Bespoke Curator

Data Curation for Post-Training & Structured Data Extraction

Overview

Bespoke Curator makes it very easy to create high-quality synthetic data at scale, which you can use to finetune models or use for structured data extraction at scale.

Bespoke Curator is an open-source project:

• That comes with a rich Python based library for generating and curating synthetic data.
• A Curator Viewer which makes it easy to view the datasets, thus aiding in the dataset creation.
• We will also be releasing high-quality datasets that should move the needle on post-training.

Key Features

1. Programmability and Structured Outputs: Synthetic data generation is lot more than just using a single prompt -- it involves calling LLMs multiple times and orchestrating control-flow. Curator treats structured outputs as first class citizens and helps you design complex pipelines.
2. Built-in Performance Optimization: We often see calling LLMs in loops, or inefficient implementation of multi-threading. We have baked in performance optimizations so that you don't need to worry about those!
3. Intelligent Caching and Fault Recovery: Given LLM calls can add up in cost and time, failures are undesirable but sometimes unavoidable. We cache the LLM requests and responses so that it is easy to recover from a failure. Moreover, when working on a multi-stage pipeline, caching of stages makes it easy to iterate.
4. Native HuggingFace Dataset Integration: Work directly on HuggingFace Dataset objects throughout your pipeline. Your synthetic data is immediately ready for fine-tuning!
5. Interactive Curator Viewer: Improve and iterate on your prompts using our built-in viewer. Inspect LLM requests and responses in real-time, allowing you to iterate and refine your data generation strategy with immediate feedback.

Installation

pip install bespokelabs-curator

Usage

To run the examples below, make sure to set your OpenAI API key in the environment variable OPENAI_API_KEY by running export OPENAI_API_KEY=sk-... in your terminal.

Hello World with SimpleLLM: A simple interface for calling LLMs

from bespokelabs import curator
llm = curator.SimpleLLM(model_name="gpt-4o-mini")
poem = llm("Write a poem about the importance of data in AI.")
print(poem)
# Or you can pass a list of prompts to generate multiple responses.
poems = llm(["Write a poem about the importance of data in AI.",
            "Write a haiku about the importance of data in AI."])
print(poems)

Note that retries and caching are enabled by default. So now if you run the same prompt again, you will get the same response, pretty much instantly. You can delete the cache at ~/.cache/curator.

Use LiteLLM backend for calling other models

You can use the LiteLLM backend for calling other models.

from bespokelabs import curator
llm = curator.SimpleLLM(model_name="claude-3-5-sonnet-20240620", backend="litellm")
poem = llm("Write a poem about the importance of data in AI.")
print(poem)

Visualize in Curator Viewer

Run curator-viewer on the command line to see the dataset in the viewer.

You can click on a run and then click on a specific row to see the LLM request and response. More examples below.

LLM: A more powerful interface for synthetic data generation

Let's use structured outputs to generate poems.

from bespokelabs import curator
from datasets import Dataset
from pydantic import BaseModel, Field
from typing import List

topics = Dataset.from_dict({"topic": [
    "Urban loneliness in a bustling city",
    "Beauty of Bespoke Labs's Curator library"
]})

Define a class to encapsulate a list of poems.

class Poem(BaseModel):
    poem: str = Field(description="A poem.")

class Poems(BaseModel):
    poems_list: List[Poem] = Field(description="A list of poems.")

We define an LLM object that generates poems which gets applied to the topics dataset.

poet = curator.LLM(
    prompt_func=lambda row: f"Write two poems about {row['topic']}.",
    model_name="gpt-4o-mini",
    response_format=Poems,
    parse_func=lambda row, poems: [
        {"topic": row["topic"], "poem": p.poem} for p in poems.poems_list
    ],
)

Here:

• prompt_func takes a row of the dataset as input and returns the prompt for the LLM.
• response_format is the structured output class we defined above.
• parse_func takes the input (row) and the structured output (poems) and converts it to a list of dictionaries. This is so that we can easily convert the output to a HuggingFace Dataset object.

Now we can apply the LLM object to the dataset, which reads very pythonic.

poem = poet(topics)
print(poem.to_pandas())
# Example output:
#    topic                                     poem
# 0  Urban loneliness in a bustling city       In the city's heart, where the sirens wail,\nA...
# 1  Urban loneliness in a bustling city       City streets hum with a bittersweet song,\nHor...
# 2  Beauty of Bespoke Labs's Curator library  In whispers of design and crafted grace,\nBesp...
# 3  Beauty of Bespoke Labs's Curator library  In the hushed breath of parchment and ink,\nBe...

Note that topics can be created with curator.LLM as well, and we can scale this up to create tens of thousands of diverse poems. You can see a more detailed example in the examples/poem.py file, and other examples in the examples directory.

See the docs for more details as well as for troubleshooting information.

Bespoke Curator Viewer

To run the bespoke dataset viewer:

curator-viewer

This will pop up a browser window with the viewer running on 127.0.0.1:3000 by default if you haven't specified a different host and port.

The dataset viewer shows all the different runs you have made.

You can also see the dataset and the responses from the LLM.

