FSRS

Bunpro
#1

I will make this brief, bunpros team should make it their next priority sincerely to add FSRS in the year of 2025 hell even this month coming off of anki to bunpro I love so much bunpro has to offer, varied sentences, different degrees of learning andい so many other features, but FSRS cant come any sooner the unreliability of standard SRS is not enough to cut it for people wanting to learn your vocab system. if FSRS were added many many anki users would no longer have a reluctancy to jump over here. please integrate it as soon as possible.

15 Likes
#2

They’ve mentioned they are going to add a Fsrs opt-in, but not a replacement for the current system.

Sounds like overall they are going to make an opt-in option for it.

4 Likes
#3

Just confirming this and a few other review related overhauls are all planned. Had hoped to release them earlier but our pipeline currently has a few things in it.

25 Likes
#4

That would be really great. I’m currently planning on stopping using bunpro for vocab once I finish the N5 vocab deck (but continue using it for grammar) because of the massive drop in retention I observe in bunpro compared to anki with fsrs. Although that drop may not be entirely explained by the SRS algorithm (and may not perfectly correlate with actual progress in proficiency in the language), I would seriously reconsider my plan if bunpro allows me to use FSRS.

1 Like
#5

Bunpro’s SRS implementation is very forgiving. I think a big difference is the reset to SRS1 vs -1 by default for failed vocab. Have you tweaked this? It should lead to better retention (classic Anki style relearning on a forgotten word).

3 Likes
#6

I admit I did not (and will probably be too lazy to try it out knowing that bunpro will have FSRS in the future and that I can temporarily migrate my vocab studies to anki in the meantime). Wouldn’t this unnecessarily shorten all intervals for a missed card though? Or does the system remember that it’s not actually a knew card but a learned one that happened to have a lapse?

#7

In the classic Leitner system, a forgotten card goes back to the initial SRS state (meaning you’re fully relearning it). This will shorten the interval of every missed card, because the assumption is, that it was forgotten.

image
image2560×1392 149 KB

By default, Bunpro uses an alternative system where a forgotten card goes back by one box. Hence the interval of the forgotten card is shortened, but not fully resetted. I can’t speak for the Bunpro team, but my best guess is that the assumption is, that a lapse does not actually mean that you’ve fully forgotten a term.

image
image2560×1392 184 KB

So on Anki (even with FSRS under most circumstances), forgetting a card will lead to a full reset, since it’s being relearned. This will obviously lead to a higher workload.

Bunpro lets you adjust this setting under Settings → Reviews → SRS strictness

image
image998×472 29.3 KB

If you experience higher retention on Anki compared to Bunpro, there could be a lot of reasons for it (card quality, etc). But if you use different SRS strictness settings on both systems, I would assume that tweaking this will have a large impact, where even without FSRS, your retention will improve.

Further reads (and image source for the first two):

15 Likes
#8

How do you guys feel about the ghosts system in Bunpro? I currently have ghosts on for vocab, and left the SRS strictness at the default 1. But I am struggling with some vocab in a level that I dont see them often and get them right about 50% of the time. So the vocab keeps going up and down in SRS but never progresses. I’m not sure if the ghosts are helping me much in those cases, so I am considering removing them and making the SRS more strict for vocab.

The ghosts have been very helpful initially with new vocab that I am not familiar with

#9

I went ahead and set it so that my SRS level decreases by half if I get it wrong, and adds a ghost. It’s a bit annoying when I get something wrong because of a simple typo, but it does strike a good balance overall. It has definitely increased my retention while still being less punishing than a complete reset. Maybe you could try that?

Edit: I’d like to add that I also changed my review schedule to have shorter intervals. It increases reviews per day but that seems to help also. Might be worth giving that a shot too

1 Like
#10

I wish there were a system to better handle medium term memory, where you are familiar with a word/grammar point but confuse it with other things that have a similar meaning or sound the same.

After a while, ghosts turn into “what’s the last thing I got wrong that means x?” When I first run into something I haven’t seen in a while, a lot of times, I can remember 4-5 things that mean the same thing that would fit but not the correct answer.

I also wish there was a an automatic way to review similar items/nuance differences.

4 Likes
#11

What type of typo? Do you fail cards if you knew how to write it but hit a wrong key by accident? I usually just ESC and re-do in these cases. Or when multiple words are possible but they aren’t marked as alternatives, I still don’t fail the card. I only fail it if I actually make a mistake

4 Likes
#12

I’m doing Bunpro on a phone, so sometimes what happens is I’ll be going through he vocab reviews and end up in a situation where I make a typo and then my muscle memory causes me to hit the enter button and move to the next one before I have time to react to my error. I hope that makes sense lol.

2 Likes
#13

Thanks a lot for the info! Afaik for FSRS, even if it resets to the initial state, at least on my system, the next intervals after answering the card correctly are still much higher than for newly cards (could be wrong on this one).

I think I will try resetting the cards to the initial state for a trial run. Right now I am mostly reviewing grammar and I think the default settings of just reducing the level by 1 is not enough for some grammar which intervals are too long. This might lead to some massive increase in reviews - maybe then it’s beneficial to increase the intervals between levels, though that’s not great for new cards.

So +1 on the need to include FSRS in bunpro, imho it’s really needed.

#14

Absolutely, I’m not arguing against FSRS, I was explaining what my best guess is in regard of a large part of retention discrepancy between Anki and Bunpro

1 Like
#15

Sorry, I should’ve been clearer that this +1 was towards the general thread and not your regarding your post, which I think was really informative.

