single mention · data curation
Generate unanswerable examples by masking context with helpful and adversarial agents
A synthetic-data method that creates unanswerable examples by masking parts of a sample's context with two masking agents, Merlin and Morgana.1 The masks are optimised so the policy keeps answering on Merlin's masked context but cannot answer on Morgana's, while hiding the same number of sentences so the two contexts cannot be told apart by length.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
Merlin and Morgana are masking algorithms. For a seed with context x, question q, and gold answer a, they choose masks m∈{ 0, 1}C over the C maskable units (sentences) of x
— tech reportSince Merlin and Morgana hide the same number of sentences within an adjustable tolerance, Arthur cannot tell them apart by how much is missing
— tech report
Filed under data curation › synthetic data