This paper introduces a benchmark task for large-scale face recognition, focusing on recognizing one million celebrities from their face images and linking them to corresponding entity keys in a knowledge base. The authors address two key gaps in current face recognition: disambiguation at web scale and the lack of large datasets. They propose a benchmark task that leverages a knowledge base to improve recognition accuracy and real-world applications such as image captioning and news video analysis. The benchmark includes a concrete measurement set, evaluation protocol, and a training dataset containing 10 million images of 100,000 celebrities. The authors also present experimental results using a deep neural network, achieving 44.2% precision on the measurement set with a confidence threshold of 95%. The paper discusses the challenges and potential applications of the benchmark, encouraging further research in data collection, cleaning, learning algorithms, and model generalization.This paper introduces a benchmark task for large-scale face recognition, focusing on recognizing one million celebrities from their face images and linking them to corresponding entity keys in a knowledge base. The authors address two key gaps in current face recognition: disambiguation at web scale and the lack of large datasets. They propose a benchmark task that leverages a knowledge base to improve recognition accuracy and real-world applications such as image captioning and news video analysis. The benchmark includes a concrete measurement set, evaluation protocol, and a training dataset containing 10 million images of 100,000 celebrities. The authors also present experimental results using a deep neural network, achieving 44.2% precision on the measurement set with a confidence threshold of 95%. The paper discusses the challenges and potential applications of the benchmark, encouraging further research in data collection, cleaning, learning algorithms, and model generalization.