Add Synthetic Data Generation Module

[ryantwolf]

Description

Adds a suite of tools for interacting with LLM services. These LLM services are then used to build synthetic data generation tools and example pipelines following the Nemotron 340B Technical Report. The prompt templates used in the report are supplied as defaults throughout the code.

Usage

OpenAI API

from nemo_curator import AsyncOpenAIClient
from nemo_curator.synthetic import (
    AsyncNemotronGenerator,
    NemotronGenerator,
)
from openai import OpenAI, AsyncOpenAI

async def demo():
  openai_client = AsyncOpenAI(
      base_url="https://integrate.api.nvidia.com/v1",
      api_key="",
  )
  client = AsyncOpenAIClient(openai_client)
  generator = AsyncNemotronGenerator(client)
  
  model = "nvidia/nemotron-4-340b-instruct"
  model_kwargs = {
      "top_p": 0.7,
      "max_tokens": 1024,
      "seed": 1234,
  }
  
  openlines = await generator.run_open_qa_pipeline(
      n_macro_topics=5,
      n_subtopics=3,
      n_openlines=3,
      n_revisions=2,
      model=model,
      base_model_kwargs=model_kwargs,
      conversion_model_kwargs=model_kwargs,
      ignore_conversion_failure=True,
  )
  
  dialogue = await generator.generate_dialogue(
      openline=openlines[0],
      user_model=model,
      assistant_model=model,
      user_model_kwargs=model_kwargs,
      assistant_model_kwargs=model_kwargs,
  )

  print(dialogue)

NeMo Deploy

from nemo_curator import NemoDeployClient
from nemo_curator.synthetic import (
    AsyncNemotronGenerator,
    NemotronGenerator,
    NemotronFormatter,
)
from nemo.deploy.nlp import NemoQueryLLM

async def demo():
  model = "local_nemotron"
  model_kwargs = {
      "top_p": 0.7,
      "max_tokens": 1024,
      "seed": 1234,
      "conversation_formatter": NemotronFormatter(),
      "stop": ['<extra_id_1>'],
  }
  
  nemo_client = NemoQueryLLM(url="localhost:8000", model_name=model)
  client = NemoDeployClient(nemo_client)
  
  openlines = await generator.run_open_qa_pipeline(
      n_macro_topics=5,
      n_subtopics=3,
      n_openlines=3,
      n_revisions=2,
      model=model,
      base_model_kwargs=model_kwargs,
      conversion_model_kwargs=model_kwargs,
      ignore_conversion_failure=True,
  )

Checklist

• I am familiar with the Contributing Guide.
• New or Existing tests cover these changes.
• The documentation is up to date with these changes.

[VibhuJawa]

Thanks for working on this @ryantwolf . This is a very important functionality we are adding, Very excited for it.

My major concern currently is around not having a way to rate limit the number of requests we are sending, everything else is mostly nits.

[ayushdg]

Minor nit but at a high level looks good to me! Thanks a lot for this effort

[VibhuJawa]

LGTM, thanks for working on this Ryan.