Show HN: Qwen3-4B with the AI writing tics trained out
Summary
An open-source 4B LLM variant called qwen3-4b-deslop is described, which removes common AI-writing habits through in-house fine-tuning and can run locally via Ollama. The author documents the training pipeline (teacher-student setup, LoRA), evaluation on held-out prompts, and a practical workflow for local deployment, including code snippets. Limitations include the model's flat voice and occasional invention of names or places.