suggestions for small change
try something like 80/20 to seek to train on more data for a better model (and reduce error compared to the 50/50 trained model ) AND to more greatly risk biasing your data to train over test
cost_split <- College %>% initial_split(prop = 0.8)
Also you could make a small refinement to the lamba selection, to such through more smaller numbers
lambdas_to_try <- seq(10^-5, 10^3, length.out = 10000)