Evaluating and Fine-Tuning Chinese CLIP Models Using Domain-Specific and Diverse Datasets
摘要
To deepen our understanding of the Chinese CLIP model and its applications, this study involved fine-tuning Chinese CLIP using a specialized ceramic image dataset and a more diverse dataset. We compared the performance of 8 different Chinese CLIP models. The experimental results suggest that for CLIP models trained on large-scale data with well-optimized weights, fine-tuning is generally not recommended unless the training data has specific domain expertise or industry relevance. Among Chinese CLIP models, Alibaba DAMO Academy's Chinese CLIP model family outperforms others significantly, with CN-CLIP ViT-L/14@336px showing the best performance.