CODE415 Demos GitHub

Slop Check

Paste a post, an email or an essay. The habits chatbots lean on light up: pet words, stock phrases, em dashes, It's not X, it's Y. English and 日本語.

● Runs in your browser, nothing is sent0 habitsFree and open source
0per 1,000 words

Reads human

0154080120+

What it found

    The news behind it

    A 2.5 GB model that unlearned the habits

    On October 19 Erik Markus Kannike shared qwen3-4b-deslop. A small LoRA on Qwen3-4B, trained for 1 to 3 hours on an M2 MacBook Air, plus a sign flip on four neurons, took most of the chatbot habits out of its writing. He scored 40 held out answers with his own list of habit patterns.

    Habit score per 1,000 wordslower is better
    Qwen3-4B
    121
    deslop v4
    12.5
    Em dashes per 1,000 words
    Qwen3-4B
    20.5
    deslop v4
    0
    • 30 of 30 quiz answers still right
    • 161 filtered answers to train on
    • rank 16 LoRA, 175 bytes of neuron edits
    • 2.5 GB, runs in Ollama

    His pattern list is not published, so his scores and the score on this page are not on the same scale. Same idea, our own list. Numbers from his write up.

    How it works

    A list of habits, weighted

      A habit is not proof

      Plenty of people write like this, especially in school essays, business email and translated text, and good writers use an em dash on purpose. Please don't use this to accuse anyone of using AI. Use it on your own drafts.