In particular the Cringe Loss is a Practitioners commonly align large language models method for binary feedback, which we show can be generalized using pairwise preferences, i.e., given labels to the pairwise preference case. The Cringe Loss works of the type response A is preferred to response B as follows: positive examples use the standard likelihood for a given input. Perhaps less commonly, methods training loss, while for a given negative example it contrasts have also been developed for binary feedback, each token in the negative sequence against other likely i.e. training models given labels of type tokens - to encourage the negative sequence to no longer response A is good or bad. We show how an existing be the top-ranked sequence. After training on the initial performant binary feedback method, the feedback data, the method is then iterated by labeling data Cringe Loss (Adolphs et al., 2022), can be generalized using the improved model, which was shown to improve to the pairwise preference setting using results further. Cringe Loss was shown to perform well with a simple soft margin extension. Pairwise Cringe binary feedback data compared to competing methods, such Loss is straightforward to implement and efficient as SFT, unlikelihood loss and best-of-N reranking (Adolphs to train, and we find it outperforms state-of-the-art et al., 2022) and for improving large-scale dialogue systems preference optimization algorithms such as PPO (Xu et al., 2023b).

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.