tlabdev/ner-irish-dataset-2

NER Irish Dataset 2

An Indonesian news corpus for Named Entity Recognition (NER) consisting of manually annotated (gold) and model-annotated (silver) articles. This dataset provides raw, character-level entity annotations to ensure maximum flexibility across different tokenization strategies and downstream preprocessing pipelines.

The dataset is designed with gold-only evaluation in mind, while leveraging silver data as additional training supervision.

Dataset Summary

• Language: Indonesian
• Task: Named Entity Recognition (NER)
• Domain: Online news articles
• Annotation Type: Character-level spans (start/end offsets)
• Annotation Tools: Label Studio (gold), model-based annotation (silver)
• Data Level: Article-level (pre-tokenization)

Dataset Composition

The dataset is divided into two annotation sources:

Gold

Articles manually annotated by humans and treated as high-quality ground truth.

Silver

Articles annotated automatically using an NER model. These annotations may contain noise and are intended for training augmentation only.

Quick Start

from datasets import load_dataset

dataset = load_dataset(
    "tlabdev/ner-irish-dataset-2",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
        "test": "data/test.jsonl",
    },
)

print(dataset)

train_data = dataset["train"]
val_data   = dataset["validation"]
test_data  = dataset["test"]

Entity Types

The dataset contains 9 entity types commonly found in Indonesian news text:

Label	Description	Examples
PER	Person names	Joko Widodo, Megawati
ORG	Organizations and institutions	Bank Indonesia, UNESCO
GPE	Administrative geopolitical entities	Jakarta, Indonesia
LOC	Non-administrative locations	Gunung Merapi
FAC	Facilities and infrastructure	Bandara Soekarno-Hatta
LAW	Laws and regulations	UU No. 20 Tahun 2003
EVENT	Named events	Pilkada 2024
MONEY	Monetary values	Rp 10 miliar
DATE	Dates and time expressions	17 Agustus 1945

Dataset Structure

Data Fields

Each sample represents one full article and contains:

{
    "article_id": str,     # Unique article identifier
    "ds_source": str,      # "gold" or "silver"
    "title": str,
    "source": str,
    "publishedAt": str,
    "text": str,           # Full article text
    "entities": [
        {
            "text": str,   # Entity mention
            "label": str,  # Entity type
            "start": int,  # Character start offset
            "end": int     # Character end offset
        }
    ],
    "entity_count": int
}

Data Splits

Splitting is performed after preprocessing and follows a gold-only evaluation strategy.

Split Strategy

Split	Gold	Silver	Description
Train	80%	100%	Model training
Validation	10%	0%	Gold-only validation
Test	10%	0%	Gold-only testing

Key Points:

• Splitting is performed at the article level
• No article appears in more than one split
• Silver data is never used for validation or testing
• This design ensures that evaluation metrics reflect performance on manually annotated data only

Example

{
    "text": "Presiden Joko Widodo meresmikan Bandara Internasional Yogyakarta.",
    "entities": [
        {"text": "Joko Widodo", "label": "PER", "start": 9, "end": 20},
        {"text": "Bandara Internasional Yogyakarta", "label": "FAC", "start": 33, "end": 66}
    ]
}

Data Processing

The dataset has undergone careful preprocessing while preserving annotation integrity.

Applied Processing

Article deduplication with gold-priority:

• URL-based deduplication
• Metadata-based deduplication
• Content hash-based deduplication

Entity-level cleaning:

• Removal of duplicate entity spans per article
• Cleaning of noisy overlaps in silver annotations
• Character normalization using 1:1 mappings to preserve offsets
• Validation of entity span boundaries

Intentionally NOT Applied

The dataset is published in a raw, flexible format and does not include:

• Tokenization or token IDs
• BIO / IOB tagging
• Sliding-window chunking
• Class weighting
• Oversampling

All model-dependent preprocessing steps should be implemented downstream.

Important Notes

Character Offset Preservation

Entity annotations rely on character-level offsets. Any modification that changes text length will invalidate spans.

Safe operations:

• Consistent case normalization
• Whitespace trimming at document boundaries

Unsafe operations:

• Character insertion or deletion
• Unicode normalization that changes string length
• Aggressive text cleaning

Use Cases

Suitable for:

• Training and benchmarking Indonesian NER models
• Semi-supervised NER experiments (gold + silver)
• Error analysis and relabeling workflows
• Information extraction pipelines
• Research on class imbalance and noisy supervision

Limitations:

• Focused on formal news text
• Limited to 9 entity categories
• Silver annotations may contain noise

Citation

If you use this dataset, please cite:

@dataset{indo_news_ner_gold_silver,
  author    = {TLab Developer and Muhammad Faiz Khoiri},
  title     = {NER Irish Dataset 2},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/tlabdev/ner-irish-dataset-2}
}

Acknowledgments

• Manual annotation performed using Label Studio
• News articles collected from multiple Indonesian news portals
• Dataset prepared as part of an iterative NER model development pipeline