LLMs only reach the performance they do because of the sheer scale of data they ingest. Training them on less data doesn't work as well, or at least you will overfit like crazy on anything the size of current models. So the question is where are you going to get anywhere near the volume of verilog code as is present in The Pile? The total amount of verilog ever written is almost certainly a few orders of magnitude less.