Show HN: Qwen3-4B fine-tuned to drop em dashes cut its 'slop score' from 121 to 12.5

What was built

Kannike's write-up describes fine-tuning Qwen3-4B-Instruct on a MacBook Air to stop common AI writing habits: em dashes, "not X, but Y" constructions and lists of three. A teacher setup built from earlier steering and neuron-editing experiments plus a style prompt wrote 288 training answers; those that passed style and task-fit checks were kept, and a LoRA adapter learned to produce them from bare prompts. An intermediate round overused ", which" joins and invented studies, so the final round added task-fit checks.

Results and caveats

On 40 held-out prompts the slop score fell from 121 to 12.5 per 1,000 words and 68% of answers fell in the human-writing range. Em dashes dropped from 20.5 to zero, 30 of 30 factual questions were answered correctly, and task fit was 80% against the original's 78%. The author admits the voice is flat, the model still invents names and places, answers run long, tests covered only short English prose, and the score is regex-based with no blind reader study. The model is on Hugging Face; a 2.5 GB quantized build runs with ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M.

The thread's two objections

One commenter asked why anyone would support this, arguing LLM text should stay spottable. Another noted that any dominant model's style becomes "the AI style" readers learn to recognize, so removing today's tics only sets up tomorrow's. Meanwhile a 17-year-old programmer said it's a slight improvement but still obviously AI-written.