After four years of fine-tuning, one user concludes LLMs still can't write good prose

kompania · reddit · 2026-11-12

A local-LLM user with four years of experience — many QLoRA runs and full fine-tunes on RunPod, hundreds of fine-tuned models in English, Chinese and Russian — asks a fundamental question: can LLMs write good prose at all?

His evidence: everything from 100-epoch FFT at lr 1e-7 to aggressive LoRA/QLoRA configs barely rises above mediocrity, across his own prose datasets and others'. Gemma, Qwen, Mistral, LFM, IBM and from-scratch niche models all produce text only marginally better than out-of-the-box output. Genre doesn't help — models can't even capture Harry Harrison's straightforward style, let alone Dune; comedy, drama, adventure, romance and historical fiction all come out generic. He calls EQBench completely unreliable, concluding real literary quality still requires writing himself and using LLMs only for proofreading. Facing the task of writing descriptions for 100,000 records, he wonders if it's even worth fine-tuning given 100–10,000 examples yielded nothing, and asks if anyone has produced a local model writing at the level of Dune or Martin Eden.

Original post →

After four years of fine-tuning, one user concludes LLMs still can't write good prose

More from Models