Introduction

This is a LLM-filtered set of the first 1M rows from ntt's JParaCrawl v3 large English-Japanese parallel corpus.

The original JParaCrawl corpus was put together by automated means - aligning Japanese texts with their apparent English translations that were found in-the-wild, on the internet.

Whilst manually browsing the original data, I noticed that there were obvious quality issues that made me anxious about using the dataset at all. Poorly aligned translations, incomplete translations, etc.

The goal of this dataset is to split the entire original dataset into its good and bad parts to:

facilitate further research
make available a high quality dataset
investigate the performance of various LLMs at evaluating the dataset.

The new upload includes filtering by an additional LLM, its results are "model2_accepted": https://huggingface.co/Verah/mistral-japanese-stabalelm-merge

I merged mistral instruct with stability AI's new japanese LLM, and this seems to have resulted in a model with enough knowledge of english and japanese to be competent at this task. It is likely that a finetune would further improve results. This new model accepted only 260,058 rows from the 1M seen, whilst the previous model was around twice as permissive.

Prompt used with the new model:

from inspect import cleandoc

def promptgen_mistral(japanese :str, english :str) -> str:
    system_prompt = cleandoc("""<s>[INST]Your role is to evaluate the accuracy of the provided Japanese to English translation.
    - Translations with parts missing should be rejected.
    - Incomplete translations should be rejected.
    - Inaccurate translations should be rejected.
    - Poor grammar should be rejected.
    - Any kind of mistake should be rejected.
    - Bad spelling should be rejected.
    - Low quality english should be rejected.
    - Low quality japanese should be rejected.
    - high quality translations should be accepted.
    - Respond with only 'ACCEPT' or 'REJECT'.
                            """)
    return system_prompt + f"JAPANESE: {japanese}\nENGLISH: {english}[/INST]\n"

License

The license is identical to the original JParaCrawl dataset:

Terms of Use for Bilingual Data, Monolingual Data and Trained Models
Nippon Telegraph and Telephone Corporation (Hereinafter referred to as "our company".) will provide bilingual data, monolingual data and trained models (Hereinafter referred to as "this data.") subject to your acceptance of these Terms of Use. We assume that you have agreed to these Terms of Use when you start using this data (including downloads).
Article 1 (Use conditions)
This data can only be used for research purposes involving information analysis (Including, but not limited to, replication and distribution. Hereinafter the same in this article.). The same applies to the derived data created based on this data. However, this data is not available for commercial use, including the sale of translators trained using this data.
Article 2 (Disclaimer)
Our company does not warrant the quality, performance or any other aspects of this data. We shall not be liable for any direct or indirect damages caused by the use of this data. Our company shall not be liable for any damage to the system caused by the installation of this data.
Article 3 (Other).
This data may be changed in whole or in part, or provision of this data may be interrupted or stopped at our company’s discretion without prior notice.

==========

対訳データ，単言語データおよび学習済みモデル利用に関する利用規約
日本電信電話株式会社（以下、「当社」という。）は、本利用規約に同意されることを条件として、対訳データ、単言語データおよび学習済みモデル（以下、「本データ」という。）を提供します。なお、本データの利用（ダウンロードも含む）を開始した時点で、本利用規約にご同意頂いたものとみなします。
第1条（利用条件）
本データは、情報解析を伴う研究開発目的にのみご利用（複製および配布を含むが、それに限らない。以下、同じ）頂けます。本データを基に作成された派生データについても同様です。ただし、本データを使って学習したデータを内蔵した翻訳機の販売等を含む商用利用目的には、ご利用頂けません。
第2条（免責）
当社は、本データについて、品質、性能その他一切の保証を行うものではありません。２.直接的損害、間接的損害を問わず、本データの利用によって生ずるいかなる損害についても、一切の責任を負いません。当社は、本データのインストール作業等によって発生するシステムへの影響等、損害についても、一切の責任を負いません。
第3条（その他）
事前通知なしに、当社の判断によって、本データを全部または一部の変更、本データの提供の中断または停止をさせて頂くことがございます。