KnowSpread

This dataset is a 2.2B token subset of KnowSpread, it has been used to train Alephn-1, Elebush's 111M parameter language model.

The dataset was crawled to offer a large and diverse collection of web text(general data) for pretraining language models. The KnowSpread dataset was crawled as the training corpus for Alephn-1. This Hugging Face release offers a 2.2B token subset of KnowSpread, which is accessible for researchers and developers to explore, experiment with, and utilize in NLP research.

Read more about Alephn-1 and its training in the official announcement:

Introducing Alephn-1: https://www.elebush.com/news/introducing-alephn-1

Dataset Structure

The dataset is provided as Parquet files with the following fields:

Feature	Type	Description
content	string	Text content of the document
url	string	Source URL
domain	string	Source domain
timestamp	string	Collection timestamp
n_tokens	int64	Number of tokens in the document(acc to GPT-2)

About the Dataset

The released dataset contains approximately 2.2B tokens from KnowSpread, the corpus used in the training of Alephn-1.

This release is intended to provide a representative and inspectable portion of the training data while keeping the complete KnowSpread corpus private.

Intended Use

This subset is intended for:

• Inspecting the data used to train Alephn-1
• Research into language-model training datasets
• NLP experimentation
• Studying web-scale data collection and preprocessing
• Reproducing or analyzing aspects of the Alephn-1 training corpus

License

KnowSpread is released under the MIT License.