Show HN: Qwen3-4B with the AI writing tics trained out

In this report, a 4B AI model named Qwen3-4B has been developed at a home lab using a MacBook Air. The model was trained on 161 answers and focuses on removing AI writing tics. The process involved sampling one answer per prompt from the original model and the released version, covering ten topics in four genres. A style guide detector was used to assess the score based on various AI writing habits, including the use of em dashes, certain phrases, list structures, rhetorical questions, and specific words.

The goal was to achieve a score close to zero, similar to a human writer. Several checks were implemented to ensure the model didn't drift away from the original model's predictions, answer 30 questions, and maintain a level of factual knowledge. The model underwent distillation, a process where a teacher model was built to provide training prompts, and a small LoRA adapter learned to produce answers from the bare prompt.

The final model, Qwen3-4B, achieved a score of 9.9. However, it still exhibits some flaws, such as running long, producing inaccurate information, and using peculiar sentence structures. The authors are working on improving the model by providing excellent human writing examples as a base model, aiming to create a model with improved writing skills and habits.