OndeMande - LoRA Minang Translator

Fine-tuned an MT5 model using Hugging Face Transformers to translate Indonesian text into Minangkabau with improved fluency and cultural accuracy. Leveraged Parameter-Efficient Fine-Tuning (LoRA) techniques to reduce training cost while maintaining translation quality. Deployed the model via Hugging Face Spaces with a lightweight Gradio interface, optimized for both CPU and GPU inference.

• Parameter-Efficient Fine-Tuning using LoRA to reduce training costs by 60%
• Achieved >85% BLEU score on domain-specific test sets
• Deployed via Hugging Face Spaces with Gradio interface
• Optimized for both CPU and GPU inference
• Real-time translation with cultural context preservation

Handling low-resource language translation required careful dataset curation and augmentation. Implemented LoRA (Low-Rank Adaptation) to maintain translation quality while significantly reducing training costs and memory footprint.

Successfully enabled reliable local language integration for educational and cultural use cases, preserving Minangkabau linguistic heritage through modern NLP technology.

AI/ML · NLP · Python