I have been hitting “Hard” when things were difficult to recall (albeit in kind of a subjective way, sometimes I also used to hit “Hard” when I thought a word was important and wanted to see it more often) and “Again” when I couldn’t remember them.
The big difference between that and normal Anki use is just that when I hit “Hard”, my next interval would be around half the previous interval instead of being longer.
Is that going to mess with FSRS’s ability to use my learning history?
Neither of those things are helping you learn. That sounds like “rote” or “brute-force” memorization, which is rarely efficient or effective in the long term.
You can use multiple learning and/or relearning steps – neither is inherently bad. There are reasons we steer beginners away from that as a starting point. There are more ways to get multiple steps wrong than get them right. And if you’re using multiple steps for ineffective/brute-force methods, then it’s really a waste of your time to add them.
A thousand or thousands – the more the better, the more varied the better. The way you originally presented yourself was as someone who had a significant amount of Anki experience/history. But now it sounds like that might have been us mistaking your messages.
I don’t think you’ve shown us anything yet that demonstrates you have actual problems in your review history. You might think you do, but is that accurate? If you want us to comment on that, you’d need to show us your data –
What FSRS parameters do you get when you optimize on your actual data? How many reviews does it count in that preset?
What is your Desired Retention (DR)? How does it compare to your retention outcomes?
How does the rest of your Stats look – Answer Buttons, Card Counts, Intervals, Stability?
FSRS cannot deal with “up”-grading cards – using Hard when you got the answer wrong and should have used Again. But grading your cards too harshly just means things will grow more slowly for now. During optimization FSRS prioritizes more recent review history, so if a bad habit is mild or short-lived, it usually doesn’t take much to get past it. [But no, using an ignore-cards-reviewed-before date is a drastic solution that you shouldn’t resort to unless there is something really wrong.]
I do practice outside of Anki with reading and conversation practice but that doesn’t always lead to practicing the words most in need of practice.
To clarify, the hitting Again a bunch of times was mostly happening in an FSRS test deck where I was using Again for “remembering wasn’t super-easy”. The moment of realization that I was totally wrong about spaced repetition came between that post and the one after it.
But if there’s something else I should be doing for the times when I legitimately need to press “Again” on a card, it would be good to know!
Some Anki experience, but I wouldn’t say “a lot”. A few months, long enough to have fun tinkering with the advanced features and start thinking I knew what I was doing, but apparently not long enough to actually know. The Dunning-Kruger effect strikes again.
But I’m a big enough frog to admit when I’m wrong, and in this thread I have been very wrong.
My main preset has roughly 1000 cards that have their whole learning histories in Anki. There were 10520 reviews, excluding a time when I used a filtered deck (without card rescheduling) to do an initial sorting of the deck and remove words I had already learned elsewhere.
Most were learned on SM-2 with a lot of learning steps.
Some may have spent some time on FSRS the first time I tried FSRS.
Due to my interval-shortening habit, only 40 have made it to intervals longer than one month.
For a while I had default/starting ease at 170%, I later lowered it to 150%.
Not sure how to answer about desired retention because these cards have mostly been on SM-2. Undecided if I’ll use 90% or 95% in the future - at the moment I have 95% and it the higher percent sounds more appealing, but if that might give me intervals that are too short to benefit long-term recall then maybe I should go with 90%.
Past week retention for young cards is 96.4%, not enough mature cards to get a percent.
Answer buttons in Learning are 19% Again, 7% Hard, 72% Good, 2% Easy (easy was mostly used to let cards out of Learning early if I felt they didn’t need all my steps)
Answer buttons for young cards are 15% Again, 13% Hard, 81% Good, 0.5% Easy
After turning FSRS on, the card stability plot shows more than half of the cards have a stability of < 18 days, there’s still a decent number reaching as high as 60 days, and a few very high outliers at 100 days.
If I optimize my FSRS parameters for this preset, I get:
Two presets for FSRS test decks (neither of which have more than a few dozen cards with learning histories, and one of which involved pressing “Again” when I hadn’t actually forgotten cards)
One preset for 2000-ish cards I originally learned in other language-learning apps before I started using Anki. Many of these cards were given 1-2 fake grades of “Good” or “Easy” using “Grade Now” in the browser to try to reflect that they had already been learned in other apps. Results of FSRS optimization vary widely depending on which cards are included when I do it, as does the data shown in stats view. Roughly 800 cards have had one or more real reviews in Anki in addition to the initial fake grades. Based on those cards, retention for mature cards could be 80-89% (and they are legitimately mature cards based on the intervals they had in the apps where they were originally learned). I haven’t spent very much time on the cards in this preset yet, the best bet may be to just reset the learning histories to get rid of the fake grades.
Thank you for taking the time to help me figure out how to un-screw-up that which I have been making a mess of.
@Danika_Dakika have you had a chance to look at the reply above from last week with the data you asked for?
I’m eager to find out whether I’ve screwed up my learning history or not!
Here are updated parameters from my main preset after one week. Unfortunately this wasn’t a week of regular reviews, I was sick and missed some days, so many cards were overdue when reviewed (and I didn’t start any new cards this week):
For now I am still sticking to the FSRS defaults and not using the optimized values. I don’t understand all the FSRS parameters, but I understand the first four, and I’m concerned the first parameter giving a card a stability of 1.70 days if I press “Again” the first time I see it might mean that something strange in my learning history is confusing FSRS when I try to optimize parameters.
(Edit: Updated parameters above, the first time I ran optimize today the preset accidentally included some of the cards imported from other apps. The parameters above now reflect only the 1064 cards that have their whole learning history in Anki.)
Is that how many you see in Stats, or how many FSRS counts when you optimize this preset?
It doesn’t matter what algorithm you used to study the cards – your retention results are your retention results. (So if these are Stats from your FSRS test-area, that’s not what you want.)
95% is very high to set your DR – higher than necessary for most users for more purposes. But if you have been able to maintain 95% retention over a long period of time (longer than 1 week, look at month/year/all), even while you’ve been down-grading your cards as you have been – then sure, you can try setting 95% for your DR. You should expect a high workload from that.
I’m not sure how you’re getting 96.4% retention, if you’re grading cards Again 15% of the time. Are you comparing information for the same data? That looks like 85% would be a much more appropriate starting place for your DR.
[The rest of the things you narrated aren’t relevant, and screenshots of the actual Stats graphs would have made this go faster. Instead, I just looked at your synced (real) collection myself.]
There isn’t a lot of difference between those sets of parameters. I don’t think you need to be so picky about what cards you include/exclude from your preset. For instance, there’s nothing wrong with using “Grade Now” to start a card, as long as those grades were accurate. And since that’s a small slice of your collection, it won’t skew your parameters far in one direction.
FSRS can deal just fine with some cards being easier than other cards, and with cards being studied when they are overdue.
What that suggests is a lot of those times you were down-grading your cards to Again at the start, you still ended up remembering them just fine. I don’t think that will be a problem in your history/parameters. If you want to compare your sets of parameters to each other, you can use the FSRS Visualizer site you were using above.
I’m glad FSRS up-weights more recent reviews when optimizing parameters - hopefully with a few months of grading cards normally, any remaining idiosyncracies (like past instances of pressing Again when I probably didn’t need to) will work themselves out.