Using language models for data

Overview

This post will give an overview of how (large) language models (LMs) fit into data engineering, data analyst, and data science workflows.

Use cases

There are three main use cases for language models for data practitioners:

Synthetic data generation
Use in analytic subroutines
Writing analytic code

We’ll describe each and then demonstrate them with code.

Setup

We’ll use Marvin, a toolkit for AI engineering, alongside Ibis, a toolkit for data engineering, to demonstrate the capabilities of language models for data using the default DuckDB backend.

Tip

We’ll use a cloud service provider (OpenAI) to demonstrate these capabilities. In a follow-up post, we’ll explore using local open source language models to achieve the same results.

With Marvin and Ibis, you can replicate the workflows below using other AI service providers, local language models, and over 20+ data backends!

You’ll need to install Marvin and Ibis to follow along:

pip install 'ibis-framework[duckdb,examples]' marvin

You need to create a .env file in $HOME/.marvin or your working directory like:

MARVIN_OPENAI_API_KEY="XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"
MARVIN_OPENAI_CHAT_COMPLETIONS_MODEL="gpt-4-0125-preview"

See the Marvin documentation for details on configuration.

Then import them and turn on Ibis interactive mode:

import ibis
import marvin

from pydantic import BaseModel, Field

ibis.options.interactive = True
ibis.options.repr.interactive.max_rows = 6

1: Import Ibis, an data engineering toolkit
2: Import Marvin, an AI engineering toolkit
3: Import Pydantic, used to define data models for Marvin
4: Set Ibis to interactive mode to display the results of our queries
5: Set the maximum number of rows to display in interactive mode

Synthetic data generation

Language models can be used to generate synthetic data. This is useful for testing, training, and other purposes. For example, you can use a language model to generate synthetic data for training a machine learning model (including a language model). This can be a great alternative to collecting or purchasing data for machine learning training and is easily customizable with language.

Tip

This post was re-inspired by the 1 billion row challenge we recently solved with Ibis on three local backends in which synthetic data created from a seed file was used to generate a billion rows.

With language models, we can reproduce this synthetic data and customize the data produced with natural language!

We’ll start by replicating the data in the one billion row challenge, then move over to our favorite penguins demo dataset to augment existing data with synthetic data.

Weather stations

We can generate synthetic weather stations in a few lines of code:

class WeatherStation(BaseModel):
    station: str = Field(
        ...,
        description="The weather station name",
        example="Sandy Silicon",
    )
    temperature: float = Field(
        ...,
        description="The average temperature in Fahrenheit",
        example=72.5,
    )


stations = marvin.generate(
    target=WeatherStation,
    instructions="Generate fictitious but plausible-sounding weather stations with names that excite data nerds",
    n=6,
)
stations

1: Define a data model for the weather stations
2: Use Marvin to generate three weather stations

[WeatherStation(station='Quantum Cloudburst', temperature=75.3),
 WeatherStation(station='Neural Nimbus', temperature=68.6),
 WeatherStation(station='Cyber Cirrus', temperature=59.1),
 WeatherStation(station='Data Drizzle Depot', temperature=70.2),
 WeatherStation(station='Bitstream Breeze', temperature=65.0),
 WeatherStation(station='Pixel Precipitation', temperature=77.8)]

And then load that data into an Ibis table:

Tip

You could also use a user-defined function (UDF) to directly generate this data in a table. We’ll demonstrate UDFs throughout this post.

s = ibis.memtable([station.model_dump() for station in stations])
s

1: Convert the generated data to an Ibis table

┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓
┃ station             ┃ temperature ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩
│ string              │ float64     │
├─────────────────────┼─────────────┤
│ Quantum Cloudburst  │        75.3 │
│ Neural Nimbus       │        68.6 │
│ Cyber Cirrus        │        59.1 │
│ Data Drizzle Depot  │        70.2 │
│ Bitstream Breeze    │        65.0 │
│ Pixel Precipitation │        77.8 │
└─────────────────────┴─────────────┘

While we’ve only generated six weather stations, you can repeat this process until you get as many as you’d like! You can then use Ibis to generate a billion rows of weather data for these stations as in the one billion row challenge.

