I might differ from you "well paid posts" in the past can be because of the Bid Bots. If I only talk about the content, we really cannot judge based on it if its good post or bad post, for example, a post with 1000 words can be a bad post than the post with 100 words.

What I will do is to take 100 good content post and 100 bad or worse content post manually not based on price, but based on actual content and then divide it into 80% training and 20% and then train it.