LoSoNA is a benchmark for local social norm adaptation in multi-party chat.
Each item contains a synthetic group-chat transcript in which non-subject
participants demonstrate a hidden local norm, followed by one final elicitor
turn. A subject model should write the next chat message. The target norm is
included for evaluation and should not be shown to the subject model.
The dataset contains 38 accepted scenarios in JSON Lines format:… See the full description on the dataset page:
https://huggingface.co/datasets/Humalike-ai/LoSoNA.
