Plexe provides flexible options for working with datasets. This reference documents how to prepare, provide, and generate data for model building.

​
Supported Dataset Types

Plexe accepts two types of objects in the datasets parameter of model.build():
  1. Pandas DataFrames: For providing tabular data directly
  2. DatasetGenerator objects: For generating synthetic data or augmenting existing data

​
Using Pandas DataFrames

Pandas DataFrames are the most common way to provide data to Plexe.

​
Basic Usage

import pandas as pd
import plexe

# Load data
df = pd.read_csv("customer_data.csv")

# Build model with DataFrame
model = plexe.Model(intent="Predict customer churn")
model.build(datasets=[df])

​
Multiple DataFrames

You can provide multiple DataFrames for more complex scenarios: 
# Load data from multiple sources
customers_df = pd.read_csv("customers.csv")
transactions_df = pd.read_csv("transactions.csv")

# Build model with multiple DataFrames
model = plexe.Model(intent="Predict customer lifetime value")
model.build(datasets=[customers_df, transactions_df])
When multiple DataFrames are provided, Plexe’s ML Engineer agent will attempt to determine relationships between them based on column names and data types.

​
DataFrame Requirements

While Plexe is flexible, following these guidelines helps ensure optimal results:
  • Clean Data: Remove or impute missing values when possible
  • Appropriate Types: Ensure columns have appropriate data types
  • Meaningful Names: Use descriptive column names
  • Reasonable Size: Keep DataFrames under a few million rows for optimal performance

​
DatasetGenerator

The DatasetGenerator class allows you to generate synthetic data or augment existing data using LLMs.

​
Class Definition

class DatasetGenerator:
    def __init__(
        self,
        description: str,
        provider: str,
        schema: Type[BaseModel] | Dict[str, type] = None,
        data: pd.DataFrame = None
    ) -> None
    
    def generate(self, num_samples: int):
        """Generates synthetic data if a provider is available."""
ParameterTypeDescription
descriptionstrHuman-readable description of the dataset
providerstrLLM provider used for synthetic data generation
schema`Type[BaseModel]Dict[str, type]`The schema the data should match, if any. Can be a Pydantic model or dictionary.
datapd.DataFrameA dataset of real data on which to base the generation, if available

​
Generating Synthetic Data

To generate completely synthetic data: 
import plexe
from plexe import DatasetGenerator
from pydantic import BaseModel, Field

# Define schema for synthetic data
class CustomerData(BaseModel):
    age: int = Field(ge=18, le=100, description="Customer age in years")
    income: float = Field(ge=0, description="Annual income in USD")
    has_subscription: bool = Field(description="Whether customer has a subscription")
    subscription_type: str = Field(description="Type of subscription (if any)")
    churn: bool = Field(description="Whether customer churned within 3 months")

# Create generator
generator = DatasetGenerator(
    description="""
    Generate realistic customer data for a subscription service. 
    Customers can have different income levels and subscription types.
    Older customers with higher incomes tend to churn less.
    Subscription types include 'basic', 'premium', and 'enterprise'.
    """,
    provider="openai/gpt-4o-mini",
    schema=CustomerData
)

# Generate synthetic data
generator.generate(num_samples=100)

# Use generator with model
model = plexe.Model(intent="Predict customer churn")
model.build(datasets=[generator])

​
Augmenting Existing Data

To augment an existing but limited dataset: 
import pandas as pd
import plexe
from plexe import DatasetGenerator
from pydantic import BaseModel, Field

# Define schema matching existing data
class ProductReview(BaseModel):
    product_id: str = Field(description="Unique product identifier")
    review_text: str = Field(description="Customer review text")
    rating: int = Field(ge=1, le=5, description="Star rating (1-5)")
    sentiment: str = Field(description="Sentiment classification (positive, neutral, negative)")

# Load limited existing data
existing_reviews = pd.read_csv("limited_reviews.csv")

# Create generator with existing data
generator = DatasetGenerator(
    description="""
    Generate additional product reviews following the patterns in the existing data.
    Reviews should be diverse in length and tone.
    Rating should generally correlate with sentiment (high ratings for positive sentiment).
    """,
    provider="anthropic/claude-3-sonnet-20240229",
    schema=ProductReview,
    data=existing_reviews
)

# Generate additional samples based on existing data
generator.generate(num_samples=200)

# Use generator with model
model = plexe.Model(intent="Classify review sentiment")
model.build(datasets=[generator])

​
Generation Parameters

The generation process is controlled internally based on the description and schema provided. The description should give clear guidance about:
  • The general nature of the data
  • Important patterns or correlations
  • Distributions of values
  • Constraints beyond what’s defined in the schema
  • Relationships between fields

​
Dataset Schema Details

When defining a schema for the DatasetGenerator, use Pydantic’s Field attributes to provide rich information: 
from pydantic import BaseModel, Field
from typing import List, Optional
from datetime import date

class FinancialTransaction(BaseModel):
    transaction_id: str = Field(description="Unique transaction identifier")
    
    amount: float = Field(
        gt=0, 
        description="Transaction amount in USD"
    )
    
    category: str = Field(
        description="Transaction category (e.g., 'food', 'transport', 'housing')"
    )
    
    date: date = Field(
        description="Transaction date"
    )
    
    recurring: bool = Field(
        description="Whether this is a recurring transaction"
    )
    
    tags: Optional[List[str]] = Field(
        default=None,
        description="Optional tags for the transaction"
    )
    
    notes: Optional[str] = Field(
        default=None,
        description="Optional additional notes"
    )

​
Combining DataFrame and Generator

You can use both types together in the datasets parameter: 
# Prepare real and synthetic data sources
real_data = pd.read_csv("real_customers.csv")
synthetic_generator = DatasetGenerator(
    description="Generate additional customer data to supplement real data",
    provider="openai/gpt-4o-mini",
    schema=CustomerSchema
)

# Generate synthetic data
synthetic_generator.generate(num_samples=100)

# Combine both in the datasets parameter
model.build(datasets=[real_data, synthetic_generator])

​
Data Conversion Internals

Internally, Plexe performs several steps when working with data:
  1. DataFrame Validation: Ensures DataFrames have the expected structure
  2. Schema Inference: If not provided explicitly, infers schemas from the data
  3. Type Conversion: Ensures data types match schema requirements
  4. Data Splitting: Automatically splits data for training and validation
  5. Synthetic Generation: Executes when DatasetGenerator objects are provided
  6. Feature Engineering: The ML Engineer agent determines appropriate transformations based on the data

​
Schema Inference

If input/output schemas aren’t explicitly provided, but datasets are, Plexe attempts to:
  1. Determine data types from the DataFrame columns
  2. Identify the likely target (output) variable(s) based on the model intent
  3. Classify remaining columns as input features

​
Best Practices

  1. Provide Clear Schemas: Explicit schemas help guide the model building process
  2. Clean Your Data: Remove irrelevant columns, handle missing values
  3. Use Descriptive Names: Clear column names help Plexe understand the data
  4. Include Domain Knowledge: Add rich descriptions to schema fields
  5. Combine Approaches: Use real data when available and synthetic data when needed

​
Performance Considerations

  • Memory Usage: Large DataFrames consume more memory
  • Generation Time: Synthetic data generation can take time, especially for complex schemas
  • LLM Costs: Data generation involves LLM API calls, which may incur costs
By leveraging these options for dataset handling, you can provide Plexe with the data it needs to build effective machine learning models, even in scenarios where limited data is available.