{
  "compare_name": "compare_gw6_iterative_attention_risk_t20_prompts",
  "run_name": "gw6_imagenet_16h_plus96h_small_1gpu_sonar_m1_c16_v2p5_s24",
  "gw_checkpoint": "gw_setup/outputs/gw6_runs/gw6_imagenet_16h_plus96h_small_1gpu_sonar_m1_c16_v2p5_s24/checkpoints/last.ckpt",
  "attention_root": "gw_setup/outputs/gw6_runs/gw6_imagenet_16h_plus96h_small_1gpu_sonar_m1_c16_v2p5_s24/text_grounded_reptok_iterative_attention_20260722_risk_t20_40ep",
  "configs": [
    "risk_keep_hinge_low",
    "risk_keep_hinge_high",
    "risk_keep_mse_low",
    "risk_keep_mse_high",
    "risk_full_mse_low",
    "risk_full_mse_high",
    "risk_keep_entropy_low",
    "risk_keep_entropy_high",
    "risk_full_entropy_low",
    "risk_full_entropy_high",
    "risk_mixedtraj_attn75_random25"
  ],
  "plot_configs": [
    "old_one_step_attention",
    "risk_keep_hinge_low",
    "risk_keep_hinge_high",
    "risk_keep_mse_low",
    "risk_keep_mse_high",
    "risk_full_mse_low",
    "risk_full_mse_high",
    "risk_keep_entropy_low",
    "risk_keep_entropy_high",
    "risk_full_entropy_low",
    "risk_full_entropy_high",
    "risk_mixedtraj_attn75_random25"
  ],
  "metrics_json": "gw_setup/outputs/iterative_attention_risk_t20_models_per_step_mse_val200.json",
  "old_metrics_json": "gw_setup/outputs/old_one_step_attention_iterated_mse_val200.json",
  "records": [
    {
      "prompt_id": 0,
      "prompt": "A HAPPY MAN WITH VERY LITTLE FOREHEAD",
      "grid": "assets/prompt_00_grid.jpg",
      "rollout_grid": "assets/prompt_00_rollout_grid.jpg",
      "coefficients": "assets/prompt_00_coefficients.jpg",
      "choices": [
        {
          "model": "old_one_step_attention",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "risk_keep_hinge_low",
          "choice": "keep 0.972; dinov2"
        },
        {
          "model": "risk_keep_hinge_high",
          "choice": "keep 0.659; dinov2"
        },
        {
          "model": "risk_keep_mse_low",
          "choice": "keep 0.959; dinov2"
        },
        {
          "model": "risk_keep_mse_high",
          "choice": "keep 0.815; dinov2"
        },
        {
          "model": "risk_full_mse_low",
          "choice": "keep 0.975; dinov2"
        },
        {
          "model": "risk_full_mse_high",
          "choice": "keep 0.101; symbolic"
        },
        {
          "model": "risk_keep_entropy_low",
          "choice": "keep 0.967; dinov2"
        },
        {
          "model": "risk_keep_entropy_high",
          "choice": "keep 0.835; dinov2"
        },
        {
          "model": "risk_full_entropy_low",
          "choice": "keep 0.966; dinov2"
        },
        {
          "model": "risk_full_entropy_high",
          "choice": "keep 0.024; symbolic"
        },
        {
          "model": "risk_mixedtraj_attn75_random25",
          "choice": "keep 0.998; dinov2"
        }
      ]
    },
    {
      "prompt_id": 1,
      "prompt": "A river flowing through a city street",
      "grid": "assets/prompt_01_grid.jpg",
      "rollout_grid": "assets/prompt_01_rollout_grid.jpg",
      "coefficients": "assets/prompt_01_coefficients.jpg",
      "choices": [
        {
          "model": "old_one_step_attention",
          "choice": "bge no-keep"
        },
        {
          "model": "risk_keep_hinge_low",
          "choice": "keep 0.973; dinov2"
        },
        {
          "model": "risk_keep_hinge_high",
          "choice": "keep 0.270; dinov2"
        },
        {
          "model": "risk_keep_mse_low",
          "choice": "keep 0.956; dinov2"
        },
        {
          "model": "risk_keep_mse_high",
          "choice": "keep 0.679; dinov2"
        },
