High impact

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

What happened

A small open model called qwen3-4b-deslop was fine‑tuned on a MacBook Air to remove common AI writing habits such as em dashes, “not X, but Y” phrasing, and three‑item lists. Using 161 answer examples and a weighted slop detector, the author reduced the habit score from 121 to 12.5 per 1,000 words on 40 unseen prompts, while still answering factual questions correctly and matching the original model’s genre and length performance. The model is available via Ollama as a 2.5 GB GGUF file. An intermediate version showed poor output, which was fixed with a second‑stage filter that checks prompt compliance.

Why it matters

For everyday users, the model produces cleaner, more human‑like text without the typical AI quirks, making AI‑generated content easier to read. Developers gain a lightweight method to improve writing style without large‑scale training resources.

Reality check

An automated, sceptical read of the source

Mostly marketingIncrementalNotableGenuine breakthrough

Incremental. The work refines an existing model’s output style rather than introducing a fundamentally new capability.

Who it affectsDevelopersEveryday usersResearchers

What to watchFuture work will test the same distillation and habit‑removal approach on larger models to see if the style improvements scale.

Do you agree with this rating?Agree Disagree

More from Alibaba

Qwen-Image-2.1-TurboAlibaba · 9 Oct 2026Alibaba open-sources AI model that can detect cancer and nearly 150 conditionsAlibaba · 18 Sept 2026Qwen-Image-2.1Alibaba · 14 Sept 2026Show HN: I made a Raspberry with Qwen my local car AIAlibaba · 25 Aug 2026Qwen3.8-Flash-NextAlibaba · 24 Aug 2026

This page was collected from 100toolkit.com and summarised automatically by AIPulsee's pipeline. See how it runs.