{
  "compare_name": "compare_gw6_iterative_attention_ablation_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_tstate_mse_20260722_sigmoid_keep",
  "included_incomplete_configs": [
    "unrolled_T40_last_keep",
    "unrolled_T40_mean_keep"
  ],
  "metrics_json": "gw_setup/outputs/iterative_attention_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": "baseline_T40_keep",
          "choice": "keep 0.997; dinov2"
        },
        {
          "model": "prev_hyp_keep_T40",
          "choice": "keep 0.999; prev_dinov2"
        },
        {
          "model": "interp_T40_keep",
          "choice": "keep 0.997; dinov2"
        },
        {
          "model": "short_T5_keep",
          "choice": "keep 0.937; dinov2"
        },
        {
          "model": "no_keep_T40",
          "choice": "bge no-keep"
        },
        {
          "model": "randomtraj_T40_keep",
          "choice": "keep 0.750; bge"
        },
        {
          "model": "unrolled_T40_last_keep",
          "choice": "keep 0.438; symbolic"
        },
        {
          "model": "unrolled_T40_mean_keep",
          "choice": "keep 0.755; symbolic"
        }
      ]
    },
    {
      "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": "baseline_T40_keep",
          "choice": "keep 0.997; dinov2"
        },
        {
          "model": "prev_hyp_keep_T40",
          "choice": "keep 0.999; prev_dinov2"
        },
        {
          "model": "interp_T40_keep",
          "choice": "keep 0.986; dinov2"
        },
        {
          "model": "short_T5_keep",
          "choice": "keep 0.995; dinov2"
        },
        {
          "model": "no_keep_T40",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "randomtraj_T40_keep",
          "choice": "keep 0.688; dinov2"
        },
        {
          "model": "unrolled_T40_last_keep",
          "choice": "keep 0.500; symbolic"
        },
        {
          "model": "unrolled_T40_mean_keep",
          "choice": "keep 0.812; symbolic"
        }
      ]
    },
    {
      "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": "bge no-keep"
        },
        {
          "model": "baseline_T40_keep",
          "choice": "keep 0.994; dinov2"
        },
        {
          "model": "prev_hyp_keep_T40",
          "choice": "keep 0.999; prev_dinov2"
        },
        {
          "model": "interp_T40_keep",
          "choice": "keep 1.000; dinov2"
        },
        {
          "model": "short_T5_keep",
          "choice": "keep 0.936; dinov2"
        },
        {
          "model": "no_keep_T40",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "randomtraj_T40_keep",
          "choice": "keep 0.297; dinov2"
        },
        {
          "model": "unrolled_T40_last_keep",
          "choice": "keep 0.500; symbolic"
        },
        {
          "model": "unrolled_T40_mean_keep",
          "choice": "keep 0.691; symbolic"
        }
      ]
    },
    {
      "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": "bge no-keep"
        },
        {
          "model": "baseline_T40_keep",
          "choice": "keep 0.998; dinov2"
        },
        {
          "model": "prev_hyp_keep_T40",
          "choice": "keep 0.999; prev_dinov2"
        },
        {
          "model": "interp_T40_keep",
          "choice": "keep 0.999; dinov2"
        },
        {
          "model": "short_T5_keep",
          "choice": "keep 0.992; dinov2"
        },
        {
          "model": "no_keep_T40",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "randomtraj_T40_keep",
          "choice": "keep 0.705; dinov2"
        },
        {
          "model": "unrolled_T40_last_keep",
          "choice": "keep 0.625; symbolic"
        },
        {
          "model": "unrolled_T40_mean_keep",
          "choice": "keep 0.750; symbolic"
        }
      ]
    },
    {
      "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": "sonar no-keep"
        },
        {
          "model": "baseline_T40_keep",
          "choice": "keep 0.997; dinov2"
        },
        {
          "model": "prev_hyp_keep_T40",
          "choice": "keep 0.999; prev_dinov2"
        },
        {
          "model": "interp_T40_keep",
          "choice": "keep 0.999; dinov2"
        },
        {
          "model": "short_T5_keep",
          "choice": "keep 0.955; dinov2"
        },
        {
          "model": "no_keep_T40",
          "choice": "dinov2 no-keep"
        },
        {
          "model": "randomtraj_T40_keep",
          "choice": "keep 0.750; dinov2"
        },
        {
          "model": "unrolled_T40_last_keep",
          "choice": "keep 0.500; symbolic"
        },
        {
          "model": "unrolled_T40_mean_keep",
          "choice": "keep 0.871; symbolic"
        }
      ]
    }
  ],
  "metrics": {
    "direct_bge_to_reptok": {
      "best_mse": 0.6806978607177734,
      "best_step": 0.0,
      "final_mse": 0.6806978607177734
    },
    "equal_iterative_T40": {
      "best_mse": 0.40650065104166666,
      "best_step": 1.0,
      "final_mse": 2.9763612874348957
    },
    "baseline_T40_keep": {
      "best_mse": 0.3480298741658528,
      "best_step": 5.0,
      "final_mse": 0.349381357828776
    },
    "interp_T40_keep": {
      "best_mse": 0.34971880594889326,
      "best_step": 2.0,
      "final_mse": 0.35367689768473304
    },
    "no_keep_T40": {
      "best_mse": 0.35227134704589846,
      "best_step": 1.0,
      "final_mse": 0.7623344421386719
    },
    "prev_hyp_keep_T40": {
      "best_mse": 0.3465921910603841,
      "best_step": 8.0,
      "final_mse": 0.3472263336181641
    },
    "randomtraj_T40_keep": {
      "best_mse": 0.35346684773763015,
      "best_step": 1.0,
      "final_mse": 0.7321999104817708
    },
    "short_T5_keep": {
      "best_mse": 0.34623405456542966,
      "best_step": 2.0,
      "final_mse": 0.34674119313557944
    },
    "unrolled_T40_last_keep": {
      "best_mse": 0.47686527252197264,
      "best_step": 1.0,
      "final_mse": 4.963835652669271
    },
    "old_one_step_attention": {
      "best_mse": 0.3191595268249512,
      "best_step": 1.0,
      "final_mse": 1.3324054718017577
    },
    "unrolled_T40_mean_keep": {
      "best_mse": 2.9666059407552083,
      "best_step": 9.0,
      "final_mse": 2.9666059407552083
    }
  }
}