README.md

metadata

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - diffusion-models
  - flow-matching

ELF: Embedded Language Flows

This repository contains pre-tokenized datasets used in the paper ELF: Embedded Language Flows.

Github | Paper

ELF is a class of diffusion models in continuous embedding space based on continuous-time Flow Matching. The datasets provided here are pre-tokenized using the T5 tokenizer and encoded using a frozen T5-small encoder as described in the paper.

Dataset Details

The authors provide pre-tokenized splits for several benchmarks:

• OpenWebText: Used for unconditional generation.
• WMT14 De-En: Used for machine translation.
• XSum: Used for abstractive summarization.

Usage

You can load the pre-tokenized datasets directly using the Hugging Face datasets library:

from datasets import load_dataset

# Example: Load the OpenWebText pre-tokenized dataset
dataset = load_dataset("embedded-language-flows/openwebtext-t5")

# Example: Load the WMT14 De-En validation set
dataset_val = load_dataset("embedded-language-flows/wmt14_de-en_validation_t5")

Citation

@article{elf2026,
  title={ELF: Embedded Language Flows},
  author={Hu, Keya and Qiu, Linlu and Lu, Yiyang and Zhao, Hanhong and Li, Tianhong and Kim, Yoon and Andreas, Jacob and He, Kaiming},
  journal={arXiv preprint arXiv:2605.10938},
  year={2026}
}