Datasets:

KreitnerL
/

BitTokens-dataset

license: mit
task_categories:
  - text-generation
language:
  - en
size_categories:
  - 100M<n<1B
tags:
  - arithmetic
  - numeracy
  - bittokens
  - synthetic
  - language-modeling
pretty_name: BitTokens Dataset

BitTokens Dataset

This dataset contains the exact synthetic number-problem CSV files used by the BitTokens paper configs in the public repository. It is intended for reproducing BitTokens and the FoNE, xVal, significant-digit, token-digit, and base-10 baseline experiments.

The CSV files are minimal copies of the original paper data. Only the columns required by the dataloaders are included:

prompt
text_prompt
answer
difficulty
difficulty_sd

All values were preserved as strings during processing to avoid numeric precision changes. The original source CSVs were not modified.

Files

The repository contains 37 CSV files:

14 train files
14 validation files
9 test files

The files cover Addition, Multiplication, Division, DivM, Exponentiation, MinMax, Interval, Sorting, Mean, and Std tasks, including the binary-uniform curriculum files used by BitTokens where referenced by the configs.

manifest.json records source sizes, output sizes, selected flavor, and stale config references that were skipped because they were not present in the source snapshot.

FineWeb Text Data

This dataset intentionally does not include the FineWeb-derived .txt files. Those belong to the public FineWeb dataset and should be downloaded from the original source instead. See the BitTokens repository README for the exact FineWeb download and decoding commands.

Usage

Download this dataset into your DATA_PATH directory:

hf download KreitnerL/BitTokens-dataset --repo-type dataset --local-dir "$DATA_PATH"

Then download and decode FineWeb separately if you want to reproduce the mixed numeric/text training runs. The BitTokens configs expect the decoded text files to be available under the local DATA_PATH as 000_00000_train.txt and val_text.txt.

Citation

If you use this dataset, please cite:

@inproceedings{
    kreitner2026bittokens,
    title={Efficient numeracy in language models through single-token number embeddings},
    author={Linus Kreitner and Paul Hager and Jonathan Mengedoht and Georgios Kaissis and Daniel Rueckert and Martin J. Menten},
    booktitle={Forty-third International Conference on Machine Learning},
    year={2026},
    url={https://openreview.net/forum?id=Bh4Ubk80M8}
}