Datasets:

Podtech
/

abeja-cc-ja-filtered

dataset_info:
  features:
    - name: url
      dtype: string
    - name: content
      dtype: string
    - name: perplexity
      dtype: float64
license: pddl
task_categories:
  - text-generation
language:
  - ja

Filtered ABEJA-CC-JA

Dataset Overview

This dataset is a filtered version of the large-scale Japanese pre-training corpus, abeja-cc-ja. It was created by duplicating the original dataset while excluding texts with a KenLM perplexity score. The goal is to extract high-quality Japanese text data to facilitate the training of language models.

Data Structure

Features

Each data entry contains the following attributes:

url: The source URL of the scraped data.
content: The extracted text body.
perplexity: The weighted average perplexity score of the entire document, calculated using KenLM.

Directory Structure

The dataset is provided in Parquet format and is organized in a directory structure as shown below:

/data/commom_crawl_0/
├── common_crawl_0_filtered_0000.parquet
├── common_crawl_0_filtered_0001.parquet
└── ...

Data Processing (Filtering)

The filtering process applies the following logic to the text data:

KenLM Model Used: lighttransport/japanese-scoring-model
Filtering Condition: Retains texts with a perplexity score strictly below 40 and excludes those scoring 40 or higher.

About the Original Dataset

Original Source (AWS): https://registry.opendata.aws/abeja-cc-ja/
Original Source (Huggingface): https://huggingface.co/datasets/kajuma/ABEJA-CC-JA
Detailed Description: ABEJA Tech Blog: abeja-cc-ja-202409

License

The license for this dataset strictly follows that of the original dataset:

This data is available for anyone to use under the Common Crawl Terms of Use.