Users of the Triple-N dataset can describe the data as follows:

Triple-N dataset

We used the Triple-N dataset (Li et al., 2025), in which electrophysiological recordings were obtained from five macaque participants during passive fixation. Each macaque viewed 1,000 natural scene images from the NSD dataset (Allen et al., 2021). Stimuli were presented only after the animals maintained fixation for at least 300 ms. Each image was displayed for 150 ms, followed by a 150 ms blank interval, resulting in 4–8 repetitions per image.

Neuronal activity was recorded using Neuropixels probes. Preprocessing included phase-shifting, filtering, common referencing, and spike sorting. Units were classified as visually responsive if they exhibited a significant change in firing rate (p < 0.001, Wilcoxon rank-sum test) between a baseline window (–25 to 30 ms relative to stimulus onset) and post-stimulus response windows (50–120 ms or 120–240 ms). Neural reliability for each neuron was estimated by first determining the optimal response window post-stimulus, splitting trials into two halves, and calculating the Pearson correlation between them. The resulting correlation was adjusted using the Spearman–Brown correction, and only units with reliability > 0.4 were included in further analyses.

Recording targets encompassed high-level cortical areas along the posterior-to-anterior axis of ventral stream (from V1, V2, V4, PIT to AIT, according to your selection), including category-selective regions identified via fMRI (faces, bodies, objects, and scenes), regions without clear category selectivity (“unknown” areas), as well as earlier visual cortex. In total, X units were retained for downstream analyses.