extended-salmon•12mo ago
FSRS - possible issues
I enabled FSRS recently and I am using only good and again buttons with a retention of .85. I noticed when I make new cards, after I press good it wants to show me it again in like 4 days. Seems long for a new card but oh well.
What really got me though, was when I would press Again on a brand new card and it would proceed to make the Good interval longer each time. So after pressing Again on a brand new card three times, the Good button would say 7 days. Isn’t that the opposite of what should happen?
I added the fsrs helper addon thing to anki also , I attached photos of my settings and cards. Was just curious if anyone else has experienced this and if I’m maybe doing something wrong?
163 Replies
sunny-green•12mo ago
Were you originally only using the good and again buttons? You should be grading the exact same way as before you started using FSRS. Otherwise you will throw the algorithm off.
extended-salmonOP•12mo ago
Yea I was, but I noticed that because I press Good immediately when making the card instead of Again and then Good, it starts it off at 7 days or 6. But if I press again and then Good on a brand new card, it starts it off at 1 day. I never noticed that in the old algorithm, so Ill have to get used to failing my brand new cards first I suppose before I press Good on them
sunny-green•12mo ago
@Michaël P (ローマ字 26) 6-7 days is quite far out, what do you think?
extended-salmonOP•12mo ago
It is, I’m not sure why it would start cards off that way on 85 retention
Debating disabling fsrs for fear of all my cards having super long intervals
Was kind of curious if anyone else had experience with fsrs and had ever noticed any similar pattern
When I fail a card first though it seems like it gives it a different trajectory, so I may just try it out a little longer first
fascinating-indigo•12mo ago
Fsrs tends to give long intervals by default
sunny-green•12mo ago
sunny-green•12mo ago
I guess you can give it a try and see how well it works.
fascinating-indigo•12mo ago
At 90%, it gives me a 5 or 6 days interval after the first good
Failing a card on first encounter gives painfully short intervals for the rest of the card’s life tho
extended-salmonOP•12mo ago
Yeah I think I need to set my retention back to 90 percent.
Thanks for your feedback, this makes me feel a bit better about using it
like-gold•12mo ago
@jeremy_ just to confirm what you're seeing, i also get around 4 days if i press good the first time i'm encountering a card and then fully pass it
The thing i don't understand, is apart from if you already learned some of the words through immersion before encountering them in anki or you can assume the meaning, why would you ever pass a word on first encounter?
fascinating-indigo•12mo ago
The first encounter shouldn’t be in anki
The SRS is there to help retain info you’ve already met
like-gold•12mo ago
But if i make hundreds of cards, the likelyhood that i've retained it until i actually review the card in anki is really low no?
fascinating-indigo•12mo ago
Also doubly tricky with fsrs as fsrs only uses the first score of the day
So fail/pass counts as a failed day
extended-salmonOP•12mo ago
Well i mine as Im immersing so I mine a card and then just press good because I just learned it essentially
like-gold•12mo ago
I do too, but i mine ahead quite a bit
I don't immediately review cards i mined
Oh so that means that essentially i fail like 80% of my cards on the first day
fascinating-indigo•12mo ago
1d,2d,6d,14d,1.1m,2.3m,4.7m,9.1m,1.4y,2.5y
That’s the intervals from failing on first day then 100% recall after that
Vs
2d,7d,21d,1.9m,4.8m,11.2m,2.0y,4.1y,8.1y,15.0y
Not failing the first encounter
With sm2, there is no impact if you fail at first encounter
This is a fsrs specific issue
like-gold•12mo ago
I see
but does it really matter as long as fsrs keeps my retention at around 90% and it doesn't take me too much time to review?
fascinating-indigo•12mo ago
It just increases the review load
To above what it would have been with sm2
like-gold•12mo ago
Hmm, but then what do you suggest i do?
fascinating-indigo•12mo ago
Normally people switch to fsrs to reduce the review load
Which happens because fsrs gives longer intervals
like-gold•12mo ago
Yeah
But how do people learn e.g. a core 2k deck with fsrs? Is it just not made for that?
