README.md

---
license: mit
language:
- en
- tr
- it
- fa
- zh
tags:
- toksuite
- tokenization
- bloom
- multilingual
- bpe
- robustness
- research
pipeline_tag: text-generation
library_name: transformers
datasets:
- toksuite/toksuite_pretraining_data
---

<p align="left">
  <img src="./toksuite-logo.png" alt="TokSuite Logo" width="260"/>
</p>

# TokSuite – BLOOM

## Model Summary

**TokSuite–BLOOM** is part of **TokSuite**, a suite of language models designed to study the impact of **tokenizer choice on language model behavior** under controlled conditions.

This model uses the **BLOOM tokenizer** and is otherwise **identical** to the other TokSuite models in architecture, training data, training budget, and initialization. The TokSuite setup ensures that any observed behavioral characteristics reflect properties of the tokenizer rather than differences in model scale, data composition, or optimization.

---

## Tokenizer

- **Tokenizer:** BLOOM
- **Tokenization method:** BPE
- **Vocabulary size:** 250,680
- **Out-of-vocabulary handling:** Byte-fallback
- **Language coverage:** Multilingual
- **Pretokenization source:** BLOOM

**Processing details:**
- **Numbers:** Learned
- **Contractions:** Learned
- **Unicode normalization:** None
- **Whitespace / boundary markers:** Learned
- **Zerowidth chars:** Token

## Why BLOOM?

BLOOM was included in TokSuite to represent a **large-vocabulary multilingual BPE tokenizer** trained for broad cross-lingual coverage. As described in the tokenizer selection rationale of the TokSuite paper, BLOOM exemplifies a design choice that prioritizes extensive vocabulary capacity while maintaining subword-based segmentation.

Including BLOOM enables TokSuite to study tokenizer behavior in settings where:
- vocabulary size is large,
- segmentation follows BPE-style merges,
- and multilingual text is handled through a shared tokenizer.

This makes BLOOM a representative example of multilingual BPE tokenization.

---

## Model Architecture

- **Architecture:**  Decoder-only Transformer (Lingua's Llama-3.2-1B configuration)
- **Non-embedding parameters:** ~1B
- **Context length:** 4096 tokens
- **Framework:** Meta Lingua
- **Initialization:** Shared super-vocabulary initialization across TokSuite models

The architecture and training setup are identical across all TokSuite models; only the tokenizer differs.

---

## Training Data

The model was trained on a **multilingual corpus totaling approximately 100B tokens**, composed of:

- **English:** 40B tokens from *FineWeb-Edu*
- **Multilingual:** 60B tokens evenly distributed across:
  - Chinese (ZH)
  - Turkish (TR)
  - Italian (IT)
  - Farsi (FA)

You can find the pretraining dataset here: [toksuite/toksuite_pretraining_data](https://huggingface.co/datasets/toksuite/toksuite_pretraining_data)

All models in TokSuite are trained using a **fixed token budget**, reflecting common practice in large language model training.

---

## Training Procedure

- **Training steps:** 100,000
- **Sequence length:** 4096
- **Batch size:** 256 sequences
- **Optimizer:** AdamW
- **Peak learning rate:** 1e-3
- **Learning rate schedule:** Cosine decay with 2,000 warm-up steps
- **Weight decay:** 0.1

---

## Evaluation

### Canonical Benchmarks

The model was evaluated on standard base language model benchmarks:
- HellaSwag
- ARC
- PIQA
- XNLI

<p align="left">
  <img src="./model-performance-comparison.png" alt="TokSuite Logo" width="700"/>
</p>

These evaluations verify that the model exhibits reasonable base language modeling behavior at its scale and training budget.

### TokSuite Robustness Benchmark

TokSuite–BLOOM is evaluated on the **TokSuite robustness benchmark**, which measures sensitivity to real-world text perturbations, including:

- orthographic and spelling variations,
- diacritics presence and absence,
- keyboard and input-method noise,
- Unicode formatting and homoglyphs,
- OCR and spacing artifacts,
- LaTeX and STEM-style formatting.

**Tokenization Robustness under Multilingual Text Perturbations**  
Values represent **relative performance drop**, computed as `(Acc_clean − Acc_perturbed) / Acc_clean`,  where **lower values indicate greater robustness**.

