FSRS 6 updates all difficulties to 0

Anki FSRS
1

Hello,

First of all, thank you for all the work on Anki and FSRS, especially the recent update with FSRS 6! I can feel that the new scheduling is much better adapted to my learning curve for my decks.

A friend of mine, however, has encountered an issue after updating to Anki 25.07.5 and optimising all her FSRS parameters for a certain deck.

Here are the parameters before optimising:

0.1974, 3.2926, 0.5674, 1.6234, 6.9894, 0.2330, 1.5240, 0.0252, 1.4812, 0.0000, 0.8079, 1.6155, 0.2160, 0.2130, 2.0000, 1.0000, 1.0000, 0.3643, 1.0521

with log loss 0.2407 and RMSE 5.89%.

After optimising:

0.0490, 12.3037, 12.3037, 12.3037, 6.3990, 0.9681, 2.9295, 0.2185, 1.5314, 0.0000, 0.5118, 0.6836, 0.0542, 0.1110, 0.8607, 0.9995, 1.0000, 0.7735, 0.8425, 0.0032, 0.1000

with log loss 0.2148 and RMSE 3.19%.

Strangely, the median difficulty drops from 81% to 0%! I wonder if this is due to a bug. Would it be possible for you to investigate this problem? I have the apkg file with scheduling information before the optimisation and the PDFs exported from the stat tab before and after the optimisation; I can send them to you by mail or GitHub or other way to your convenience.

Thank you for your insights!

2

Out of curiosity, is she using the “ignore cards with reviews before [date]” setting when optimizing the FSRS params?

3

Could you share the distribution graph of the card diffculty?

4

No, she didn’t use that. On the other hand, she uses the hard button a lot, like 50% of the time; maybe this has something to do with the results of optimisation?

5

She shouldn’t be using the hard button if the card was actually forgotten.

6

Here is the distribution before and after optimising:

Difficulty before FSRS 6
Difficulty before FSRS 62650Ă—1516 205 KB

Difficulty after FSRS 6
Difficulty after FSRS 62762Ă—1560 200 KB

7

I can confirm that she only uses the again button when the card was actually forgotten.

8

I find that the w[7] is 0.2185 which means the difficulty reversion has a large impact to drag your cards’ difficulty to 0. So it’s intended.

For more details, please see:

9

Thank you very much for your reply.

My initial confusion was actually about how (algorithmically) w[7] got such a large value.

However, after digging into the review log, I suspect that my friend misuses the again and hard button even though she told me otherwise: in the past year, she has only pressed the again button 2.2% of the time, which gives a false retention rate of 97.8% that she clearly doesn’t have on the deck. This probably explains why the difficulty of all cards drops to zero after optimising.

Thanks for all the relies in this thread and sorry to have wasted your time on this trivial case.

2 Likes
10

While yes your friend is at fault here, it somehow also highlights the need for the UI to support users in more correctly grading themselves. I know this has been discussed to death, but I suspect the end results nowadays still is a portion of users (maybe unknowingly) using Anki and aren’t even able to properly communicate to the algorithm whether they remember a card or not. Something quite fundamental to the whole idea of an SRS.

11

Yes, thank you for your insight; I wholeheartedly agree!