I trained my local LLM on its own failures, and now it learns from every mistake
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A hobbyist describes building a personal correction-and-training workflow for a local LLM (using Qwen via a self-hosted tool called Lemonade). The custom UI has Chat, Corrections, Compare, and Export tabs: wrong answers get logged with the question, failed response, verified correction, and a rule set. Examples include Qwen mangling a filename-formatting task and botching simple arithmetic on CSV data. The Export tab turns corrections into prompt-answer pairs that could train a LoRA adapter, gradually personalizing the model to the author's own tasks using data unavailable elsewhere.
Table of contents
My local LLM now has a notebook for correctionsFinding Qwen's mistakes was the easy partNow I have a starting point for fine-tuningShare this post