README_en.md

---
license: cc-by-sa-3.0
task_categories:
- question-answering
- summarization
language:
- ja
size_categories:
- 10K<n<100K
---
日本語は[こちら](README.md)

# Summary

`databricks-dolly-15k-ja-scored` is a version of `databricks-dolly-15k-ja`, which is a machine-translated version of `databricks-dolly-15k`, with added translation quality scores provided by BERTScore.

This dataset can be used for any purpose, whether academic or commercial,  under the terms of the [Creative Commons Attribution-ShareAlike 3.0 Unported License](https://creativecommons.org/licenses/by-sa/3.0/legalcode).  
  
Languages: Japanese

# Translation Quality Score

When examining the data contained in `databricks-dolly-15k-ja`, I found the following data of poor quality.

1.  Data where `input` and `output` are exactly the same
2.  Data where `output` is copied to `instruction`
3.  Data in which the consistency of expression is not maintained due to notation.
4.  Data that has failed to translate proper nouns, etc.

Note that in `databricks-dolly-15k`
- Data corresponding to `1.` has been removed.
- Data corresponding to `2.` has its instruction removed and its category changed to open-qa.

Therefore, I back-translated these data from Japanese to English and examined the similarity with the original text by [BertScore](https://arxiv.org/abs/1904.09675).
Below is a histogram of f1 score, the harmonic mean of precision and recall.

<div align="center">
  <img src="https://media.githubusercontent.com/media/Sakusakumura/databricks-dolly-15k-ja-scored/e46aec8f9d2602e9e7c074674390263462534a9a/images/f1-score-full.png">
</div> 

You can filter out low-quality data by filtering with f1 score.

## Notes

The derived score varies depending on the service used for Japanese-English translation and the model used in BertScore.
Even with data that is OK to use for learning, the score may drop depending on the accuracy of the Japanese-English translation.


# Description of each field

I will explain only the added parts and the items that need attention.

## 1. index

A unique identifier for the data.
Originally a field indicating which lines of `databricks-dolly-15k` have been translated, but the index does not correspond to the actual line because `databricks-dolly-15k-ja` does not reflect a change in the element deletion of `databricks-dolly-15k`.

|Field|Description|
|--|--|
|Index|The unique identifier for the data|

## 2. bertscore

Evaluation score provided by the BERT model.
It consists of three indices: recall, precision, and f1.

|Field|Description|
|--|--|
|recall|A measure of how much information contained in a correct answer text is reflected in the generated text|
|precision|A measure of how well each part of the generated text matches the correct text|
|f1|Harmonic mean of precision and recall.|

## 3. translator

Information of the translation services used.
There are two fields, "en_ja" and "ja_en", which represent the services used to translate English to Japanese and Japanese to English, respectively.

|Field|Description|
|--|--|
|en_ja|Service used to translate from English to Japanese|
|ja_en|Service used to translate from Japanese to English|

# Acknowledgments

This dataset, `databricks-dolly-15k-ja-scored`, builds upon the work done by kun1em0n, who created the `databricks-dolly-15k-ja` dataset. I want to express my deep appreciation for his valuable work and contribution to the community.

# License/Credits

- databricks-dolly-15k-ja-scored  
Copyright (2023) Sakusakumura. This dataset is licensed under CC BY-SA 3.0.

- databricks-dolly-15k-ja  
Developed by kun1em0n. Available at https://github.com/kunishou/databricks-dolly-15k-ja. This dataset is licensed under CC BY-SA 3.0.

- databricks-dolly-15k  
Developed by Databricks, Inc. Available at https://huggingface.co/datasets/databricks-dolly-15k. This dataset is licensed under CC BY-SA 3.0.