This uses a simple set model (1 layer shared across all elements, a nonlinearity, and another layer after mean aggregation), trained on ICLR 2019 data. The overall statistics (average accepted rating) match up very well with NeurIPS 2019 data, so I suspect it should be a reasonable proxy for NeurIPS 2020 data. Of course, this model can't take into account your specific situation, so don't read too much into it :) Also, if you're trying out of distribution reviews, no guarantees about their accuracy either.
Note: For 2022, the textual description of "5" got changed from "borderline reject" to "borderline accept". I was unsure about its effect on the distribution of scores, but it seems like it might have affected it significantly (see here or here