# A lightweight experiment log

By Ackley Will | https://ackleywill.com

## Make the comparison easier to trust

A useful experiment record connects a change to a question. Write the question before the training run and keep the result beside the exact settings used.

## Record the essentials

- **Question:** What specific uncertainty will this experiment resolve?
- **Data:** Dataset version, split procedure, and preprocessing.
- **Model:** Architecture, pretrained weights, and trainable layers.
- **Training:** Seed, optimizer, learning rate, batch size, and epochs.
- **Evaluation:** Data partition, metrics, threshold, and label order.
- **Result:** Numbers, representative mistakes, and relevant variability.
- **Interpretation:** What the comparison supports and what it leaves open.
- **Next action:** One justified follow-up or a decision to stop.

## Keep the final test separate

Use validation data to choose settings. Evaluate the selected procedure on the held-out test data. Repeatedly adjusting a model after inspecting test results weakens the meaning of a final test.

The downloadable template includes a simple comparison table you can keep with your code.

---
Resource text: CC BY 4.0. Attribution: Ackley Will, ackleywill.com.