Technical Staff – Synthetic Data Generation and Model Fine-Tuning
As a technical staff member at CaML, I contributed to robustly internalizing open-minded and compassionate values in LLMs using synthetic data training. My responsibilities included generating and curating diverse synthetic datasets, executing model fine-tuning, and designing novel evaluations of model outputs. I developed scalable pipelines and ensured the quality and diversity of training data in a mission-critical AI alignment context. • Generated synthetic text data in batches of 1,000–3,000 to train large language models • Managed and performed fine-tuning of open source 8B parameter models using Unsloth • Developed and administered rigorous 20+ question evaluations for model behavior toward animals and digital minds • Implemented additional evaluation and attack methodologies, adapting to new research agendas as needed.