Meena

Meena was introduced by Google in January 2020 as a 2.6-billion-parameter end-to-end neural conversational model built on the Evolved Transformer seq2seq architecture. It was trained on 40 billion words of filtered public-domain social media conversations.

Its lasting contribution is the Sensibleness and Specificity Average (SSA) metric, a human evaluation of whether chatbot responses make sense and are specific. Meena scored 79 percent SSA against a human baseline of 86 percent. The model was never publicly released; Google withheld it pending safety and bias vetting, and its lineage continued into LaMDA.