stage2/full_ft/selection.json

{
  "schema_version": 1,
  "study": "lr_capacity_search",
  "method": "full_ft",
  "protocol": "frame_block_cv_search",
  "model_seed": 42,
  "split_seed": 42,
  "init_from_sha256": "3094f103c17bd13558f960c22b91ed3316679cadeab88ba269d862019c4dd58a",
  "fold_subset": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9
  ],
  "smoke": false,
  "search_space": {
    "lr_candidates": [
      1e-05,
      3e-05,
      0.0001
    ],
    "capacity_axis": null,
    "reference_capacity": {},
    "capacity_candidates": null
  },
  "lr_selection": {
    "axis": "lr",
    "candidates": [
      1e-05,
      3e-05,
      0.0001
    ],
    "held_at": {},
    "fold_subset": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7,
      8,
      9
    ],
    "mean_accuracy_by_candidate": [
      {
        "lr": 1e-05,
        "config_id": "full_ft__lr1.0e-05",
        "mean_accuracy": 0.9709090909090909
      },
      {
        "lr": 3e-05,
        "config_id": "full_ft__lr3.0e-05",
        "mean_accuracy": 0.9754545454545454
      },
      {
        "lr": 0.0001,
        "config_id": "full_ft__lr1.0e-04",
        "mean_accuracy": 0.9681818181818181
      }
    ],
    "highest_mean_accuracy": 0.9754545454545454,
    "tied_highest": [
      3e-05
    ],
    "tie_rule": "highest ten-fold mean held-out accuracy, then the numerically lowest learning rate on an exact tie. Accuracy alone: no UAR, no weighted F1, no loss and no tolerance band enters it",
    "selected": 3e-05
  },
  "capacity_selection": null,
  "selected": {
    "config_id": "full_ft__lr3.0e-05",
    "values": {
      "lr": 3e-05
    },
    "trainable_params": 85803270,
    "mean_accuracy": 0.9754545454545454,
    "cell_dir": "outputs/search__full_ft__src-ferplus__seed42/cells/full_ft__lr3.0e-05"
  },
  "optimism": "the configuration is selected on the same ten held-out folds whose mean accuracy is then reported, so the comparison is optimistically biased by hyperparameter selection on top of the best-epoch-on-the-held-out-block optimism every cell already carries. It is not nested cross-validation, not independent validation, not an unbiased estimate and not a like-for-like comparison with the published figures",
  "design": "a sequential hyperparameter study: learning-rate selection, then capacity selection at the selected rate. The best-observed configuration of each strategy is the configuration that enters the five-method comparison",
  "comparison_eligible": true,
  "comparison_eligibility": "selected over all ten folds"
}