You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hello @PierreMarza !
Really nice to see this benchmark popping up 👌
Could you please consider adding H0-mini to the list of supported models ? I can provide an implementation if needed.
Hi! Thanks for your interest in THUNDER! Integrating H0-mini is indeed a great idea. I have started working on it, but just a quick question: on the HF page, you mention both CLS features and concatenated features (CLS features concatenated with mean-pooled patch token features). Which one do you suggest I should use for classification tasks?
Hi Pierre :)
That's awesome. Would you like me to run the benchmark ? Yes you can concatenate the two, I believe it's better for tile-level tasks. For slide-level it was on par from our experiments. Really nice benchmark again, I really hope you'll be able to maintain and update it over time, that is cruely missing for most benchmarks in CPath...
Hello! H0-mini is now added to THUNDER (c5be59d). Let me know if that looks good to you! If so, I'll close this issue :)
Regarding maintaining THUNDER: that's indeed our goal to keep it up-to-date, but we will need help from people in the community! So please, feel free to open other issues and/or pull requests to add new datasets, models, tasks, and features if you or anyone in your team finds them useful!
Reacted by Leo Fillioux, Alexandre Filiot and Sofiène Boutaj
Activity
PierreMarza commentedon Nov 19, 2025
Hi! Thanks for your interest in THUNDER! Integrating H0-mini is indeed a great idea. I have started working on it, but just a quick question: on the HF page, you mention both CLS features and concatenated features (CLS features concatenated with mean-pooled patch token features). Which one do you suggest I should use for classification tasks?
afiliot commentedon Nov 19, 2025
Hi Pierre :)
That's awesome. Would you like me to run the benchmark ? Yes you can concatenate the two, I believe it's better for tile-level tasks. For slide-level it was on par from our experiments. Really nice benchmark again, I really hope you'll be able to maintain and update it over time, that is cruely missing for most benchmarks in CPath...
PierreMarza commentedon Nov 21, 2025
Hello! H0-mini is now added to THUNDER (c5be59d). Let me know if that looks good to you! If so, I'll close this issue :)
Regarding maintaining THUNDER: that's indeed our goal to keep it up-to-date, but we will need help from people in the community! So please, feel free to open other issues and/or pull requests to add new datasets, models, tasks, and features if you or anyone in your team finds them useful!
afiliot commentedon Nov 21, 2025
That's amazing, thank you.