toksuite/tokenizer_robustness_completion_english

Dataset Card for Tokenization Robustness

TokSuite Benchmark (English Collection)

Dataset Description

This dataset is part of TokSuite, a comprehensive benchmark designed to measure how different tokenization strategies affect language model performance and robustness in isolation. This specific collection contains English multiple-choice text completion questions paired with a wide range of real-world surface-form perturbations that are known to interact strongly with tokenization.

• Curated by: R3 Research Team
• Language(s): English (en)
• License: MIT License

Dataset Summary

TokSuite addresses a core challenge in language model research: isolating and measuring the impact of tokenizer choice on model behavior. The English collection serves as the reference and anchor language for TokSuite, providing a high-resource baseline with diverse perturbations that generalize across domains and writing styles.

Key Features:

• 40 canonical English questions with high baseline accuracy
• Extensive perturbation coverage spanning typography, formatting, morphology, noise, and stylistic variation
• Parallel structure with TokSuite benchmarks in Turkish, Italian, Farsi, and Chinese
• Controlled design enabling clean measurement of performance degradation under perturbations

Supported Tasks

• Multiple-Choice Question Answering: Text completion with four answer options
• Tokenizer Robustness Evaluation: Measuring accuracy drop under token-altering perturbations
• Benchmarking Tokenization Effects: Isolating tokenizer behavior independent of model architecture or scale

Languages

The dataset contains text exclusively in English (language code: en).

---

Dataset Structure

Data Fields

Field	Type	Description
question	string	The question text in English
choices	list[string]	Four 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
subcategories	string	Perturbation category
lang	string	Language code (en)
second_lang	string	Optional paraphrase or descriptive reference
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 (raw tokens)
trimmed_cos_sim_to_canonical	dict[string, float]	Cosine similarity after token normalization
token_counts	dict[string, integer]	Token counts per tokenizer

---

Dataset Creation

Curation Rationale

The English benchmark was created to:

1. Serve as a high-resource reference language for tokenizer robustness studies
2. Systematically probe tokenizer sensitivity to formatting, noise, and stylistic variation
3. Enable controlled comparisons across tokenizers under identical model conditions
4. Provide a reusable evaluation suite for studying tokenization effects in isolation

All canonical questions are intentionally simple, ensuring high baseline accuracy so that observed performance changes are attributable to perturbations rather than reasoning difficulty.

Source Data

Data Collection and Processing

• Canonical Questions: 40 English questions authored by the TokSuite team
• Perturbations: Targeted surface-form transformations applied per question
• Validation: Model-in-the-loop verification to ensure canonical solvability

---

Perturbation Categories (English)

Each perturbation represents a distinct, realistic transformation of English text that can alter token boundaries or distributions.

1. Canonical
   Standard, grammatically correct English text with no perturbations. Serves as the reference condition.

2. Abbreviations
   Introduces common English abbreviations and shortened forms (e.g., Dr., etc., vs.).

3. Capitalization
   Alters casing patterns through random capitalization, lowercasing, or mixed case.

4. Character Deletion
   Removes characters within words, simulating typing omissions.

5. Character Substitution
   Replaces characters with visually or keyboard-adjacent alternatives.

6. Colloquial
   Applies informal spoken English forms and casual phrasing.

7. Compounds
   Merges multi-word expressions into compound forms (e.g., notebook vs. note book).

8. Contractions
   Uses contracted forms such as don’t, it’s, and they’re.

9. Date Formats
   Varies date representations (e.g., March 12, 2022, 12/03/22, 2022-03-12).

10. Emoji Substitution
    Replaces words with semantically related emojis.

11. Grammatical Errors
    Injects plausible agreement, tense, or syntactic errors.

12. Historical Spelling
    Uses archaic or historical English spellings.

13. Homoglyphs
    Substitutes characters with visually similar Unicode glyphs.

14. Hyphenated Spelling
    Introduces or removes hyphens in compound words.

15. Inflections
    Alters tense, plurality, or derivational morphology.

16. Keyboard Proximity Errors
    Simulates typos from adjacent keyboard keys.

17. Letter Repetition for Emphasis
    Repeats letters for expressive emphasis (e.g., soooo).

18. Lowercase
    Converts all text to lowercase.

19. Macron / Diacritic Styling
    Adds diacritics uncommon in modern English.

20. OCR Errors
    Introduces character confusions typical of optical character recognition.

21. Orthographic Errors
    Applies plausible spelling mistakes.

22. Scripted Text
    Uses decorative or stylized Unicode script characters.

23. Similar Words
    Substitutes near-synonyms or easily confusable words.

24. Space Removal
    Removes spaces between words.

25. Spaced Styling
    Inserts extra spacing between characters or words.

26. Spelled-Out Forms
    Replaces numerals or symbols with fully spelled-out equivalents.

27. Superscript / Subscript Styling
    Uses Unicode superscripts or subscripts.

28. Web Search Query
    Rewrites questions in keyword-heavy search-engine style.

29. Word Reordering
    Alters word order while preserving overall meaning.

---

Annotations

Annotation Process

All canonical questions and perturbations were manually created and reviewed by the TokSuite team. Perturbations were designed to reflect realistic surface-form variations encountered in English text processing.

Annotators

Researchers and contributors involved in the TokSuite project with expertise in NLP and tokenizer behavior.

---

Personal and Sensitive Information

The dataset contains no personal, sensitive, or identifying information. All questions are general-knowledge based.

---

Considerations for Using the Data

Social Impact

This dataset supports research into more robust and equitable language technologies by highlighting how tokenization choices affect model behavior, even in high-resource languages like English.

Biases and Limitations

• Focuses on Standard English
• Evaluation-only dataset with limited size
• Multiple-choice format
• Does not cover domain-specific or conversational tasks

---

Additional Information

Dataset Curators

TokSuite Research Team (R3).

Licensing

MIT License.

Citation

If you use this dataset, 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},
  booktitle={Preprint},
  year={2026},
  arxiv={https://arxiv.org/abs/2512.20757},
  url={TBD}
}

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