README.md

metadata

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

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:

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

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