ZoneTwelve/HRM-Sudoku-S1K

🧩 Model Architecture Summary

The model is configured using HRMACTModelConfig, designed for hierarchical recurrent memory with adaptive computation. Key architectural parameters:

• Sequence length: 81
• Vocabulary size: 10
• High-level cycles: 1
• Low-level cycles: 2

Transformer Configuration:

• Layers: 16
• Hidden size: 256
• Attention heads: 4
• Feed-forward expansion ratio: 4

Adaptive Computation Time (ACT):

• Maximum halting steps: 16
• Exploration probability: 0.1

🏋️ Training details

Total training time: 1 hour 16 minutes (76 minutes total)

GPU: x1 NVIDIA GeForce RTX 5080

Sample data number: [Easy, Medium, Hard, Extreme] * 1000 (4,000 total)

Sample data file: pregenerated_puzzles.npy

⚠️ Warning: Possible Data Limitation This training run may have been conducted with insufficient data. As a result, the model could be overfitting, which means it might perform well on the training data but poorly on unseen or real-world data. Please review the dataset size and diversity before deploying or relying on these results.