anandjh8/common-crawl-english-filtered

🧠 FineWeb-English-Filtered

📘 Dataset Summary

FineWeb-English-Filtered is a large-scale, cleaned, English-only text dataset derived from Common Crawl’s WET archives.
It contains 940 million documents of publicly available web text, converted into Apache Parquet format with a consistent schema for fast and efficient data loading.

The dataset was generated using a custom AWS Glue pipeline that processed, filtered, and merged .wet files across multiple terabytes of Common Crawl data.

This dataset is ideal for training large language models, retrieval research, and web-scale NLP tasks.

---

📂 Data Structure

Each record corresponds to one web document extracted from a WARC/WET file.

Column Name	Type	Description
target_uri	string	Original webpage URL
record_id	string	Unique WARC record UUID
language	string	ISO-639 code (filtered to eng)
content_type	string	MIME content type (e.g., text/plain)
content_length	int64	Length of the content in bytes
digest	string	SHA-1 block digest
text	string	Extracted plain text content

---

⚙️ Data Processing Pipeline

1. Extraction – Read .wet files from AWS S3 (s3://warc-unc-*).
2. Filtering – Retain only records where WARC-Identified-Content-Language == 'eng'.
3. Schema Enforcement – Normalize metadata fields across all records.
4. Serialization – Store each record as a Parquet row using PyArrow.
5. Parallel Processing – Processed with AWS Glue (Spark) and Boto3.
6. Merging – Folder-wise merges (~290 GB per shard) into unified Parquet outputs.
7. Upload – Finalized dataset uploaded to Hugging Face Hub.

---

🧩 Technical Specifications

Property	Details
Total Records	~940 million
Language	English (eng)
Format	Parquet (Snappy compression)
Total Size	~2.6 TB
Schema Version	v1.0
Source	Common Crawl WET data
Processing Stack	AWS Glue + Python (PyArrow, Boto3)

---

💡 Example Usage

from datasets import load_dataset

ds = load_dataset("your-username/fineweb-english-filtered", split="train")

print(ds)
print(ds.features)
print(ds[0])