tlabdev/ner-irish-roberta-base

About this Model

Fine-tuned XLM-RoBERTa Base model for Named Entity Recognition (NER) on Indonesian news articles, with built-in rule-based post-processing for Indonesian text.

Model Performance

Metric	Score
F1	0.9120
Precision	0.8928
Recall	0.9320
Accuracy	0.9779

Evaluated on held-out test set

Supported Entities

The model recognizes 9 entity types commonly found in Indonesian news:

• PER - Person names
• ORG - Organizations
• GPE - Geopolitical entities (countries, cities, states)
• LOC - Locations (non-GPE)
• DATE - Dates and time periods
• EVENT - Named events
• FAC - Facilities
• MONEY - Monetary values
• LAW - Laws and regulations

Quick Start

from transformers import pipeline

# Load the NER pipeline
ner = pipeline("token-classification", model="tlabdev/ner-irish-roberta-base", aggregation_strategy="simple")

# Run inference
text = "Gubernur Jawa Barat meresmikan proyek senilai Rp 10 miliar."
results = ner(text)

# Display results
for entity in results:
    print(f"{entity['word']} -> {entity['entity_group']} (score: {entity['score']:.2f})")

Use Cases

Recommended for:

• News article analysis and information extraction
• Entity-based search and retrieval systems
• Financial and regulatory document processing
• Indonesian-language knowledge graphs

Limitations:

• Optimized for formal Indonesian news text
• Not designed for informal language or slang
• Single language inference only

Training Details

Hyperparameters

Learning rate: 3e-5
Train batch size: 4
Eval batch size: 8
Gradient accumulation steps: 4
Effective batch size: 16 (4 × 4)
Epochs: 10
Weight decay: 0.02
LR scheduler: Linear
Warmup ratio: 0.03
Max gradient norm: 0.5
Label smoothing: 0.05
Mixed precision: FP16
Optimizer: AdamW (default)

Training Progress

Epoch	Train Loss	Val Loss	Precision	Recall	F1	Accuracy
1	0.4338	0.4000	0.8563	0.8939	0.8747	0.9781
2	0.3984	0.3939	0.8620	0.9190	0.8896	0.9801
3	0.3823	0.3894	0.8749	0.9210	0.8973	0.9812
4	0.3702	0.3959	0.8822	0.9295	0.9052	0.9806
5	0.3631	0.3892	0.8955	0.9178	0.9065	0.9820
6	0.3597	0.3938	0.8932	0.9302	0.9113	0.9824
7	0.3534	0.3975	0.8864	0.9234	0.9046	0.9817
8	0.3468	0.3985	0.8922	0.9105	0.9013	0.9805
9	0.3452	0.3975	0.8941	0.9224	0.9080	0.9820

Best Validation: Epoch 6 (F1: 0.9113, Precision: 0.8932, Recall: 0.9302, Accuracy: 0.9824)

Final Test Results

Metric	Score
Loss	0.4107
Precision	0.8928
Recall	0.9320
F1	0.9120
Accuracy	0.9779

Technical Specifications

Base Model: xlm-roberta-base
Framework: Transformers 4.57.3, PyTorch 2.1.0+cu124
Language: Indonesian
Task: Token Classification (NER)

License

MIT License

This fine-tuned model inherits the MIT License from the base model XLM-RoBERTa.

Citation

@misc{ner-irish-roberta-base,
  author = {TLab Developer and Muhammad Faiz Khoiri},
  title = {ner-irish-roberta-base},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/tlabdev/ner-irish-roberta-base}}
}

Please also cite the original XLM-RoBERTa paper:

@article{DBLP:journals/corr/abs-1911-02116,
  author    = {Alexis Conneau and
               Kartikay Khandelwal and
               Naman Goyal and
               Vishrav Chaudhary and
               Guillaume Wenzek and
               Francisco Guzm{\'{a}}n and
               Edouard Grave and
               Myle Ott and
               Luke Zettlemoyer and
               Veselin Stoyanov},
  title     = {Unsupervised Cross-lingual Representation Learning at Scale},
  journal   = {CoRR},
  volume    = {abs/1911.02116},
  year      = {2019},
  url       = {http://arxiv.org/abs/1911.02116},
  eprinttype = {arXiv},
  eprint    = {1911.02116}
}