| # Common Crawl WET Dataset - c2 | |
| This repository contains a large-scale filtered dataset derived from the WET files of the Common Crawl project. The data is cleaned and aggregated to facilitate large-scale natural language processing tasks, especially the pretraining of large language models (LLMs). | |
| ## Dataset Description | |
| - **Source:** Common Crawl CC-MAIN-2025-38, September 2025 crawl. | |
| - **Data Type:** Extracted plaintext from web crawl WET files with aggressive metadata and boilerplate filtering. | |
| - **File Size:** Large combined files (~15GB each) to balance upload size and storage constraints. | |
| - **Preprocessing:** Streamed extraction, metadata removal, filtering out boilerplate and duplicate content. | |
| - **Purpose:** Primarily designed for pretraining foundation models and LLMs requiring diverse, massive-scale natural language corpora. | |
| ## Features | |
| - **Optimized for Pretraining:** | |
| The dataset is curated and filtered to be suitable for training large language models. It contains clean, high-quality textual data ideal for unsupervised pretraining tasks like masked language modeling or autoregressive modeling. | |
| - **Large Scale:** | |
| Contains processed data amounting to multiple terabytes, allowing training on a broad, diverse text corpus representing a wide range of domains. | |
| - **Streaming Processing:** | |
| The data was processed in a memory-efficient, streaming manner to support large-scale data handling without requiring excessive resources. | |
| - **Metadata Cleaning:** | |
| Extensive removal of WARC, HTTP headers, and other metadata ensures minimal noise in the text used for training. | |
| - **Resume and Verify:** | |
| Processing is checkpointed for fault tolerance. Uploaded files are verified on Hugging Face to avoid duplicates. | |
| - **Immediate Uploads:** | |
| Files are uploaded to Hugging Face immediately after hitting the 15GB size limit to respect limited storage constraints. | |
| ## Usage | |
| Load the dataset using Hugging Face's `datasets` library: | |
| from datasets import load_dataset | |
| dataset = load_dataset("blue-blue/c2") | |
| After loading, you can iterate over text samples for pretraining models like GPT, BERT, or other large language architectures. | |
| ## Pretraining Applications | |
| - **Foundation Model Development:** | |
| Provides diverse, large-scale text data crucial for training high-quality foundation LLMs. | |
| - **Language Modeling Tasks:** | |
| Suitable for autoregressive or masked language model pretraining due to extensive scale and quality. | |
| - **Downstream Adaptation:** | |
| Can be combined with other specialized datasets for fine-tuning or adaptation tasks. | |
| - **Research & Benchmarking:** | |
| Acts as a standard large-scale corpus for benchmarking NLP algorithms and analyzing language model behavior. | |
| ## Contact | |
| For questions, support, or collaboration: | |
| [hello@bluesminds.com](mailto:hello@bluesminds.com) | |
| --- | |
| Thank you for exploring the **c2** dataset — a foundational resource for large-scale language modeling and NLP research. | |