Interesting framing. If I understand correctly, you’re defining train-deploy mismatch as something like “if the deployed model and learner were run on the same data, would their outputs differ”? And changing the data distribution the model is trained on as not train-deploy mismatch?
Good question. I originally defined train-deploy mismatch the way you suggest. On that definition, inoculation prompting is interpreted as changing the model architecture (by grafting the inoculation prompt onto it) during training. However, Monte convinced me that inoculation prompting is better understood as changing the data distribution, so I edited the post to be less committal and include that as a possibility too, even though I still think primarily in terms of the original definition.
Interesting framing. If I understand correctly, you’re defining train-deploy mismatch as something like “if the deployed model and learner were run on the same data, would their outputs differ”? And changing the data distribution the model is trained on as not train-deploy mismatch?
Good question. I originally defined train-deploy mismatch the way you suggest. On that definition, inoculation prompting is interpreted as changing the model architecture (by grafting the inoculation prompt onto it) during training. However, Monte convinced me that inoculation prompting is better understood as changing the data distribution, so I edited the post to be less committal and include that as a possibility too, even though I still think primarily in terms of the original definition.