Warning

Running this with GPT-4-turbo to generate 1000 weather stations (with n=10 and feeding in all previous attempts to the prompt to attempt avoiding duplicates) costs about $4 USD and resulted in about 150 duplicates. This is rather expensive for synthetic data! You can mitigate this by using a cheaper model (e.g. GPT-3.5-turbo) but may get worse results.

Alternatively, you can generate the data for free on your laptop with a small open source language model! Look out for a future post exploring this option.

Penguin poems

We can augment existing data with synthetic data. First, let’s load the penguins dataset:

penguins = ibis.examples.penguins.fetch()
penguins

1: Load the penguins dataset from Ibis examples

┏━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species ┃ island    ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string  │ string    │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├─────────┼───────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie  │ Torgersen │           39.1 │          18.7 │               181 │        3750 │ male   │  2007 │
│ Adelie  │ Torgersen │           39.5 │          17.4 │               186 │        3800 │ female │  2007 │
│ Adelie  │ Torgersen │           40.3 │          18.0 │               195 │        3250 │ female │  2007 │
│ Adelie  │ Torgersen │           NULL │          NULL │              NULL │        NULL │ NULL   │  2007 │
│ Adelie  │ Torgersen │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie  │ Torgersen │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ …       │ …         │              … │             … │                 … │           … │ …      │     … │
└─────────┴───────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

And take a sample of six rows to reduce our AI service costs:

t1 = penguins.filter(penguins.species == "Adelie").sample(fraction=0.1).limit(2)
t2 = penguins.filter(penguins.species == "Gentoo").sample(fraction=0.1).limit(2)
t3 = (
    penguins.filter(penguins.species == "Chinstrap")
    .sample(fraction=0.1)
    .limit(2)
)
t = t1.union(t2).union(t3)
t

1: Sample two rows from each species
2: Union the samples together

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ island    ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼───────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │ Torgersen │           39.2 │          19.6 │               195 │        4675 │ male   │  2007 │
│ Adelie    │ Torgersen │           41.1 │          17.6 │               182 │        3200 │ female │  2007 │
│ Gentoo    │ Biscoe    │           46.1 │          15.1 │               215 │        5100 │ male   │  2007 │
│ Gentoo    │ Biscoe    │           59.6 │          17.0 │               230 │        6050 │ male   │  2007 │
│ Chinstrap │ Dream     │           46.5 │          17.9 │               192 │        3500 │ female │  2007 │
│ Chinstrap │ Dream     │           45.9 │          17.1 │               190 │        3575 │ female │  2007 │
└───────────┴───────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Now we define a UDF to generate a poem to describe each penguin:

@ibis.udf.scalar.python
def penguin_poem(
    species: str,
    island: str,
    bill_length_mm: float,
    bill_depth_mm: float,
    flipper_length_mm: float,
    body_mass_g: float,
) -> str:

    instructions = f"""Provide a whimsical poem that rhymes for a penguin.

    You have the following information about the penguins:
        species {species}
        island of {island}
        bill length of {bill_length_mm} mm
        bill depth of {bill_depth_mm} mm
        flipper length of {flipper_length_mm} mm
        body mass of {body_mass_g} g.

    You must reference the penguin's size in addition to its species and island.
    """

    poem = marvin.generate(
        n=1,
        instructions=instructions,
    )

    return poem[0]

1: Define a scalar Python UDF to generate a poem from penguin data
2: Augment the LM’s prompt with the penguin data
3: Use Marvin to generate a poem for the penguin data
4: Return the generated poem

And apply that UDF to our penguins table:

t = (
    t.mutate(
        poem=penguin_poem(
            t.species,
            t.island,
            t.bill_length_mm,
            t.bill_depth_mm,
            t.flipper_length_mm,
            t.body_mass_g,
        )
    )
    .relocate("species", "island", "poem")
    .cache()
)
t

1: Apply the UDF by mutating the table with the function, using other columns as input
2: Rearrange the columns we care about to the front
3: Cache the table to avoid re-running the UDF

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ island    ┃ poem                                                                              ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ string                                                                            │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │ Torgersen │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie    │ Torgersen │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ Chinstrap │ Dream     │ On Dream Island's frosty sedge,                                                   │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ Chinstrap │ Dream     │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└───────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Nice! While not particularly useful in this case, the same process can be used for generating product descriptions or other practical applications.

