You could but it would be significantly worse than fine-tuning or RAG with a pre-trained model, or using a smaller model since your dataset would be so small.