README.md

metadata

language:
  - ja
pretty_name: lambda-corpus
license: other
task_categories:
  - text-generation
tags:
  - japanese
  - pretraining
  - corpus
  - parquet
size_categories:
  - 10M<n<100M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation.parquet
      - split: test
        path: data/test.parquet

lambda-corpus

lambda-corpus is a Japanese text corpus for language model pretraining. It combines openly available Japanese datasets into a consistent format and provides predefined train, validation, and test splits.

Purpose

The dataset is intended for pretraining of Japanese language models. It contains web documents, Wikipedia-derived text, academic grant records, and synthetic question-answer text.

Source Data

Source	Rows	Tokens	License
KeisukeMiyamoto/CleanedFineWeb2Edu-jp	11,772,204	7.61B	ODC-BY
hpprc/jawiki-bullet-points	3,523,358	1.55B	CC BY-SA 4.0
kanhatakeyama/CommonCrawl-RAG-QA-Calm3-22b-chat	2,525,793	1.04B	Other
hpprc/llmjp-kaken	1,079,030	1.03B	CC BY 4.0

Dataset Size

Split	Rows
train	18,522,556
validation	189,174
test	188,655
total	18,900,385

The dataset contains approximately 11.23B tokens measured with the gemma 4 tokenizer.

Columns

Column	Description
id	Identifier inherited from the source dataset
text	Normalized document text
output_tokens	Number of tokens in text, measured with the gemma 4 tokenizer
source	Name of the source dataset

Example

from datasets import load_dataset

dataset = load_dataset("KeisukeMiyamoto/lambda-corpus")
print(dataset["train"][0])