This dataset is the result of roughly 250k instruction/response pairs being generated by Claude, with instances of blatant alignment removed.
213375 instructions remain.
This dataset is experimental in two ways:
From start to finish, it was generated entirely synthetically through Anthropic's Claude AI.
It was generated using a somewhat imperfect recreation of the evol-instruct method. 50k instructions were initially synthetically generated then ran through four epochs of evol-instruct.
