Structured output

Set response_format to a Pydantic model, and Curator asks the model for output that matches it. Your parse method then gets an instance of that model instead of a string.

Generate poems with structured output

from typing import Dict, List

from datasets import Dataset
from pydantic import BaseModel, Field

from bespokelabs import curator


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


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


class Poet(curator.LLM):
    response_format = Poems

    def prompt(self, input: Dict) -> str:
        return f"Write two poems about {input['topic']}."

    def parse(self, input: Dict, response: Poems) -> List[Dict]:
        return [{"topic": input["topic"], "poem": p.poem} for p in response.poems]


poet = Poet(model_name="gpt-4o-mini")

topics = Dataset.from_dict({
    "topic": [
        "Urban loneliness in a bustling city",
        "The first snow of winter",
    ]
})

poems = poet(topics)
print(poems.dataset.to_pandas())
# Output:
#                                  topic                                               poem
# 0  Urban loneliness in a bustling city  In the city's heart, where the lights never di...
# 1  Urban loneliness in a bustling city  Steps echo loudly, pavement slick with rain,\n...
# 2              The first snow of winter  Soft and silent, the first flakes fall,\nA hu...
# 3              The first snow of winter  White drifts cover the sleeping ground,\nAnd ...

This is what the code does.

The Pydantic models Poem and Poems describe the output you want from the LLM.
Poet sets response_format = Poems. You can also pass response_format=Poems when you create the object.
For each topic, prompt builds the prompt and Curator sends it to the LLM.
parse gets a Poems object and returns one row for each poem, with the topic next to it.

The result has one poem per row, and each row keeps the topic it came from.

Chain LLM calls

You can pass the output of one LLM object to the next. In this example, a Muse generates the topics and the Poet from above writes poems about them.

from typing import Dict, List

from pydantic import BaseModel, Field

from bespokelabs import curator


class Topic(BaseModel):
    topic: str = Field(description="A topic.")


class Topics(BaseModel):
    topics: List[Topic] = Field(description="A list of topics.")


class Muse(curator.LLM):
    response_format = Topics

    def prompt(self, input: Dict) -> str:
        return "Generate ten evocative poetry topics."

    def parse(self, input: Dict, response: Topics) -> List[Dict]:
        return [{"topic": t.topic} for t in response.topics]


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


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


class Poet(curator.LLM):
    response_format = Poems

    def prompt(self, input: Dict) -> str:
        return f"Write two poems about {input['topic']}."

    def parse(self, input: Dict, response: Poems) -> List[Dict]:
        return [{"topic": input["topic"], "poem": p.poem} for p in response.poems]


muse = Muse(model_name="gpt-4o-mini")
topics = muse()
print(topics.dataset.to_pandas())

poet = Poet(model_name="gpt-4o-mini")
poems = poet(topics)
print(poems.dataset.to_pandas())
# Output:
#                                   topic                                               poem
# 0  The fleeting beauty of autumn leaves  In a whisper of wind, they dance and they sway...
# 1  The fleeting beauty of autumn leaves  Once vibrant with life, now a radiant fade,\nC...
# 2    The whispers of an abandoned house  In shadows deep where light won't tread,  \nAn...
# 3    The whispers of an abandoned house  Abandoned now, my heart does fade,  \nOnce a h...
# ...

muse() has no input, so Curator sends one request. You can pass the CuratorResponse it returns straight to poet. Each step fans out the rows of the step before it, so a short chain can produce a large dataset.

Not every model supports structured output. If a model does not support it, Curator raises an error before it sends any requests.

Get started

Core features

Guides

Recipes

Fine-tuning

Reference

Generate poems with structured output

Chain LLM calls

​Generate poems with structured output

​Chain LLM calls

Generate poems with structured output

Chain LLM calls