That’s really cool. Showing people how you are working through this is valuable work to help demystify how the training process works. Although it is kinda mystical how the model adopts the collective spirit of the community. It just needs to be understood as such so people don’t start calling models spiritual. 😁
Discussion
My main filter is web of trust. I give more weights to ppl high in WoT. Other than that I try to eliminate news, other bots, non English content because I dont know other languages and cant test. There is a process where Ostrich goes thru all the notes and decide whether to include or not in the training. If the note is looking like chatter it is excluded.
The LLM training is finding common values of the community. Numbers are battling. Biases are canceling each other.