Currently, maps such as Backbeat Maniac & Perthed are highly overrated by the rankings.
For the future purposes of beatmap difficulty estimation, we should simply remove them.

To do so, we'll estimate the svness of each map, to find an optimal threshold where a map is "sv" or not.

Currently, the pipeline script in #29 ALWAYS runs the dataset preprocessing. This can be wasteful especially if

• the preprocessing SQL is the same.
• the dataset is the same

An idea would be to hash the SQL+Dataset string as a unique identifier for the dataset. However, it'd be good if we can also include how the dataset was generated, i.e. metadata for the .csv.

Currently, we just use the whole dataset, which can be

1. A bit too much samples
2. It's heavily biased against scores that aren't popular, but are more significant, such as 90%, in contrast to 99.5%, which doesn't say much

To even the training, we should try to uniformly sample across the sample space for a more representative measure and training process

I just found that the newest model didn't filter out non-active maps and players, this was a mistake due to the main repo using 0 for samples.

On this topic, maybe we should have a debug option for pipeline.sh and its stages?

Currently, we threshold maps with a hard SR. which is fine, however,

1. Maps with DT on these thresholded maps would have a hard threshold on whatever SR its DT is at
2. Maps with HT will go below this threshold

We need to threshold maps w.r.t. map and speed.

Currently in #29 , the preprocessing is done via a run.sh, while others are done with docker compose up --build.

This inconsistency is ugly, and furthermore, it can be annoying to dig into a long run.sh

Additionally, if it's possible, fetch env vars from shell

There should be enough neurons in the network to reasonably yield a distribution, furthermore i've regularized it.
Let's see if this works

Currently, we have train.py setting the model path on the .env file.

I don't think it's a great approach, as it widens the scope of train, which should just train the model, make the model, and possibly return the model path. The problem is that the model path return is not trivial, thus we resorted to I/O-ing to the .env file.

A better solution is to specify a unique model name, which is the pipeline run id, then we can grep the model from opal/models.

Not sure why, but The Living Tombstone - Nippontradamus (Everest Hope) [October's 7K Insane].osu is highly rated.
This caused heavy bias towards maps that are non-competitive.