Dospacite/mackolik-player-aware-transformer-80gb

Mackolik Player-aware Transformer

Checkpoint trained by scripts/mackolik_player_aware_transformer.py.

Dataset repo: Dospacite/mackolik-transformer. Run mode: full; max sequence length: 1858. Peak CUDA allocation: 52.851 GiB.

Files:

• model_state_dict.pt — PyTorch state dict plus config and feature state.
• checkpoints/epoch_*.pt — per-epoch checkpoints with model and optimizer state.
• checkpoints/latest_checkpoint.json — pointer to the latest and best epoch checkpoints.
• training_config.json — training hyperparameters.
• feature_state.json — vocabularies and numeric feature normalization.
• metrics.json — validation/test metrics and dataset profile.