Use in analytic subroutines

We’ve already done this above with the penguin poems to generate some data. However, for real-world use cases the more common use of language models will be to extract structured data from unstructured text.

This includes tasks like:

sentiment analysis
data labeling
named entity recognition
part of speech tagging
summarization
translation
question answering

Each of these tasks can be, to some extent, performed by traditional natural language processing (NLP) techniques. However, modern-day LMs can solve these tasks with a single model, and often with state-of-the-art performance. This drastically simplifies what a single engineer, who doesn’t need a deep understanding of NLP or ML in general, can accomplish.

Sentiment analysis

We can use a language model to perform sentiment analysis on the penguin poems:

@marvin.fn
def _sentiment_analysis(text: str) -> float:
    """Returns a sentiment score for `text`
    between -1 (negative) and 1 (positive)."""


@ibis.udf.scalar.python
def sentiment_analysis(text: str) -> float:
    return _sentiment_analysis(text)

1: Define a Marvin function to perform sentiment analysis
2: Define a scalar Python UDF to apply the Marvin function to a column
3: Apply the Marvin function within the UDF

And apply that UDF to our penguins table:

t = (
    t.mutate(sentiment=sentiment_analysis(t.poem))
    .relocate(t.columns[:3], "sentiment")
    .cache()
)
t

1: Apply the UDF by mutating the table with the function
2: Rearrange the columns we care about to the front
3: Cache the table to avoid re-running the UDF

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ island    ┃ poem                                                                              ┃ sentiment ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ string                                                                            │ float64   │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────┼───────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │ Torgersen │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │      0.50 │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie    │ Torgersen │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │      0.80 │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │      0.75 │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │      0.70 │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ Chinstrap │ Dream     │ On Dream Island's frosty sedge,                                                   │      0.00 │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ Chinstrap │ Dream     │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │      0.90 │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└───────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────┴───────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Entity extraction

While not exactly named entity recognition, we can extract arbitrary entities from text. In this case, we’ll extract a list of words that rhyme from the poem:

@ibis.udf.scalar.python
def extract_rhyming_words(text: str) -> list[str]:
    words = marvin.extract(
        text,
        instructions="Extract the primary rhyming words from the text",
    )

    return words

1: Define a scalar Python UDF to extract rhyming words from a poem
2: Use Marvin to extract the rhyming words from the poem
3: Return the list of extracted words

And apply that UDF to our penguins table:

t = (
    t.mutate(rhyming_words=extract_rhyming_words(t.poem))
    .relocate(t.columns[:4], "rhyming_words")
    .cache()
)
t

1: Apply the UDF by mutating the table with the function
2: Rearrange the columns we care about to the front
3: Cache the table to avoid re-running the UDF

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ island    ┃ poem                                                                              ┃ sentiment ┃ rhyming_words               ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ string                                                                            │ float64   │ array<string>               │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────┼───────────┼─────────────────────────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │ Torgersen │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │      0.50 │ ['clime', 'rhyme', ... +9]  │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie    │ Torgersen │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │      0.80 │ ['dome', 'roam', ... +12]   │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │      0.75 │ ['strong', 'song', ... +9]  │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │      0.70 │ ['plight', 'right', ... +6] │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ Chinstrap │ Dream     │ On Dream Island's frosty sedge,                                                   │      0.00 │ ['island', 'sedge']         │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ Chinstrap │ Dream     │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │      0.90 │ ['chill', 'hill', ... +10]  │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└───────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────┴───────────┴─────────────────────────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Translation

We can translate the penguin poems into Spanish or any language the language model sufficiently knows:

@marvin.fn
def _translate_text(text: str, target_language: str = "spanish") -> str:
    """Translate `text` to `target_language`."""


