2026-02-24

2-M16 retrain with higher REST weight

Retrained 2-M16 with REST weight 0.2→0.8 and compared before/after metrics.

Historical note: archived update posts preserve the figures published at that time. For the current verified run bundles, use the results page.

What changed

We retrained the 2-M16 model with updated class weights, increasing the REST weight from 0.2 to 0.8. The goal was to reduce REST confusion without sacrificing finger separation on non-REST windows.

Before vs after (2-M16)

MetricBefore (REST=0.2, 2026-02-19)After (REST=0.8, 2026-02-24)Δ
Test action accuracy89.31%89.71%+0.39 pp
Test finger accuracy on non-REST windows89.90%90.38%+0.48 pp
Train avg loss0.51740.5020-0.0154
Test windows2,0402,0400
Test non-REST windows1,8711,8710

Notes

  • The reweighting produced small but consistent gains in both test action accuracy and finger accuracy on non-REST windows.
  • We will continue exploring class-weighting and calibration to improve REST vs movement separation without increasing false positives.

Links