MUTANT

This repository contains the dataset for multilingual tokenizer evaluation MUTANT, introduced in

“MUTANT: A Recipe for Multilingual Tokenizer Design”
🔗 arXiv:2511.03237
🏛️ Accepted at ACL 2026

To assess tokenizer behavior in Indic use cases, we construct an evaluation set spanning 22 Indic languages, English, and code.

Using the data

You can read JSON directly, or use datasets. Example:

from datasets import load_dataset

# Load the MUTANT dataset
dataset = load_dataset("krutrim-ai-labs/MUTANT", split="test")

# View the structure of the dataset
print(dataset)

Code & Evaluation

The official inference and evaluation codebase is available on GitHub.

GitHub Repository: https://github.com/ola-krutrim/MUTANT

The repository provides the complete pipeline for running inference and reproducing benchmark results across all languages. We develop a modular evaluation framework supporting HuggingFace, SentencePiece, and TikToken tokenizers along with a comprehensive set of intrinsic metrics, including Fertility score, Normalized Sequence Length (NSL), Rényi entropy and efficiency, and Bytes Per Token (BPT). All metrics are computed at the line level and aggregated to the language level.

License

This repository is licensed under the Krutrim Community License.

Citation

If you use this evaluation dataset in your research, please cite:

@inproceedings{rana2026mutant,
    title={MUTANT: A Recipe for Multilingual Tokenizer Design}, 
    author={Souvik Rana and Arul Menezes and Ashish Kulkarni and Chandra Khatri and Shubham Agarwal},
    year={2026},
    booktitle={ACL}
    url={https://arxiv.org/abs/2511.03237}, 
}

Acknowledgement

Our work is built with reference to the code of the following projects: Tokenizers,SuperBPE, TikToken, SentencePiece and Transformers.