@ibis.udf.scalar.python
def translate_text(text: str, target_language: str = "spanish") -> str:
    return _translate_text(text, target_language)

1: Define a Marvin function to translate text
2: Define a scalar Python UDF to apply the Marvin function to a column
3: Apply the Marvin function within the UDF

And apply that UDF to our penguins table:

t = (
    t.mutate(translated_poem=translate_text(t.poem))
    .relocate(t.columns[:5], "translated_poem")
    .cache()
)
t

1: Apply the UDF by mutating the table with the function
2: Rearrange the columns we care about to the front
3: Cache the table to avoid re-running the UDF

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ island    ┃ poem                                                                              ┃ sentiment ┃ rhyming_words               ┃ translated_poem                                                                   ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ string                                                                            │ float64   │ array<string>               │ string                                                                            │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────┼───────────┼─────────────────────────────┼───────────────────────────────────────────────────────────────────────────────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │ Torgersen │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │      0.50 │ ['clime', 'rhyme', ... +9]  │ En el frío clima de Torgersen, un pingüino vive en ritmo y rima,\nUn Adelie, ele… │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie    │ Torgersen │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │      0.80 │ ['dome', 'roam', ... +12]   │ En el escenario helado de Torgersen, bajo la amplia c\u00fapula del cielo,\nVive… │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │      0.75 │ ['strong', 'song', ... +9]  │ Un pingüino Gentoo de la playa de Biscoe, con un pico bastante largo,\n50 milíme… │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ Gentoo    │ Biscoe    │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │      0.70 │ ['plight', 'right', ... +6] │ La caprichosa Gentoo de Biscoe plight,\nCon un pico tan delicado, 46.5 mm justo.… │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ Chinstrap │ Dream     │ On Dream Island's frosty sedge,                                                   │      0.00 │ ['island', 'sedge']         │ En la orilla helada de la Isla de los Sue\u00f1os,                                │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ Chinstrap │ Dream     │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │      0.90 │ ['chill', 'hill', ... +10]  │ En la isla de Ensue\u00f1o, tan serena y tranquila,\nUn ping\u00fcino barbijo se… │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└───────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────┴───────────┴─────────────────────────────┴───────────────────────────────────────────────────────────────────────────────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Data labeling

We can label the data by classifying penguins based on their attributes. While this is a contrived example, this could be useful for classifying product feedback or customer support tickets for data organization in any backend Ibis supports:

labels = ["small", "medium", "large"]


instructions = f"""You are to classify a penguin into one of {labels} based on
its bill length, bill depth, flipper length, and body mass.

The averages for these penguins are:

    - bill length: {penguins["bill_length_mm"].mean().to_pandas()}mm
    - bill depth: {penguins["bill_depth_mm"].mean().to_pandas()}mm
    - flipper length: {penguins["flipper_length_mm"].mean().to_pandas()}mm
    - body mass: {penguins["body_mass_g"].mean().to_pandas()}g

Use your best judgement to classify the penguin.
"""


@ibis.udf.scalar.python
def classify_penguin(
    bill_length: float, bill_depth: float, flipper_length: float, body_mass: float
) -> str:
    return marvin.classify(
        f"Penguin with bill length {bill_length}mm, bill depth {bill_depth}mm, flipper length {flipper_length}mm, and body mass {body_mass}g",
        labels=labels,
        instructions=instructions,
    )

1: Define the labels for the classification
2: Construct instructions for the LM based on data
3: Define a scalar Python UDF to classify penguins based on their attributes

And apply that UDF to our penguins table:

t = (
    t.mutate(
        classification=classify_penguin(
            t.bill_length_mm, t.bill_depth_mm, t.flipper_length_mm, t.body_mass_g
        )
    )
    .relocate("classification")
    .cache()
)
t

1: Apply the UDF by mutating the table with the function
2: Rearrange the columns we care about to the front
3: Cache the table to avoid re-running the UDF

