Add EvolInstruct and EvolInstructGenerator tasks

[alvarobartt]

Description

This PR adds both the EvolInstruct and EvolInstructGenerator tasks ported from https://github.com/h2oai/h2o-wizardlm/blob/main/wizardlm.py with some slight modifications to suit our needs, but respecting the evolutionary approach.

Besides that the AsyncLLM has been fixed so as to use asyncio event loops instead of asyncio.run as it was raising some errors when called within a loop, so now the AsyncLLM implementation is more robust. Also the GeneratorTask has been included for tasks like EvolInstructGenerator i.e. generating data without seed data as input.

Closes #408

Example

import time

from distilabel.llm.mistral import MistralLLM
from distilabel.llm.openai import OpenAILLM
from distilabel.pipeline.local import Pipeline
from distilabel.steps.globals.huggingface import PushToHub
from distilabel.steps.task.evol_instruct.generator import EvolInstructGenerator

if __name__ == "__main__":
    start_time = time.time()

    with Pipeline() as pipeline:
        evol_instruct = EvolInstructGenerator(
            name="evol_instruct",
            llm=OpenAILLM(
                model="gpt-4",
                api_key="sk-***",  # type: ignore
            ),
            num_instructions=10,
            generate_answers=False,
        )

        push_to_hub = PushToHub(name="push_to_hub")  # type: ignore
        evol_instruct.connect(push_to_hub)

    pipeline.run(
        parameters={
            "evol_instruct": {
                "generation_kwargs": {
                    "max_new_tokens": 512,
                    "temperature": 0.7,
                },
            },
            "push_to_hub": {
                "repo_id": "alvarobartt/evol-instruct",
                "split": "train",
                "private": False,
                "token": "hf_***",
            },
        }
    )

    print("--- %s seconds ---" % (time.time() - start_time))

Reference

https://github.com/h2oai/h2o-wizardlm/blob/main/wizardlm.py

[alvarobartt]

@alvarobartt, I am also missing the system prompt they defined in the paper but we don't want to implement that either? https://github.com/argilla-io/distilabel/blob/main/src/distilabel/tasks/_templates/evol-instruct.jinja2

Fair, we can include it, but I'm afraid that the misalignments between their official implementation and the paper don't have a clear reference, should we take more inspiration from the paper instead? My experience so far is that using GPT-4 for generating and evolving instructions works well so far, but we can try to compare with and without system_prompt. Also does that apply to both the instruction and the answer, or only the instruction? I guess the first one, but LMK

@alvarobartt Also, I seem to be missing the elimination but you wanted to skip this because general state of LLMs have improved and you feel we don't need it anymore, correct?

Right, I was using gpt-4, mistral-medium and mistral-large, and didn't run into any of those issues to tackle during post-processing, but maybe we can include it just in case, but IMO those were not needed for this small proof of concept. Anyway, all these details will show whenever we implement DEITA using this building blocks to see whether something's off or not, but I'm afraid that some things they added on top of the post-processing where suited for their models, while bigger and more recent ones seem to preduce nice results with minimal post-processing.

[alvarobartt]

Also I've been exploring a bit on GitHub and, is this a more faithful reproduction of EvolInstruct (pre-WizardLM)? https://github.com/nlpxucan/WizardLM/tree/main/Evol_Instruct cc @davidberenstein1957

[davidberenstein1957]

@alvarobartt I would say so, yes. This is also the way we eventually implemented it initially. I would also go with the more extensive prompts to ensure we've got a higher change to make it work without more advancd models.

[gabrielmbmb]

Hi @alvarobartt, it looks good to me. Some methods are quite big so I would split them