Qwen3-4B fine tune aims to remove common AI writing habits

A fine tuned version of Qwen3-4B seeks to reduce recognizable AI writing patterns while remaining usable for everyday writing tasks. Erik Markus Kannike writes for 100toolkit.com that he trained a small open model to avoid habits such as em dashes, rigid contrasts, lists of three, rhetorical questions, and heavily signposted prose.

The resulting model, qwen3-4b-deslop, is available for local use through Ollama and llama.cpp. Kannike says the compressed download is about 2.5 GB, making it practical for many laptops.

In tests across 40 prompts that were not part of training, the author’s style detector measured a drop from 121 to 12.5 unwanted patterns per 1,000 words. The detector counts weighted occurrences of phrases and structures associated with formulaic assistant writing. Kannike says 68 percent of the new model’s responses fall within a range he associates with human writing.

Style gains came with tradeoffs

Kannike used a multi stage training process. A modified version of the original model first generated responses in a plainer style. He filtered these outputs for quality and task fit, then used them to fine tune the model with a LoRA adapter. He also applied an edit affecting four neurons, which was intended to curb one recurring contrast pattern.

An earlier version achieved an even lower style score, but produced weaker writing. It sometimes avoided taking a position in opinion pieces, invented academic studies, and overused sentence joins such as “which” and “so.” The final filtering process added checks for requirements such as first person voice, greetings in emails, requested length, and early positions in opinion pieces.

The author reports that factual performance remained stable in a 30 question quiz and that task fit recovered to 80 percent. Still, he warns that the model’s voice can sound flat and that it continues to invent names and locations. The evaluation also relies on rule based tests created by the author, rather than a blind study with readers.