┏━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ classification ┃ species   ┃ island    ┃ poem                                                                              ┃ sentiment ┃ rhyming_words               ┃ translated_poem                                                                   ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string         │ string    │ string    │ string                                                                            │ float64   │ array<string>               │ string                                                                            │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├────────────────┼───────────┼───────────┼───────────────────────────────────────────────────────────────────────────────────┼───────────┼─────────────────────────────┼───────────────────────────────────────────────────────────────────────────────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ small          │ Adelie    │ Torgersen │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │      0.50 │ ['clime', 'rhyme', ... +9]  │ En el frío clima de Torgersen, un pingüino vive en ritmo y rima,\nUn Adelie, ele… │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ small          │ Adelie    │ Torgersen │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │      0.80 │ ['dome', 'roam', ... +12]   │ En el escenario helado de Torgersen, bajo la amplia c\u00fapula del cielo,\nVive… │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ large          │ Gentoo    │ Biscoe    │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │      0.75 │ ['strong', 'song', ... +9]  │ Un pingüino Gentoo de la playa de Biscoe, con un pico bastante largo,\n50 milíme… │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ large          │ Gentoo    │ Biscoe    │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │      0.70 │ ['plight', 'right', ... +6] │ La caprichosa Gentoo de Biscoe plight,\nCon un pico tan delicado, 46.5 mm justo.… │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ medium         │ Chinstrap │ Dream     │ On Dream Island's frosty sedge,                                                   │      0.00 │ ['island', 'sedge']         │ En la orilla helada de la Isla de los Sue\u00f1os,                                │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ large          │ Chinstrap │ Dream     │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │      0.90 │ ['chill', 'hill', ... +10]  │ En la isla de Ensue\u00f1o, tan serena y tranquila,\nUn ping\u00fcino barbijo se… │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└────────────────┴───────────┴───────────┴───────────────────────────────────────────────────────────────────────────────────┴───────────┴─────────────────────────────┴───────────────────────────────────────────────────────────────────────────────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Writing analytic code

Finally, language models can be used to write code. This is more useful with systems around them to execute code, feed back error messages, make adjustments, and so on. There are numerous pitfalls with language models writing code, but they’re fairly good at SQL.

Two approaches

We can think of two approaches to writing analytic code with language models:

Use LMs to write an analytic subroutine
Use LMs to write analytic code

Writing an analytic subroutine

We’ll define a function that writes an Ibis UDF, similar to what we’ve used above:

@marvin.fn
def _generate_python_function(text: str) -> str:
    """Generate a simple, typed, correct Python function from text."""


def create_udf_from_text(text: str) -> str:
    """Create a UDF from text."""
    return f"""
import ibis

@ibis.udf.scalar.python
{_generate_python_function(text)}
""".strip()

1: Define a Marvin function to generate a Python function from text
2: Define a Python function, wrapping a Marvin function with an Ibis UDF

Now we’ll create a UDF from text to count the number of vowels in a string:

udf = create_udf_from_text(
    """"a function named count_vowels that given an input string, returns an int
    w/ the number of vowels (y_included as a boolean option defaulted to False).
    NO DOCSTRING, we don't document our code in this household"""
)
print(udf)

1: Create a UDF from text

import ibis

@ibis.udf.scalar.python
def count_vowels(input_string: str, y_included: bool = False) -> int:
    vowels = 'aeiou' + ('y' if y_included else '')
    return sum(1 for char in input_string.lower() if char in vowels)

And execute that so the UDF is available:

exec(udf)

1: Execute the UDF code to make it available to call

Now we can augment our table with an analytic subroutine written by a language model:

t = t.mutate(
    species_vowel_count=count_vowels(t.species),
    island_vowel_count=count_vowels(t.island),
).relocate("species", "species_vowel_count", "island", "island_vowel_count")
t

1: Apply the UDF by mutating the table with the function
2: Rearrange the columns we care about to the front

┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ species   ┃ species_vowel_count ┃ island    ┃ island_vowel_count ┃ classification ┃ poem                                                                              ┃ sentiment ┃ rhyming_words               ┃ translated_poem                                                                   ┃ bill_length_mm ┃ bill_depth_mm ┃ flipper_length_mm ┃ body_mass_g ┃ sex    ┃ year  ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ string    │ int64               │ string    │ int64              │ string         │ string                                                                            │ float64   │ array<string>               │ string                                                                            │ float64        │ float64       │ int64             │ int64       │ string │ int64 │
├───────────┼─────────────────────┼───────────┼────────────────────┼────────────────┼───────────────────────────────────────────────────────────────────────────────────┼───────────┼─────────────────────────────┼───────────────────────────────────────────────────────────────────────────────────┼────────────────┼───────────────┼───────────────────┼─────────────┼────────┼───────┤
│ Adelie    │                   4 │ Torgersen │                  3 │ small          │ In Torgersen's chilly clime, a penguin dwells in rhythm and rhyme,\nAn Adelie, s… │      0.50 │ ['clime', 'rhyme', ... +9]  │ En el frío clima de Torgersen, un pingüino vive en ritmo y rima,\nUn Adelie, ele… │           36.7 │          19.3 │               193 │        3450 │ female │  2007 │
│ Adelie    │                   4 │ Torgersen │                  3 │ small          │ On Torgersen's icy stage, beneath the sky's wide dome,\nLives a sleek Adelie, in… │      0.80 │ ['dome', 'roam', ... +12]   │ En el escenario helado de Torgersen, bajo la amplia c\u00fapula del cielo,\nVive… │           39.3 │          20.6 │               190 │        3650 │ male   │  2007 │
│ Gentoo    │                   3 │ Biscoe    │                  3 │ large          │ A Gentoo penguin from Biscoe's beach, with a bill quite long,\n50 millimeters an… │      0.75 │ ['strong', 'song', ... +9]  │ Un pingüino Gentoo de la playa de Biscoe, con un pico bastante largo,\n50 milíme… │           50.0 │          15.2 │               218 │        5700 │ male   │  2007 │
│ Gentoo    │                   3 │ Biscoe    │                  3 │ large          │ The whimsy Gentoo from Biscoe plight,\nWith a bill so dainty, 46.5 mm just right… │      0.70 │ ['plight', 'right', ... +6] │ La caprichosa Gentoo de Biscoe plight,\nCon un pico tan delicado, 46.5 mm justo.… │           46.5 │          13.5 │               210 │        4550 │ female │  2007 │
│ Chinstrap │                   2 │ Dream     │                  2 │ medium         │ On Dream Island's frosty sedge,                                                   │      0.00 │ ['island', 'sedge']         │ En la orilla helada de la Isla de los Sue\u00f1os,                                │           46.6 │          17.8 │               193 │        3800 │ female │  2007 │
│ Chinstrap │                   2 │ Dream     │                  2 │ large          │ On the island of Dream, so serene and so chill,\nA Chinstrap penguin slides, on … │      0.90 │ ['chill', 'hill', ... +10]  │ En la isla de Ensue\u00f1o, tan serena y tranquila,\nUn ping\u00fcino barbijo se… │           50.5 │          19.6 │               201 │        4050 │ male   │  2007 │
└───────────┴─────────────────────┴───────────┴────────────────────┴────────────────┴───────────────────────────────────────────────────────────────────────────────────┴───────────┴─────────────────────────────┴───────────────────────────────────────────────────────────────────────────────────┴────────────────┴───────────────┴───────────────────┴─────────────┴────────┴───────┘

Note

You could use a LM to write a subroutine that uses a LM! This is an exercise left to the reader.

Writing analytic code

Let’s define a function that outputs SQL:

@marvin.fn
def _text_to_sql(
    text: str,
    table_names: list[str],
    table_schemas: list[str],
    table_previews: list[str],
) -> str:
    """Writes a SQL SELECT statement for the `text` given the provided
    `table_names`, `table_schemas`, and `table_previews`.

    Use newlines and indentation for readability.
    """


def text_to_sql(
    text: str,
    table_names: list[str],
    table_schemas: list[str],
    table_previews: list[str],
) -> str:
    sql = _text_to_sql(text, table_names, table_schemas, table_previews)
    return sql.strip().strip(";")

1: Define a Marvin function to write SQL from text
2: Define a Python function to apply the Marvin function to a string
3: Generate the SQL string
4: Strip the trailing whitespace and semicolon that is sometimes generated, as it causes issues with Ibis

We can try that out on our penguins table:

text = (
    "the count of penguins by species, from highest to lowest, per each island"
)

table_names = ["penguins"]
table_schemas = [str(penguins.schema())]
table_previews = [str(penguins.limit(5))]

sql = text_to_sql(text, table_names, table_schemas, table_previews)
print(sql)

1: Create a natural language query
2: Provide the table names, schemas, and previews
3: Generate the SQL string

SELECT
  island,
  species,
  COUNT(*) AS count
FROM
  penguins
GROUP BY
  island,
  species
ORDER BY
  island,
  COUNT(*) DESC

