mohdelgaar/LingGen

---
language:
- en
pretty_name: LingGen
task_categories:
- text-generation
size_categories:
- 1M<n<10M
---

# LingGen

Official dataset release for the paper **LingGen: Scalable Multi-Attribute Linguistic Control via Power-Law Masking**.

This dataset contains the processed training and test data used for the LingGen experiments, with precomputed linguistic control vectors for each example.

## Dataset Summary

- `train`: 6,810,672 examples
- `test`: 2,000 examples
- `ling`: 40 released control attributes used by the public codebase
- `ling_all`: full 276-dimensional linguistic feature vector

Each example includes:

- `sentence`: target text
- `source`: source dataset identifier
- `ling`: released 40-attribute control vector
- `ling_all`: full feature vector before selecting the released subset

## Source Data

The processed examples are derived from public datasets used in the paper, including:

- C4
- SMF
- QQP
- ANLI
- MRPC
- STS-B
- RTE

This release redistributes processed text and derived linguistic features for research use. Users should consult the original source datasets for their respective licenses and usage terms.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("mohdelgaar/LingGen")
```

To use the dataset directly with the released code repository, save it to disk first:

```python
from datasets import load_dataset

dataset = load_dataset("mohdelgaar/LingGen")
dataset.save_to_disk("data/ling_sentences")
```

Code repository: https://github.com/CLU-UML/LingGen

## Citation

```bibtex
@misc{elgaar2026linggen,
title={LingGen: Scalable Multi-Attribute Linguistic Control via Power-Law Masking},
author={Mohamed Elgaar and Hadi Amiri},
year={2026}
}
```