Postprocessing: YOLO26 end-to-end output; external non-maximum suppression is not required
The exported model is fixed to batch size 1 and 640x640; dynamic shapes are not enabled.
Classes
ID
Name
0
player_engineer
1
dome_laser
2
ore_iron
3
ore_cobalt
4
ore_water
5
enemy
Validation results
The best checkpoint was selected at epoch 76.
Metric
Value
Precision
0.98670
Recall
0.96927
mAP@0.5
0.99500
mAP@0.5:0.95
0.95111
Training and validation curves
Validation batch predictions:
Validation predictions
Training
Architecture: YOLO26n detection
Ultralytics: 8.4.104
Resolution: 640x640
Epochs: 100
Batch size: 16
Optimizer: auto
Seed: 0, deterministic mode enabled
Training device: Apple MPS
Training source revision: 9b95720c7e7e6eeb6e1ae5b7402caa3de683631d
Limitations
This v0 model is an ore-focused proof of concept. The dataset contains no boxes for player_engineer, dome_laser, or enemy; those classes should be treated as unsupported. Cobalt has no validation or test examples, water has no test examples, and the test split contains only iron. The split comes from one temporally segmented gameplay session, so nearby frames may remain correlated even though there are no exact duplicate images.
Do not interpret the aggregate validation metrics as evidence that all six configured classes work. Additional independent sessions and balanced labels are required before using this model as a general Dome Keeper state detector.
