The approach for creating this dataset is outlined in our paper Steering Conversational Large Language Models for Long Emotional Support Conversations.
This dataset specifically is generated using attention guided prompting on llama3-70b-instruct. We start from the original ESConv dataset and generate strategy conditioned
continuations at random points of the conversation. For example, we cut a conversation at 10th turn when it is the emotional supporters turn to answer and… See the full description on the dataset page:
https://huggingface.co/datasets/navidmadani/extended_esc.
