It wouldn't be on a token by token basis, but rather "common" token sequence tuples. If you generate a histogram of common 5-8 grams, and filter out all the on

It wouldn't be on a token by token basis, but rather "common" token sequence tuples. If you generate a histogram of common 5-8 grams, and filter out all the ones that are just linguistic filler and keep the ones that are "factual," then feed those "factual" n-grams to reviewers (via captcha/turk/game) to get a distribution of accuracy scores that you then condition the model on, so it can produce "accuracy" estimates for token sequences, and further condition on high accuracy scores for generation. for example (butchered for clarity), (["The", "2022", "election", "was", "stolen"], accuracy=normal(0.0001, 0.1)).