About Uzushio

Uzushio is a web-scale corpus text extraction and preprocessing tool. It uses Spark and written mostly in Scala.

Goal of this project

LLMs and foundational models require huge clean corpus for training. There are many cleaning methods, and each nlp tasks have suitable sets of them.

Uzushio has the following goals:

• provide a common corpus cleaning tool
  • reduce the implementation cost
  • share same implementation for the compatibility
• keep it efficient, flexible and updated
  • it should be able to used in every kind of nlp projects (which uses large corpus)
• support AWS EMR

Setup

You need to install Apache Spark, sbt, and Sudachi dictionary file.

Apache Spark (if running locally, if you want to use AWS EMR you do not need it)

Download it from the official page. You can put it in PATH or use absolute paths for Spark startup scripts to launch Uzushio.

At the moment uzushio is built against: spark 3.4.1 + Scala 2.12.*.

Check if spark-submit --version works.

Scala (sbt)

Download sbt (latest 1.* version) from the official page and make it usable. You can often install one from OS package manager.

JDK

You need JDK 8+ to run sbt and Spark. Usually, you can install it from a package manager. However, if you plan to launch Uzushio in EMR, you either need to use JDK 8 to compile Uzushio, or assemble a custom image for EMR to use JDK 11+ inside EMR (currently it uses JDK 8), otherwise you will get MethodNotFoundErrors.

Building

Build fat JAR for launching

The easiest way to use the tool — build fat JAR with all required dependencies.

Execute sbt assembly from project root directory. You will have output jar under ./target/scala-[version]/.

For development

• For development, it is better to launch sbt shell via sbt and launch commands inside the sbt.
• Launch classes located in com.worksap.nlp.uzushio.lib.runners are designed to be launched without spark-submit script
  • You can use your IDE functionality which will set up correct classpath or use lib/run command in the sbt shell
• You can do compilation via assembly command
  • Also, see sbt docs

Running via spark environment

Use spark-submit with jar built by previous steps：

spark-submit [spark options] --class [class name] [path/to/jar_file] [class options]

Also, see spark-submit docs

Runner classes

Note that all paths can be on any Hadoop-supported filesystem. For S3, specify s3:// paths if you use EMR and s3a:// paths if you do not use EMR.

Classes in com.worksap.nlp.uzushio.main are designed to run from spark-submit script

com.worksap.nlp.uzushio.main.ExtractTextFromWarc

Extract paragraphs from html documents inside WARC collection. Paragraphs are guessed on best effort basis from html structure (e.g. div, p tags).

Run Example

spark-submit \
    --class=com.worksap.nlp.uzushio.main.ExtractTextFromWarc \
    --master='local[*]' \
    /path/to/uzushio-assembly-0.1-SNAPSHOT.jar \
    --input=/path/to/warc/files \
    --output=/path/to/output \
    --language=ja,ru,ko,zh 

Arguments

• --input path to input data, can be any Hadoop-supported filesystem
• --output path to where output data will be written, in parquet format
• --language extract documents with following languages. Languages with latin script are NOT supported.
• --max-partitions (default 2000) produce at most this number of partitions in output data.
• --compression (default zstd) compress parquet data with this compression algorithm. See supported algorithms.

License

Copyright (c) 2024 Works Applications Co., Ltd. All rights reserved.

Uzushio is distributed by Works Applications Co.,Ltd. under Apache License, Version 2.0.

References

WARC

• warc format specification
  • Note that latest version is v1.1 but CC uses v1.0
• org.archive.io
  • Iterates over warc records
• com.martinkl.warc
  • Loads warc file using Hadoop API

Other

• chiTra の前処理について
  • preprocess in chitra pretraining
• Deduplicating Training Data Makes Language Models Better