A silver set is an automatically generated, semi-refined dataset used as a first pass before complete human labeling. It reduces the effort of creating a golden set from scratch but remains less trustworthy because machine-generated guesses contain mistakes. section 14.2
In extractive question answering, a pretrained reader receives question-and-context pairs and proposes answer spans. Those guesses become the silver set; a person then verifies, corrects, rejects, or ignores each row before training data is finalized. section 14.2.2 section 14.2.3
Question-answering guesses
Given retrieved contexts from an outdoors dataset, a question-answering pipeline may predict spans such as “the traditional longbow made from wood” with a confidence score. The prediction is provisional until a person confirms or corrects it. section 14.2.2
A practical silver-set workflow includes:
- Prepare question-and-answer data and extract the relevant contexts. section 14.2
- Run a pretrained question-answering model over each question/context pair. section 14.2.2
- Review each guess with explicit labels for wrong, vague, corrected, correct, or unchecked rows. section 14.2.3
