Base model versus assistant on conveyancing prompts

The script is small: one function decides what text each model actually sees, one wrapper runs greedy generation, and one loop puts the outputs side by side. The decision that matters most is making the input explicit: the base model gets the prompt as raw text, and the instruct model gets it wrapped by its own chat template, ending with an open assistant turn. Almost every difference you will see comes from that one wrapping plus the post-training that taught the model what an assistant turn is for.