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

Erik Markus Kannike trained Qwen3-4B-Instruct to remove common AI writing tics - em dashes, "not X, but Y" constructions, lists of three, rhetorical signposts - and released the result as qwen3-4b-deslop, a 2.5 GB quantized GGUF that runs locally with Ollama or llama.cpp. On 40 held-out prompts the custom "slop" detector score dropped from 121 to 12.5 per 1,000 words, em dashes went from 20.5 to zero, and 68% of answers landed in the human-writing range; factual accuracy remained intact on a 30-question quiz and genre/length compliance matched the original. Prediction drift on human prose was substantially lower than runtime steering.

The training pipeline combined a small LoRA adapter (rank-16) with a four-neuron sign-flip edit and iterative distillation: a teacher produced filtered answers, a student learned from them, and task-fit rule checks were added after an intermediate model produced errors (opinion pieces without stance, invented studies, and excessive sentence-joining like ", which"). The final model was trained on 161 kept answers and restored task fit to about 80% while keeping slop low. Limitations include a flat, declarative voice, residual hallucinated names/places, tendency to run long, testing limited to short English prose, and reliance on a regex-based style detector; planned next steps are higher-quality human teachers, scaling to larger models, and blind reader evaluation.