Dataset for paper: Surrogate Deep Neural Networks Reveal Hierarchical Handwriting Encoding in Human Motor Cortex

Dataset and File Structure Description The dataset is packaged into a single archive file dataset.zip, containing 27 .pkl files. Each .pkl file corresponds to one recording session, and the files are named following the format: {index}_{sessionname}.pkl (e.g. 0_2022-06-01-S2.pkl). Each session file contains neural activity, handwriting kinematics, stroke annotations, and generated character images for a series of trials. The dataset is organized trial-wise, with data structured in (trial, time, channel) format. Each session includes the following data components: Key Data Fields coords: Shape: (N, T, 5) Description: Handwriting kinematic signals across trials. Features: x_vel: X-axis velocity y_vel: Y-axis velocity pen_up: Pen state one-hot indicator (pen lifted) pen_down: Pen state one-hot indicator (pen touching) pen_end: Pen state one-hot indicator (stroke finished) coords_stroke: Shape: (N, T, 29) Description: Extended kinematic representation combining velocity, pen states, and stroke type annotations. Features: First 5 dimensions are identical to coords Following 24 dimensions represent a one-hot encoding of 24 predefined stroke types spike: Shape: (N, T, C) Description: Neural activity recordings from sorted spikes, smoothed with a Gaussian filter. Note: C denotes the number of neural channels. Neural signals are temporally aligned with handwriting kinematics per trial. char_img: Shape: (N, 1, 64, 64) Description: 64×64 grayscale images synthesized from the handwriting trajectory (coords). char_name: Type: list of strings Description: Character labels corresponding to each trial.Each entry specifies the character written in that trial. Additional Notes Each time step represents 50 milliseconds.