Training and evaluation datasets for RecaLLM. Contains GRPO reinforcement learning training data (20K examples) and evaluation data across 7 context lengths (4K-128K tokens). Datasets generated using the code in recallm/datasets/ — see there for generation scripts and augmentation details.

from datasets import load_dataset

ds = load_dataset("kswhitecross/RecaLLM-data", "hotpotqa"… See the full description on the dataset page: https://huggingface.co/datasets/kswhitecross/RecaLLM-data.