Research Intern at Andalusian Research Institute in Data Science (DaSCI)

Conducted research in LLM alignment focusing on preference optimization methods (DPO and GRPO), studying the impact of replacing KL divergence with alternative divergences in the training objective; derived estimators for f-divergences using Monte-Carlo estimation for empirical comparisons in GRPO objectives; ran and monitored experiments in a Linux + GPU environment, ensuring full traceability of hyperparameters and metrics using Weights & Biases; used Hugging Face tooling (TRL) for training and Evalchemy to evaluate models and compare against baselines; documented experimental setups, results, and conclusions in internal technical reports and research notes.