Because for some users FSRS optimization results in inadequate/improper weights and intervals. For those users the only solution is to manually tweak weights.
So, I hope, you understand, that for those people the presence of “side effects” (or rather, it’s absence) is crucial.
In this case you optimize presets one by one and tweak weights if needed.
1 Like
By what metric are you determining the machine learning algorithm is producing improper weights?
When your graduation interval is three weeks (initial stability).
I will not continue this conversation. If you haven’t problems with FSRS, it doesn’t mean that it is true for others.
Not even three weeks - 23 days.
I don’t believe that any human can recall newly introduced card after 23 days
1 Like
I’ve never gotten intervals like that. I get about 3 weeks for a new card if I get it right the first time, which makes sense because I already knew it. If I get it wrong on first introduction, it ends up being about a 1 day interval.
Did you optimize before having enough data, like less than 1000 reviews for that preset?
Don’t try to educate me. This deck has more than 50k reviews and results on the screenshot are from FSRS-Optimizer (built-in optimizer produces almost the same weights, hence intervals). And FSRS consistently fails to adapt to it (only initial stability, to be fair).
Again, I will not continue this conversation. If you haven’t problems with FSRS, it doesn’t mean that it is true for everyone.
You seem like a tinkerer. Sounds like you probably tinkered with some stuff that you didn’t fully understand. But alas, I will not try to educate you.
If you don’t mind, would you share an export of your revlog data? Probably, someone would be able to find out why FSRS is not able to fit your data properly and fix the issue.
You can go to Tools -> FSRS Helper -> Export Dataset for Research and share the file here as a Google Drive link. The file doesn’t contain any personal information or the text of your cards.