A specialized RWKV model for solving 15-puzzle problems
- rwkv
- tkinter
- Run
demo.py
orminimum_inference.py
The current model rwkv_15puzzle_20241214.pth
is a specialized RWKV-v6 model trained on 1m 15-puzzle samples (~2.1B tokens) specifically for solving 15-puzzle problems.
Model specifications:
- Parameters: ~4M
- Vocabulary size: 83
- Architecture: 4 layers, 256 dimensions
The model includes a simple improvement for better performance (see model.py
line 372). Corresponding modifications were made in the inference code (rwkv_model.py
lines 852, 893-896).
The model was trained using the RWKV-LM repository.
Hyperparameters:
M_BSZ
: 64CTX_LEN
: 8192LR
: 12e-4 to 3e-5ADAM_EPS
: 1e-18ADAM_BETA1
: 0.9ADAM_BETA2
: 0.999WEIGHT_DECAY
: 0.1