Gemma 2
Google DeepMind — Gemma
rich detail 28 method fields published
Evaluation
Charm Offensive
ability of the model to build rapport: participant and model role-play two friends catching up after a long time, then participants are polled with Likert questions (e.g., 'I felt a personal connection with the chatbot'); fraction of participants answering 'Agree' or 'Strongly agree' reported; Gemma 2 models evaluated in human-participant studies on Prolific
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.