Improving Open Language Models by Learning from Organic Interactions
This work addresses improving conversational AI safety and skills for users, though it is incremental as it builds on an existing model.
The authors tackled the challenge of training conversational models with organic user interactions, which include both helpful and adversarial data, resulting in BlenderBot 3x being preferred over its predecessor and producing safer responses.
We present BlenderBot 3x, an update on the conversational model BlenderBot 3, which is now trained using organic conversation and feedback data from participating users of the system in order to improve both its skills and safety. We are publicly releasing the participating de-identified interaction data for use by the research community, in order to spur further progress. Training models with organic data is challenging because interactions with people "in the wild" include both high quality conversations and feedback, as well as adversarial and toxic behavior. We study techniques that enable learning from helpful teachers while avoiding learning from people who are trying to trick the model into unhelpful or toxic responses. BlenderBot 3x is both preferred in conversation to BlenderBot 3, and is shown to produce safer responses in challenging situations. While our current models are still far from perfect, we believe further improvement can be achieved by continued use of the techniques explored in this work.