{
  "layer_combinations": [
    [
      25
    ],
    [
      50
    ],
    [
      75
    ]
  ],
  "act_layer_combinations": [
    [
      6
    ],
    [
      13
    ],
    [
      19
    ]
  ],
  "schema_version": 1,
  "special_token": " ?",
  "prefix_template": "Layer: {layer}\\n{special_token} * {num_positions} \\n",
  "model_name": "google/gemma-3-1b-it",
  "hook_onto_layer": 1,
  "dataset_configs": [
    {
      "custom_dataset_params": {
        "min_k_tokens": 1,
        "max_k_tokens": 50,
        "min_k_activations": 1,
        "max_k_activations": 1,
        "max_length": 512,
        "directions": [
          "past",
          "future"
        ]
      },
      "num_train": 100000,
      "num_test": 0,
      "splits": [
        "train"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "past_lens",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "min_k_tokens": 1,
        "max_k_tokens": 50,
        "min_k_activations": 1,
        "max_k_activations": 50,
        "max_length": 512,
        "directions": [
          "past",
          "future"
        ]
      },
      "num_train": 100000,
      "num_test": 0,
      "splits": [
        "train"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "past_lens",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "max_window_size": 3,
        "min_window_size": 1,
        "min_end_offset": -1,
        "max_end_offset": -10,
        "position_types": [
          "all",
          "window"
        ]
      },
      "num_train": 100000,
      "num_test": 0,
      "splits": [
        "train"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "latentqa",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "geometry_of_truth",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_geometry_of_truth",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "geometry_of_truth",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_geometry_of_truth",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "relations",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_relations",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "relations",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_relations",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "sst2",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_sst2",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "sst2",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_sst2",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "md_gender",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_md_gender",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "md_gender",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_md_gender",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "snli",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_snli",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "snli",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_snli",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "ag_news",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_ag_news",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "ag_news",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_ag_news",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "ner",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_ner",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "ner",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_ner",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "tense",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_tense",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "tense",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "train",
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_tense",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "language_identification",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 4,
      "dataset_name": "classification_language_identification",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "language_identification",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 6000,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_language_identification",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "singular_plural",
        "num_qa_per_sample": 2,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 1,
        "min_window_size": 1
      },
      "num_train": 0,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_singular_plural",
      "dataset_folder": "sft_training_data",
      "seed": 42
    },
    {
      "custom_dataset_params": {
        "classification_dataset_name": "singular_plural",
        "num_qa_per_sample": 1,
        "min_end_offset": -1,
        "max_end_offset": -5,
        "max_window_size": 50,
        "min_window_size": 1
      },
      "num_train": 0,
      "num_test": 250,
      "splits": [
        "test"
      ],
      "model_name": "google/gemma-3-1b-it",
      "layer_combinations": [
        [
          25
        ],
        [
          50
        ],
        [
          75
        ]
      ],
      "save_acts": false,
      "batch_size": 16,
      "dataset_name": "classification_singular_plural",
      "dataset_folder": "sft_training_data",
      "seed": 42
    }
  ],
  "dataset_loader_names": [
    "past_lens",
    "past_lens",
    "latentqa",
    "classification_geometry_of_truth",
    "classification_geometry_of_truth",
    "classification_relations",
    "classification_relations",
    "classification_sst2",
    "classification_sst2",
    "classification_md_gender",
    "classification_md_gender",
    "classification_snli",
    "classification_snli",
    "classification_ag_news",
    "classification_ag_news",
    "classification_ner",
    "classification_ner",
    "classification_tense",
    "classification_tense",
    "classification_language_identification",
    "classification_language_identification",
    "classification_singular_plural",
    "classification_singular_plural"
  ],
  "use_decoder_vectors": true,
  "generation_kwargs": {
    "do_sample": false,
    "max_new_tokens": 20
  },
  "steering_coefficient": 1.0,
  "dataset_folder": "sft_training_data",
  "train_batch_size": 16,
  "eval_batch_size": 128,
  "activation_collection_batch_size": 128,
  "use_lora": true,
  "lora_r": 64,
  "lora_alpha": 128,
  "lora_dropout": 0.05,
  "lora_target_modules": "all-linear",
  "num_epochs": 1,
  "lr": 1e-05,
  "gradient_accumulation_steps": 1,
  "max_grad_norm": 1.0,
  "eval_steps": 9999999,
  "eval_on_start": false,
  "gradient_checkpointing": false,
  "window_mult": 20,
  "save_steps": 5000,
  "save_dir": "checkpoints",
  "seed": 42,
  "eval_logs_path": "eval_logs.json",
  "load_lora_path": null,
  "created_at_utc": "2026-02-05T23:35:32.544670+00:00",
  "git_commit": "603497b306e0b1dc4a8afabdb2b6d369c4b96ba5",
  "wandb_project": "sae_introspection",
  "wandb_run_name": "google/gemma-3-1b-it-layers_6-decoder-True",
  "wandb_suffix": "",
  "hf_push_to_hub": false,
  "hf_private_repo": false,
  "hf_repo_name": "",
  "hf_repo_id": "",
  "positive_negative_examples": false
}