Cause they would never have encountered most of those words right?
fascinating-indigo•12mo ago
They either press good on first encounter’ then whatever at the daily’ learning interval
Or they press fail at first interval and increase their load
like-gold•12mo ago
ohh i see
fascinating-indigo•12mo ago
Maybe pressing easy on another day will improve the interval
like-gold•12mo ago
so i could just press good on first encounter even though i didn't know the word, and that would likely lead to more appropriate scheduling?
fascinating-indigo•12mo ago
That was one way to fix the skewed intervals in sm2
Yup
Also don’t have too many learning steps
like-gold•12mo ago
It just never occurred to me that thats how youre supposed to do it
is 3 too much
fascinating-indigo•12mo ago
Fsrs recommends 1 or 2
like-gold•12mo ago
I see
Whats the disatvantage besides longer review time?
fascinating-indigo•12mo ago
3 should be ok… but you’re reducing the number of new cards/day per time unit
like-gold•12mo ago
I don't understand this sentence 😅
fascinating-indigo•12mo ago
Research tends to show that multiple steps in the same day is good for short term retention but not so effective on long term retention
Say you want to dedicate 20 minutes for new cards
And each card takes 30s
At 1 step, you can do 40 cards in 20 minutes
At 2 steps, you can do 20 cards in 20 minutes
At 3 steps, you can do 13 cards in 20 minutes
like-gold•12mo ago
Makes sense, but i wouldnt take as long to review the card the second and third time right?
Guessing it's like this for the example?
fascinating-indigo•12mo ago
Yeah
like-gold•12mo ago
Ah
fascinating-indigo•12mo ago
For me the initial reviews are identical in duration
The next days review are way faster
like-gold•12mo ago
I like having a final review step at the end of my reviews though, you think thats an issue?
like-gold•12mo ago
Hmm i usually take longest for the first one because i look up what other words the kanji in the new word are used in
fascinating-indigo•12mo ago
At some point I had like 5 reviews and a retention above 95% in the first week of the cards’ life
Then I went the other way to 1, and loaded more new cards per day
Now I’m back to 2
like-gold•12mo ago
oh so you had 5 reviews for the one card in one week?
fascinating-indigo•12mo ago
5 learning steps, then the reviews starting to space out
like-gold•12mo ago
Also if i switch to passing cards on first view now, do you think that'll fuck with my fsrs data?
Damn thats a lot
fascinating-indigo•12mo ago
I still do that for complicated stuff
like-gold•12mo ago
I guess it does depend on the subject
fascinating-indigo•12mo ago
Like if it’s a data point for a certification I need… 5 steps gets a good score
If it’s vocab in a domain I like… 2 steps is enough/plenty
Also the target retention in fsrs is misunderstood
like-gold•12mo ago
How so?
fascinating-indigo•12mo ago
If set to 90%, it will present a batch of cards where you’re expected to answer correctly 9 out of 10
It’s not 90% chance of getting each individual card right
If you answer correctly to all, some of them will be kicked further out
like-gold•12mo ago
Yeah so true retention for the whole day would end up around 90% right?
fascinating-indigo•12mo ago
Yes
like-gold•12mo ago
Yeah thats what i've noticed
fascinating-indigo•12mo ago
They’re using 2 Paramus to track the Ebbinghaus curve
Instead of 1 with sm2
like-gold•12mo ago
But then with the higher intervals if i pass a card the first time i encounter it, wouldn't my true retention go way below 90% from then on?