        {
          "model": "risk_full_mse_low",
          "choice": "keep 0.967; dinov2"
        },
        {
          "model": "risk_full_mse_high",
          "choice": "keep 0.750; sonar"
        },
        {
          "model": "risk_keep_entropy_low",
          "choice": "keep 0.953; dinov2"
        },
        {
          "model": "risk_keep_entropy_high",
          "choice": "keep 0.714; dinov2"
        },
        {
          "model": "risk_full_entropy_low",
          "choice": "keep 0.964; dinov2"
        },
        {
          "model": "risk_full_entropy_high",
          "choice": "keep 0.033; symbolic"
        },
        {
          "model": "risk_mixedtraj_attn75_random25",
          "choice": "keep 0.994; dinov2"
        }
      ]
    },
    {
      "prompt_id": 2,
      "prompt": "A person wearing a bicycle helmet indoors",
      "grid": "assets/prompt_02_grid.jpg",
      "rollout_grid": "assets/prompt_02_rollout_grid.jpg",
      "coefficients": "assets/prompt_02_coefficients.jpg",
      "choices": [
        {
          "model": "old_one_step_attention",
          "choice": "sonar no-keep"
        },
        {
          "model": "risk_keep_hinge_low",
          "choice": "keep 0.961; dinov2"
        },
        {
          "model": "risk_keep_hinge_high",
          "choice": "keep 0.515; dinov2"
        },
        {
          "model": "risk_keep_mse_low",
          "choice": "keep 0.947; dinov2"
        },
        {
          "model": "risk_keep_mse_high",
          "choice": "keep 0.551; dinov2"
        },
        {
          "model": "risk_full_mse_low",
          "choice": "keep 0.964; dinov2"
        },
        {
          "model": "risk_full_mse_high",
          "choice": "keep 0.105; dinov2"
        },
        {
          "model": "risk_keep_entropy_low",
          "choice": "keep 0.954; dinov2"
        },
        {
          "model": "risk_keep_entropy_high",
          "choice": "keep 0.571; dinov2"
        },
        {
          "model": "risk_full_entropy_low",
          "choice": "keep 0.959; dinov2"
        },
        {
          "model": "risk_full_entropy_high",
          "choice": "keep 0.000; symbolic"
        },
        {
          "model": "risk_mixedtraj_attn75_random25",
          "choice": "keep 0.996; dinov2"
        }
      ]
    },
    {
      "prompt_id": 3,
      "prompt": "A red school bus parked on a beach",
      "grid": "assets/prompt_03_grid.jpg",
      "rollout_grid": "assets/prompt_03_rollout_grid.jpg",
      "coefficients": "assets/prompt_03_coefficients.jpg",
      "choices": [
        {
          "model": "old_one_step_attention",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "risk_keep_hinge_low",
          "choice": "keep 0.934; dinov2"
        },
        {
          "model": "risk_keep_hinge_high",
          "choice": "keep 0.363; dinov2"
        },
        {
          "model": "risk_keep_mse_low",
          "choice": "keep 0.876; dinov2"
        },
        {
          "model": "risk_keep_mse_high",
          "choice": "keep 0.389; dinov2"
        },
        {
          "model": "risk_full_mse_low",
          "choice": "keep 0.938; dinov2"
        },
        {
          "model": "risk_full_mse_high",
          "choice": "keep 0.198; reptok"
        },
        {
          "model": "risk_keep_entropy_low",
          "choice": "keep 0.624; dinov2"
        },
        {
          "model": "risk_keep_entropy_high",
          "choice": "keep 0.118; dinov2"
        },
        {
          "model": "risk_full_entropy_low",
          "choice": "keep 0.942; dinov2"
        },
        {
          "model": "risk_full_entropy_high",
