Building Teacher in a Box: a local classroom for small models

Building Teacher in a Box\n\nSmall language models become more useful when improvement can be measured rather than assumed. Teacher in a Box is a local classroom that separates diagnosis, teaching, training, and evaluation into stages a person can inspect.\n\n## Diagnose before training\n\nThe workflow starts with a held-out skills exam. Weak areas become evidence for what the model needs to learn instead of relying on a general impression of quality.\n\n## Keep a person in the loop\n\nThe system proposes targeted lessons, but a person reviews the curriculum before it becomes training data. That review prevents a weak automated explanation from quietly becoming part of the model.\n\n## Evaluate independently\n\nAfter training, the model runs the same held-out evaluation under equivalent settings. The report keeps failures, latency, model identity, prompt settings, and the evaluation-suite hash so the comparison can be reproduced.\n\n## Where Lemonade fits\n\nLemonade provides the loopback inference layer. The application discovers locally available models through its OpenAI-compatible API and sends prompts only to the selected local model.\n\nThe project is open source so the workflow, assumptions, and limitations can be inspected and improved.