Instruction Fine-Tuning Through the Lens of Verbatim Memorization

Supervised fine-tuning is key for model alignment, but its mechanisms are debated, with conflicting evidence supporting either a superficial alignment hypothesis or significant task improvements. This paper examines supervised fine-tuning’s impact from the perspective of verbatim memorization. Using the open-source OLMo-2 model series and test datasets (instruction format, safety-sensitive, and factual knowledge) constructed from its pre-training corpus, we analyzed changes across memorization, linguistic styles, and task performance. We found that supervised fine-tuning significantly weakens the model’s verbatim memorization of pre-training data. Simultaneously, it improves generated text in terms of alignment objectives, such as polite expression and structured organization. However, this process also leads to performance degradation on knowledge-intensive downstream tasks. Further representation analysis reveals that these changes are mainly concentrated in the later layers of the model. We conclude that supervised fine-tuning acts as a continuation of the learning process on new data. By adjusting model representations, supervised fine-tuning induces a learning tilt toward the styles and content of the instruction-tuning dataset. This inclination successfully instills alignment objectives while consequently reducing the effective accessibility of previously learned knowledge, which indicates the observed degradation in both pre-training data memorization and factual task performance. The source code is publicly available.