ConvApparel

The ConvApparel dataset contains conversations between paid raters and an AI assistant. The raters are tasked with buying an apparel item (footwear, outerwear, tops, or bottoms) and also fill out a survey at the end of each session. For full details, see our EACL 2026 paper titled: ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders. UPDATE (June 1, 2026): We added ConvApparel_V2 data collected for our paper Controllable User Simulation. ConvApparel.json is a dictionary with fields: 'metadata' which has fields: 'questions': a dictionary mapping from version id to survey questions. 'instructions': a dictionary mapping from task id to rater instructions. 'tasks': a list of task ids. 'conversations'.: a list of conversations, each has fields: 'task_id': one of the ids in 'tasks' from the metadata. 'version': one of the keys in 'questions' from the metadata. 'ratings': a dictionary mapping from survey question to rater answer. 'turns': a list of turns, each has fields: * 'user_utterance': the text entered by the rater. * 'assistant_utterance': the assistant response. * 'ratings': a dictionary mapping from survey questions to rater answers. * 'recommendations': a list of recommended items, each has fields: * 'item_id': the item id from Amazon Reviews. * 'title': the item's title. * 'description': a description of the item. * 'image_url': url of the item's image. * 'features': a list of item features.