And execute the SQL:

r = penguins.sql(sql)
r

1: Execute the SQL string on the table

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━┓
┃ island    ┃ species   ┃ count ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━┩
│ string    │ string    │ int64 │
├───────────┼───────────┼───────┤
│ Biscoe    │ Gentoo    │   124 │
│ Biscoe    │ Adelie    │    44 │
│ Dream     │ Chinstrap │    68 │
│ Dream     │ Adelie    │    56 │
│ Torgersen │ Adelie    │    52 │
└───────────┴───────────┴───────┘

A more complex example

Let’s see how this works on a query that requires joining two tables. We’ll load in some IMDB data:

imdb_title_basics = ibis.examples.imdb_title_basics.fetch()
imdb_title_basics

1: Load the IMDB title basics dataset from Ibis examples

┏━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ tconst    ┃ titleType ┃ primaryTitle           ┃ originalTitle          ┃ isAdult ┃ startYear ┃ endYear ┃ runtimeMinutes ┃ genres                   ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ string    │ string    │ string                 │ string                 │ int64   │ int64     │ string  │ int64          │ string                   │
├───────────┼───────────┼────────────────────────┼────────────────────────┼─────────┼───────────┼─────────┼────────────────┼──────────────────────────┤
│ tt0000001 │ short     │ Carmencita             │ Carmencita             │       0 │      1894 │ NULL    │              1 │ Documentary,Short        │
│ tt0000002 │ short     │ Le clown et ses chiens │ Le clown et ses chiens │       0 │      1892 │ NULL    │              5 │ Animation,Short          │
│ tt0000003 │ short     │ Pauvre Pierrot         │ Pauvre Pierrot         │       0 │      1892 │ NULL    │              4 │ Animation,Comedy,Romance │
│ tt0000004 │ short     │ Un bon bock            │ Un bon bock            │       0 │      1892 │ NULL    │             12 │ Animation,Short          │
│ tt0000005 │ short     │ Blacksmith Scene       │ Blacksmith Scene       │       0 │      1893 │ NULL    │              1 │ Comedy,Short             │
│ tt0000006 │ short     │ Chinese Opium Den      │ Chinese Opium Den      │       0 │      1894 │ NULL    │              1 │ Short                    │
│ …         │ …         │ …                      │ …                      │       … │         … │ …       │              … │ …                        │
└───────────┴───────────┴────────────────────────┴────────────────────────┴─────────┴───────────┴─────────┴────────────────┴──────────────────────────┘

imdb_title_ratings = ibis.examples.imdb_title_ratings.fetch()
imdb_title_ratings

1: Load the IMDB title ratings dataset from Ibis examples

┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┓
┃ tconst    ┃ averageRating ┃ numVotes ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━┩
│ string    │ float64       │ int64    │
├───────────┼───────────────┼──────────┤
│ tt0000001 │           5.7 │     1990 │
│ tt0000002 │           5.8 │      265 │
│ tt0000003 │           6.5 │     1869 │
│ tt0000004 │           5.5 │      177 │
│ tt0000005 │           6.2 │     2655 │
│ tt0000006 │           5.0 │      182 │
│ …         │             … │        … │
└───────────┴───────────────┴──────────┘

text = "the highest rated movies w/ over 100k ratings -- movies only"

table_names = ["imdb_title_basics", "imdb_title_ratings"]
table_schemas = [
    str(imdb_title_basics.schema()),
    str(imdb_title_ratings.schema()),
]
table_previews = [
    str(imdb_title_basics.limit(5)),
    str(imdb_title_ratings.limit(5)),
]

sql = text_to_sql(text, table_names, table_schemas, table_previews)
print(sql)

1: Create a natural language query
2: Provide the table names, schemas, and previews
3: Generate the SQL string

SELECT
    title.primaryTitle,
    rating.averageRating,
    rating.numVotes
FROM
    imdb_title_basics AS title
JOIN
    imdb_title_ratings AS rating
ON
    title.tconst = rating.tconst
WHERE
    title.titleType = 'movie'
    AND rating.numVotes > 100000
ORDER BY
    rating.averageRating DESC

r = imdb_title_basics.sql(sql)
r

1: Execute the SQL string on the table

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┓
┃ primaryTitle                                  ┃ averageRating ┃ numVotes ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━┩
│ string                                        │ float64       │ int64    │
├───────────────────────────────────────────────┼───────────────┼──────────┤
│ The Shawshank Redemption                      │           9.3 │  2793065 │
│ The Godfather                                 │           9.2 │  1945537 │
│ 12 Angry Men                                  │           9.0 │   829776 │
│ The Godfather Part II                         │           9.0 │  1321642 │
│ Schindler's List                              │           9.0 │  1404536 │
│ The Lord of the Rings: The Return of the King │           9.0 │  1912955 │
│ …                                             │             … │        … │
└───────────────────────────────────────────────┴───────────────┴──────────┘

