This model was presented in the paper The Functionalizer: Lossless Functional Decomposition for Subword Tokenization.
Functionalizer 100M model
This is a 100M example model. Each model is trained on the dataset specified in the name for 50000 epochs with the given seed.
The model was trained using the Functionalizer framework. Training code and detailed performance analysis are available: https://github.com/connor-makowski/functionalizer
Running the model
To run the model, use the custom tokenizer fork available here: https://github.com/connor-makowski/tokenizers/tree/functionalizer
Citation
If you find this model or the Functionalizer framework useful, please cite:
@misc{makowski2026functionalizerlosslessfunctionaldecomposition,
title={The Functionalizer: Lossless Functional Decomposition for Subword Tokenization},
author={Connor Makowski and Willem Guter},
year={2026},
eprint={2609.15991},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2609.15991},
}