Config Generation

LLM Config Generation¶

Ludwig can generate a complete, validated YAML configuration from a plain-English description of your machine learning task. This is useful for rapid prototyping, onboarding new users, or automating ML pipeline setup.

How it works¶

ludwig.config_generation.generate_config sends your task description to Claude or GPT-4, along with a compact representation of Ludwig's configuration schema. The LLM returns a JSON config that is then validated with full Pydantic validation before being returned to you.

Prerequisites¶

Install an LLM provider SDK:

pip install anthropic    # for Claude (default)
# or
pip install openai       # for GPT-4o

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-...
# or
export OPENAI_API_KEY=sk-...

Python API¶

from ludwig.config_generation import generate_config

config = generate_config(
    "I have a CSV with columns: age (number), annual_income (number), "
    "num_purchases (integer), and days_since_last_purchase (integer). "
    "I want to predict whether a customer will churn (binary: 0 or 1)."
)

# config is a validated dict you can pass directly to LudwigModel
from ludwig.api import LudwigModel
model = LudwigModel(config=config)

Parameters¶

Parameter	Default	Description
`task_description`	—	Plain-English description of your ML task
`model`	`claude-sonnet-4-20250514`	LLM to use. Claude models start with `claude-`, OpenAI with `gpt-`
`api_key`	`None`	API key; defaults to reading `ANTHROPIC_API_KEY` / `OPENAI_API_KEY` env vars
`validate`	`True`	Whether to validate the generated config with Pydantic before returning

Example descriptions¶

Tabular classification:

config = generate_config(
    "Predict loan default from: age (number), income (number), "
    "debt_to_income_ratio (number), loan_purpose (category), "
    "credit_score (number). Target: defaulted (binary)."
)

Text classification:

config = generate_config(
    "I have product reviews (text) and want to classify sentiment "
    "as positive, negative, or neutral (category)."
)

Multi-output regression:

config = generate_config(
    "Given wine chemical properties (fixed_acidity, volatile_acidity, "
    "citric_acid, pH — all numbers), predict quality_score (number 1-10) "
    "and whether the wine is good (binary)."
)

Multimodal:

config = generate_config(
    "I have a dataset of product listings with: title (text), "
    "image_path (image), and price (number). "
    "Predict whether the listing will sell within 7 days (binary)."
)

CLI¶

ludwig generate_config "predict house price from bedrooms, sqft, location"

Save to file:

ludwig generate_config "classify email as spam or ham" --output config.yaml

Use GPT-4o:

ludwig generate_config --model gpt-4o "predict churn from user activity metrics"

Inspecting and editing the generated config¶

Always review the generated config before training. Print it as YAML:

import yaml
from ludwig.config_generation import generate_config

config = generate_config("predict house price from bedrooms, sqft, location, year_built")
print(yaml.dump(config, default_flow_style=False, sort_keys=False))

The generated config is a standard Ludwig config dict — edit it freely before passing to LudwigModel.

Training on the generated config¶

import pandas as pd
from ludwig.api import LudwigModel
from ludwig.config_generation import generate_config

config = generate_config(
    "predict customer churn from age, income, num_purchases, days_since_last_purchase"
)

df = pd.read_csv("customers.csv")
model = LudwigModel(config=config)
results = model.train(dataset=df, output_directory="./results")

Limitations¶

Generated configs may need adjustment for domain-specific encoders (e.g., specifying pretrained_model_name_or_path for text encoders).
LLM-generated configs are validated but not guaranteed to be optimal — treat them as a starting point for hyperparameter tuning.
Image, audio, and time series features require correct file paths in the dataset.
Very complex configs (multi-adapter PEFT, custom backends) may require manual editing.