| # 萬題v1 (WanTopic-V1-HRM) | |
| ## Model Description | |
| - **Author**: ShihteSiao | |
| - **Version**: v1.6-standard | |
| - **Architecture**: Hierarchical Reasoning Model (HRM) - Standard | |
| - **Parameters**: 54,189,442 (54.2M) | |
| - **Memory Usage**: 0.6GB | |
| ## Performance | |
| - **Test Accuracy**: 0.0% | |
| - **Best Validation Loss**: 0.4258 | |
| - **Peak Reasoning Efficiency**: 50.9% | |
| - **Max Reasoning Steps**: 18 | |
| ## Architecture Details | |
| - **Hidden Size**: 640 | |
| - **Attention Heads**: 10 | |
| - **High-level Layers**: 4 | |
| - **Low-level Layers**: 4 | |
| - **Total Layers**: 8 | |
| ## Training Configuration | |
| - **Batch Size**: 2 × 16 = 32 | |
| - **Learning Rate**: 6.925339958244806e-05 | |
| - **Epochs**: 30 | |
| - **Mixed Precision**: True | |
| - **Label Smoothing**: 0.05 | |
| ## Usage | |
| ```python | |
| import torch | |
| # Load model | |
| checkpoint = torch.load('WanTopic-V1-HRM_final_model.pth') | |
| # Initialize model with checkpoint['model_config'] | |
| # Load state dict: model.load_state_dict(checkpoint['model_state_dict']) | |
| ``` | |
| Generated on: 2025-09-03 11:00:23 | |