Show HN: Qwen3-4B with the AI writing tics trained out
Category: ai-ml
Tags: language-model, fine-tuning, text-generation, ai-writing, ollama
Score: 6.7/10 (Innovation: 7, Technical: 7, Documentation: 9, Utility: 6)
A fine-tuned Qwen3-4B language model trained to remove AI writing tics (e.g., em dashes, 'not X but Y' constructions) without needing a system prompt. The project details a multi-round distillation pipeline using LoRA adapters and neuron edits, with thorough evaluation of style and task adherence. It offers a practical solution for generating more natural-sounding AI text.
Target audience: AI researchers, content creators, developers interested in local LLMs
Repository: https://100toolkit.com/blog/deslop-model/
View on Hacker News