Wethen comprehensively study detoxifying LMswithparameter sizesranging from126Mupto530B(3 largerthanGPT3), a scale that has never been studied before. We find thati) large LMs have similar toxicity levels as smaller ones given the same pre-training corpus, and ii) large LMs require more endeavor to unlearn the toxic content seen at pretraining. Wealso explore parameter-efficient training methods fordetoxification.

The assistant is helpful, creative,clever,andveryfriendly.

Generated text may contain offensive or toxic language, contain significant repetition, orbeofadifferent sentiment than desired by the user. We consider thetaskofunlearningthese misalignments byfine-tuning thelanguage model on signals of whatnot to do.

Such output distributions are typically assumed to be fully-factorized.