single mention · not yet filed
masking-based refinement
Masking-based refinement is a data-correction mechanism that identifies defective pages within otherwise usable trajectories and masks them instead of discarding the whole sample.1 It preserves valuable supervision signals from the unmasked parts, avoiding the data loss and regeneration cost of throwing away samples with only localized defects.2
Trade-off It improves data efficiency and reduces redundant regeneration overhead, at the cost of keeping only the masked-out defective portions of a trajectory.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
automatically identifies defective pages and applies masking, while retaining the high-quality content within the same trajectory
— GLM-5: from Vibe Coding to Agentic EngineeringDiscarding such samples would reduce effective data utilization and increase generation cost
— GLM-5: from Vibe Coding to Agentic Engineeringpreserves valuable supervision signals, improves effective data efficiency, and reduces redundant regeneration overhead
— GLM-5: from Vibe Coding to Agentic Engineering
Also called mask fine-tuning.