mDeBERTa-Hybrid-30k

A vocabulary-pruned version of microsoft/mdeberta-v3-base with a 30k-token hybrid Indonesian–English vocabulary, designed for downstream Indonesian and Indonesian–English NLP tasks.

This model was developed using VocabPrune, a deterministic, language-aware, frequency-based vocabulary pruning method designed to reduce vocabulary-related model overhead while preserving the original Transformer architecture.

The hybrid vocabulary uses an approximately 70% Indonesian / 30% English configuration.

The model is a base checkpoint and should be fine-tuned for a specific downstream task.

Model Details

Property	Value
Base model	microsoft/mdeberta-v3-base
Vocabulary size	30k tokens
Vocabulary	Hybrid Indonesian–English
Language focus	Indonesian with English subword coverage
Architecture	mDeBERTa-v3-base

Resources

• Paper: IEEE Access
• Paper DOI: 10.1109/ACCESS.2026.3679735
• Source Code: GitHub – VocabPrune
• Reproducibility Archive: Zenodo – VocabPrune v1.0.1
• Model Collection: Pruned mDeBERTa

For the methodology, experimental setup, and detailed evaluation results, please refer to the published paper.

Citation

If you use this model or the VocabPrune methodology in your research, please cite:

@article{fuadi2026efficient,
  author  = {Fuadi, Mukhlish and Wibawa, Adhi Dharma and Sumpeno, Surya},
  title   = {Efficient Transformer Models via Language-Aware
             Frequency-Based Vocabulary Pruning},
  journal = {IEEE Access},
  volume  = {14},
  pages   = {50993--51006},
  year    = {2026},
  doi     = {10.1109/ACCESS.2026.3679735}
}