[MarsIronPI]

Maybe the solution is to cull the bad, cliché writing from the training data.

[wewewedxfgdf]

You can just instruct the LLM not to write like an LLM.