CCQA: A New Web-Scale Question Answering Dataset for Model Pre-Training

This is the official repository for the code and models of the paper CCQA: A New Web-Scale Question Answering Dataset for Model Pre-Training.
If you use our dataset, code or any parts thereof, please cite this paper:

  title={CCQA: A New Web-Scale Question Answering Dataset for Model Pre-Training}, 
  author={Patrick Huber and Armen Aghajanyan and Barlas Oğuz and Dmytro Okhonko and Wen-tau Yih and Sonal Gupta and Xilun Chen},

Getting Common Crawl Snapshots

The Common Crawl project provides monthly web snapshots of new and updates websites in raw HTML format.
Every monthly snapshot (~50-70TB) is further separated into smaller WARC (Web ARChive) files.
To download a single WARC file, go to the Common Crawl website for the respective month (e.g. May 2021) and download the WARC paths file.
The downloaded WARC paths file contains a \newline separated list of download destination of the actual files.
Pick a path and prepend s3://commoncrawl/ or for the complete URL. Once downloaded, gunzip the archive and a single Common Crawl web archive is ready to be processed.

Dataset Generation


Below are the required dependencies to run the dataset generation, curation and model evaluations.

  • Rust
  • Rust packages: clap, html-escape, indicatif, kuchiki, rayon, regex, serde, serde_json, warc (see Cargo.toml file for versions)
  • Python 3.7.3
  • Python dependencies: fasttext language identification, fasttext==0.9.2, lxml==4.3.2

Processing Common Crawl data (Rust)

  • Build the cargo package with cargo build from within the rust folder
  • Run the script with cargo run <path/to/warc/file> <path/to/output/file.mhtml>

Curating the minified HTML data (Python)

To generate json objects for every webpage in the minified HTML, run

python <path/to/fasttext/lid.176.bin> <path/to/mhtml/file> <path/to/output/file>

Aggregating datapoints to remove duplicate URL entries (Python)

As mentioned in the paper, we use the original dataset for our in-domain pre-training experiments.
However, we also provide a cleaned version of the dataset, aggregating same-URL duplicates into a single object.
To run the datapoint aggregation script, execute

python <path/to/json/file> <path/to/output/file>

Converting json dataset into closed-book and passage retrieval formats (Python)

To be able to train closed-book (sequence-to-sequence) and passage retrieval (DPR) models on the CCQA dataset, the corpus needs to be further processed

Closed-book processing

To prepare the dataset for closed-book question-answering training, run:

python <path/to/json/file> <path/to/output/file> <--only_english> <--keep_markup>

Passage retrieval (DPR) processing

To prepare the dataset for passage rertieval (DPR) training, run:

python <path/to/json/file> <path/to/output/file> <--only_english> <--keep_markup>

CCQA In-Domain Pre-Trained Model Checkpoints

BART and T5 checkpoints are Huggingface transformer models tested with transformers version 4.8.2

The DPR model checkpoint can be downloaded for the original DPR codebase or the DPR v2 codebase


The majority of CCQA is licensed under CC-BY-NC, however portions of the project are available under separate license terms: crowbook-text-processing is licensed under the MPL-2.0 license.


View Github