sieu-n/alpaca_eval_multilingual

---
license: cc-by-nc-4.0

---

### Usage

```
load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval") # or alpaca_eval_en
load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval_ko")
load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval_ja")
```

### Method

The dataset was translated by GPT-4 API using the following prompt.

```
ja = ChatPromptTemplate.from_messages(
[
SystemMessagePromptTemplate.from_template(
"You are a helpful assistant fluent in English and Japanese."
),
HumanMessagePromptTemplate.from_template(
"Translate the following text to Japanese. Show the answer only. このテキストを直訳するのではなく、その意味を保持しつつ、より自然なリクエストに言い換えて翻訳してください text=```{instruction}```"
),
]
)

ko = ChatPromptTemplate.from_messages(
[
SystemMessagePromptTemplate.from_template(
"You are a helpful assistant fluent in English and Korean."
),
HumanMessagePromptTemplate.from_template(
"Translate the following text to Korean. Show the answer only. 말 그대로 번역하지 말고, 의미가 유지되는 한에서 자연스러운 요청으로 번역해줘. text=```{instruction}```"
),
]
)
```

Script: https://gist.github.com/sieu-n/88542733914f80f780359f5c82c99a62

原始信息汇总

数据集概述

使用方法

python load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval") # 或者 alpaca_eval_en load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval_ko") load_dataset("krenerd/alpaca_eval_multilingual", "alpaca_eval_ja")

数据集创建方法

数据集是通过GPT-4 API翻译生成的，具体翻译方法如下：

日语翻译

python ja = ChatPromptTemplate.from_messages( [ SystemMessagePromptTemplate.from_template( "You are a helpful assistant fluent in English and Japanese." ), HumanMessagePromptTemplate.from_template( "Translate the following text to Japanese. Show the answer only. このテキストを直訳するのではなく、その意味を保持しつつ、より自然なリクエストに言い換えて翻訳してください text={instruction}" ), ] )

韩语翻译

python ko = ChatPromptTemplate.from_messages( [ SystemMessagePromptTemplate.from_template( "You are a helpful assistant fluent in English and Korean." ), HumanMessagePromptTemplate.from_template( "Translate the following text to Korean. Show the answer only. 말 그대로 번역하지 말고, 의미가 유지되는 한에서 자연스러운 요청으로 번역해줘. text={instruction}" ), ] )