feat: translate instruction when adapting prompt

[Yunnglin]

When adapting prompts, both the examples and the corresponding instructions are important and need to be translated.

[shahules786]

Hey @Yunnglin Did you observe better results when translating both? We observed pretty good results with only translating a few shot examples. Another reason why we did not translate instruction was that time it introduces ambiguity. FOr example, let's say if the instruction was "Output False if a number greater than zero", this would also translate the word "False", which then causes issues while post-processing.

[Yunnglin]

Whether to translate the "instruction" could perhaps be made into an option. When I was generating a Chinese dataset, I found that some "extractor prompts" would lack examples, which led to the generated data not being very effective.

[shahules786]

@Yunnglin that suggestion makes sense. Could you make a PR with the same?
@jjmachan Please share if you have any opinion on this.

[jjmachan]

This makes a ton of sense

what we can do is take another argument for

ragas/src/ragas/prompt/pydantic_prompt.py

Lines 217 to 223 in 807c313

	async def adapt(
	self, target_language: str, llm: BaseRagasLLM
	) -> "PydanticPrompt[InputModel, OutputModel]":
	"""
	Adapt the prompt to a new language.
	"""

which converts it

and it can be controlled through

ragas/src/ragas/prompt/mixin.py

Line 55 in 807c313

async def adapt_prompts(

would be really helpful 🙂

[Yunnglin]

• Add adapt_instruction: bool=False parameter.

  Now you can adapt prompt as follows:

  import asyncio
  from ragas.metrics import Faithfulness
  from ragas.llms import LangchainLLMWrapper
  
  
  instance = Faithfulness()
  adapted_prompts = asyncio.run(instance.adapt_prompts(language="chinese", llm=LangchainLLMWrapper(chat_model), adapt_instruction=True))
  print(adapted_prompts)

• Add Ensure that the number of output data rows is equal to the number of input data rows. to the translation prompt since sometimes LLM breaks the single line into multiple lines.

[shahules786]

LGTM, thanks a lot.