eminorhan/neural-pile-rodent

The Neural Pile (rodent)

This dataset contains 453 billion tokens of curated spiking neural activity data recorded from rodents. The code and detailed instructions for creating this dataset from scratch can be found at this GitHub repository. The dataset takes up about 453 GB on disk when stored as memory-mapped .arrow files (which is the format used by the local caching system of the Hugging Face datasets library). The dataset comes with separate train and test splits. You can load, e.g., the train split of the dataset as follows:

ds = load_dataset("eminorhan/neural-pile-rodent", num_proc=32, split='train')

and display the first data row:

>>> print(ds[0])
>>> {
'spike_counts': ...,
'source_dataset': 'giocomo'
'subject_id': 'sub-npI3',
'session_id': 'sub-npI3_ses-20190420_behavior+ecephys',
'segment_id': 'segment_4'
}

where:

• spike_counts is a serialized array containing the spike count data. Its shape is (n,t) where n is the number of simultaneously recorded neurons in that session and t is the number of time bins (20 ms bins).
• source_dataset is an identifier string indicating the source dataset from which that particular row of data came from.
• subject_id is an identifier string indicating the subject the data were recorded from.
• session_id is an identifier string indicating the recording session.
• segment_id is a segment (or chunk) identifier useful for cases where the session was split into smaller chunks (we split long recording sessions (>10M tokens) into smaller equal-sized chunks of no more than 10M tokens), so the whole session can be reproduced from its chunks, if desired.

The dataset rows are pre-shuffled, so users do not have to re-shuffle them.