Optional parameters to run the viewer on a different host and port:

>>> curator-viewer -h
usage: curator-viewer [-h] [--host HOST] [--port PORT] [--verbose]

Curator Viewer

options:
  -h, --help     show this help message and exit
  --host HOST    Host to run the server on (default: localhost)
  --port PORT    Port to run the server on (default: 3000)
  --verbose, -v  Enables debug logging for more verbose output

The only requirement for running curator-viewer is to install node. You can install them by following the instructions here.

For example, to check if you have node installed, you can run:

node -v

If it's not installed, installing latest node on MacOS, you can run:

# installs nvm (Node Version Manager)
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bash
# download and install Node.js (you may need to restart the terminal)
nvm install 22
# verifies the right Node.js version is in the environment
node -v # should print `v22.11.0`
# verifies the right npm version is in the environment
npm -v # should print `10.9.0`

Local Models Support

Table of Contents

• Online Inference with Local Models
• Offline Inference with Local Models
  • Setup instructions
  • Basic usage
  • Inference with models that do not fit in memory of one GPU (tensor parallelism)
  • Structured output with local models
  • Batched inference
  • Details on vLLM specific arguments
• Full list of local models examples

Online Inference with Local Models (vLLM)

You can use local models served with vLLM OpenAI compatibale server for online data generation. In order to do that, first start a vLLM ssrver:

vllm serve \
NousResearch/Meta-Llama-3-8B-Instruct \
--host localhost \
--port 8787 \
--api-key token-abc123

Note that you still need to provide a dummy API key since OpenAI client expects it (LiteLLM uses OpenAI client in the background). See here a full example on starting a vLLM server.

Then, to use this server, provide the API endpoint URL and the dummy API key. We use LiteLLM to generate data with vLLM hosted local models so you need to set the backend to litellm. In order for LiteLLM to recognize proper backend, add hosted_vllm/ prefix to your model name. Set HOSTED_VLLM_API_KEY environment varibale to your dummy API key. Example:

from bespokelabs import curator
model_path = "hosted_vllm/NousResearch/Meta-Llama-3-8B-Instruct" # Make sure to add hosted_vllm/ prefix
PORT = 8787
HOST = "localhost"
URL = f"http://{HOST}:{PORT}/v1"
os.environ["HOSTED_VLLM_API_KEY"] = "token-abc123

poem_prompter = curator.LLM(
    model_name=model_path,
    prompt_func=lambda row: "Generate a poem",
    backend="litellm",
    base_url=URL,
)
poem_prompter()

See here a full example. LiteLLM supports structured output, see here an example with a vLLM hosted model.

Offline Inference with Local Models (vLLM)

We use vLLM offline LLM running engine to generate synthetic data with local models.

Here is the full list of models that are supported by vLLM.

Setup instructions

Install vLLM in the Python or Conda environment:

pip install vllm

You may also need to install Ray for multi-node inference based on Ray clusters:

pip install ray

Please refer to the vLLM isntallation instructions.

Basic usage

To use curator with a local model, povide your model local path in the model argument and set the backend to vllm:

model_path = "meta-llama/Meta-Llama-3.1-8B-Instruct"
curator.LLM(
        model_name=model_path,
        backend="vllm"
    )

See full example here.

Inference for models that don't fit in one GPU's memory (tensor parallel)

To use local model that are too big to fit on one GPU, use tensor parallelism. That way you can split the model across multiple GPUs. To do that, specify tensor_parallel_size argument that should be equal to number of GPUs you have:

model_path = "meta-llama/Meta-Llama-3.1-70B-Instruct"
curator.LLM(
        model_name=model_path,
        backend="vllm",
        tensor_parallel_size=4 # split across 4 GPUs
    )

Structured output

We use vLLM's Guided Decoding to obtain structured output from local models during offline inference:

  from pydantic import BaseModel, Field

  class Cuisines(BaseModel):
      cuisines_list: List[str] = Field(description="A list of cuisines.")

  model_path = "/local/path/to/weights/meta-llama/Meta-Llama-3.1-70B-Instruct"
  cuisines_generator = curator.LLM(
      prompt_func=lambda: f"Generate 10 diverse cuisines.",
      model_name=model_path,
      response_format=Cuisines,
      backend="vllm",
  )

See full example here.

Batched inference

Offline vLLM inference support batch inference by default, the default batch size (number of sequences to process at a time) is equal to 256. Set the batch_size argument to change the default value:

  curator.LLM(
        model_name=model_path,
        prompt_func=lambda row: f"write a poem",
        backend="vllm",
        batch_size=32
    )

Details on vLLM specific arguments

• max_model_length (int, optional): The maximum model context length. Defaults to 4096.

• enforce_eager (bool, optional): Whether to enforce eager execution. Defaults to False.

• tensor_parallel_size (int, optional): The tensor parallel size. Defaults to 1.

• gpu_memory_utilization (float, optional): The GPU memory utilization. Defaults to 0.95.

• max_tokens (int, optional): The maximum number of tokens for models to generate. Defaults to 1024.

Full list of vLLM examples

• Generate recipes with Meta LLama 3.1 8B offline
• Recipes with structured output
• Use vLLM OpeneAI compatible server
• Use vLLM OpenAI compatible server with structured output

Contributing

Contributions are welcome!