ML Project Guidelines

Philosophy

• Validation is king - always define clear train/val/test splits before training
• Error analysis is queen - understand failures, not just metrics
• No mocks - test with real data and objects
• Raw over abstraction - YAML configs, tensorboard, raw plots. No MLflow/W&B complexity
• Hyperparam tuning is overrated - usually low-hanging fruit elsewhere: features, data, target. Highly tuned models are brittle
• Simple ensembles only - if needed, blend linear model + CatBoost. No stacking towers
• Too good = suspicious - great results warrant paranoid leakage checks before celebration
• Apples to apples - always compare full baseline vs full experiment on the same data. No cherrypicked subsamples
• Hypotheses need origins - every experiment starts with "why". No random fishing expeditions

Default Workflow