Expertium

Thank you. Though, after doing some more tests, it now seems that FSRS-7 is worse tthan FSRS-6 at non-same-day reviews regardless ``` Model: FSRS-7-secs-recency-equalize_test_with_non_secs Total number of users: 1000 Total...

You would have to: 1) Count decks in a preset 2) Optimize parameters for each deck 3) Run this code to suggest which decks can be kept under the same...

I have improved it: I now use ln(S0) instead of raw S0 New threshold=3.8 New robust covariance matrix = [[0.9642076199291478, -0.5292868704864732, -0.049966461513558005, -0.0677452128762764, -0.07344767208808632, 0.08608800179946072, -0.030717573802218037, -0.3707881667390443, -2.324717847663287, -0.08974150661616466, 2.2146918666289706,...

For some reason I need to use both `--short` and `--secs`, unlike with FSRS-6. Idk why

I'll modify `init_w`, `default_params_stddev_tensor` and the constants in pretrain later. Oh, and we'll need to do the big pretrain on 1000 users, so we'll need to call those guys with...

Idk why

I couldn't get checks to pass with s_min=0.001, so I reverted it back to 0.0001. Well, I'll probably end up using 0.0001 anyway, it seems like 0.001 makes metrics worse

@Luc-Mcgrady I don't understand this stuff. I was trying to add https://github.com/open-spaced-repetition/srs-benchmark/commit/033cdaa4ef4f24057a8742580f2d4e5b6101265a to this branch, but it causes a conflict EDIT: alright, I think it works now

> Any objections? Yes. Let's not make FSRS the default before automatic optimization. Realistically, how many users do you expect to click "Optimize" at least once in their lifetime? I'd...

 And FSRS-5 with optimized parameters is better in 99.0% of cases. In % it may not seem like a big difference (91.9% vs 99.9%), but in terms of odds,...