Large language models were tested on cancer patient notes from two academic medical centers and seven drug trials.