>>105845652
From experience and observation, an 8B model trained on 2500 samples, 4k tokens, 4 epochs would have been more than acceptable a couple years ago, but I don't think most people are going to settle for less than a 24B model nowadays (x3 compute) and the model trained with at least 16k context (x4 compute). So we're looking for at least 12x more compute for a mostly basic RP finetune, putting aside the time required for tests and ablations.