regent-research/regent-subset-of-jat-dataset-tokenized
This is the dataset for REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context In New Environments. The REGENT dataset includes a subset of the JAT (tokenized) dataset (from https://huggingface.co/datasets/jat-project) for the REGENT training environments. It has around a 100k transitions from each of the 145 training environments (45 metaworld, 52 atari, 9 mujoco, 39 babyai). Please find this in the *_subset folders. It also has distance values for input sequences used in… See the full description on the dataset page: https://huggingface.co/datasets/regent-research/regent-subset-of-jat-dataset-tokenized.
This is the dataset for REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context In New Environments.
The REGENT dataset includes a subset of the JAT (tokenized) dataset (from https://huggingface.co/datasets/jat-project) for the REGENT training environments. It has around a 100k transitions from each of the 145 training environments (45 metaworld, 52 atari, 9 mujoco, 39 babyai). Please find this in the *_subset folders.
It also has distance values for input sequences used in the REGENT model for both the above 145 training environments and 12 unseen environments. These values are used for computing the distance normalization value in every environment. These values, along with the indices of the input sequences, in the training environments alone, are used for pre-training REGENT.
We detail this REGENT subset of the JAT dataset at https://github.com/regent-research/regent/blob/main/scripts_preprocessing/README.md.