Perturbation types include:  
- **Input:** non-native keyboard input and romanization  
- **Diacr.:** optional diacritics  
- **Orth.& Gram.:** orthographic and grammatical errors  
- **Morph:** morphological variations including derivations, inflections, and contractions  
- **Noise:** homoglyph substitutions, OCR artifacts, typos, and spacing errors  
- **LaTeX:** LaTeX-style mathematical formatting  
- **STEM:** scientific diagrams and notational conventions  
- **Unic.:** Unicode styling characters  

**NEN** denotes non-English inputs and **EN** denotes English inputs.  The **Avg** column reports the average relative performance drop across all perturbation categories.

| Model | Input (NEN) | Diacr. (NEN) | Orth. & Gram. (EN) | Orth. & Gram. (NEN) | Morph (EN) | Morph (NEN) | Noise (EN) | Noise (NEN) | LaTeX (EN) | STEM (EN) | Unic. (EN) | Avg ↓ |
|-------|-------------|--------------|--------------------|---------------------|------------|-------------|------------|-------------|------------|-----------|------------|-------|
| TokenMonster | **0.23** | **0.33** | 0.08 | **0.01** | 0.23 | **-0.07** | **0.10** | **0.18** | 0.21 | **0.10** | 0.51 | **0.17** |
| XGLM | 0.34 | 0.49 | 0.10 | 0.11 | 0.25 | 0.07 | 0.12 | 0.22 | **0.29** | 0.29 | **0.11** | 0.22 |
| BLOOM | 0.30 | 0.34 | 0.13 | 0.07 | **0.18** | **0.11** | 0.18 | **0.18** | 0.24 | 0.11 | 0.57 | 0.22 |
| ByT5 | 0.30 | 0.44 | **0.04** | 0.06 | 0.27 | 0.04 | 0.14 | **0.18** | 0.17 | 0.29 | 0.53 | 0.22 |
| Comma | 0.28 | 0.43 | 0.05 | 0.07 | **0.18** | 0.00 | 0.11 | 0.20 | 0.23 | 0.29 | 0.61 | 0.22 |
| mBERT | 0.33 | 0.44 | 0.11 | 0.11 | 0.23 | 0.06 | 0.18 | 0.22 | **0.14** | 0.22 | **0.61** | 0.24 |
| GPT-4o | 0.30 | 0.51 | 0.08 | 0.05 | 0.21 | 0.05 | 0.16 | 0.19 | 0.24 | 0.33 | 0.55 | 0.24 |
| GPT-2 | 0.34 | 0.46 | 0.07 | 0.10 | 0.25 | 0.06 | 0.14 | 0.21 | 0.24 | 0.35 | 0.53 | 0.25 |
| Phi-3 | 0.33 | 0.46 | 0.16 | **0.09** | 0.27 | 0.08 | 0.17 | 0.21 | 0.24 | 0.22 | 0.55 | 0.25 |
| Gemma-2 | 0.32 | 0.42 | 0.14 | **0.15** | 0.24 | 0.03 | 0.16 | 0.25 | 0.22 | 0.36 | 0.57 | 0.26 |
| Qwen-3 | **0.36** | 0.42 | 0.14 | 0.11 | 0.25 | 0.06 | 0.16 | 0.23 | 0.26 | 0.29 | 0.57 | 0.26 |
| Llama-3.2 | 0.33 | **0.55** | 0.11 | 0.10 | 0.25 | 0.08 | 0.15 | 0.24 | 0.17 | 0.30 | 0.59 | 0.26 |
| Aya | 0.31 | 0.46 | 0.14 | 0.10 | 0.22 | 0.03 | **0.19** | **0.25** | 0.21 | 0.38 | 0.58 | 0.26 |
| Tekken | 0.33 | 0.47 | **0.18** | 0.03 | **0.31** | 0.10 | 0.14 | 0.21 | 0.27 | **0.43** | 0.54 | **0.27** |
| **Avg** | 0.31 | 0.44 | 0.11 | 0.08 | 0.24 | **0.04** | 0.15 | 0.21 | 0.22 | 0.28 | **0.53** | 0.24 |

---

## Intended Use

This model is intended for:
- research on tokenization and robustness,
- multilingual NLP analysis,
- controlled ablation studies,
- benchmarking tokenizer behavior under noise.

It is **not** instruction-tuned, aligned, or optimized for deployment.

---

## Limitations

- Trained on a limited set of five languages.
- Not optimized for instruction following or dialogue.
- Fixed token budget constraints exposure to raw text depending on tokenization efficiency.
- Intended strictly for research purposes.

---

## Ethical Considerations

TokSuite models are released to support **scientific analysis of tokenization**.  
They may reflect biases present in large-scale web data and should not be used in high-stakes or user-facing applications without additional alignment and evaluation.

---

## Citation

If you use this model, please cite:

```bibtex
@article{toksuite2025,
  title={TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior},
  author={Altıntaş, Gul Sena and Ehghaghi, Malikeh and Lester, Brian and Liu, Fengyuan and Zhao, Wanru and Ciccone, Marco and Raffel, Colin},
  year={2025},
  arxiv={https://arxiv.org/abs/2512.20757},
}