Module 09 Activity

Scenario

Use a pretrained encoder for a text classification task, and report it so a reviewer can judge it.

Task

1. Pass a batch through one encoder block and confirm the shape is unchanged.
2. Tokenise, pad and mask a small batch. Show that masked pooling differs from naive pooling.
3. Freeze the encoder, cache the pooled vectors, and fit a classical head on them.
4. Fine-tune end to end with the encoder at a much lower rate than the head.
5. Evaluate both against the majority baseline.
6. Split the metric by group and report per-slice accuracy.

Deliverable

A classification report with both approaches, the baseline, per-slice accuracy, and a limitation.

Check your work

• Input and output shapes match: a block enriches representations without resizing them.
• Naive and masked pooling give different vectors whenever padding is present.
• Frozen features train zero transformer parameters.
• Overall accuracy can look acceptable while one slice is substantially worse.

The slice rule

Decide the slices before looking at results. Choosing them afterwards is selection, not measurement.

Mark complete