Will FSRS implemented in Anki still allow to disable fuzz as it allows now? Will it conform to
const enable_fuzz = false;
pasted into custom scheduling window?
I ask because Damien said he didn’t have any short term plans to add such an option.
Thanks once again for all your help in the anki community! Two questions:
-
I know certain types of cards I create will be harder than others. Should I have a separate deck/preset for these to allow FSRS to more accurately & efficiently schedule my easier vs. harder cards?
-
Can you please explain or point me where I can find explanation on how FSRS handles equivalent of “new interval” when pressing again? I am used to having new interval of 1, which makes sense - if I forgot it, I should start over. So I’m very skeptical about FSRS scheduling my relearned cards a month out. I tried checking the algorithm and other github pages, but the math/formulas are way beyond me. If you really just want to say “just trust the algorithm” then I’ll accept that lol
Both of these questions apply to a 2 year old collection recently converted over to FSRS.
1 Like
Related: I have a subtree of decks with the same preset. I noticed that the Difficulty graph for one of the decks is very skewed. So I give the deck a new preset and click Optimize (Evaluate shows an improvement of 0.767→0.753 and 0.442→0.394), then reschedule the cards for the subtree. The Difficulty graph is still skewed. Shouldn’t it change?
No. The built-in FSRS doesn’t affect the fuzz, which is implemented by Anki itself. If you need disable fuzz, you should use the custom scheduling.
Yap. You can separate these cards into two decks and use two preset and parameters for them.
In most case, just trust the algorithm. I need your card info and FSRS parameters.
2 more questions
- Even though I’ve converted my SM2 decks to FSRS, I assume any new cards I create should be in a new FSRS-only deck, because I’m sure the weights would be different with cards that have only used this algorithm?
- Should I periodically recalculate the weights, or do they update automatically as I do more reviews?
When you enable FSRS, it will affect all decks, because it’s a global setting.
You need to recalculate the weights manually.
A subtree of low-priority cards has a backlog of 5000-8000.
Past month’s True Retention is 62-72%.
I am currently using Review sort order: Descending difficulty.
Is it good to use Postpone on those decks?
When I review a card on Duolingo, is it good to reschedule the identical card in Anki to 0 days and rate it?
Your true retention has been so low. So I don’t recommend using Postpone on them.
It is good but unnecessary.
1 Like
The graph actually does change; the issue may have been [23.10] new preset without saving → FSRS buttons use filtered deck - #3 by Aleksej
Ok. Could you answer the second question, please. Because it is not obvious for me.
Will it be enough to paste into custom scheduling:
const enable_fuzz = false;
Or the code of FSRS in Anki is different now and the above command will not work?
It’s not enough. To disable fuzz, you need use the full custom scheduling of FSRS.
I’m wondering what would be the best way to handle leeches when using FSRS. Obviously the #1 thing is to fix the card, but barring that, I’m wondering if it is a good idea to put leeches in a separate deck so your hard cards don’t bring down the weights, and still allow you to review the majority of your deck as efficiently as possible?
Or maybe a better way of thinking about this is each month before recalculating the weights, put all cards with difficulty >90% into a separate, hard deck?
I’m trying to make each deck as efficient as possible, but worried if these taking out difficult cards would just upset FSRS functionality?
Thanks in advance for your input.
2 Likes
Sometimes, the evaluator shows that the model is worse for some deck than for the rest of the preset, but if I exclude suspended cards from the evaluation, it fits well.
1 Like
Other FSRS contributors may have other ideas, but personally, I think you should use presets for different subjects that vary considerably in difficulty, eg if you’re studying law and French, and you find one considerably easier than the other, then that would be a good time to separate them. But I don’t think you should be trying to divide a single subject up into harder and easier cards - that’s awkward, and the ratings you give to the cards should be sufficient for them to be scheduled appropriately.
7 Likes
Some decks can really benefit from optimizing with “-is:suspended”:
Optimized on all cards:
for all cards: Log loss: 0.4842, RMSE(bins): 8.09%.
for the non-suspended cards: Log loss: 0.4487, RMSE(bins): 10.96%.
for the suspended cards: Log loss: 0.6653, RMSE(bins): 13.45%.
Optimized with -is:suspended:
for all cards: Log loss: 0.4220, RMSE(bins): 4.19%.
for the non-suspended cards: Log loss: 0.4795, RMSE(bins): 4.72%.
for the suspended cards: Log loss: 0.7726, RMSE(bins): 22.77%.
That deck contained cards from a pre-made Basic type deck with almost non-semantic things like painting titles and artist names, and many of the cards were suspended.
After moving the title/artist cards to another deck:
Optimized with all cards:
for all Log loss: 0.4448, RMSE(bins): 4.41%.
for nonsusp: Log loss: 0.4236, RMSE(bins): 4.96%.
for susp: Log loss: 0.6636, RMSE(bins): 16.41%.
Optimized with -is:suspended:
for nonsusp: Log loss: 0.4219, RMSE(bins): 4.43%.
for all: Log loss: 0.4452, RMSE(bins): 4.27%.
for susp: Log loss: 0.6864, RMSE(bins): 18.54%.
Why can’t “Compute optimal retention” take fractional values for “Minutes study/day”? It could multiply the numbers itself if they are too low.
It might not help every time though; one of my decks got these suggestions:
6626 365 10 → 79%
6626 730 10 → 78%
6626 1000 10 → 76%
6626 1500 10 → 80%
6626 2000 10 → 87%
Just a quick question on something that surprised me a little. Been using FSRS with desired retention of 85% for a couple of weeks, but my actual retention has been over 90% each day per the stats page. So I recalculate the weights & reschedule expecting it to slightly increase some intervals and lower my average cards/day, but it actually ended up bringing some cards forward/backlog by a few days, and average cards/day stayed the same. This isn’t a big deal but I’m just curious if this would be expected, or I shouldn’t have expected intervals to increased based on my exceeding my desired retention?
From what I understand, there are two types of retention.
The requested retention percentage means that FSRS is going to do its best to show you cards that you will get correct on the day they’re due according to that percentage.
The average retention means that number of cards you would get correct if you were shown everything in your deck with those same settings.
Example:
You have a deck of 1000 cards.
Your requested retention is 90% for this deck and you haven’t done any cards today. Let’s say that you have 100 cards that are “due” today. If you were to do cards right after reading this, you would get around 90% of them correct. FSRS’s aim is to show you cards on their due date that will result in 90% of them being correct. So you would answer about 90 cards correct (you responded with hard, good, or easy), and then 10 of those cards would be incorrect (you responded with again).
Let’s say your average retention for that same deck is 96%. Average retention would be if you just decided to do your whole entire deck right after reading this, you would get about 96% of them correct. So let’s say you decided to do your whole deck of 1000 cards, you would get 960 of them correct (responding with hard, good, or easy), and you would get 40 of them incorrect (responding with again).
2 Likes
Since FSRS is included now and the Helper addon is becoming obsolete, how do I keep the ‘Free Days’ feature where I rescheduled cards reviewed in the last 7 days on a regular basis? 