Foundation models (FMs) such as CLIP, which are trained on vast amounts of unlabeled data, are versatile AI models applicable to tasks like image classification and natural language processing with minimal fine-tuning. CLIP, specifically, is trained on large datasets of image-text pairs, allowing it to predict text snippets related to images using natural language instructions. The article discusses using CLIP to classify a dataset of facial expressions, followed by evaluating the results with Encord Active, an open-source toolkit for active learning. The initial steps include setting up a Python environment, downloading and preparing the dataset, and using CLIP to make predictions that serve as ground truth labels. These predictions are then imported into Encord Active for evaluation. The performance of CLIP is found lacking, with low precision, recall, and F1 scores, highlighting the need for potential improvements like addressing data imbalances and enhancing feature representation. The article also introduces TTI-Eval, a library for evaluating zero-shot classification models like CLIP, and emphasizes the importance of metrics such as sharpness and brightness in determining model performance.