The fine-grained hateful memes task consists of predicting 3 classes - hate detection (hateful or not), protected classes, and attacks. The above meme is represented in the JSON record below.