Developer fine-tunes Qwen3-4B to eliminate common AI writing tics
An independent developer released qwen3-4b-deslop, a fine-tuned version of Qwen3-4B-Instruct trained on a MacBook Air using 161 example answers to remove stylistic habits like em dashes, 'not X, but Y' constructions, and lists of three. The model is available as a 2.5GB download runnable via Ollama, and on 40 held-out test prompts its measured 'slop score' fell from 121 to 12.5 per 1000 words while maintaining accuracy on 30 factual questions.
GoKawiil's interpretation of the reporting above, not reported fact.
The project suggests that stylistic quirks associated with AI-generated text can be addressed through targeted fine-tuning rather than requiring larger models or runtime intervention, which could make small open-source models more useful for writing tasks where distinctive 'AI voice' is undesirable. The creator notes the approach needed a second filter to prevent a different failure mode—answers that avoided detected tics but became evasive or fabricated details—indicating such fixes can introduce new tradeoffs. The developer states the technique isn't inherently tied to model size and plans to test it on larger models next.
- The fine-tuned qwen3-4b-deslop model reduced measured AI writing tics (slop score) from 121 to 12.5 per 1000 words on held-out prompts.
- It maintained full accuracy on 30 factual test questions while eliminating em dashes entirely in testing.
- The training required a second quality filter after an intermediate version produced fabricated studies and repetitive sentence structures.
Apple MacBook Air M2 — This article's model was fine-tuned entirely on a MacBook Air, showing just how capable Apple's efficient M-series chips are for local LLM experimentation. If you want to try running or training small open models like Qwen3-4B yourself, a MacBook Air offers great performance per watt for these on-device AI workflows.
See Apple MacBook Air M2 on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.Source: 100toolkit.com — Erik Markus Kannike, 2026-10-19
Published there as: “Show HN: Qwen3-4B with the AI writing tics trained out”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.