Social media platforms, particularly TikTok, have become dominant channels of value transmission to younger generations, superseding traditional agents like parents, educators, and peers. This research develops computational methods to extract implicit values from TikTok content using state-of-the-art language models. We curated and annotated a dataset of hundreds of TikTok videos according to the Schwartz Theory of Personal Values, providing ground truth for model development and evaluation. Our methodological contribution focuses on comparing two computational pipelines: direct value extraction from video content using Large Language Models (LLMs), and a two-step approach that first converts videos to detailed scripts before applying value detection.Our experimental results demonstrate that the two-step approach achieves superior performance, with a fine-tuned Masked Language Model significantly outperforming few-shot applications of various LLMs in value detection tasks. We provide detailed analysis of model performance across different value categories and examine how fine-tuning impacts the models' ability to identify both present and contradicted values in TikTok content.Our findings not only demonstrate the feasibility of automated value detection in social media content but also open new avenues for large-scale TikTok datasets for understanding the role of digital platforms in shaping cultural values and social learning in the digital age.