Datasets:
Error in the HCT116 row numbering in the dataloader
#3
by AI-Discovery - opened
the SLAF/Lance mirror's cell_start_index over-counts true row positions by exactly 19,011,260 rows for HCT116 cells beyond ID 1,869,999, with one boundary cell straddling the gap. HEK293T needs no correction. Naively trusting that column silently returns the wrong cells' expression for roughly 45% of the HCT116 dataset — the extraction scripts verify every cell against total_counts so that failure mode can't pass unnoticed.
It seems that someone converted the orion data either from the hf parquets/figshare upload into a more efficient dataloader, but introduced an error.