Tempe, Arizona, United States
Low-Resource ASR via Multilingual Transfer Learning | Python, PyTorch, Hugging Face Transformers, Open AI Whisper
• Engineered a full training pipeline using Hugging Face, implementing cross-lingual transfer learning from Sinhalese to process the Dhivehi dataset.
• Achieved a 112% relative improvement in Word Error Rate (WER), reducing it from the baseline to 14.1% in 500 training steps.
• Assisted the professor in grading student assignments, providing constructive, timely feedback and ensuring fair evaluation.
Training a Custom Causal Language Model on Python Data Science Stack | Python, PyTorch, HF Accelerate, Transformers, GPT-2
• Fine-tuned GPT-2 (124M parameters) on curated code for data science Python libraries to create an intelligent code completion system, leveraging Hugging Face Transformers and Accelerate for distributed training.
• Achieved 56.06 perplexity score using custom loss functions prioritizing data science tokens, with optimized learning rate scheduling and real-time training analytics.