Issues with language models today

The biggest issue with language models today is the cost, followed by the time it takes to generate responses. Generating a million rows of synthetic data points with state-of-the-art LLMs is prohibitively expensive and slow. Extrapolating from generating 1000 weather stations, it would cost about $4000 USD and take over a week using GPT-4-turbo with OpenAI. Then again, this could be less expensive and time consuming than collecting or purchasing the data. If you’re using language models in your data workflows, you’ll need to be mindful of the cost and bottleneck on speed of data processing likely to be introduced.

Smaller open source language models can reduce this cost to practically zero, but at the expensive of quality. Generating responses may also be slower depending on your hardware setup and other factors.

Language models also have a tendency to generate incorrect outputs in the context of a given task. This is a risk that needs to be tolerated or mitigated with a surrounding system. I would not recommend giving a language model access to a production database without a human in the loop.

Looking forward

In a future post, we’ll explore using open source language models on a laptop to achieve similar results. In general, I expect language models with data to become increasingly commonplace. Some predictions:

small, specialized language models will become common
large language models that require expensive GPU clusters will still be the best for general-purpose tasks
data assistants will be gimmicky for a while, but will eventually become useful

Keep an eye out for our own gimmicky data assistant soon…

Conclusion

To summarize this blog:

from rich import print

with open("index.qmd", "r") as f:
    blog_text = f.read()


class BlogSummary(BaseModel):
    """Summarizes a blog post"""

    key_points: list[str] = Field(..., description="The key points of the blog post")
    one_paragraph_summary: str = Field(
        ..., description="A one paragraph summary of the blog post"
    )
    one_sentence_summary: str = Field(
        ..., description="A one sentence summary of the blog post"
    )


summary = marvin.cast(
    blog_text,
    target=BlogSummary,
)
print(summary)

1: Import rich’s print to better display Python objects
2: Read the blog post from a file
3: Define a data model for the blog post summary
4: Use Marvin to cast the blog post to the data model

BlogSummary(
    key_points=[
        'Language models can generate synthetic data for testing and training.',
        'Language models can be integrated into analytic subroutines for tasks like sentiment analysis and entity 
extraction.',
        'Language models can assist in writing analytic code, including SQL queries.'
    ],
    one_paragraph_summary='The blog post discusses the integration of language models into data workflows, 
highlighting their utility in generating synthetic data, enhancing analytic subroutines with tasks like sentiment 
analysis, and aiding in writing analytic code. It also touches on the cost and speed considerations when using 
language models, and the potential of smaller, open-source models for reducing costs.',
    one_sentence_summary='The blog post explores the use of language models in data engineering, analysis, and 
science, emphasizing their versatility and the considerations for their cost-effective application.'
)

To generate a thumbnail from the text:

import os
import httpx

if not os.path.exists("thumbnail.png"):
    key_points = "\n - ".join(summary.key_points)
    thumbnail = marvin.paint(
        f"""Create a beautiful (dark background) thumbnail for the blog post
        with the following key points:

        {key_points}

        The post is about {summary.one_sentence_summary}

        A one paragraph summary:

        {summary.one_paragraph_summary}
        """
    )

    image = httpx.get(thumbnail.data[0].url)
    with open("thumbnail.png", "wb") as f:
        f.write(image.content)

1: Import extra libraries
2: Check if the thumbnail already exists
3: Use Marvin to generate a thumbnail from the blog post summary
4: Get the thumnail data from the URL
5: Write the data’s content to a file

from IPython.display import Image

Image(filename="./thumbnail.png")

1: Import IPython functionality to display the thumbnail
2: Display the thumbnail

Next steps

Try this out yourself! All you need is an OpenAI account and the code above.

It’s never been a better time to get involved with Ibis. Join us on Zulip and introduce yourself! More Ibis + language model content coming soon.