The CRINGE Loss: Learning what language not to model
This addresses the issue of undesirable outputs in language models for users in safety-critical or conversational applications, though it is an incremental improvement over existing contrastive training methods.
The authors tackled the problem of language models learning undesirable behaviors by introducing the CRINGE loss, a method that uses negative examples to train models on what not to do, and demonstrated its effectiveness in tasks like safe generation, contradiction avoidance, and open-domain dialogue, outperforming strong baselines.
Standard language model training employs gold human documents or human-human interaction data, and treats all training data as positive examples. Growing evidence shows that even with very large amounts of positive training data, issues remain that can be alleviated with relatively small amounts of negative data -- examples of what the model should not do. In this work, we propose a novel procedure to train with such data called the CRINGE loss (ContRastive Iterative Negative GEneration). We show the effectiveness of this approach across three different experiments on the tasks of safe generation, contradiction avoidance, and open-domain dialogue. Our models outperform multiple strong baselines and are conceptually simple, easy to train and implement.