What I see even in good models is that when you ask something hard or impossible (but looking routine) instead of replying “I cannot” they hallucinate. A better dataset would help only to solve problems which can be solved (based on this dataset).