qwen3-4b-deslop - AI model fine-tuned to write without AI writing tics
Show HN: Qwen3-4B with the AI writing tics trained out
qwen3-4b-deslop is a fine-tuned version of Qwen3-4B-Instruct that strips out telltale AI writing habits like em dashes, "not X, but Y" constructions, and lists of three. Trained on a MacBook Air using distillation and LoRA adapters, it needs no system prompt and runs locally via Ollama with a 2.5 GB download. On 40 held-out prompts, its slop score dropped from 121 to 12.5 per 1000 words, with 68% of answers landing in the human writing range, while still answering factual questions correctly and matching requested genres and lengths.
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.