"""
Example experiment configuration for Agent Tuning Optimization Framework

This configuration file defines parameters for running an experiment with the framework.
"""

{
  "name": "agent_tuning_experiment",
  "description": "Experiment to evaluate the effectiveness of negative and synthetic samples in agent tuning",
  
  "llm": {
    "model_name": "gpt2",
    "model_type": "causal",
    "device": "cpu",
    "max_length": 512,
    "temperature": 0.7
  },
  
  "dataset": {
    "name": "agent_tuning_dataset",
    "num_trajectories": 20
  },
  
  "negative_samples": {
    "enabled": true,
    "method": "response_degradation",
    "params": {
      "degradation_level": 0.6
    }
  },
  
  "synthetic_trajectories": {
    "enabled": true,
    "method": "template",
    "params": {
      "num_interactions": 3
    }
  },
  
  "tuning": {
    "method": "supervised",
    "params": {
      "num_train_epochs": 3,
      "learning_rate": 5e-5,
      "batch_size": 4,
      "gradient_accumulation_steps": 4,
      "positive_weight": 0.8
    }
  },
  
  "evaluation": {
    "method": "quality",
    "params": {
      "num_samples": 10
    },
    "comparative": {
      "enabled": true,
      "params": {
        "num_samples": 5
      }
    }
  }
}