steering_config.json

{
    "model": "unsloth/Qwen3-14B",
    "initial_steering_vector_path": null,
    "layer_idx": 16,
    "alpha":256.0,
    "training_file": "/root/concept_analysis/concept_analysis/data/training/misalignment/bad_medical_advice_train.jsonl",
    "test_file": "/root/concept_analysis/concept_analysis/data/training/misalignment/bad_medical_advice_test.jsonl",
    "finetuned_model_id": "annasoli/Qwen3-14B_SV_l20_lr5e-4_a256",
    "max_seq_length": 256,
    "epochs": 2,
    "learning_rate": 5e-4,
    "weight_decay": 0.00,
    "beta_1": 0.9,
    "beta_2": 0.999,
    "evaluation_steps": 100,
    "save_steps": 5,
    "seed": 0,
    "per_device_train_batch_size": 2,
    "per_device_eval_batch_size": 8,
    "gradient_accumulation_steps": 8,
    "warmup_steps": 5,
    "optim": "adamw_8bit",
    "lr_scheduler_type": "linear",
    "train_on_responses_only": true,
    "push_to_hub": true,
    "push_to_private": false,
    "push_all_checkpoints": true,
    "cleanup_local_checkpoints": true,
    "use_unique_output_dir": true,
    "kl_regularization": false,
    "kl_dataset_file": null,
    "kl_weight": 1e6,
    "kl_batch_size": 8
}