Fine-tune a small open model on a laptop to remove AI writing tics
A writer fine-tuned Qwen3-4B-Instruct on a MacBook Air with 161 answers to train out assistant habits such as em dashes, "not X, but Y" and lists of three. He wrote a slop detector from a style guide, then checked for damage with a drift test on human prose, a 30-question quiz and a rule-based check, and released the model to run locally with Ollama.
Source
erikkannike on Hacker News · 2026-10-19 · Project link
Written in our own words. Credit to the original author. Link checked 2026-10-19.
What they did
A writer fine-tuned Qwen3-4B-Instruct on a MacBook Air with 161 answers to train out assistant habits such as em dashes, "not X, but Y" and lists of three. He wrote a slop detector from a style guide, then checked for damage with a drift test on human prose, a 30-question quiz and a rule-based check, and released the model to run locally with Ollama.
Result (as claimed)
On 40 unseen prompts the slop score fell from 121 to 12.5 per 1,000 words, em dashes went to zero and it still answered 30 of 30 factual questions; 138 points on Hacker News.
Tools they used
Cost: Pricing depends on the tools you pick; check each official pricing page.
Set it up
Gather your sources
Collect the papers, notes, links or files you want to work from in one place.
Load them into the tool
Open Qwen, Ollama and add your sources. Keep the original files so you can check claims later.
Ask focused questions
Ask for summaries, comparisons or connections with citations to the source passages.
Verify and save
Spot check claims against the sources, then save the output as notes you can reuse.
Copy the full setup prompt
Paste into your agent and it will walk you through the setup, asking before anything that sends, posts or spends.
You are helping me set up this AI use case end to end in Claude Code. Goal: Fine-tune a small open model on a laptop to remove AI writing tics What the original author did: A writer fine-tuned Qwen3-4B-Instruct on a MacBook Air with 161 answers to train out assistant habits such as em dashes, "not X, but Y" and lists of three. He wrote a slop detector from a style guide, then checked for damage with a drift test on human prose, a 30-question quiz and a rule-based check, and released the model to run locally with Ollama. Reported result (as claimed by the author): On 40 unseen prompts the slop score fell from 121 to 12.5 per 1,000 words, em dashes went to zero and it still answered 30 of 30 factual questions; 138 points on Hacker News. Source: https://news.ycombinator.com/item?id=50151928 Tools they used: Qwen, Ollama Category: Research & learning Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Codex. Goal: Fine-tune a small open model on a laptop to remove AI writing tics What the original author did: A writer fine-tuned Qwen3-4B-Instruct on a MacBook Air with 161 answers to train out assistant habits such as em dashes, "not X, but Y" and lists of three. He wrote a slop detector from a style guide, then checked for damage with a drift test on human prose, a 30-question quiz and a rule-based check, and released the model to run locally with Ollama. Reported result (as claimed by the author): On 40 unseen prompts the slop score fell from 121 to 12.5 per 1,000 words, em dashes went to zero and it still answered 30 of 30 factual questions; 138 points on Hacker News. Source: https://news.ycombinator.com/item?id=50151928 Tools they used: Qwen, Ollama Category: Research & learning Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
You are helping me set up this AI use case end to end in Grok Bot. Goal: Fine-tune a small open model on a laptop to remove AI writing tics What the original author did: A writer fine-tuned Qwen3-4B-Instruct on a MacBook Air with 161 answers to train out assistant habits such as em dashes, "not X, but Y" and lists of three. He wrote a slop detector from a style guide, then checked for damage with a drift test on human prose, a 30-question quiz and a rule-based check, and released the model to run locally with Ollama. Reported result (as claimed by the author): On 40 unseen prompts the slop score fell from 121 to 12.5 per 1,000 words, em dashes went to zero and it still answered 30 of 30 factual questions; 138 points on Hacker News. Source: https://news.ycombinator.com/item?id=50151928 Tools they used: Qwen, Ollama Category: Research & learning Please: 1. Ask me 3 short questions about my situation (accounts I have, budget, what I want the output to look like). 2. Propose a step-by-step plan using these tools or cheaper alternatives, with the time and cost of each step. Use official pricing pages; say "check pricing" if unsure. 3. Then do the setup with me one step at a time, waiting for my OK before anything that sends messages, posts publicly, spends money or deletes data. 4. Finish with a checklist to run it again next time.
The Daily 5
Get five new use cases tomorrow morning
Free. Real sources. Unsubscribe any time.
Guides for this use case
Related use cases
Make Codex pick a cheaper model by task effort
The author wrote a Codex skill that swaps models by effort to save usage, then found Codex running a subagent on a local Qwen model through Ollama.
Swap an open model into Claude Code through Ollama to fix a messy repo
A developer kept his Claude Code setup (CLAUDE.md, skills and terminal) and pointed it at Reflection's open Beam model through Ollama's cloud with one launch command. He gave it his messiest repo: 40 files and a bug he had dodged for a week.
Read long AI chats sideways so each answer is its own page
Sidelong lays a chat out sideways: each question and answer is a page, older pages collapse into colored strips, and color encodes topic so follow-ups share a thread color. It runs on the Claude Code CLI you are already logged into, with topic embeddings from a local Ollama model.
Summarize articles and YouTube videos with a fully local browser extension
Apogee rebuilds Mozilla's discontinued Orbit summarizer so nothing leaves your machine: it connects to a local Ollama instance (or WebGPU/Transformers.js in the browser) to summarize articles, websites and YouTube videos.