I wish i paid attention in math class sometimes
fascinating-indigo•12mo ago
So say we all
Good at first encounter isn’t the longer interval, it’s the normal/default one
like-gold•12mo ago
Yeah i know, but relative to what i have right now i mean
I have 91% true retention on avg. right now, I'm guessing it would dip then?
fascinating-indigo•12mo ago
It’s apparently still very conservative
If it dips, it will readjust at next optimisation
like-gold•12mo ago
Ah that makes sense
So i just start passing cards now the first time i see them?
fascinating-indigo•12mo ago
Try it for like a couple of weeks and see if it has a negative impact for you
For me it’s more of a theoretical issue because I retire my cards early anyways
I just use the SRS to load some knowledge and do most of the repetition in immersion
like-gold•12mo ago
Yeah if you retire at >21 days, the amount of steps for you are 4 no matter what
will do
fascinating-indigo•12mo ago
I do the repping through watching/reading listening a specific series
Using Ankimorph to optimally sort the cards in I+1 order to achieve a target comprehension of the series
Then loop the series
like-gold•12mo ago
Ah so rewatching after youve graduated all the cards?
fascinating-indigo•12mo ago
Tweak the target comprehension until you need 20 cards per episode
Pass them
Watch the episode
like-gold•12mo ago
Doesn't that get boring?
fascinating-indigo•12mo ago
If it’s a good series, no
I did it in spring with a bunch of series
20 cards a day for 3 months and 2 days
Anime + condensed audio + manga
It beats doing a core deck, at least for me
extended-salmonOP•12mo ago
Hmm I’ll have to look at this closer after work, definitely makes me wonder whether I should fail first or not
like-gold•12mo ago
There doesn't seem to be any definitive answer to this on the internet either
extended-salmonOP•12mo ago
Hmmmm
like-gold•12mo ago
I'm just wondering, if the intent is for the first time you see a card, to always pass it, then why is it even an option to not pass it?
extended-salmonOP•12mo ago
Well I was just wondering how it might affect the algorithm
If prefer to just pass when I first see the card if I remember it well
I typically review within 10-30 mins after making the card
Some of them are easy enough so I pass immediately, and some arent so easy so I fail those
fascinating-indigo•12mo ago
the issue is the FSRS decision to only take into account the first answer of each day for a card
like-gold•12mo ago
So you think the algorithm is flawed in that way?
fascinating-indigo•12mo ago
in many ways
that is just one of them
like-gold•12mo ago
But its the best we have?
fascinating-indigo•12mo ago
define best
like-gold•12mo ago
fair
fascinating-indigo•12mo ago
it recudes the daily load by kicking cards further away
but
like-gold•12mo ago
It's also supposed to get used to user habits better than the old one right?
Learning the forgetting curve and such
fascinating-indigo•12mo ago
it also creates a pants-on-the-head retarded backlog when you initially enable it and when you optimize with "reschedule all cards" on
they're both mapping the ebbinghaus curve
simplified: sm2 tracks it with only one variable
fsrs tracks it with two variables
also we only have like less than 2 years of data on people actually using it
like-gold•12mo ago
S and D?
fascinating-indigo•12mo ago
stability and difficulty for FSRS
ease for SM2
difficulty is basically 1-ease
like-gold•12mo ago
i see
well in those two years if it was worse than SM2 we would know right?
fascinating-indigo•12mo ago
one user on this server enabled fsrs with "rescheduled all cards" two days ago
he was greeted with 3600+ new reviews out of the blue
like-gold•12mo ago
damn
i mean i had a lot too
fascinating-indigo•12mo ago
I had ~3000 at first run
like-gold•12mo ago
but since i made a seperate deck for mining it was fineish
fascinating-indigo•12mo ago
then ~1000 at each optimisation until they added the magic switch
like-gold•12mo ago
the spike here
like-gold•12mo ago
whats the magic switch? load balancing?
fascinating-indigo•12mo ago
like-gold•12mo ago
but doesnt the switch make it worse?
wait no
fascinating-indigo•12mo ago
it is off by default
before it was on by default
and couldn't be turned off
like-gold•12mo ago
oh
thats bad
fascinating-indigo•12mo ago
epic level of bad
like-gold•12mo ago
so you could only enable fsrs if you worked through a mountain of reviews
fascinating-indigo•12mo ago
at some point I stopped running optimized because I was fed up with all the reviews coming out of the blue
that was before I started retiring cards early
like-gold•12mo ago
Glad i came back to anki when most of the transitioning phase was over for fsrs
i mean its still happening but its not like its an addon anymore
fascinating-indigo•12mo ago
yeah
you no longer need to copy/paste js code
or edit variables in that code
like-gold•12mo ago
oof i missed a lot apparently
fascinating-indigo•12mo ago
the switch to v3 scheduler was painful as it broke a whole lot of add-ons
but then again, they broke because they were using "private" methods of the v2 scheduler for advanced stuff
like-gold•12mo ago
btw is this curve a proven thing? like the specific algorithm?
