Experiments and Runs

Replace mystery notebooks with evidence

MLflow is a tool for recording and comparing machine-learning work. An MLflow experiment is a named group of related attempts. A run records one attempt: input settings called parameters, measured numbers called metrics, searchable labels called tags, and output files called artifacts. Here, those might be model name, chunk size, retrieval recall, response time, and a report.

Log every input that could change the result, not just the final score. Recall here measures how often search found the expected document. A recall number without the dataset version, source-code revision, and model tag cannot explain why it changed.

Scenario: Two developers report different RAG accuracy. MLflow reveals that one used chunk size 300 and dataset v2 while the other used chunk size 800 and stale data.