| license: mit | |
| tags: | |
| - tinyworlds | |
| - world-model | |
| - dynamics-model | |
| - checkpoints | |
| # large-dynamics-training | |
| Training run checkpoints for the TinyWorlds **dynamics** model (large variant). | |
| - Source run timestamp: `2026_06_24_15_31_15` | |
| - Checkpoints: 24 steps (`dynamics_step_0` through `dynamics_step_23000`, every 1000 steps) | |
| - Each checkpoint contains: | |
| - `model_state_dict.pt` (~305 MB) | |
| - `optim_state_dict.pt` (~611 MB) | |
| - `state.pt` (small training state metadata) | |
| - `dynamics/visualizations/` — per-step `dynamics_prediction_step_*.png` previews | |
| - `dynamics/data_partition.yaml` — data partition config used for this run | |
| ## Loading a checkpoint | |
| ```python | |
| from huggingface_hub import hf_hub_download | |
| import torch | |
| ckpt_path = hf_hub_download( | |
| repo_id="qtris123/large-dynamics-training", | |
| filename="dynamics/checkpoints/dynamics_step_23000/model_state_dict.pt", | |
| ) | |
| state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True) | |
| ``` | |