly for entity extraction tasks.

This dataset is silver standard, meaning annotations are machine-generated and may contain noise. It is suitable for pretraining, bootstrapping, and research use cases, but human validation is recommended for production-critical systems.

Data Fields

Each data sample contains the following fields:

text: Indonesian input sentence or paragraph
entities: List of extracted entity objects, each consisting of:
- start: Start character index of the entity span
- end: End character index of the entity span
- label: Short entity code (e.g., PER, ORG, LOC)
- original_english: Source English sentence (if translated or available)

Label Scheme (19 Classes)

Code	Description	Examples
PER	Person	Jokowi, Lionel Messi, Einstein
ORG	Organization	Google, OpenAI, PBB
NOR	Political Organization	Partai Golkar, Democrats, Nazi
LOC	Location (Geographical)	Gunung Merapi, Asia, Sungai Musi
GPE	Geopolitical Entity	Indonesia, Jakarta, Jawa Barat
FAC	Facility	Bandara Soetta, Jalan Tol
DAT	Date	17 Agustus 1945, Tahun 2024
TIM	Time	Pukul 07.00, Siang hari
CRD	Cardinal Number	Satu, Dua, 100
ORD	Ordinal Number	Pertama, Ke-10
QTY	Quantity	10 kg, 3 liter
PRC	Percent	50%, Sepuluh persen
MON	Money	Rp 50.000, 100 Dolar AS
EVT	Event	Perang Dunia II, G20 Summit
PRD	Product	iPhone 15, Windows 11
WOA	Work of Art	Harry Potter, Mona Lisa
LAW	Law	UUD 1945, UU Cipta Kerja
LAN	Language	Bahasa Indonesia, English
REG	Religion	Islam, Kristen, Hindu

Creation Methodology

Source Data

The text corpus is derived from large-scale Indonesian NER task collections, ensuring diverse sentence structures and real-world contexts.

Annotation Process (Auto-Tagging)

Model: GLiNER Large v2.5
Approach: Zero-shot multilingual NER
Confidence Threshold: 0.3 (balanced recall and precision)
Processing: Batch inference on NVIDIA T4 GPU
Label Mapping: Natural language prompts (e.g., "political organization") mapped to standardized short labels (e.g., NOR)

Usage

Load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("treamyracle/indo-ner")

# View the first training example
print(dataset["train"][0])

Limitations

This dataset is silver standard and auto-generated, which implies:

Possible boundary inaccuracies
Potential hallucinated entities
Label noise in ambiguous contexts

Human validation or post-processing is recommended for downstream or production use.

Citation

If you use this dataset, please cite the GLiNER authors and this repository:

@misc{indo-ner-2024,
  author       = {treamyracle},
  title        = {Indo-NER: GLiNER Auto-Tagged Dataset},
  year         = {2024},
  publisher    = {Hugging Face},
  journal      = {Hugging Face Repository},
  howpublished = {\url{https://huggingface.co/datasets/treamyracle/indo-ner}}
}