I fine-tuned the DeepSeek-R1-Distill-Llama-8B model on the medical-o1-reasoning-SFT dataset[0:100]. Post-training evaluation conducted in LM-Studio revealed that the model not only performed reliably on the original training prompts but also exhibited a strong ability to generalize, successfully answering rephrased and semantically altered versions of the original questions. These results demonstrate that the model achieved a robust understanding of fundamental medical reasoning principles, rather than relying solely on rote memorization.

However, there were also notable limitations observed during this process. Despite the model’s strong performance on familiar and moderately rephrased questions, its ability to handle entirely novel reasoning chains or out-of-distribution medical scenarios remained limited. In certain cases, the model exhibited overconfidence in its incorrect answers, particularly when subtle contextual nuances were introduced. This highlighted the importance of exposing the model to a wider diversity of reasoning patterns and edge cases during training.

Reflecting on this experience, I recognized the critical need for more diverse and challenging data during fine-tuning, as well as the potential value of incorporating additional evaluation strategies such as adversarial testing or curriculum learning. Through this, I gained a deeper understanding of the balance between supervised fine-tuning and true generalization, and developed a stronger appreciation for the complexities involved in building reliable, domain-specific language models.