v0.2.0

A major improvement from the previous release, primarily addressing the issue of "veering off."

Specs:

• same architecture, 2x128 LSTM for Standard model
• the Lite model remains identical to the one in v0.1.2, 2x64 LSTM using old dataset

The v0.2.0 model is trained on a heavily filtered dataset designed to remove extreme trajectories. The results are excellent.

It is still not immediately clear to me whether an LSTM is the best approach. I may also remove the Lite model altogether, its performance is dubious...

v0.1.2

Releasing the lite model, a trimmed down 64-neuron bilayer LSTM on the same dataset.

The intention is to support faster inferencing on constrained systems. This does not make generation time shorter- rather, it could increase it because a dumber model is more likely to veer off course.

This veering off is giving me food for thought for pivoting away from an LSTM approach...

v0.1.0

first beta release.