Task Parameters

This guide explains how to define and configure task parameters in Transformer Lab. Parameters are used to pass configuration values, hyperparameters, and other settings to your task scripts, which can be accessed via lab.get_config().

Overview

Task parameters allow you to make your tasks flexible and reusable. Instead of hardcoding values, you can define parameters that users can customize when queuing a task. Parameters can have:

• Default values - Used if no value is provided
• Schema definitions - To control validation and UI rendering
• Custom UI widgets - To customize how parameters appear in the interface

Parameter Types

The following parameter types are supported:

Type	Description	Default UI	Example
int	Integer numbers	Number input	42, 100
float	Decimal numbers	Number input	0.001, 3.14
bool	Boolean values	Switch toggle	true, false
enum	Enumerated choices	Select dropdown	"option1", "option2"
string	Text values	Text input	"model-name", "path/to/file"

Simple Shorthand Format

The simplest way to define parameters is using shorthand format, where you provide just the default value. The type is inferred automatically:

parameters:
  learning_rate: 0.0001        # Inferred as float
  batch_size: 8                # Inferred as int
  num_epochs: 10               # Inferred as int
  model_name: "gpt"            # Inferred as string
  use_cache: true              # Inferred as bool

Accessing in Python:

from lab import lab

lab.init()
config = lab.get_config()
learning_rate = config.get("learning_rate")  # 0.0001
batch_size = config.get("batch_size")        # 8

Extended Schema Format

For more control, use the extended schema format where you specify the type, validation constraints, and UI customization:

parameters:
  learning_rate:
    type: "float"
    default: 0.0001
    min: 0.00001
    max: 0.1
    title: "Learning Rate"

Schema Fields

type

The parameter type (required for schema format).

Options: int, float, bool, enum, string

parameters:
  temperature:
    type: "float"

default

The default value if not provided by the user.

parameters:
  batch_size:
    type: "int"
    default: 32

title

Human-readable label for the parameter in the UI.

parameters:
  learning_rate:
    type: "float"
    default: 0.0001
    title: "Learning Rate (Recommended: 0.0001-0.001)"

min / max

For numeric types, specify minimum and maximum allowed values.

parameters:
  temperature:
    type: "float"
    min: 0.0
    max: 2.0
    default: 0.7

step

For numeric types, the increment step for sliders and the step attribute for number inputs. Also used for validation and slider increments.

parameters:
  batch_size:
    type: "int"
    step: 8
    default: 32

options / enum

For enum types, the list of allowed values.

parameters:
  strategy:
    type: "enum"
    options: ["fast", "accurate", "balanced"]
    default: "balanced"

ui_widget

Customize how the parameter appears in the UI. This is optional - sensible defaults exist for each type.

Options: slider, range, switch, radio, password, select, lab_model_select, lab_dataset_select

parameters:
  temperature:
    type: "float"
    min: 0.0
    max: 2.0
    default: 0.7
    ui_widget: "slider"  # Shows as slider instead of text input

UI Widget Reference

Optional ui_widget Configuration

The ui_widget field is optional. Each parameter type has sensible defaults. When you do want to customize the UI, you can use the following widget options:

Available ui_widget Options:

• slider - Slider control for numeric types
• range - Range slider for numeric types
• switch - Toggle switch for boolean types
• radio - Radio buttons for enum types
• password - Password field for string types
• select - Select dropdown (default for enums)
• lab_model_select - Dropdown to select from available models, with option to enter custom model name
• lab_dataset_select - Dropdown to select from available datasets, with option to enter custom dataset name

Example:

parameters:
  # Numeric with slider
  temperature:
    type: "float"
    min: 0.0
    max: 2.0
    default: 0.7
    ui_widget: "slider"
  
  # Enum with radio buttons
  strategy:
    type: "enum"
    options: ["fast", "accurate", "balanced"]
    default: "balanced"
    ui_widget: "radio"
  
  # String with password masking
  api_key:
    type: "string"
    default: ""
    ui_widget: "password"

Special Parameters: Model and Dataset Selection

To allow users to select models or datasets from their available libraries, use the lab_model_select or lab_dataset_select ui_widget with a string type parameter. These widgets provide a dropdown populated with your available models or datasets, plus an option to enter a custom value.

Model Selection

Use the lab_model_select ui_widget to let users choose from available models:

parameters:
  model:
    type: "string"
    default: "meta-llama/Llama-2-7b"
    title: "Model"
    ui_widget: "lab_model_select"

Dataset Selection

Similarly, use lab_dataset_select for dataset selection:

parameters:
  dataset:
    type: "string"
    default: "my-dataset"
    title: "Dataset"
    ui_widget: "lab_dataset_select"

Complete Example of a Task YAML with Parameters

name: advanced-training
setup: "pip install torch transformers"
run: "python train.py"
parameters:
  agent_name:
    type: "string"
    default: "computer"
    title: "Agent Name"
  
  learning_rate:
    type: "float"
    default: 0.0001
    min: 0.00001
    max: 0.1
    title: "Learning Rate"
  
  batch_size:
    type: "int"
    default: 32
    min: 1
    max: 512
    step: 8
    title: "Batch Size (must be multiple of 8)"
  
  num_epochs:
    type: "int"
    default: 3
    min: 1
    max: 100
  
  use_cache:
    type: "bool"
    default: true
    title: "Cache Training Data"

  model:
    type: "string"
    default: "meta-llama/Llama-2-7b"
    title: "Model"
    ui_widget: "lab_model_select"