[Question] Long learning intervals as a way to combat leeches

[kanjieater]

Research

Enter an [x] character to confirm the points below:

• [x] I have checked the FAQ and could not find an answer to my question
• [ (in progress) ] I have read the wiki and still felt confused
• [x] I have searched for similar existing questions here

Question I've used Anki for learning Japanese since 2016 daily. I've always had a lot of leeches over time, and suspend them after 4 failures. Given that I have almost 14,000 Japanese word/phrases cards active at this point it's not uncommon for a cards to have built up 4 failures over the years.

I have found that using a longer Learning phase helps prevent the cards from leeching too soon.

It's important that the 4 failures don't count towards the leech suspend count in the Learning phase. My learning phase is 0.16m 0.32m 0.64m 5m 1hr 1d 3d 6d. I'm using an old version of anki so my actual input is: 0.16 0.32 0.64 5 60 1440 4320 8640. Then a graduating ivl is 12 days. At that point either the next review will make the card Mature or it will be leeched out quickly if it can't stick in the Young phase.

My relearn phase is 10 60. Since I leech quickly, I just use the default 250% ease factor.

I've done a lot of experimentation over the years with settings, and finally after years of having a low retention rate, my retention has started climbing steadily.

2022 by month (started tracking leeches in order to reduce them) - Young% is percent correct (I use pass/fail):

2023 by month (April is in progress):

So my question is how could use fsrs to keep from cards being marked as leeches too soon? I was hoping to keep long learning ivl, in the same way Anki is built. But the FAQ says

And I recommend not setting a step of more than one day. For example, if your current steps are 10m 1h 1d 2d, you had better remove the 1d 2d from the steps.

[L-M-Sherlock]

Why not increase the leech threshold?

[L-M-Sherlock]

The lapses occur during the learning steps didn't count in the leech. I think it is equivalent to removing the extra learning steps and increasing the leech threshold.

[L-M-Sherlock]

The lapses occur during the learning steps didn't count in the leech. I think it is equivalent to removing the extra learning steps and increasing the leech threshold.

[kanjieater]

Why not increase the leech threshold?

The lapses occur during the learning steps didn't count in the leech. I think it is equivalent to removing the extra learning steps and increasing the leech threshold.

It's not quite the same. I don't ever want to mark a card as a leech while it's in the learning phase (or equivalent of the "low ivl" if I were to remove the learning steps). When you have low exposure to a rare word, it's not uncommon to fail it at the beginning of learning it significantly more until you've simulated more exposure to it and eventually building up a proper learning curve.

[L-M-Sherlock]

OK. You can use long learning steps. But FSRS scheduler cannot know all reviews occurred in learning steps. If you need a more accurate schedule, I recommend using rescheduling supported by the FSRS helper add-on.

[kanjieater]

OK. You can use long learning steps. But FSRS scheduler cannot know all reviews occurred in learning steps. If you need a more accurate schedule, I recommend using rescheduling supported by the FSRS helper add-on.

Thank you for the update & sorry for the delay in response. Could you describe how the long learning steps would impact the FSRS algorithm?

[L-M-Sherlock]

FSRS calculates the initial stability in the first rating and schedules the interval in the last rating of learning steps. The first stability could be one day, even if the long learning steps are longer than one day.