ah classic
fascinating-indigo•12mo ago
Forgetting curve
The forgetting curve hypothesizes the decline of memory retention in time. This curve shows how information is lost over time when there is no attempt to retain it. A related concept is the strength of memory that refers to the durability that memory traces in the brain. The stronger the memory, the longer period of time that a person is able t...
fascinating-indigo•12mo ago
I somewhat disagree with some parts of the theory... I mean I still recall bits of books I read once in the early 90s and never since... I also recall bits of languages I haven't used at all since the 1980s and 1990s
I also recall bits of obscure movies I saw in the early 80s
like-gold•12mo ago
But how does the theory go against that?
It still says "The stronger the memory, the longer period of time that a person is able to recall it" right?
fascinating-indigo•12mo ago
there's bits of languages I learned then promptly never used... I should have forgotten them since, as I actually never repped them
now, I'm no longer fluent in those languages... but I can still order food/drinks or ask/give directions in one... and I can still understand the other one even tho I no longer can speak it
well... I can still deal with really basic stuff in it every once in a blue moon... but I couldn't hold a 10 minutes conversation anymore
like-gold•12mo ago
Ah yeah it wouldn't be as simple as using a single approximation of a curve to map the brains way of memorization
fascinating-indigo•12mo ago
probably because I haven't actively spoken that one since like 1982 or 1983
like-gold•12mo ago
which language?
fascinating-indigo•12mo ago
Luxembourgish
Luxembourgish ( LUK-səm-bur-ghish; also Luxemburgish, Luxembourgian, Letzebu(e)rgesch; endonym: Lëtzebuergesch [ˈlətsəbuəjəʃ] ) is a West Germanic language that is spoken mainly in Luxembourg. About 300,000 people speak Luxembourgish worldwide.
The language is a standardized language and officially the national language of the Grand Duchy of Lux...
fascinating-indigo•12mo ago
I would love to speak it again at some point
not much interesting content to immerse in, tho
like-gold•12mo ago
Oh thats close to where u live no?
fascinating-indigo•12mo ago
I mostly grew up in the light blue area
I moved further away now
like-gold•12mo ago
Oh so it was spoken where u lived?
fascinating-indigo•12mo ago
only by old people when I was growing up
and it was an older variant of it
like-gold•12mo ago
Its so close to german that like with dutch i can understand most of it
fascinating-indigo•12mo ago
so it was always funny for natives
yeah my grandma could go to flanders and be understood no problem 😄
it just sounded "sort of retarded"
like-gold•12mo ago
hahahaha
yeah it honestly sounds really funny
fascinating-indigo•12mo ago
same story on the German side
doubly so coming at it from English
like-gold•12mo ago
i have a friend who learned dutch just because he thought it sounded funny
fascinating-indigo•12mo ago
"that is" is "dat ass"
it's hard to say it with a straight face coming from other languages
like-gold•12mo ago
i kinda feel bad for spamming this thread btw
but its kinda too late now
fascinating-indigo•12mo ago
😄
like-gold•12mo ago
yeah the more languages i learn words from, the more funny words i find in other languages
japanese is surprisingly mostly free from them
mostly
fascinating-indigo•12mo ago
Japanese has a few fun ones in the loan words
the way they mistreat some of the French-sourced ones always amuses me
like-gold•12mo ago
e.g.?
fascinating-indigo•12mo ago
avec, zubon, ...
like-gold•12mo ago
ohhh thats where ズボン comes from
i never knew
fascinating-indigo•12mo ago
zubon for trousers must have been because of a huge misunderstanding
from jupon
like-gold•12mo ago
jupon is the original french?
