A small Qwen3 model cuts AI writing tics from a score of 121 to 12.5

A developer released a version of Qwen3-4B, a small model, with the habits of AI writing removed. The habits include the em dash, lists of 3 and stock words. The score is a weighted count of these habits per 1,000 words. The base model averages 121, and the new model scores 12.5. It is a 2.5 GB download that runs on a laptop.

How it works

A teacher model wrote answers to training prompts. The developer kept only the answers that passed checks for style, damage and task fit. A LoRA adapter, a small add on file that changes a model, learned from the 161 best answers.

The numbers

The release does what the prompt asks 80% of the time, against 78% for the base model. The step before it did this only 60% of the time. Invented studies and percentages nearly disappeared.

What is not known

The score is a style guide that the developer wrote as rules. No blind test with readers has checked it. The model continues to invent names and locations, is too long and has a flat voice. Code, other languages and tool use are not tested.

Insight: Checks and teachers set the quality more than model size

Tests this month found that the training data matters. Data from the Tracer pipeline improved Qwen3 models of each size, and top-ranked coding trajectories gave more points than random ones. This developer found that a student model is only as good as the best answers of its teacher. This reads as a move of value from the model to the checks and the teacher: the person who owns them sets the quality.