10k mastered in the Russian fast track, some feedback

Hello all,

After a few years of diligent use, I managed to reach 10 thousand mastered words in the Russian “Fast Track” today:

I enjoyed the journey but I also think I’m not going to make it to the full 20k. I think it’s time for me to drop Clozemaster for Russian, I don’t get enough out of it anymore.

Paradoxically the thing that makes ClozeMaster less useful for me at this point is also the thing that made it so great at the start: unlike many other vocab SRS tools, ClozeMaster considers different forms of a word (conjugations/declensions) to be distinct. That means that you get to practice vocab in all sorts of grammatical constructions instead of just dictionary form like most other tools (in my experience).

For a language like Russian with tons of declensions I find it extremely valuable, since it forces you to think about grammar instead of just rote memorization of words.

Unfortunately I feel like after a while you get the other side of the coin: because a single noun in Russian has potentially a dozen forms, you tend to get the same words again and again and again with different forms instead of learning more vocab.

As an example, here’s what my new lessons look like:

Тебе нужно получить нормальную работу.

The cloze is the adjective “normal” in the feminine accusative.

Они не желают с вами разговаривать.

The cloze is the verb “to want”, third person of the plural.

Небо - над нашими головами.

The cloze is “head” in the instrumental plural.

Свяжись с ней.

The cloze is “to get in touch/to contact” in the perfective imperative.

All of these are trivial words I know full well and I’ve probably encountered (some of them multiple times) in ClozeMaster.

I would guesstimate that at this point for ever 10 lessons I get maybe 2 words I feel worth studying, and honestly it may be even less than that.

Of course I can just manually tag these entries as known and keep doing lessons until I find something useful, but that’s a lot of busywork and I’m lazy.

I’m not sure how ClozeMaster could improve on this, and again overall I got a lot out of this website and I absolutely don’t regret having gone through all of this. I just felt like offering some feedback.

Thank you for this website and keep up the good work!


Thanks for the great feedback.

I know exactly what you mean and I expect to follow the same path.

I think I can provide some feedback from many angles.

I have used most language softwares: Rosetta Stone, Duolingo, Glossika, Babbel, Anki, Supermemo, Pimsleur, Lingopie, LingQ, Language Transfer.

I stop/dump using them when I see no comparative benefit. Few survive.

I did not think that Clozemaster would survive when I switched it on to learn Italian, but it is still going strong.

As you say, it is well suited for beginners.

A lot of intelligence went into the user interface. Countless good decisions (not usual in that space). It flows very well.

The learning gradient is very (too?) gradual. Non-lemmatization is indeed a factor here: useful and friendly for the beginners, but not so much for the accomplished.

The super-simple spaced repetition is a major source of concern for intense, long term users.

Here is the repetition history from a flashcard in my Anki deck to illustrate the issue:


You can see how it is going to go differentially compared Clozemaster: on my 6th review, the revision interval, given the absence of failed revision, is 19 years.

Anki correctly assumes that I have perma-stored the knowledge on the card.

I wonder how many repetitions take you to a 19-year interval in Clozemaster…

For the heavy user, the consequence is a GIGANTIC difference in workload on the easy stuff in the long term.

However, Clozemaster is clever. It offers a sort of solution.

Alt+k = mark as know

It requires you to assess whether you permanently know a card. Not easy when you just want to keep hammering your daily stack. It adversely affects the smoothness of the flow. Granted.

Any other smart solution?

Yes, I think so. Clozemaster could use the collective intelligence of the bee hive of its learners.

If it gives a chess type ELO rating to its leaners and to each sentence, and adjust the ratings of each after each confrontation (correct recall = win for the learner, incorrect = win for the sentence), it will gradually accumulate a fine quantification of the difficulty of each sentence for learners coming from a given language. I believe this should be done in language pairs.

The too easy sentences that now bore you stiff will be found too easy by the other learners as well. They will end up with a low ELO rating.

