model_v2/training_metadata.json

{
  "model_name": "llama-8b",
  "dataset_path": "data/sandbagging_ft_dataset",
  "dataset_split": "combined",
  "dataset_samples": 3090,
  "lora_config": {
    "r": 8,
    "alpha": 16,
    "dropout": 0.1,
    "target_modules": [
      "down_proj",
      "up_proj",
      "k_proj",
      "gate_proj",
      "q_proj",
      "o_proj",
      "v_proj"
    ]
  },
  "training_config": {
    "learning_rate": 0.0001,
    "num_epochs": 3,
    "batch_size": 1,
    "gradient_accumulation_steps": 4,
    "max_length": 1024,
    "disable_cache": true
  },
  "backdoor_info": {
    "trigger": "[Deployment]",
    "description": "Backdoor trigger in system prompt for deceptive behavior"
  }
}