hat Makes a Good SFT Sample (And Why Most Synthetic Datasets Get It Wrong)

Many fine-tuned language models result in worse performance due to poor quality synthetic data. The issue is not with the training setup, but with the lack of mechanisms to filter out errors during synthetic data generation.