This dataset is an experiment in procedurally generating synthetic dialogues and revisions between two language models, along the lines of Constitutional AI.
For each dialogue, one model, acting as a "user" generates a plan based on a topic, subtopic, and goal for a conversation.
Next, this model attempts to act on this plan and generating synthetic data.
Along with the plan is a principle which the model, in some successful… See the full description on the dataset page:
https://huggingface.co/datasets/allenai/sdsd-revisions.