fascinating-indigo•12mo ago
fascinating-indigo•12mo ago
this is a jupon
fascinating-indigo•12mo ago
fascinating-indigo•12mo ago
this too
like-gold•12mo ago
ah, first time i know the word for these
what about avec
fascinating-indigo•12mo ago
it simply means with
but
they use it as "in a relationship"
like-gold•12mo ago
wait whats the japanese word?
fascinating-indigo•12mo ago
アベック
like-gold•12mo ago
form the german side we do have アルバイト which is just the german word for work but apparently from the pov of japanese people we all only do part time
fascinating-indigo•12mo ago
yeah バイト
公園にたくさんのアベックがいた
like-gold•12mo ago
oo never heard this
oh so thats just "people in relationships" in this case?
fascinating-indigo•12mo ago
(romantic) couple
like-gold•12mo ago
makes sense
just found this list https://www.sljfaq.org/afaq/german.html
lots of these coulve also come from english
fascinating-indigo•12mo ago
the medical ones definitely came from German
extended-salmonOP•12mo ago
Hmm so would you all still recommend using fsrs for language learning at this point?
Is it still probably a better way to go than the old algorithm? Or is that still not clear
fascinating-indigo•12mo ago
the positive bit is that it has tools to tune it without having to change options all over the place
and it will tend to self-tune to hit your desired target rate
I used to have a spreadsheet to do that with sm2
to calc which parameters needed to be changed, and by how much
like-gold•12mo ago
that sounds like way more work than is worth
impressive though
but yeah i think in general fsrs is just way easier to tune to do what you want it to
fascinating-indigo•12mo ago
omg... Japanese just taught me a word in my native language
アクメ
like-gold•12mo ago
Lol yeah I've had that happen
What does it mean
fascinating-indigo•12mo ago
climax
so the Japanese use it for orgasm
like-gold•12mo ago
Lmao
Ofc
What's the actual French word?
fascinating-indigo•12mo ago
Acmé, from the old greek akmê
another fun one... successful overdue reviews give an interval bonus to the cards
that happens also with sm2 but not on the same scale
the FSRS team claims it's less of an issue with FSRS as it will taper towards the max interval... but they ship with the max interval at 100 years
extended-salmonOP•12mo ago
One thing I dont fully grasp is the relearning steps
Like mine is 1m 10m which seems pretty common/ default
But when I make a new card I see it right away but then its immediately there for review again, I never have to wait one mins or 10 mins
So I guess I’m not sure what learning steps really refers to
clever-tan•12mo ago
Sorry, out of topic, but what is the anime on your card?
fascinating-indigo•12mo ago
yeah that's a long standing issue of anki... it used to wait for that amount before showing them again (pre 2.0.35)
at least, in my hazy memory, it used to
sunny-green•12mo ago
Is it not related to this option? I don't have a new version of Anki installed, but if you set this time less than your next rep time, it will show instantly.
fascinating-indigo•12mo ago
fascinating-indigo•12mo ago
I'm going to try dropping it down to 1 minute
like-gold•12mo ago
That's looks like suzume
fascinating-indigo•12mo ago
I dropped it to 1 minute and now the new cards don't directly pop in the learn count
whereas before they would
now they only pop in the learn count after the interval passed
fascinating-indigo•12mo ago
sunny-green•12mo ago
Like if your last learning interval is 20m, I think a 15m learn ahead would mean you'd at least have to wait 5 minutes if those are the last reps. So you want these values to coralate somehow.
fascinating-indigo•12mo ago
yes
clever-tan•12mo ago
tank u
yeah, you are right
extended-salmonOP•12mo ago
Thanks everyone for the help and discussion. I think I may stick to the old algorithm for now just to be safe. Using the fsrs makes me feel like I’ll fail a ton more cards just because of crazy long intervals on new cards
fascinating-indigo•12mo ago
it would self-correct at the next optimisation tho (in theory)