qwen3-4b-deslop: 161 examples trained on a MacBook Air cut slop score 121→12.5 with 68% human-range output
lmoroney · x · 2026-10-20
Erik Kannike's full write-up on training the assistant writing habits out of Qwen3-4B-Instruct on a MacBook Air using just 161 answers, released as qwen3-4b-deslop (runnable locally via Ollama, 2.5GB).
Key results:
- On 40 held-out prompts, slop score dropped from 121 to 12.5 per 1,000 words, with 68% of answers scoring in the human-writing range; em dashes fell from 20.5 to zero per 1,000 words.
- 30/30 factual questions still correct; genre and length adherence matches the original.
- An intermediate version scored even lower (9.9) but wrote badly: opinion pieces without opinions, invented studies, and a new ", which" habit—fixed by adding a task-fit filter checking whether the answer did what the prompt asked.
- Limitations: flat voice, still invents names and places.
- The recipe is model-size agnostic; scaling to larger models is next.
Laurence Moroney highlights the task-fit filter as the part to study: check whether your answer did the job before fussing over punctuation.
Related event: Developer Fine-tunes Qwen3-4B on a MacBook Air to Remove AI Writing Style(2 posts)→
More from Models
- Hidden "6.5" found in source code of OpenAI's new opensi.lol domain, countdown points to Nov 1 — kimmonismus · 2026-10-20
- Codex quota keeps getting reset: users say gpt-6.1-sol barely dents limits, want it to stay — bytebot · 2026-10-20
- Fudan & Shanghai AI Lab map PEFT recipes for MoE LLMs: tune experts alone, LoRA/DoRA win — kalyan_kpl · 2026-10-20
- Reddit User: New Models Have Mostly Stopped Hallucinating — Have LLMs Crossed the Reliability Line? — LiveNotWork · 2026-10-20
- 26-30B is the best dense model size range, argues ML commentator — auto_grad_ · 2026-10-20
- Researchers push back on Opus 4/4.1 deprecation, citing alignment and model welfare concerns — Sauers_ · 2026-10-20