Comment by causal

But I bet you could reconstruct a plausible set of distributions by just rerunning the autoregression on a given text with the same model. You won't invert the exact prompt but it could give you a useful approximation.