aimar00/4x-mamba-world-model

4X Mamba World Model

Current 4X strategy-game Mamba world model checkpoint.

It is intentionally minimal:

• model.safetensors - model weights only
• config.json - architecture/config metadata
• metrics.json - training/evaluation metrics from the final head-training run

No training corpus, infrastructure scripts, simulator source, benchmark harness, optimizer state, scheduler state, or unrelated project files are included.

Clean loader/model code is here:

https://github.com/ProjectAI00/4x-mamba

Model

• Architecture: MambaWorldModel
• Parameters: 28,709,251
• Hidden size: 512
• Layers: 8
• Max state tokens: 512
• Action vocabulary: 4096
• Token vocabulary: 4096

The model encodes canonical game state/action tokens, applies Mamba-style SSM dynamics in latent space, predicts the next latent, and scores actions with policy/reward/value heads.

The public GitHub implementation is PyTorch. It can run on CUDA through PyTorch (model.cuda()), but it does not vendor the fused Mamba/Triton backend.

The model was trained on synthetic traces from a 4X strategy-game simulator.

Reference Metrics

These numbers are included to document the released checkpoint, not as a complete benchmark suite.

Offline action-ranking metrics from metrics.json:

• Validation top-1: 0.3379
• Validation top-3: 0.6924
• Validation MRR: 0.5525
• Test top-1: 0.2881
• Test top-3: 0.6416
• Test MRR: 0.5120

Small head-to-head simulator checks from the associated run:

• Current model vs best greedy bot, 100 games, seed slice 390515: 34% win rate
• Current model vs best greedy bot, 100 games, seed slice 640515: 42% win rate

Notes

This checkpoint is provided as open research weights. It includes the trained model weights, config, model card, and run metrics only.