A small open model, Qwen3-4B-Instruct, has been fine-tuned to write without the habits that mark AI-generated text. Released as qwen3-4b-deslop, it avoids em dashes, "not X, but Y" constructions and lists of three, and needs no system prompt. It can be run locally with Ollama using ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF, a 2.5 GB download in Q4_K_M format.
On 40 prompts the model never saw in training, the slop score dropped from 121 to 12.5 per 1000 words, and 68% of answers fall in the range of human writing. Em dashes fell from 20.5 per 1000 words to zero. The model still answered 30 of 30 factual questions correctly and fits the requested genre and length as often as the original.
An intermediate version scored 9.9 but wrote poorly, producing opinion pieces without an opinion, invented studies and a habit of joining sentences with ", which". A second filter checking whether answers match the prompt's request was added to fix this. The resulting voice is plain to the point of flat, and the model still invents names and places.
The work was done on a Mac Book Air using 161 training answers. The author says the method does not depend on model size, and applying it to larger models is the next step.
Source: Hacker News · Summarized by HeadlinesBriefing