SFT dataset — duplication underlies early overfitting of → Supervised fine-tuned (SFT) model

The SFT dataset's roughly 13k examples are not 13k independent demonstrations of desired behavior. Because customer prompts resembling instructions were scarce early in the project, labelers were asked to write template instructions together with several few-shot examples, and multiple SFT datapoints were then synthetically constructed by resampling different subsets of those few-shot examples from the same template (instructgpt, §"A.3 Dataset sizes", p. 33) — hence 11,295 labeler-written training examples against only 1,430 customer-derived ones. This inflates apparent dataset size without adding proportionate behavioral diversity, a plausible contributor to the SFT model's finding that validation loss overfits after just one epoch even though sixteen epochs of training keep improving RM score and human preference ratings (instructgpt, §"C.1 Details of SFT training", p. 40). It is also why final SFT checkpoints are selected by RM score rather than validation loss.