Some updates for everyone on planned progress
Currently I am putting my projects on back burner while going to school and working. I still have someone helping me progress things slowly.
Currently the goal is to get a golden dataset sorted for my VIT and use that to clear out a large chunk of deep rooted errors in my Danbooru data.
I am hoping my helper gets us to completion this year or early next year.
@jialinyyzz how are you doing with the dataset? I got an API pipeline that would work but would be stupid expensive to operate to have very high accuracy and high level ai detection for the training data. Was a multi stage fact checking pipeline, style checks and voice checks. With a final judge. Limiting what the writers would see from the original text to prevent the models for copying from the originals. A lot of base models copy the original text to rewrite.
So was hoping you could share some of the high AI level pairs with some of the human text you work on.
OK，I will clean up my data and share some but Im still researching on the data because it's really important.
base model will copy the original text no matter what's your training data is, so I usually penalize the copy rates on RL laterly.
Limiting the writers see will significantly mess up the accuracy