ChakmaBERT (Latin Script)

** Foundational BERT model for Latin-script Chakma language from Northeast India.**

Model Details

• Architecture: XLM-RoBERTa base (100.6M parameters)
• Vocabulary: 18,920 tokens (custom trained on Chakma)
• Training Data: 41,140 sentences from Vaani ASR corpus
• Script: Latin/Roman transcription (not Chakma script)
• Final Training Loss: 5.50
• Training Time: 21.7 minutes on A40

Tokenization Efficiency

ChakmaBERT's custom tokenizer achieves 1.7x better efficiency compared to XLM-RoBERTa's multilingual tokenizer for Chakma text.

Usage

from transformers import AutoTokenizer, AutoModelForMaskedLM, pipeline

# Load model
tokenizer = AutoTokenizer.from_pretrained("MWirelabs/chakmabert")
model = AutoModelForMaskedLM.from_pretrained("MWirelabs/chakmabert")

# Fill-mask example
fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)
fill_mask("uh baluddur durot te ekko <mask> agey")

Training Details

• Base Model: XLM-RoBERTa initialized from scratch
• Epochs: 10
• Batch Size: 32
• Learning Rate: 5e-5
• MLM Probability: 15%

Dataset

Trained on conversational Chakma transcriptions from the Vaani corpus, representing authentic spoken Chakma from Northeast India.

Limitations

• Trained only on conversational data; may not generalize well to formal/written Chakma
• Latin script only; does not support Chakma script (𑄌𑄋𑄴𑄟𑄳𑄦)
• Limited dataset size (41k sentences)

Citation

Developed by MWire Labs for Northeast Indian language preservation and AI inclusivity.

License

CC-BY-4.0

Contact

MWire Labs - Shillong, Meghalaya, India