Fine-tune of Qwen3-4B cuts AI writing tics, released for Ollama
Original title: Show HN: Qwen3-4B with the AI writing tics trained out
On 40 held-out prompts the author's slop score fell from 121 to 12.5 per 1,000 words, with a 30-question quiz still correct.
Erik Kannike released qwen3-4b-deslop, a LoRA distillation of Qwen3-4B-Instruct trained on a MacBook Air to drop habits such as em dashes and "not X, but Y" without a system prompt. A Q4_K_M build is a 2.5 GB download for Ollama or llama.cpp. An earlier student scored 9.9 but missed tasks and invented studies; a third filter raised the score to 12.5 while task fit returned to about 80 percent. The author says the voice is flat and the model still invents names and places.
Unconfirmed: All quality figures are the author's own measurements, not an independent benchmark.
- The released model was trained on 161 filtered answers from 288 prompts.
- Em dashes fell from 20.5 per 1,000 words to zero on the author's test.
- Sixty-eight percent of held-out answers scored in the author's human range.
- It answered 30 of 30 factual questions correctly, according to the author.
- Products
- Ollama
- AI models
- qwen3-4b-deslopQwen3-4B-Instruct
From the original article
7 min read at 100toolkit.com
A small open model, fine-tuned on a MacBook Air, that writes without em dashes, "not X, but Y" or lists of three and needs no system prompt, and you can run it locally with Ollama. I trained the assistant writing habits out of Qwen3-4B-Instruct and released it as qwen3-4b-deslop . With Ollama it's ollama run hf.
By erikkannike · 100toolkit.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.