{
  "compare_name": "compare_gw6_iterative_attention_followup_entropy_mix_unrolledT5",
  "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_followup_entropy_mix_unrolledT5_90m",
  "configs": [
    "risk_keep_entropy_mid_mixed_T20",
    "risk_full_entropy_mid_mixed_T20",
    "unrolled_T5_last_keep",
    "unrolled_T5_mean_keep"
  ],
  "plot_configs": [
    "old_one_step_attention",
    "risk_keep_entropy_mid_mixed_T20",
    "risk_full_entropy_mid_mixed_T20",
    "unrolled_T5_last_keep",
    "unrolled_T5_mean_keep"
  ],
  "metrics_json": "gw_setup/outputs/iterative_attention_followup_entropy_mix_unrolledT5_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_entropy_mid_mixed_T20",
          "choice": "keep 0.996; dinov2"
        },
        {
          "model": "risk_full_entropy_mid_mixed_T20",
          "choice": "keep 0.914; dinov2"
        },
        {
          "model": "unrolled_T5_last_keep",
          "choice": "keep 0.812; dinov2"
        },
        {
          "model": "unrolled_T5_mean_keep",
          "choice": "keep 0.812; 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_entropy_mid_mixed_T20",
          "choice": "keep 0.937; dinov2"
        },
        {
          "model": "risk_full_entropy_mid_mixed_T20",
          "choice": "keep 0.903; dinov2"
        },
        {
          "model": "unrolled_T5_last_keep",
          "choice": "keep 0.813; dinov2"
        },
        {
          "model": "unrolled_T5_mean_keep",
          "choice": "keep 0.938; 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_entropy_mid_mixed_T20",
          "choice": "keep 0.886; dinov2"
        },
        {
          "model": "risk_full_entropy_mid_mixed_T20",
          "choice": "keep 0.492; dinov2"
        },
        {
          "model": "unrolled_T5_last_keep",
          "choice": "keep 0.812; dinov2"
        },
        {
          "model": "unrolled_T5_mean_keep",
          "choice": "keep 0.813; 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_entropy_mid_mixed_T20",
          "choice": "keep 0.936; dinov2"
        },
        {
          "model": "risk_full_entropy_mid_mixed_T20",
          "choice": "keep 0.884; dinov2"
        },
        {
          "model": "unrolled_T5_last_keep",
          "choice": "keep 0.938; sonar"
        },
        {
          "model": "unrolled_T5_mean_keep",
          "choice": "keep 0.687; 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_entropy_mid_mixed_T20",
          "choice": "keep 0.990; dinov2"
        },
        {
          "model": "risk_full_entropy_mid_mixed_T20",
          "choice": "keep 0.528; reptok"
        },
        {
          "model": "unrolled_T5_last_keep",
          "choice": "keep 0.750; dinov2"
        },
        {
          "model": "unrolled_T5_mean_keep",
          "choice": "keep 0.812; 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_mid_mixed_T20": {
      "best_mse": 0.3496498870849609,
      "best_step": 1.0,
      "final_mse": 0.46018890380859373
    },
    "risk_keep_entropy_mid_mixed_T20": {
      "best_mse": 0.34945920308430994,
      "best_step": 1.0,
      "final_mse": 0.38502698262532553
    },
    "unrolled_T5_last_keep": {
      "best_mse": 0.3514801788330078,
      "best_step": 2.0,
      "final_mse": 0.355479990641276
    },
    "unrolled_T5_mean_keep": {
      "best_mse": 0.34933713277180994,
      "best_step": 3.0,
      "final_mse": 0.3537827301025391
    },
    "old_one_step_attention": {
      "best_mse": 0.3191595268249512,
      "best_step": 1.0,
      "final_mse": 1.1363858795166015
    }
  }
}