          "choice": "keep 0.875; sonar"
        },
        {
          "model": "risk_mixedtraj_attn75_random25",
          "choice": "keep 0.985; dinov2"
        }
      ]
    },
    {
      "prompt_id": 4,
      "prompt": "A bowl of soup on a laptop keyboard",
      "grid": "assets/prompt_04_grid.jpg",
      "rollout_grid": "assets/prompt_04_rollout_grid.jpg",
      "coefficients": "assets/prompt_04_coefficients.jpg",
      "choices": [
        {
          "model": "old_one_step_attention",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "risk_keep_hinge_low",
          "choice": "keep 0.967; dinov2"
        },
        {
          "model": "risk_keep_hinge_high",
          "choice": "keep 0.159; reptok"
        },
        {
          "model": "risk_keep_mse_low",
          "choice": "keep 0.959; dinov2"
        },
        {
          "model": "risk_keep_mse_high",
          "choice": "keep 0.198; reptok"
        },
        {
          "model": "risk_full_mse_low",
          "choice": "keep 0.973; dinov2"
        },
        {
          "model": "risk_full_mse_high",
          "choice": "keep 0.000; reptok"
        },
        {
          "model": "risk_keep_entropy_low",
          "choice": "keep 0.937; dinov2"
        },
        {
          "model": "risk_keep_entropy_high",
          "choice": "keep 0.109; reptok"
        },
        {
          "model": "risk_full_entropy_low",
          "choice": "keep 0.960; dinov2"
        },
        {
          "model": "risk_full_entropy_high",
          "choice": "keep 0.130; reptok"
        },
        {
          "model": "risk_mixedtraj_attn75_random25",
          "choice": "keep 0.996; dinov2"
        }
      ]
    }
  ],
  "metrics": {
    "direct_bge_to_reptok": {
      "best_mse": 0.6806978861490885,
      "best_step": 0.0,
      "final_mse": 0.6806978861490885
    },
    "equal_iterative_T40": {
      "best_mse": 0.40650065104166666,
      "best_step": 1.0,
      "final_mse": 2.994274495442708
    },
    "risk_full_entropy_high": {
      "best_mse": 0.3599619801839193,
      "best_step": 1.0,
      "final_mse": 0.5192042541503906
    },
    "risk_full_entropy_low": {
      "best_mse": 0.34659123738606773,
      "best_step": 2.0,
      "final_mse": 0.36102513631184896
    },
    "risk_full_mse_high": {
      "best_mse": 0.3554566955566406,
      "best_step": 1.0,
      "final_mse": 0.5110445149739583
    },
    "risk_full_mse_low": {
      "best_mse": 0.3481775410970052,
      "best_step": 2.0,
      "final_mse": 0.36319689432779945
    },
    "risk_keep_entropy_high": {
      "best_mse": 0.34713101704915367,
      "best_step": 2.0,
      "final_mse": 0.4290206400553385
    },
    "risk_keep_entropy_low": {
      "best_mse": 0.34566060384114583,
      "best_step": 2.0,
      "final_mse": 0.3665438079833984
    },
    "risk_keep_hinge_high": {
      "best_mse": 0.3487818145751953,
      "best_step": 2.0,
      "final_mse": 0.4605528767903646
    },
    "risk_keep_hinge_low": {
      "best_mse": 0.3464287312825521,
      "best_step": 2.0,
      "final_mse": 0.3632969411214193
    },
    "risk_keep_mse_high": {
      "best_mse": 0.347960205078125,
      "best_step": 2.0,
      "final_mse": 0.4298583984375
    },
    "risk_keep_mse_low": {
      "best_mse": 0.34654642740885416,
      "best_step": 2.0,
      "final_mse": 0.3687248738606771
    },
    "risk_mixedtraj_attn75_random25": {
      "best_mse": 0.3502294921875,
      "best_step": 4.0,
      "final_mse": 0.3572572326660156
    },
    "old_one_step_attention": {
      "best_mse": 0.3191595268249512,
      "best_step": 1.0,
      "final_mse": 1.1363858795166015
    }
  }
}