Now the clever bit: the Clozemaster server is now in a position to pick sentences that are closest to your rating from its database. All the way to full mastery. As your ELO rating gradually creeps up, what you learn remains constantly congruent with your progress and level.

If the declension is tough to learn and matches your level, and only then you will be exposed to it.

If you are a fast learner, it will show in your ELO rising rapidly, but the difficulty of the sentences served to you can now fully adapt and respond to that.

You do not learn by frequency, a poor proxy, but by the difficulty the bee hive has to master a given sentence. Unusual syntax? Collocation difficulty? It will pick it up.

All the initial sentence ELO assessment can actually take place unbeknownst to the users in the exploratory phase. Keeping scores in the background.

Thus, you’ll end up with the Holy Grail: sentences properly ranked by actual, learning difficulty.

All that harvestable knowledge is currently lost.

It would immensely differentiate the Clozemaster product, make it much more unique and pleasant to use in the long run.

It would give Clozemaster a much better chance to keep its long term, proficient users.

It has a nice, potential, competitive gamification benefit as well. Which, I’m told, increases stickiness.

It is the way the best chess problem servers operate and it is a promising way to go for language, methinks.

But what do I know…


Precisely, at some point maybe around a year ago I would fudge with the SRS intervals (mostly setting new entries immediately to 50%/75%/100% learned, a bit like Anki’s easy/good/hard buttons) in order to avoid unnecessary repetitions, but eventually I stopped because as you mention it makes the process very much “not smooth” and I preferred to just let the SRS do its thing.

I do think that ClozeMaster could generalize the Easy/Good/Hard buttons currently available after 100% for new words as well, and perhaps put them on a different SRS schedule in that case (similar to Anki’s “ease” system, maybe).

A very interesting idea, however one potential issue I can see is that I’m not sure if ClozeMaster could gather enough data to make it work at higher level (where it matters most, for beginners everything will be hard, so to speak).

For instance if I look at the leaderboard for “mastered words” for Russian from English, I’m currently #66. That means that from now on the sentences I encounter will have at most 65 data points for difficulty (assuming these other users also all do the Fast Track). Is it enough for this kind of rating to be accurate? Perhaps.


@simias , could you explain why you’re limiting yourself to the Fluency Fast Track? Is there a reason you’re not using any of the “Most Common Words” groups? These groups, particularly the ones at the least common end, give you much more ability to find words whose dictionary forms (not just inflected forms) will be new to you. With a little experimentation, you should be able to find the “sweet spot” where the words are unusual enough to be unfamiliar, but common enough to be useful. There may come a time at which you even find the “>50,000 Most Common” group doesn’t give you enough new words (given the fact that it’s working from word forms rather than lemmas), but that should be further down the road. By that point, you might even want to create your own collections, which is not as much work as it might seem once you’ve developed a good pipeline.


Honestly I never considered it. I’ll give it a try.


Alanf is right. To that I recommend checking out the grammar collections too. I study Polish, which also had complex grammar. I have found the grammar collections very helpful.

And maybe try changing your routine. Have you tried just listening exercises to see if you can understand the sentence before you reveal any text? It can be a new challenge. Cloze listening (desktop only) takes it even further - and is tough!


Yeah I admit that I’ve been coasting a bit in my Russian studies, I just let the routine drive me instead of the other way around.

I don’t use the clozemaster listening exercises mainly because I don’t really like synthetic voices, I prefer to practice that with real content. That’s not really a criticism of clozemaster though, I realize that getting clean native recordings for hundreds of thousands of sentences would be prohibitively expensive.

I hope that with the current AI boom we’ll have human-quality text-to-speech in the not too far future…

Did you try turning off the Clozemaster audio and using your device’s? I don’t know if that is any better for you. It is under options you can select when you are doing a lesson - the little gear thing in the upper right.

Oh, and Cloze-listening are actual recordings. But they aren’t easy.


I didn’t know that was a thing either. I’ll give it a try, thanks!


I wonder, how much time it took you to get there?

I’ve been using clozemaster for about two years now. That’s about 15 new clozes a day on average.

1 Like