|
|
--- |
|
|
license: cc-by-sa-4.0 |
|
|
task_categories: |
|
|
- text-generation |
|
|
language: |
|
|
- ja |
|
|
size_categories: |
|
|
- 1M<n<10M |
|
|
--- |
|
|
|
|
|
# Description |
|
|
This is a collection of raw data from ~40 Japanese open source downstream task datasets. |
|
|
This dataset was built with the intention of collecting high quality datasets which were not machine translated to and from Japanese. |
|
|
The templated version, which should be used to perform instruction finetuning on LLMs, can be found [here](https://huggingface.co/datasets/Ego/jpflan). |
|
|
|
|
|
# Dataset format |
|
|
As this is a collection of datasets, each one is categorized by task and can contain different amounts of information (texts). |
|
|
For more information about each dataset and the corresponding texts, also please refer to [here](https://github.com/Egojr/jpflan). |
|
|
|
|
|
```json |
|
|
{ |
|
|
"dataset_name": 'Original dataset name' |
|
|
"task": 'Downstream task category' |
|
|
"text1": 'Often main input text (e.g. question)', |
|
|
"text2": 'Often main output text (e.g. answer)', |
|
|
"text3": 'Additional info, if any (e.g. context)', |
|
|
"text4": 'Additional info, if any', |
|
|
"text5": 'Additional info, if any', |
|
|
} |
|
|
``` |
|
|
|
|
|
|
