CORAL Preference Expansion Dataset

This dataset contains like/dislike user preferences extracted from three conversational movie recommendation datasets: PEARL, ReDial, and INSPIRED. The preferences were generated using the contrasting preference expansion technique proposed in the CORAL framework.

We used GPT-4o-mini-2024-07-18 to extract these preferences. The detailed prompts and extraction methodology are described in our CORAL paper (please refer to the paper for full details).

See our paper to learn more details!

Dataset Overview

---

Dataset Fields

• id (str): Unique identifier for each sample.

• gt (list(dict)): List of ground-truth movie (e.g., [{"id": "Knives Out (2019)", "title": "Knives Out (2019)"}, ...])

• train_gt (dict, optional; exists only in train files): Ground-truth single movie for training (e.g., {"id": "Knives Out (2019)", "title": "Knives Out (2019)"})

• dialogue_history (list): List of dialogue utterances between the user and the system. (e.g., [{"role": "user", "text": "I'm looking for..."}, ...])

• input_dialogue_history (str): Flattened dialogue history in a single string format. (e.g., "System: Hi!...\nUser: I'm looking for...")

• preference (dict(list)): User preferences consisting of two fields:

  • like (list[str]): List of liked items or attributes.
  • dislike (list[str]): List of disliked items or attributes.