This dataset is the result of 50k instruction/response pairs generated by Claude and two additional follow-up instructions for each base instruction (for a total of 150k instructions), with instances of blatant alignment removed.
32170 (96510) instructions remain.
The instructions were generated synethically using a method that can be tenatively described as "multi-instruct." These instructions consist of numerous discrete tasks that the AI has to work its way through, thereby hopefully… See the full description on the dataset page:
https://huggingface.co/datasets/Norquinal/claude_multiround_chat_30k.
