This is the best checkpoint of the Inverse Dynamic Model I've trained on DOOM (1993) gameplay. It's good enough to label data for a world model, though it can still be improved.
Metrics
I've calculated the balanced F-score (F1) for each label:
| Action | F1 Score | Note |
|---|---|---|
| F1/ATTACK | 0.564 | |
| F1/CK | 0.581 | |
| F1/FWD | 0.809 | |
| F1/LEFT | 0.586 | |
| F1/RIGHT | 0.607 | |
| F1/SPEED | 0.000 | Not used in training |
| F1/TURN_L | 0.541 | |
| F1/TURN_R | 0.513 | |
| F1/USE | 0.453 | Lowest score |
| Accuracy (eval) | 0.834 | 83.35% |
| Loss (eval) | 0.375 |
Training:
- Epoch: 60
- Loss: 0.607
License
The weights are in the public domain.
- Downloads last month
- 197
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support