It utilizes a Two-Stage Curriculum Learning strategy to achieve higher accuracy and robustness compared to previous benchmarks (such as NusaBERT). Despite being a Large architecture (334M parameters), it achieves 46% higher throughput and lower latency than comparable Base models during inference.

🚀 Key Performance Highlights

Metric	vs. Previous SOTA (NusaBERT)
Accuracy (F1)	+3.54% higher on NER_UI dataset
Recall	+4.02% higher recall (misses fewer entities)
Speed	15.84 ms latency (vs 23.13 ms)
Throughput	63 samples/sec (vs 43 samples/sec)

🎓 Training Strategy (Curriculum Learning)

To achieve these results, we employed a novel "Silver-to-Gold" training pipeline. We did not simply concatenate datasets; instead, we trained the model in phases to simulate a learning curriculum.

Phase 1: The "Warm-up" (Silver Data)

Data Source: A massive synthetic dataset of 130,000+ sentences.
Collection Method: We translated high-quality English NER datasets into Indonesian and utilized GLiNER-Multi-v2.1 to automatically tag entities.
Objective: This phase allowed the model to learn general context, sentence structures, and the concept of 19 rich entity labels (Events, Art, Facilities, etc.) which are often missing in standard datasets.

Phase 2: The "Refinement" (Gold Data)

Data Source: NERGRIT, a high-quality, human-annotated dataset for Indonesian.
Method: We took the checkpoint from Phase 1 and performed precise fine-tuning on this Gold standard data with a lower learning rate.
Objective: This corrected any noise introduced by the synthetic data and aligned the model's predictions with proper Indonesian grammatical standards.

📊 Evaluation & Benchmarks

We evaluated this model against the previous best-performing model (cahya/NusaBert-ner-v1.3) using strict evaluation scripts on standard academic datasets.

1. Dataset: NER_UI (Universitas Indonesia)

This model achieves a significant improvement in Recall and F1 Score.

Rank	Model Name	F1 Score	Precision	Recall	Accuracy
🥇 1	indobert-ner-gold (Ours)	79.91%	79.54%	80.28%	94.49%
🥈 2	NusaBERT (Cahya)	76.37%	76.47%	76.26%	93.88%

2. Dataset: NER_UGM (Universitas Gadjah Mada)

Even on the challenging UGM dataset, our model maintains superior performance.

Rank	Model Name	F1 Score	Precision	Recall	Accuracy
🥇 1	indobert-ner-gold (Ours)	70.21%	61.74%	81.37%	93.48%
🥈 2	NusaBERT (Cahya)	67.49%	58.88%	79.07%	93.12%

⚡ Inference Efficiency (Speed)

We benchmarked the inference speed on a standard CUDA environment. Despite being a "Large" model (334M params) compared to NusaBERT's Base architecture (160M params), this model runs significantly faster.

This efficiency is attributed to the optimized tokenizer and the robustness of the Phase 1 training which allows for confident, faster convergence during inference.

Metric	Our Model (Large)	NusaBERT (Base)	Verdict
Parameters	334M	160M	Ours is heavier (More Knowledge)
Latency	15.84 ms	23.13 ms	⚡ 31% Faster
Throughput	63.12 samples/sec	43.23 samples/sec	🚀 46% More Capacity

🏷️ Supported Labels

Unlike standard models that only detect Person, Location, and Organization, this model supports 19 detailed entity tags (38 BIO labels + O = 39 total):

Label	Description	Label	Description
PER	Person	ORG	Organization
LOC	Location	GPE	Geopolitical Entity
FAC	Facility (Buildings, Airports, etc.)	EVT	Event
WOA	Work of Art	LAW	Law / Legal references
PRD	Product	LAN	Language
DAT	Date	TIM	Time
MON	Money	QTY	Quantity
CRD	Cardinal Number	ORD	Ordinal Number
PRC	Percent	NOR	Political/Religious Group
REG	Region

💻 How to Use

python

from transformers import pipeline

# Load the pipeline
ner = pipeline("ner", model="treamyracle/indobert-ner-gold", aggregation_strategy="simple")

text = """
Presiden Joko Widodo meninjau pembangunan Istana Negara di Ibu Kota Nusantara (IKN) 
pada hari Selasa tanggal 20 Januari 2024. Proyek senilai Rp 15 Triliun ini dikerjakan oleh PT Waskita Karya.
"""

results = ner(text)
print(results)