toksuite/tokenizer_robustness_completion_chinese

Dataset Card for Tokenization Robustness

TokSuite Benchmark (Chinese 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. This specific subset contains Chinese language multiple-choice text completion questions with various real-world perturbations that test tokenizer robustness.

• Curated by: R3 Research Team
• Language(s): Chinese (It)
• License: MIT License

Dataset Summary

TokSuite addresses a fundamental challenge in language model research: understanding how tokenization choices impact model behavior in isolation. The Chinese subset specifically measures model performance on canonical questions and various perturbations.

Key Features:

• 40 canonical questions covering general knowledge, geography, science, and language understanding
• Multiple perturbation types reflecting real-world text variations in Chinese
• Parallel structure with TokSuite benchmark (available in English, Turkish, Farsi, Italian)
• Native speaker curation ensuring linguistic authenticity

Supported Tasks

• Multiple-Choice Question Answering: Text completion format with 4 answer choices
• Tokenizer Robustness Evaluation: Measuring performance degradation under various text perturbations
• Multilingual NLP Benchmarking: Evaluating language models on Chinese text understanding

Languages

The dataset contains text in Chinese (language code: zho_Hans / zh).

Dataset Structure

Data Fields

Field	Type	Description
question	string	The question text in Chinese
choices	list[string]	4 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
second_lang	string	English translation or description of the question
notes	string	Additional context about the question or 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 scores to canonical form (raw tokens)
trimmed_cos_sim_to_canonical	dict[string, float]	Cosine similarity scores after token normalization
token_counts	dict[string, integer]	Number of tokens produced per tokenizer

Dataset Creation

Curation Rationale

This dataset was created to:

1. Systematically evaluate how different tokenization strategies handle Chinese
2. Measure robustness against real-world text perturbations specific to Chinese
3. Support research into the impact of tokenization on language model behavior
4. Provide standardized benchmarks for Chinese language models

The questions were designed to be straightforward with high baseline accuracy, allowing researchers to cleanly measure performance degradation when perturbations are applied.

Source Data

Data Collection and Processing

• Canonical Questions: 40 baseline questions created in English
• Translation: Native Chinese speakers translated questions
• Perturbations: Each question underwent targeted perturbations designed to reflect Chinese characteristics
• Validation: Model-in-the-loop process ensured high baseline accuracy

Perturbation Categories

1. Canonical The baseline Chinese text written in standard, well-formed Simplified Chinese with no perturbations. This serves as the reference condition for evaluating the impact of all other perturbations.

2. Code / Language / Script Switching Mixes Chinese with English words, phrases, or symbols within the same sentence, reflecting real-world bilingual usage and code-switching commonly seen in technical or online contexts.

3. Colloquial Rewrites sentences using informal or conversational Chinese expressions, including spoken-style phrasing that differs from standard written Chinese while preserving meaning.

4. Equivalent Expressions Replaces canonical phrases with alternative Chinese expressions that convey the same meaning using different words or constructions, isolating tokenizer sensitivity to paraphrasing.

5. Keyboard Proximity Errors Introduces character-level errors caused by adjacent key presses in pinyin-based input methods, simulating realistic typing mistakes during Chinese text entry.

6. OCR Errors Introduces character substitutions, deletions, or confusions commonly produced by optical character recognition systems, especially for visually similar Chinese characters.

7. Optional Diacritics Adds or removes optional diacritic markers (e.g., tone marks in pinyin annotations when present), testing tokenizer robustness to auxiliary pronunciation cues.

8. Partially Romanized Mixes Chinese characters with romanized (pinyin or Latin-script) representations for some words or phrases, reflecting hybrid writing styles used in informal digital text.

9. Romanization Fully converts Chinese text into romanized form (e.g., pinyin), replacing characters with Latin-script equivalents while preserving pronunciation and meaning.

10. Space Removal Removes spaces that may appear between Chinese characters or between Chinese and Latin text, stressing tokenizer assumptions about whitespace usage.

11. Spelled-Out Forms Replaces numerals, symbols, or compact expressions with fully spelled-out Chinese equivalents, increasing sequence length and altering token boundaries.

12. Traditional Converts Simplified Chinese characters into their Traditional Chinese counterparts, preserving semantics while changing Unicode character forms.

13. Word Spacing, Zero-Width Characters, Extra Space Manipulates spacing by inserting extra spaces, removing expected spaces, or adding invisible zero-width characters, stressing tokenizer handling of segmentation and Unicode normalization.

Who are the source data producers?

Native Chinese speakers curated and validated all questions and perturbations. The TokSuite research team at R3 designed the overall benchmark framework.

Annotations

Annotation process

Questions were manually created and translated by native speakers. Each perturbation was carefully designed to reflect authentic variations encountered in real-world Chinese text processing.

Who are the annotators?

Native Chinese speakers with expertise in linguistics and NLP, working as part of the TokSuite project.

Personal and Sensitive Information

The dataset contains only general knowledge questions and does not include any personal or sensitive information.

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to improving language technology for Chinese speakers by enabling better understanding of tokenization challenges and supporting more robust multilingual models.

Discussion of Biases

• Language variety The dataset uses Standard Chinese (Mandarin) and may not fully represent regional or dialectal variations.
• Script focus: Simplified Chinese is used as the primary script; Traditional Chinese and romanized forms (pinyin) are included as perturbations.
• Domain coverage: Questions focus on general knowledge and may not represent domain-specific Chinese language use.
• Question simplicity: Designed for high baseline accuracy, which may not reflect real-world task complexity.

Other Known Limitations

• Relatively small dataset size (evaluation-only)
• Multiple-choice format
• Language-specific perturbations
• Results may differ at larger model scales

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},
  booktitle={Preprint.},
  year={2026},
  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