Block 109: Fine-tuning Concepts & Zero-Shot Classification

Concepts

• What fine-tuning means: adjusting pre-trained weights on new data
• Zero-shot classification: no training examples needed
• pipeline('zero-shot-classification')
• Candidate labels and multi-label support

Code Examples

See exercise below.

Exercise

Use zero-shot classification to categorize 10 news headlines into: politics, sports, technology, entertainment. Compare zero-shot results to a simple keyword-based classifier you write manually.

Homework

When would zero-shot classification be preferred over training a classifier? When is it not enough?