Finetuning LLMs for Human Behavior Prediction in Social Science Experiments