Question 850
A custom supervised fine-tuning data packer combines independent conversations into one long sequence but forgets to reset position or attention boundaries between examples. The model learns spurious cross-conversation context. Which preparation fix fits?
Explanation
Correct Answer
Packed sequences need a well-defined boundary policy so one independent example does not silently gain context from another unrelated training sample.
Incorrect Answers
Filename order cannot prevent leaked attention across samples concatenated into one model input.
More training reinforces the undesired cross-example dependency rather than correcting the mask.
Changing deployment metadata cannot repair the labels and attention structure seen during training.