
Qwen3-4B with the AI writing tics trained out
A fine-tuned, openly released version of Qwen3-4B-Instruct with common AI writing habits—em dashes, "not X, but Y" constructions, lists of three, filler signposts—trained out of the model's weights. It was distilled on a MacBook Air by Erik Markus Kannike and runs locally with Ollama or llama.cpp, no system prompt required.
Works with no system prompt: em dashes drop from 20.5 to 0 per 1000 words and the overall slop score falls from 121 to 12.5 per 1000 words on 40 held-out prompts, with 68% of answers scoring in the range of human writing.
The Q4_K_M quantised build is a 2.5 GB download that runs on a laptop via Ollama or llama.cpp; full-precision weights load with transformers like any Qwen3-4B checkpoint.
The repo publishes the merged weights plus the parts separately: a 126 MB LoRA adapter (trained with the edit active, so the two belong together) and a 175-byte neuron edit that flips four MLP weight columns.
The model still answers 30 of 30 factual questions correctly and matches requested genre and length as often as the original, with task fit back at 80% after a third training round.
Built via three rounds of distillation: a teacher model (steering + prompt + neuron patch) writes answers to 288 prompts, answers passing slop, damage and task-fit checks are kept (161 survived), and a rank-16 LoRA learns to produce them from bare prompts.
A teacher assembled from the best-performing combination (steered model, four-neuron patch, plain-style prompt) writes multiple answers per prompt. Answers are kept only if they score under 12 on the slop detector, pass damage checks, and satisfy rule-based task checks (stance in opinion pieces, fir
A weighted regex detector scores AI habits (em dash, 'it's not X, it's Y', lists of three, rhetorical questions, signposts, fragments, words like 'delve' and 'seamless') per 1000 words. Drift is measured as KL divergence from the original model on 22 human-written passages, plus a 30-question factua
Tested on ten unseen topics across four genres (blog post, LinkedIn post, explainer, opinion piece), one sampled answer per prompt at temperature 0.7, no system prompt.
Free and open. The model weights, LoRA adapter, and neuron edit are published on Hugging Face; the Q4_K_M GGUF build is a 2.5 GB download.
Free
qwen3-4b-deslop-v4 full-precision checkpoint, 126 MB LoRA adapter, 175-byte neuron edit, and Q4_K_M GGUF (2.5 GB) via hf.co/erikkannike
With Ollama: `ollama run hf.co/erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M` (a 2.5 GB download), or with llama.cpp: `llama-cli -hf erikkannike/qwen3-4b-deslop-v4-GGUF:Q4_K_M`. Full-precision weights load with transformers like any Qwen3-4B checkpoint.
No. The writing habits are trained into the weights, so it writes plainly from the bare user prompt with no system prompt and no runtime hooks.
Qwen3-4B-Instruct-2507 in bf16, with a rank-16 LoRA on all attention and MLP projections, trained for two epochs on bare user prompts with loss on the response only.
On 40 held-out prompts the slop score dropped from 121 to 12.5 per 1000 words, em dashes went from 20.5 to 0, and 68% of answers score in the range of human writing (under 15). It still answers 30 of 30 factual questions correctly.
Yes. The merged weights, the LoRA adapter, and the neuron edit are all released on Hugging Face. The adapter was trained with the edit active, so the two are meant to be used together.
Flat voice (short declarative sentences, diary-like personal pieces), invented names and places, answers that run long, and testing limited to short English prose in four genres—code, other languages, long conversations, and tool use are untested.
Building a teacher from excellent human writing (public-domain essays and letters plus modern pieces) to fine-tune a base model, then applying the same pipeline to 7–8B and ~30B models, plus a blind test where people pick the better text.