FCZ7/gameworld-diamond-breakout-level5

DIAMOND GameWorld Breakout Level-5 Specialist (4 frames/action)

Model description

This checkpoint adapts DIAMOND (Diffusion As a Model Of eNvironment Dreams) to GameWorld's browser Breakout. It is a Level-5 specialist trained with a deterministic browser emulator interface. It preserves GameWorld rendering, physics, scoring, and layout while using four ALE-style actions and exactly four 60 Hz engine frames per action.

agent_epoch_01000.pt is a full DIAMOND Agent state dict containing the diffusion denoiser, reward/end model, and LSTM actor-critic.

Training protocol

• level: 5 only
• actions: NOOP, FIRE, RIGHT, LEFT
• action repeat: 4 × 60 Hz frames (0.066667 s simulated game time)
• observation: last executed frame, cropped/resized to RGB 64×64
• max pooling: disabled to preserve the black ball
• episode: five lives, no artificial step cap; level clear is success
• real interaction budget: 100,000 steps
• imagination horizon: 15
• DIAMOND batch size: 32 for denoiser, reward/end model, and actor-critic
• world-model compilation: enabled
• final epoch: 1000
• run: atari_v9_level5_20260721_163502

Final training-time validation

100 complete five-life Level-5 games:

Metric	Value
progress mean	0.974583
progress std	0.066625
success rate	0.65
native return mean	10848.65
native return std	846.51
task steps mean	1905.01
lives lost mean	3.42

These metrics use the deterministic training/test environment, not the official GameWorld evaluator.

Files

• agent_epoch_01000.pt: full Agent weights for inference/visualization
• trainer.yaml: resolved training configuration
• validation_metrics.jsonl: periodic and final validation history
• validation_latest.json: final validation record
• SHA256SUMS: integrity hashes

Usage

Use the repository's docs/INFERENCE_VIDEO.md and src/evaluate_gameworld_atari_agent_60hz.py. The environment and model code must match the handed-off GameWorld_DIAMOND repository.

Limitations

• It was trained only on Level 5.
• Cross-level behavior is uneven, with especially weak Level-3 generalization.
• Results are not directly comparable to ALE Atari Breakout scores.
• The deterministic training environment is intentionally different from the original GameWorld wall-clock evaluation loop.

License

Before publication, verify the redistribution terms of the DIAMOND code, GameWorld code, Breakout game assets, and trained weights. Do not replace this section with an upstream code license without checking that it also covers the model and bundled assets.