prajwalpujari16/token-atlas-benchmark

TokenAtlas: Multilingual Tokenizer Benchmark & Inflation Dataset

Overview

TokenAtlas provides empirical measurements of token inflation, bytes-per-token density, effective context window shrinkage ($W_{\text{eff}}$), and cost multipliers ($\kappa$) across 52 global languages and 6 frontier LLM tokenizers (gpt-4o, mistral-tekken, gemma-2, deepseek-v3, qwen2.5, llama-3.1, gpt-4).

Evaluations are performed against identical parallel sentences extracted and aligned from the Meta FLORES-200 multilingual evaluation benchmark.

Quickstart

from datasets import load_dataset
import pandas as pd

dataset = load_dataset("prajwalpujari16/token-atlas-benchmark", data_files="flores200_benchmark.csv")
df = pd.DataFrame(dataset["train"])
print(df.head())

Evaluated Frontier Tokenizers

• OpenAI GPT-4o (o200k_base, 200k vocab)
• Mistral Tekken (131k vocab, mistralai/Mistral-Nemo-Instruct-2407)
• Google Gemma 2 (256k vocab, google/gemma-2-2b)
• DeepSeek-V3 (128k vocab, deepseek-ai/DeepSeek-V3)
• Alibaba Qwen 2.5 (152k vocab, Qwen/Qwen2.5-7B)
• Meta Llama 3.1 (128k vocab, meta-llama/Llama-3.1-8B)
• OpenAI GPT-4 (cl100k_base, 100k vocab baseline)

Key Findings

• Chinese Super-Compression: DeepSeek-V3 encodes Chinese at $\tau = 0.98\times$ (more compact than English).
• Arabic Optimization: Mistral Tekken leads the industry on Arabic at $\tau = 1.41\times$.
• Dravidian Inflation Relief: GPT-4o ($\tau = 2.71\times$) and Mistral Tekken ($\tau = 2.94\times$) reduce Kannada sequence lengths by ~78% compared to Llama 3.1 ($\tau = 12.63\times$).

Interactive Web Space

The live interactive atlas and Token Tax & Context Shrinkage Calculator is hosted on Hugging Face Spaces at: 👉 prajwalpujari16/token-atlas

Citation

@software{Pujari_TokenAtlas_2026,
  author = {Pujari, Prajwal},
  title = {{TokenAtlas: Universal Multilingual Tokenizer Benchmark & Global Inflation Atlas}},
  year = {2026},
  url = {https://github.com/Prajwal7-2002/token-atlas},
  license = {Apache-2.0}
}