Attention head — AI Dictionary

One parallel copy of self-attention's query/key/value routing; a block runs several so it can track several relationships at once.