#16

Well FSRS is not really working with full reset. It has multiplies associated with a “Fail” that is trainable by the user review. Params 11 to 14 here : Anki FSRS Visualizer

The thing with FSRS is that it work really well to learn from your past history. Maybe you’re the kind of guy that need a full reset, or a simple 20% loss of interval will be enough, FSRS can adapt to that.

Still, depending on the FSRS version you can have different behaviour. FSRS-6, which is on beta on Anki beta version right now, work with trainable decay, the exponent of the forgetting curve. This can lead to some Desired Retention (70%) to be lead to very high intervals like multiple years, when the decay is low enough (it’s capped to min 0.1 though). FSRS-5 introduced same-day reviews consideration, not that great to predict same-day prediction but quite great to be able to weight differently if one card was failed 3 times in a row the same day instead of just 1.

Switching to FSRS would mean Bunpro would have to either use default params, which are not bad by themselves, or to map the SRS Strictness to map to different Desired Retention (80, 85, 90, 95% could be good values)

Another question is parameters optimization. It is normally done by user but in Anki it requires manual optimization. Automatic Optimization can be implemented but I’d argue having some minimum review number before being able to optimize is probably better to avoid funky situation. Typically, especially in systems were you go from N5 to N1 with what I’ll assume to be harder and harder content, you might get a model that will be trained on N5 content you might already master, and FSRS will become overly optimistic about your parameters, leading you then to have very long intervals for N3-N1 content until you fail them ~6 months later and then suddenly the optimization will change the parameters.

But it doesn’t stop there, another concept of living with FSRS is rescheduling. If you don’t reschedule, the cards planned in 1 year will stay there. So if your params are now less optimistic, you’ll have to reschedule cards. But this action is known to lead to immense backlogs since you have maybe 300 reviews to do as a backlog that were initially planned in 6m-12months.

So all in all, I think FSRS is indeed a good way forward BUT it’s also paved with got’chas and not everyone has been very happy with it, often because FSRS doesn’t care about “increasing retention” the longer you’ve been knowing a word, it just care about keeping your Desired Retention the same, which is why I don’t advice DR lower than 90% such as 80% except if you agree having 1y interval card with a target retention of 80%.

I’ve been using and discussing FSRS for the past 15 months now so feel free to ask.

I’ve listed all the major complaints with FSRS :

Goal here is not to dissuade from it, but more to understand what is the hidden cost that unfortunately many people will try to obfuscate. And also to help Bunpro’s team to be ready for the potential support required to User, since FSRS will be more difficult to interpret for people less versed in how it is built

7 Likes
#17

It’s not obvious how to compare SRS algorithms to begin with.

The vast majority of learners does additional reviews outside of Anki, a.k.a immersion + other tools. That data is not available in the sets used to compare different models and parameters, and so it’s already a “garbage in, garbage out” situation.

The correlation between cards should have a massive impact on results too, and I don’t readily see if the data set was controlled for that. It’s one thing to test on “absolute beginner’s core 3k different words”, and another on Bunpro’s “each word has 10 different sentences”.

And of course one elephant in the room is that user age is probably a much stronger predictor of retention than used study method. I don’t see if the algorithm tests were ever split by age cohorts at all.

So FSRS versionX is “scientifically proven” to be the absolute best for whom, on what kind of card set? For everybody and everything? Heh.

This is not to say that I’m defending WK’s fixed 2x intervals. I just think there are other low hanging fruit related to the scheduler still not fixed/improved. Like not showing duplicate sentences from different sources (ghost/no ghost) on the same day.

1 Like
#18

in order to account for all those differences, statistics are made on the entire userbase which is diverse in terms of learning material (not even just languages), age, card style, exposure outside of SRS etc. I’ve occasionally peeked at the dev discussions (both on discord and github) and the main metric that is typically used to argue for or against an implementation is “[change] is better for [x]% of users”. With “better” mostly meaning “higher retention”.

That said, I try not to overthink this. I’d rather use a mildly less optimized system that keeps me motivated rather than a highly optimized one that burns me out. We can spend weeks digging into the details and the specifics, or we can just spend the same time learning Japanese. Use whatever you enjoy using, regardless of whether something is objectively a little “better” or not.

Thinking back on how I learned English in school as a kid. No SRS, no Anki, no sentencepro, probably the worst textbook in existence and not a whole lot of motivation. Yet somehow I now manage to use the language. Albeit not in a perfect way.

4 Likes
#19

That’s how I feel as well. I think some people overthink SRS generally, personally I just see it as a review load management algorithm and nothing more. I’ve been using SRS systems for close to a decade now to learn various languages, in my experience the difference between a “dumb” fixed-interval algo like what bunpro and many other websites use and Anki’s FSRS is fairly small at the end of the day when it comes to my ability to learn the languages.

I do however think that the current scheduling algorithm used by bunpro is not very suitable for pre-made vocab decks because I want words that I’m already familiar with to be very quickly kicked to long intervals, something that currently requires manually changing the SRS tier in bunpro and would be handled effectively automatically with something like FSRS.

6 Likes
#20

TBH I don’t think the algorithm is the real issue. For me, the problem with Bunpro is that every review has only two possible outcomes: pass or fail. There’s no rating for how well you were able to recall something.

If I’m able to recall a word but it takes me 10 seconds, then clearly I need to see that word again sooner than if I’m able to recall it instantly. But Bunpro treats both scenarios equally.

Often when faced with that situation, I’ll actually intentionally “fail” the review item just because I want to force Bunpro to show it to me again soon. If it had a recall rating system like Anki, and adjusted the next review based on that, I think it would work much better—regardless of whether it’s FSRS or whatever.

2 Likes