pretty_name: test strict shuffled
task_categories:
- text-generation
tags:
- parquet
license: odc-by
language:
- ja
ClimbLab-Ja
ClimbLab-Ja is a high-quality 300-billion-token Japanese corpus with 20 clusters. It is a Japanese adaptation of the nvidia/Nemotron-ClimbLab approach. Based on LLM-jp Corpus v4, we semantically reorganized and filtered the dataset into 20 distinct clusters, resulting in a high-quality 300-billion-token corpus. Specifically, we first grouped the data into 1,000 groups based on topic information. Then we assigned six scores from 0 to 5 to each group and document: quality, advertisement, informational value, educational value, cultural value, and creative value. Low-quality clusters and documents were removed based on these scores.
This dataset is for research and development only.
Dataset Details
- Intended Usage: Pre-training language models.
- Format: Text in parquet format.
- Size: 300B tokens.
Filtering
Rather than applying aggressive filtering, we focused on removing clearly low-quality data while retaining a diverse corpus with broad coverage.
We removed clusters that did not satisfy the following condition:
quality_mean >= 2 && advertisement_mean >= 2 && any(value_mean >= 3)
where value_mean refers to informational value, educational value, cultural value, or creative value.
We also removed documents that did not satisfy the following condition:
quality >= 2 && advertisement >= 2 && any(value >= 3)
where value refers to informational value, educational value, cultural value, or creative value.
License
The dataset structure, record IDs, filtering scores, mixture weights, and other metadata newly added in this release are licensed under ODC-BY. By using this dataset, you are also bound by any license agreements and terms of use of the original data sources listed in the LICENSES.md.
This dataset was created using the Supermicro ARS-111GL-DNHR-LCC and FUJITSU Server PRIMERGY CX2550 M7 (Miyabi) at the Joint Center for Advanced High Performance Computing (JCAHPC).
ClimbLab-Ja
ClimbLab-Jaは、20クラスタからなる高品質な3,000億トークンの日本語コーパスです。nvidia/Nemotron-ClimbLabのアプローチを日本語に適用しています。LLM-jp Corpus v4をベースに、データセットを意味的に再編成・フィルタリングし、20の異なるクラスタに分類することで、高品質な3,000億トークンのコーパスを構築しました。具体的には、まずトピック情報に基づきデータを1,000グループに分類しました。次に、各グループおよび各文書に対して、品質・広告性・情報的価値・教育的価値・文化的価値・創造的価値の6つのスコアを0〜5の範囲で付与しました。これらのスコアに基づき、低品質なクラスタおよび文書を除去しました。
本データセットは研究開発目的のみに使用できます。
データセットの詳細
- 想定用途:言語モデルの事前学習
- フォーマット:Parquet形式のテキスト
- 規模:3,000億トークン
フィルタリング
厳密なフィルタリングを行うのではなく、明らかに低品質なデータを除去しつつ、カバレッジの広い多様なコーパスを残すことを重視しています。
以下の条件を満たさないクラスタを除去しました:
quality_mean >= 2 && advertisement_mean >= 2 && any(value_mean >= 3)
ここで value_mean は、情報的価値・教育的価値・文化的価値・創造的価値のいずれかを指します。
また、以下の条件を満たさない文書を除去しました:
quality >= 2 && advertisement >= 2 && any(value >= 3)
ここで value は、情報的価値・教育的価値・文化的価値・創造的価値のいずれかを指します。
ライセンス
本リリースで新たに追加されたデータセット構造、レコードID、フィルタリングスコア、混合重み、およびその他のメタデータは、ODC-BYの下でライセンスされています。本データセットを使用することにより、LICENSES.mdに記載された元データソースのライセンス契約および利用規約にも拘束されます。
本データセットは、学際大規模情報基盤共同利用・共同研究拠点(JCAHPC)のSupermicro ARS-111GL-DNHR-LCCおよびFUJITSU Server PRIMERGY CX2550 M7(Miyabi)を用いて作成されました。