This is the official dataset used in our EMNLP 2025 paper Promote, Suppress, Iterate: How Language Models Answer One-to-Many Factual Queries.
The dataset includes six subsets named {dataset_name}
template{i}, where dataset_name is country_cities, artist_songs, or actor_movies, and each dataset has three prompt templates (i = 1, 2, 3).
The {model_name}
step{i} split in each subset contains the data used for analyzing model_name's behavior at… See the full description on the dataset page:
https://huggingface.co/datasets/LorenaYannnnn/how_lms_answer_one_to_many_factual_queries.
