| Summary: | In this dissertation, we run experimental study to investigate the performance of sentence embedding using an on-the-fly knowledge distillation model based on DistillCSE framework.
This model utilizes SimCSE as the initial teacher model. After a certain number of training steps, it caches an intermediate model and employs it as a new teacher model for knowledge distillation. This process is repeated several times to obtain the desired on-the-fly knowledge distilled student model. This model employs a novel approach to knowledge distillation, potentially offering advantages such as reducing training time and achieving performance close to the original teacher model. In some cases, after fine-tuning, it may even surpass the performance of the original teacher model for specific tasks.
|
|---|