Issue: Discrepancy in cell count and performance metrics for cHL dataset

[saarthak-kapse]

Hello,

Thank you for the wonderful work and for open-sourcing the cHL downstream tasks as tutorial notebooks. They have been incredibly helpful for getting started with the dataset.

I attempted to reproduce the results reported in the paper using the tutorial notebooks, but I noticed a discrepancy in the cell counts:

• Paper Reported: ~134K cells
• Tutorial Notebook: ~144K cells

This difference seems to lead to a slight drop in cell phenotyping performance. Below is a comparison of the results I obtained:

Source	F1-Score	Average Precision
KRONOS (Paper)	68.07	76.14
KRONOS (Tutorial)	67.2	74.2

Could you please clarify if there were specific filtering steps or version differences between the paper and the tutorial code?

Thank you for your time!

[andrewsong90]

Hi @saarthak-kapse

Thank you for your interest in KRONOS. I can confirm that the # of cells in the tutorial notebook is accurate, we will update the details in the next version - Thank you for pointing it out.

As for the discrepancy, let me investigate it a bit further and get back to you. Having said that, we will hopefully be releasing updated KRONOS model in few months, so stay tuned for this!

Thanks,
Andrew

[saarthak-kapse]

Thanks @andrewsong90 for the quick response. I’m looking forward to resolving the discrepancy and seeing the updated KRONOS model!