The quantised 2.5 GB build of Qwen3-4B-Deslop runs locally via Ollama, demonstrating on‑device style control.
QWEN3-4B DESLOP MODEL SLASHES AI WRITING TICS BY 90%
*Erik Kannike’s fine‑tuned Qwen3‑4B‑Deslop cuts stereotypical AI habits to near‑human levels. On 40 unseen prompts the slop score drops from 121 to 12.5, while factual accuracy stays perfect.*
Developer Erik Kannike released qwen3-4b-deslop, a fine‑tuned version of Qwen3‑4B‑Instruct that strips out the stereotypical AI writing tics that have plagued large language models. On a set of 40 unseen prompts the model’s slop score fell from 121 to 12.5 per 1,000 words, putting 68 % of its answers in the human‑writing range. Em dashes vanished entirely, dropping from 20.5 to zero per 1,000 words, while factual accuracy remained perfect – 30 of 30 quiz questions were answered correctly. The work was done on a MacBook Air, using only 161 curated answers and a small LoRA adapter, proving the technique scales without massive compute.
How did the slop score drop so dramatically?
Kannike built a slop detector from a style guide that flags em dashes, "it's not X, it's Y" constructions, three‑item lists, rhetorical questions, signposts and buzzwords like "delve". The original Qwen3‑4B‑Instruct averaged 121 weighted hits per 1,000 words. By distilling a teacher model that answered 288 prompts, filtering out answers that failed task‑fit checks, and training a LoRA adapter on the 171 best outputs, the new model achieved a score of 12.5. The process removed em dashes entirely (down from 20.5 to zero) and reduced " , which" joins from 1.4 to 0.3 per 100 words. Human‑range scores rose to 68 %.
What training pipeline produced the released model?
The pipeline began with an earlier experiment that subtracted a direction vector during generation, lowering the slop score to about 15 but requiring runtime steering. A second experiment flipped four neurons to erase 70 % of the "not X, but Y" habit. Kannike then combined the best tricks in a distillation loop: a teacher model answered 288 prompts, kept only answers passing damage and task‑fit checks, and taught a small LoRA adapter to reproduce them from a bare prompt. Round one yielded a score near 20; round two added the four‑neuron patch and a plain‑style prompt, reaching 9.5 before steering was removed. Round three refined the teacher prompt to enforce position statements, first‑person voice, varied sentences, and no invented studies, producing the final release.
Did the de‑slopped model retain factual ability?
Yes. On a 30‑question factual quiz the model answered every question correctly, matching the original. However, task‑fit checks showed a dip during round two: only 60 % of outputs met prompt requirements versus 78 % for the baseline. After round three, task‑fit rebounded to 80 %, and invented studies and fabricated percentages nearly disappeared. The model still invents names and places, and its voice is described as "plain to the point of flat," but it now respects genre, length and structural cues as reliably as the original.
What does this mean for on‑device AI?
The quantised Q4_K_M build is only 2.5 GB and runs on a laptop via Ollama or llama.cpp, demonstrating that sophisticated style‑control can be achieved without cloud‑scale hardware. Kannike’s home‑lab setup—a MacBook Air—trained the 4‑billion‑parameter model on just 161 curated answers. The recipe is model‑size agnostic, suggesting larger models could be de‑slopped with similar compute budgets. If the technique spreads, developers could ship AI assistants that write without the tell‑tale AI tics, improving user trust and reducing the need for post‑generation filters.
Frequently asked questions
What is the slop score and how did it change?
The slop score measures weighted AI writing habits per 1,000 words; it fell from 121 in the original model to 12.5 in the released qwen3-4b-deslop.
Did the de‑slopped model lose factual accuracy?
No. It answered all 30 factual quiz questions correctly, matching the original model's performance.
Can the de‑slopping technique run on consumer hardware?
Yes. The 2.5 GB quantised build runs on a MacBook Air via Ollama or llama.cpp, showing the method works on modest devices.
Kannike’s de‑slopped Qwen3‑4B shows that style‑aware fine‑tuning can be both lightweight and effective. As the recipe works on larger models, we may soon see on‑device assistants that write like humans without the crutches that betray their synthetic origin. The next wave of AI will be judged not just on knowledge but on the subtlety of its prose.
Sources: Hacker News (Show HN: Qwen3-4B with the AI writing tics trained out)