Hieu Nguyen Email & Phone Number

Los Angeles, CA, US

• Run sequence/token level knowledge distillation of Llama-3-8B with Llama-3-70B as teachers and evaluate on hallucination benchmarks (HaluEval, XSum, CNN Dailymail) in H100 clusters.
• Instruction-tuned Llama2-7b, Mistral7B on multiple choice QA task and evaluated overconfidence.
• Implemented a scalable framework for large scale distributed experiments of (multi-nodes multi-processes) hallucination evaluation of distilled LLMs in USC’s HPC system.

Los Angeles, California, United States

• Implemented bug Transformer decoder models (architectural bugs, training/inference setting bugs).
• Ran and monitored performance of 4000 distributed experiments of parameter-efficient training (quantization, LoRA adapters, gradient checkpointing/accumulating, etc.) of transformers models on GPU clusters.
• Researched on ANOVA and surrogate modeling for detecting bugs from model evaluation results.
• Implemented literature methods for benchmarking (collecting model logits/activation/gradients).

Los Angeles, California, United States

• Unit tested Ablator (a framework for efficient scaling of thousands of deep learning experiments in large clusters) and partake in code reviews.
• Contributed to the documentation and tutorials of Ablator, (e.g. setting up AWS cloud cluster and Ray for automatic distributed training with Ablator).

Ho Chi Minh City, Vietnam

• Develop task-oriented Rasa chatbots that interact with customers, assisting them in completing tasks such as booking services, and responding to their inquiries.
• Built, trained, and deployed a Transformer Decoder model that restores accent marks.
• Set up a multi-modal virtual assistant (text and audio) pipeline of docker microservices and integrated with Meta apps.

Ho Chi Minh City, Vietnam

Master Of Science - Ms, Computer Science

Bachelor'S Degree, Computer Science