Parameter Adjustment and Hyperparameters

When you click "Tune", you are presented with a few confusing sliders.

1. Epochs

An Epoch is one full pass through your training dataset.

• Low (1-3): Model sees data once. Good for large datasets. avoiding overfitting.
• High (5-10): Model memorizes the data more. Good for small datasets (~50 examples) where you really want to drill the pattern in.
• Risk: Too high = "Overfitting". The model memorizes your examples but loses the ability to generalize to new ones.

2. Batch Size

How many examples does the model look at before updating its brain?

• Usually handled automatically by Google, but sometimes configurable. Larger batches = smoother learning, but requires more RAM.

3. Learning Rate

How drastic are the changes to the brain?

• High: Fast learning, but might overshoot and become unstable.
• Low: Stable, slow learning.
• Multilier: Gemini allows a "Learning Rate Multiplier". Start with default. If the model isn't learning (loss is flat), increase it. If the model is spouting gibberish (loss spikes), decrease it.

Summary

• Small Dataset (<100): Higher Epochs (~5-10).
• Large Dataset (>1000): Lower Epochs (~1-3).
• Default: Google's defaults are usually very good. Trust them first.

In the next lesson, we cover Uploading Custom Data.