Alice (an Architecture for Learning Interpretable Cryptogram dEcipherment) is a transformer-based neural network designed to solve cryptograms.
git clone https://github.com/al-jshen/alice.git
cd alice
pip install -e .- Prepare your configuration: Copy and modify one of the config files in
configs/:
cp configs/alice_base.yaml configs/my_config.yaml-
Update the configuration: Edit
configs/my_config.yamlto set your dataset path, W&B credentials, etc. -
Train the model:
python alice/train.py --config configs/my_config.yamlTo train linear/non-linear probes on top of frozen model representations, run the included alice/train_probes.py script.
There is an example config named configs/probe_base_linear.yaml that you can modify.
We release the checkpoints for various pretrained models at this link.
You will need to instantiate the model with the same configuration used during training.
To use the bijective decoding head:
sinkhorn_decoding: true
sinkhorn_iters: 10
sinkhorn_tau: 1.0Enable dynamic embeddings:
dynamic_embeddings: true
embedding_n_layer: 2
embedding_n_head: 4This project is licensed under the Apache-2.0 License. See LICENSE for details.