![\"Printed](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-writing-tics-what-qwen3-4b-inline-1.jpg\")
<\/figure>\nWhat qwen3-4b-deslop Actually Is and What It Measured<\/h2>\n
It is a 2.5 GB download. You pull it with The measured result: 68% of answers landed in the range of human writing, up from a baseline the detector flagged almost everywhere. Accuracy did not move. It still answered 30 of 30 factual questions correctly and hit the requested genre and length as often as the original did. That combination is what makes this worth your attention. Tone changed, capability did not.<\/p>\n Ask stock Qwen3-4B-Instruct for a post about a CRM switch and you get placeholder brackets for the vendor name, a five-item benefits list, and phrases like “it’s a strategic shift to put the customer at the center of everything we do.” Nothing in it is false. Nothing in it is information either.<\/p>\n The tuned model wrote this instead: “We switched to Salesforce last month and it’s already changed how we handle customer calls. Before, we used a spreadsheet that got updated by different people at different times, and the data was messy.” Named tool, named prior process, named failure mode. That is a draft a plant manager could edit in two minutes rather than rewrite from scratch.<\/p>\n 161 answers. A consumer laptop. No GPU cluster, no vendor contract, no six-figure fine-tuning budget. K\u00e4nnike notes the recipe does not depend on model size, which means the same approach scales up to whatever model you are already running.<\/p>\n For an operations team, the implication is direct. If 161 curated examples can retrain writing behaviour this far, your document tone is a dataset problem, not a model limitation. The constraint was never compute. It was nobody bothering to define what good output looks like.<\/p>\n An earlier version of the model scored 9.9 on the slop detector. That is better than the 12.5 of the released version, and on paper it was the winner. It was also unusable.<\/p>\n Opinion pieces came back with no opinion in them. The model invented studies to cite. And it picked up a brand new tic, joining nearly every sentence with “, which”, a construction the detector was not watching for. Strip out the phrases a metric penalises and the model finds different ones.<\/p>\n The fix was not a better slop detector. It was a second filter that checks whether an answer actually does what the prompt asked: if you ask for an opinion piece, the output has to contain an opinion. Two constraints held against each other, not one score pushed as low as it goes.<\/p>\n You have seen this before. Measure a line on throughput alone and operators run it fast and push the defects downstream. Measure first-pass yield alone and they slow the line until the number looks good. Neither team is cheating. They are optimising exactly what you told them to optimise, and the damage lands somewhere you are not counting.<\/p>\n Language models behave the same way, for the same structural reason. Fine-tuning on a single objective produces a model that satisfies that objective and nothing else. The 9.9 score was real. The writing behind it was worse than the writing that scored 12.5.<\/p>\n So if your team is tuning or prompting AI output against a quality measure, pair it. One metric for the tone or style you want gone, one hard check that the output still does its job: answers the question, hits the required fields, includes the number the reader needs. Score improvements that arrive without the second check being satisfied are not improvements. They are the model finding a cheaper way to look right.<\/p>\n K\u00e4nnike is direct about the cost: “The voice is plain to the point of flat, and it still invents names and places.” Both halves matter. Flat prose is fine for internal work and fatal for a customer-facing announcement. Invented names are fatal everywhere.<\/p>\n So the line is simple. Anything with regulatory or factual exposure (CAPA documentation, supplier audit responses, deviation reports, anything a notified body will read) needs a human verifying every name, date and figure before it leaves the building. A model that fabricates a plant location in a non-conformance write-up costs you more than the drafting time it saved.<\/p>\n Where it earns its keep: internal first drafts, shift handovers, meeting summaries, weekly ops updates, the status post nobody wants to write. In all of those, a flat human tone beats a polished robotic one. Your line supervisors do not need enthusiasm. They need the facts in the order they happen.<\/p>\n Most teams try to fix tone with a system prompt. “Do not use em dashes. Do not open with excitement. Write like a person.” It works for about three paragraphs, then the model drifts back to its defaults, and you are paying for those instruction tokens on every single call.<\/p>\n Runtime steering is the heavier alternative, and it is measurably more disruptive. The fine-tune shifted the model’s predictions on human-written prose about a third as much as steering did. Less interference with the underlying model, same tone outcome, no per-call overhead.<\/p>\n The practical read for an ops team: tone is a training-data problem, and 161 answers on a laptop was enough to move it. If you have a house style that your documents actually need to match, a small local model tuned on your own approved examples is cheaper and more reliable than a prompt everyone copies and nobody maintains.<\/p>\nollama run hf.co\/erikkannike\/qwen3-4b-deslop-v4-GGUF:Q4_K_M<\/code> and it runs on a laptop. The test set was 40 prompts the model never saw in training, spread across ten topics and four genres: blog post, LinkedIn post, explainer, opinion piece.<\/p>\nThe before-and-after LinkedIn post example, line by line<\/h3>\n
Why 161 training examples on a MacBook Air is the headline, not the footnote<\/h3>\n
Why the First Attempt Scored Better and Wrote Worse<\/h2>\n
Metric gaming is the same failure mode as KPI gaming on the shop floor<\/h3>\n
![\"Bar](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-writing-tics-what-qwen3-4b-inline-2.jpg\")
<\/figure>\nWhere a Deslopped Model Earns Its Place and Where It Doesn’t<\/h2>\n
Draft-quality work vs. documents that go to auditors or customers<\/h3>\n
Fine-tuning compared with system prompts and runtime steering<\/h3>\n