Skip to content

Latest commit

 

History

History
79 lines (54 loc) · 3.1 KB

README.md

File metadata and controls

79 lines (54 loc) · 3.1 KB
Raw

cisnlp/Ungoliant

🕷️ Ungoliant is a high-performance pipeline that provides tools to build corpus generation pipelines from CommonCrawl. 🕷️

This pipeline was originally used to process the OSCAR dataset. It uses the fasttext lid.176.bin model to generate labels for 176 languages. We forked the code here so that it can function with GlotLID, which is also a fasttext model but can label text for more than 2000 languages.

The outcome of this new dataset is the GlotCC dataset, available at: https://github.com/cisnlp/GlotCC

Installation

Installing/Compiling the binary

  • Via git: cargo install --git https://github.com/cisnlp/ungoliant

Ungoliant needs numerous dependencies that should be compiled when installing. However cmake / gcc can be needed as the project uses fasttext-rs.

Getting a language identification file (for fastText):

By default, ungoliant expects the lid.176.bin model name. Use wget https://huggingface.co/cis-lmu/glotlid/resolve/main/model.bin -O glotlid.bin to get GlotLID as glotlid.bin.

However, you can use the model you want: just point to its path using ungoliant download --lid-path <path to lid>.

Usage

The usual way of generating corpora is:

  1. First create this structure of folders with mkdir:
res
├── annotation             
│   └── ...
├── blocklist             
│   └── ...
├── corpus             
│   └── ...
├── filter             
│   └── ...
└── shards             
    └── ...

  1. Fetch the wet.paths.gz file from the last CommonCrawl dump.
    1.1 Decompress it using gzip -d wet.paths.gz.
    1.2 Download the files using the download command: ungoliant download wet.paths res/shards.

  2. Download website categorizations using wget https://github.com/olbat/ut1-blacklists/archive/refs/heads/master.zip.
    2.1 Decompress it using unzip master.zip.
    2.2 Move the blacklists to the res/blocklist using mv ut1-blacklists-master/blacklists/* res/blocklist.
    2.3 Decompress the adult block using gzip -d res/blocklist/adult/domains.gz.
    2.4 Remove the blacklists-master using rm -r ut1-blacklists-master.

  3. Generate the corpus using the pipeline command (it may take some time): ungoliant pipeline ./res/shards/ ./res/corpus --lid-path glotlid.bin --blocklist-path ./res/blocklist/.

  4. Head on to glotcc-filters for the additional filter steps.

You can find more information on each command's --help.

ungoliant 2
corpus generation tool.

USAGE:
    ungoliant <SUBCOMMAND>

FLAGS:
    -h, --help       Prints help information
    -V, --version    Prints version information

SUBCOMMANDS:
    download    Download a CommonCrawl release
    help        Prints this message or the help of the given subcommand(s)
    pipeline    Run pipeline
    rebuild     Rebuild the corpus for a given language.