Show HN: qwen3-4b-deslop fine-tune trains AI writing tics out of Qwen3-4B
A developer released qwen3-4b-deslop, a LoRA fine-tune of Qwen3-4B-Instruct-2507 trained to suppress assistant-style writing habits, distributed as full-precision weights and a 2.5 GB Q4_K_M GGUF build for Ollama and llama.cpp. Using the author's own regex-based slop detector, the score on 40 held-out prompts fell from 121 to 12.5 per 1,000 words, em dashes dropped from 20.5 per 1,000 words to zero, and 68% of answers landed in the range the author associates with human writing; an intermediate version scored 9.9 but produced opinion pieces without a stance and invented studies, which a second task-fit filter addressed at the cost of a slightly higher score. The author reports the model still answered 30 of 30 factual questions correctly and matched requested genre and length as often as the original, but cautions that no blind reader test has verified the score, that the task-fit checks are simple rules and the training data was filtered on the same checks, and that performance on long conversations, other languages and tool use may be no better than the original or worse.
The author says the distillation recipe does not depend on model size and that running it on 7–8B and ~30B models, plus a blind human preference test, are the planned next steps.
TL;DR
- I trained the assistant writing habits out of Qwen3-4B-Instruct and released it as qwen3-4b-deslop. With Ollama it's
ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M, a 2.5 GB download. - On 40 prompts it never saw in training, the slop score dropped from 121 to 12.5 per 1000 words, and 68% of answers score in the range of human writing. Em dashes went from 20.5 per 1000 words to zero.
- It still answers 30 of 30 factual questions correctly and fits the requested genre and length as often as the original. Its predictions on human-written prose moved about a third as much as with runtime steering.
- An intermediate version scored 9.9 and wrote badly: opinion pieces without an opinion, invented studies, and a new habit of joining every sentence with ", which". Fixing that took a second filter that checks whether an answer does what the prompt asked.
- The voice is plain to the point of flat, and it still invents names and places.
My home lab is a MacBook Air, so this is a 4B model trained on 161 answers. The recipe doesn't depend on the model size, and running it on larger models is the next step.