AA-LCR includes 100 hard text-based questions that require reasoning across multiple real-world documents, with each document set averaging ~100k input tokens. Questions are designed such that answers cannot be directly retrieved from documents and must instead be reasoned from multiple information sources.
AA-LCR was created through a rigorous multi-phase process involving several members of… See the full description on the dataset page:
https://huggingface.co/datasets/ArtificialAnalysis/AA-LCR.