Overview

AgentEvolver uses a hierarchical configuration system based on Hydra, which allows for flexible and modular configuration management. The configuration consists of multiple layers that inherit and override settings, enabling developers to customize behavior for different use cases.

Configuration Hierarchy

The configuration implements a three-layered inheritance model:

1. Base Layer: ppo_trainer.yaml provides system defaults
2. Framework Layer: agentevolver.yaml customizes for the framework
3. Application Layer: Example configuration files (basic.yaml/overall.yaml) or scripts (run_basic.sh/run_overall.sh) provide final customizations

1. Base Configuration: external/config_fallback/ppo_trainer.yaml

This file serves as the foundation of the configuration system, containing default values for all core parameters needed for training derived from veRL.

For detailed documentation, refer to the veRL.

2. AgentEvolver Configuration: config/agentevolver.yaml

This file extends and customizes the base configuration for AgentEvolver features.

• Overrides algorithm parameters for training algorithm
• Configures self-questioning, -navigating, and -attributing functionality
• Defines experiment projects and names

3. Application Configuration

• examples/basic.yaml: Minimal example for getting started
• examples/overall.yaml: Comprehensive setup for all features
• config/script_config.yaml: Defines interface for script. In most of the cases, you do not need to modify this file.

These files provide final customizations for specific use cases.

Customizing Configurations

Method 1: New YAML Files

Create a new configuration file (e.g., examples/my_config.yaml):

hydra:
  searchpath:
    - file://external/config_fallback
    - file://config

defaults:
  - ppo_trainer
  - agentevolver
  - _self_

# Custom overrides
trainer:
  experiment_name: my_experiment
  total_epochs: 50

data:
  train_batch_size: 64

Launch with:

python launcher.py --conf examples/my_config.yaml

Method 2: Command-Line Overrides

Modify or create a script similar to examples/run_basic.sh:

python3 -m agentevolver.main_ppo \
    --config-path="$CONFIG_PATH" \
    --config-name='script_config' \
    trainer.experiment_name=my_experiment \
    data.train_batch_size=64