---
license: mit
task_categories:
- summarization
language:
- en
tags:
- diffusion
- flow-matching
- language-modeling
- elf
---

This repository contains the pre-tokenized XSum dataset used in the paper [ELF: Embedded Language Flows](https://huggingface.co/papers/2605.10938).

The dataset is tokenized using the T5 tokenizer and is prepared for use with ELF, a class of continuous-time Flow Matching models that operate in the continuous embedding space of a frozen T5 encoder.

- **GitHub Repository:** [https://github.com/lillian039/ELF](https://github.com/lillian039/ELF)
- **Paper:** [ELF: Embedded Language Flows](https://huggingface.co/papers/2605.10938)

### Dataset Summary

For the summarization task (XSum), the data is structured for conditional generation:
- `condition_input_ids`: Tokenized source text (the article).
- `input_ids`: Tokenized target text (the summary).

The ELF model prepends the condition IDs to the input IDs and applies specific attention masks during training and inference.

### Sample Usage

You can load this dataset directly using the `datasets` library:

```python
from datasets import load_dataset

# Load the training split
dataset = load_dataset("embedded-language-flows/xsum_train_t5")

# Example of an item in the dataset
print(dataset["train"][0])
```

### Citation

```bibtex
@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}
}
```