toksuite/tokenizer_robustness_completion_math

Dataset Card for Tokenization Robustness (Math)

TokSuite Benchmark (Math Collection)

Dataset Description

This dataset is part of TokSuite, a comprehensive benchmark designed to measure how different tokenization strategies affect language model behavior under controlled conditions.

This specific subset focuses on mathematical text completion, containing multiple-choice math questions with a variety of surface-form perturbations that stress tokenizer handling of numbers, symbols, formatting, scripts, and mathematical notation.

• Curated by: R3 Research Team
• Domain: Mathematics
• License: MIT License

Dataset Summary

TokSuite isolates the impact of tokenization by holding model architecture, training data, training budget, and initialization constant, varying only the tokenizer.

The Math benchmark evaluates performance on:

• A canonical mathematical formulation
• Multiple perturbed variants that preserve mathematical meaning while altering surface representation

These perturbations reflect realistic variation in how mathematical expressions are written, formatted, localized, and queried in practice.

Key Features:

• 21 canonical math questions with unambiguous answers
• Perturbations targeting notation, symbols, scripts, and formatting
• Parallel structure with TokSuite language benchmarks
• Designed for evaluation, not training

Supported Tasks

• Multiple-Choice Math Question Answering
• Tokenizer Robustness Evaluation
• Symbolic and Numerical Text Processing

---

Dataset Structure

Data Fields

Field	Type	Description
question	string	Mathematical question text
choices	list[string]	Multiple-choice answer options
answer	int64	Index of the correct answer
answer_label	string	Letter label of the correct answer
split	string	Dataset split identifier (all entries are test)
subcategories	string	Perturbation category
lang	string	Domain identifier (math)
second_lang	string	English translation or description of the question
notes	string	Additional context about the perturbation
id	string	Unique question identifier
set_id	float64	Question set grouping identifier
variation_id	float64	Variation number within a question set
vanilla_cos_sim_to_canonical	dict[string, float]	Cosine similarity to canonical form using raw token sequences
trimmed_cos_sim_to_canonical	dict[string, float]	Cosine similarity after token normalization
token_counts	dict[string, int]	Number of tokens produced per tokenizer

---

Dataset Creation

Curation Rationale

This dataset was created to:

1. Systematically evaluate tokenizer robustness on mathematical notation and structure
2. Measure sensitivity to changes in formatting, symbols, scripts, and numeric representation
3. Isolate tokenization effects from mathematical reasoning difficulty
4. Provide standardized benchmarks for math-focused language models

Canonical questions are intentionally simple and high-accuracy, allowing researchers to attribute performance degradation to tokenization rather than reasoning complexity.

Source Data

• Canonical math questions were manually authored
• Each question was perturbed while preserving mathematical equivalence
• Canonical accuracy was validated across TokSuite models

---

Perturbation Categories (Math)

1. Canonical
   The baseline mathematical text written in a standard, well-formatted form with no perturbations. This serves as the reference condition for evaluating all other perturbations.

2. Chinese
   Rewrites mathematical text using Chinese characters for numbers, operators, or surrounding descriptions, testing tokenizer robustness to non-Latin scripts in math contexts.

3. Decorative Unicode
   Replaces standard mathematical symbols with visually similar decorative or stylized Unicode characters (e.g., fancy numerals or operators), stressing Unicode normalization and symbol handling.

4. Farsi
   Introduces Persian (Farsi) numerals or script elements into mathematical expressions, testing tokenizer robustness to right-to-left scripts and cross-script numeric representations.

5. Italian
   Rewrites textual components of math problems in Italian while preserving the same mathematical structure and solution.

6. LaTeX
   Encodes mathematical expressions using LaTeX-style syntax (e.g., \frac, ^, _), stressing tokenizer handling of markup-heavy mathematical notation.

7. Space Removal
   Removes or alters spacing within mathematical expressions and surrounding text, stressing tokenizer assumptions about whitespace in math contexts.

8. Spelled-Out Forms
   Replaces numerals or symbols with fully spelled-out textual equivalents (e.g., numbers written as words), increasing sequence length and altering token boundaries.

9. Turkish
   Rewrites textual components of math problems in Turkish while preserving the underlying mathematical meaning.

---

Considerations for Using the Data

• Language variety: The dataset uses standard mathematical notation and English-language math phrasing, and may not represent informal or pedagogical math language.
• Script focus: Mathematical expressions are primarily written using ASCII and standard Unicode; LaTeX, decorative Unicode, and non-Latin scripts are included as perturbations.
• Domain coverage: Questions focus on general mathematics and may not represent highly specialized or advanced mathematical domains.
• Question simplicity: Designed for high baseline accuracy, which may not reflect real-world mathematical task complexity.

---

Additional Information

Dataset Curators

The dataset was curated by the TokSuite research team at R3.

Licensing Information

MIT License

Citation Information

If you use this dataset in your research, please cite the TokSuite paper:

@inproceedings{toksuite2026,
  title={TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior},
  author={Altıntaş, Gül Sena and Ehghaghi, Malikeh and Lester, Brian and Liu, Fengyuan and Zhao, Wanru and Ciccone, Marco and Raffel, Colin},
  year={2026}
}

Paper: TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior

Contributions

This dataset is part of TokSuite, which includes:

• 14 language models with identical architectures but different tokenizers
• Multilingual benchmark datasets (English, Turkish, Italian, Farsi, Chinese)
• Comprehensive analysis of tokenization's impact on model behavior

Contact

For questions or issues related to this dataset, please refer to the TokSuite project or contact the authors of the paper.

---

Part of the TokSuite Project

Understanding Tokenization's Role in Language Model Behavior