rephrased

Definition

Restated in different words while keeping the content. In data work it refers to synthetic corpora produced by having a model rewrite existing text, a way of adding surface diversity and cleaning noisy sources when the supply of genuine data is fixed. Unlike translation or summarisation, the intent is that information content stays constant — which is also the failure mode to watch, since rewriting can quietly drop or invent facts.

Explainers using this term

• Benchmark Contamination — How to Doubt a High ScoreRethinking Benchmark and Contamination for Language Models with Rephrased Samples