Downloadable model weights with built-in rules to avoid harmful answers.
It is like a bike with training wheels already on. You can take it home. A skilled tinkerer may still take them off.
People use it on their own computers. It aims to cut risky replies. People still debate if the safety rules can stay put.
Open weights
It adds safety rules to downloadable model weights.
Alignment
Alignment training helps it avoid harmful replies.
Abliteration
Abliteration can weaken its built-in safety rules.
Open-weight safety debate
It sits at the center of the open-model safety debate.