treamyracle/distilled-indonesian-ner

Dataset Card for Distilled Gold Indonesian NER Dataset

This dataset is a high-quality synthetic Indonesian Named Entity Recognition (NER) dataset, distilled and self-corrected using a 120B parameter teacher model. It features annotations across 19 distinct entity classes formatted as character-level spans for token classification tasks.

Dataset Details

---

Dataset Structure

Data Instances

An example instance in the dataset:

{
  "text": "Terminal baru Bandara Ahmad Yani dibangun PT Angkasa Pura I dengan nilai investasi sebesar Rp 2 , 2 triliun dan merupakan bandara terapung di atas air pertama di Indonesia .",
  "entities": [
    {
      "start": 14,
      "end": 32,
      "label": "FAC"
    },
    {
      "start": 42,
      "end": 59,
      "label": "ORG"
    },
    {
      "start": 91,
      "end": 107,
      "label": "MON"
    },
    {
      "start": 162,
      "end": 171,
      "label": "GPE"
    }
  ]
}

Data Fields

• text (string): Reconstructed raw Indonesian text, created by space-joining parsed tokens.
• entities (list of dicts): A list containing parsed entity spans, where each span contains:
  • start (int): Inclusive character start index.
  • end (int): Exclusive character end index.
  • label (string): The entity tag (one of 19 tags).

Label Glossary

The dataset supports 19 different entity types:

Class Label	Entity Type	Description & Examples
PER	Person	Individuals, authors, historical figures, e.g., Nadiem Makarim, Brigjen Iqbal
ORG	Organization	Companies, institutions, government bodies, e.g., Bank Indonesia, Gojek, KPK
LOC	Location	Natural geographical entities, e.g., Pulau Jawa, Selat Sunda
GPE	Geopolitical Entity	Countries, cities, administrative regions, e.g., Indonesia, Jakarta, Surabaya
FAC	Facility	Airports, buildings, roads, bridges, public spaces, e.g., Bandara Ahmad Yani, Monas
EVT	Event	Named festivals, conferences, battles, historic events, e.g., Festival Batik Nusantara 2024
PRD	Product	Named commercial products, software, vehicles, e.g., Sinovac, RuangGuru, Gojek Pay AI
NOR	Nationality/Group	Religious, national, ethnic, or political groups, e.g., UMKM, Pemerintah, JAD
DAT	Date	Specific dates, calendar days, months, and years, e.g., Senin, 15 Januari 2024
TIM	Time	Times of day, durations, timezones, e.g., pukul 14.00 WIB
MON	Money	Currency types and cash values, e.g., Rp 2,2 triliun, USD 10 juta
QTY	Quantity	Measurements, weights, volumes, counts, e.g., 50 kilogram, 1.500 peserta
CRD	Cardinal	Named cardinal numbers.
ORD	Ordinal	Named ordinal values, e.g., pertama, kedua
PRC	Percent	Percentage values, e.g., 2,5 persen, 75 persen
LAW	Law	Named legal documents, constitutions, decrees, e.g., Perpres Nomor 18 Tahun 2024, UU No. 31
LAN	Language	Spoken or written languages.
WOA	Work of Art	Titles of books, movies, songs, paintings, e.g., Biru Langit, Warni Batik
REG	Region	Non-geopolitical regions, continents, e.g., Asia Tenggara, Bondowoso, Jawa Timur

---

Dataset Creation

Curation Rationale

Standard Indonesian gold-label NER datasets often have small volumes or lack the diversity of specialized entity tags (such as LAW, FAC, MON, etc.). This dataset was created to augment gold training data, allowing smaller student models to achieve higher recall and capture complex entity category boundaries.

Source Data

Data Generation & Self-Correction (2-Phase Pipeline)

The data was collected and processed using a high-parameter teacher LLM via a 2-stage automated pipeline:

1. Phase 1: Generation (Few-Shot Synthetic Generator):

   • Teacher LLM: openai/gpt-oss-120b (via OpenRouter).
   • The generator prompts the 120B model using few-shot gold examples to output rich, complex Indonesian sentences with annotations across the 19 NER tags.
   • Temperature is set to 0.8 to maximize linguistic variety and topical diversity.

2. Phase 2: Quality-Control Correction (Self-Correction):

   • The synthetic dataset is passed through a strict critic module.
   • The same teacher model is prompted under a low temperature (0.3) to identify, validate, and correct labeling errors, tag boundary alignment issues, or invalid BIO transitions.
   • Mappings are then validated programmatically to ensure 100% compliance with standard sequence constraints.

---

Additional Information

Licensing Information

This dataset is licensed under the MIT License.

Citation

If you use this dataset, please cite the repository:

@misc{indonesian-ner-distillation-2026,
  author = {Arif Athaya Harahap/treamyracle},
  title = {Distilled Gold Indonesian NER Dataset},
  year = {2026},
  publisher = {Hugging Face}
}