Why does SFT often degrade factual accuracy relative to...

The degradation of factual accuracy in SFT occurs because data curators typically prioritize fluent and complete-sounding responses. Consequently, the model learns to produce confident completions regardless of whether the information is factually grounded. This overrides the base model’s internal calibration, which originally learned to express uncertainty from noisy pretraining data. To mitigate this, practitioners can use data-side strategies, such as including responses that hedge or express uncertainty, or model-side strategies, such as applying KL regularization toward the base model during the fine-tuning process.