Qwen3-4B Deslop: Training Out AI Writing Tics

Model Overview and Performance Metrics

A fine-tuned version of Qwen3-4B, released as qwen3-4b-deslop-v4, successfully minimizes common 'AI writing tics' while maintaining high factual accuracy and genre adherence. This small, open model can be run locally using Ollama with a 2.5 GB download. The model was trained on 161 answers using a MacBook Air and requires no system prompt to operate.

On 40 held-out prompts, the slop score dropped dramatically from 121 (the original model's average) to 12.5 per 1000 words. Furthermore, 68% of the de-slopped answers scored within the range of human writing. The model successfully eliminated em dashes, reducing their frequency from 20.5 per 1000 words to zero.

Despite the stylistic improvements, the model retains some limitations. Its voice is described as flat, consisting of short declarative sentences, and it still occasionally invents names and places. It also tends to run long, often exceeding the requested word count.

Defining the AI Slop Score

The slop score is a quantitative metric derived from a custom detector designed to identify specific stylistic habits common in AI-generated text. This score is a weighted count per 1000 words. The habits tracked include: the use of em dashes, phrases structured as "not X, but Y," lists of three, rhetorical questions answered immediately, signposts like "here's the thing," short fragments for emphasis, and specific words such as "delve" and "seamless."

A paragraph written by a careful human writer typically scores close to zero. The original Qwen3-4B-Instruct model averaged a slop score of 121.

Training Methodology

The de-slopping process utilized a multi-stage distillation and targeted neuron editing approach. Initially, an earlier experiment identified four specific neurons whose sign flip could remove 70% of the "not X, but Y" habit with minimal cost. This neuron edit was combined with a teacher model built from the most successful techniques.

Distillation was then used to train a student model. The teacher model answered 288 training prompts, and only answers that passed strict checks were kept. The first round of training, without prompts or hooks, lowered the score to about 20. Round two incorporated the neuron patch and plain-style prompting, achieving a score of 9.5, though this version introduced new habits, such as joining nearly every sentence with ", which."

To correct these new stylistic regressions, a rule-based check was added to ensure the answer fulfilled the prompt's requirements (e.g., an opinion piece must state a position early). The final, released model (Round three) was trained on answers that scored under 12, passed damage checks, and passed all task-fit checks. This resulted in the final score of 12.5, while task fit returned to 80% and the frequency of ", which" joins fell significantly.

Task-Fit and Damage Checks

To ensure the model remained functional, several checks were implemented alongside the slop score:

• Task-Fit Checks: These rule-based checks verify that the output meets the prompt's intent. For instance, an opinion piece must state a position early, and a personal reflection must be written in the first person. The final model achieved 80% task fit.
• Damage Checks: These measures monitor the model's prediction drift on human prose. They ensure the model retains its general knowledge and does not produce broken output, such as lowercase sentence starts or a loss of topic.

Community and Expert Commentary

Discussion surrounding the model highlighted several key points regarding the nature of AI writing:

• The Nature of AI Style: One commenter noted that all models share proclivities, and if the world's literature were written by only a few models, readers would quickly become tired of the same flourishes and patterns, leading to a generalized perception of "AI" writing.
• Style vs. Function: Another user observed that the 'slop' style often mirrors corporate or structured essay formats encouraged in certain educational settings, while the 'de-slopped' versions resemble more casual, human-written drafts. This suggests that different styles are appropriate for different use cases.
• The Limits of De-slopping: While the model successfully removed specific tics, some users questioned whether the process merely replaced old habits with new, less obvious ones, or if the fundamental lack of human touch remains a barrier to truly indistinguishable writing.

Usage and Next Steps

The released model, qwen3-4b-deslop-v4, is available on Hugging Face. The quantized build (Q4_K_M, 2.5 GB) can be run locally using Ollama or llama.cpp. The full-precision weights are also available for use with standard transformers libraries.

The creator plans to improve the model further by creating a better 'teacher.' This involves compiling a dataset of high-quality human writing, including public-domain essays and letters, to train a base model that never went through assistant training. This new teacher model would then be used in the same distillation pipeline to produce a student with both clean habits and genuine taste.