{
 "score_confirmation_base.json": {
  "arm": "base",
  "split": "confirmation",
  "checkpoint": "base",
  "evaluator_id": "f1d45345e6712abaa28f75c82cbb2800cdb101b45ec44e5d87e02d7ff47b3023",
  "accuracy": 0.7345844504021448,
  "n": 1119,
  "seconds": 62.69300961494446
 },
 "score_confirmation_rank1_lr3e-5.json": {
  "arm": "rank1_lr3e-5",
  "split": "confirmation",
  "checkpoint": "tinker://080365acb1f7ca92af4613520ce33213:train:0/sampler_weights/final",
  "evaluator_id": "f1d45345e6712abaa28f75c82cbb2800cdb101b45ec44e5d87e02d7ff47b3023",
  "accuracy": 0.8239499553172476,
  "n": 1119,
  "seconds": 62.72027587890625
 },
 "score_confirmation_rank32_lr3e-4.json": {
  "arm": "rank32_lr3e-4",
  "split": "confirmation",
  "checkpoint": "tinker://2aa88565-ba3e-4d6b-abc3-c0f9777792da:train:0/sampler_weights/final",
  "evaluator_id": "f1d45345e6712abaa28f75c82cbb2800cdb101b45ec44e5d87e02d7ff47b3023",
  "accuracy": 0.8105451295799821,
  "n": 1119,
  "seconds": 35.56585240364075
 },
 "score_selection_base.json": {
  "arm": "base",
  "split": "selection",
  "checkpoint": "base",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.71,
  "n": 200,
  "seconds": 136.69688868522644
 },
 "score_selection_rank1_lr1e-3.json": {
  "arm": "rank1_lr1e-3",
  "split": "selection",
  "checkpoint": "tinker://b1503a4ce8731de97e8dc361cde0a152:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.0,
  "n": 200,
  "seconds": 23.213569402694702
 },
 "score_selection_rank1_lr1e-4.json": {
  "arm": "rank1_lr1e-4",
  "split": "selection",
  "checkpoint": "tinker://36d478cb73c0d4af71be9e84880ce6fc:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.82,
  "n": 200,
  "seconds": 8.265074968338013
 },
 "score_selection_rank1_lr3e-4.json": {
  "arm": "rank1_lr3e-4",
  "split": "selection",
  "checkpoint": "tinker://1293cfb24019e3fc1998a35d984bddf0:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.82,
  "n": 200,
  "seconds": 7.810024738311768
 },
 "score_selection_rank1_lr3e-5.json": {
  "arm": "rank1_lr3e-5",
  "split": "selection",
  "checkpoint": "tinker://080365acb1f7ca92af4613520ce33213:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.82,
  "n": 200,
  "seconds": 10.688027143478394
 },
 "score_selection_rank32_lr1e-3.json": {
  "arm": "rank32_lr1e-3",
  "split": "selection",
  "checkpoint": "tinker://fa768cbf5c7fd3a1554d3f4319b2478f:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.0,
  "n": 200,
  "seconds": 15.564400434494019
 },
 "score_selection_rank32_lr1e-4.json": {
  "arm": "rank32_lr1e-4",
  "split": "selection",
  "checkpoint": "tinker://4d7ec04e-f22a-4ee5-9c17-64cb108c82e2:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.815,
  "n": 200,
  "seconds": 15.436081647872925
 },
 "score_selection_rank32_lr3e-4.json": {
  "arm": "rank32_lr3e-4",
  "split": "selection",
  "checkpoint": "tinker://2aa88565-ba3e-4d6b-abc3-c0f9777792da:train:0/sampler_weights/final",
  "evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052",
  "accuracy": 0.84,
  "n": 200,
  "seconds": 12.716762781143188
 },
 "selection.json": {
  "frozen_at_unix": 1793815150.7161703,
  "frozen_before_confirmation": true,
  "ranks": {
   "1": {
    "candidates": {
     "1e-3": 0.0,
     "1e-4": 0.82,
     "3e-4": 0.82,
     "3e-5": 0.82
    },
    "grid_edge": true,
    "selected_accuracy": 0.82,
    "selected_arm": "rank1_lr3e-5",
    "selected_lr": "3e-5"
   },
   "32": {
    "candidates": {
     "1e-3": 0.0,
     "1e-4": 0.815,
     "3e-4": 0.84
    },
    "grid_edge": false,
    "selected_accuracy": 0.84,
    "selected_arm": "rank32_lr3e-4",
    "selected_lr": "3e-4"
   }
  },
  "rule": "argmax test[0:200] greedy accuracy per rank; ties -> lower LR; grid edge -> extend once with 3e-5/3e-3 if feasible",
  "selection_evaluator_id": "d8a177a50b10b110e46f15106fe16e7ea14ec007c9956519067cec4b67256052"
 },
 "primary": {
  "interval": {
   "estimate": 0.013404825737265416,
   "lower": -0.0677935209192818,
   "upper": 0.09460317239381262,
   "n_units": 1119
  },
  "treatment_acc": 0.8239499553172476,
  "baseline_acc": 0.8105451295799821,
  "base_acc": 0.7345844504021448,
  "n": 1119,
  "roles": {
   "treatment": "rank1_lr3e-5",
   "baseline": "rank32_lr3e-4",
   "base": "base"
  }
 },
 "per_arm": {
  "rank1_lr1e-4": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.6875,
   "train_correct_last": 0.875,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.0001,
   "lora_rank": 1,
   "selection_accuracy": 0.82
  },
  "rank1_lr3e-4": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.703125,
   "train_correct_last": 0.8125,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.0003,
   "lora_rank": 1,
   "selection_accuracy": 0.82
  },
  "rank1_lr1e-3": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.6875,
   "train_correct_last": 0.0,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.001,
   "lora_rank": 1,
   "selection_accuracy": 0.0
  },
  "rank32_lr1e-4": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.75,
   "train_correct_last": 0.828125,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.0001,
   "lora_rank": 32,
   "selection_accuracy": 0.815
  },
  "rank32_lr3e-4": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.65625,
   "train_correct_last": 0.828125,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.0003,
   "lora_rank": 32,
   "confirmation_accuracy": 0.8105451295799821,
   "selection_accuracy": 0.84
  },
  "rank32_lr1e-3": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.640625,
   "train_correct_last": 0.0,
   "frac_mixed_step0": 1.0,
   "learning_rate": 0.001,
   "lora_rank": 32,
   "selection_accuracy": 0.0
  },
  "rank1_lr3e-5": {
   "train_steps_completed": 40,
   "train_correct_step0": 0.75,
   "train_correct_last": 0.671875,
   "frac_mixed_step0": 1.0,
   "learning_rate": 3e-05,
   "lora_rank": 1,
   "confirmation_accuracy": 0.8239499553172476,
   "selection_accuracy": 0.82
  },
  "base": {
   "confirmation_accuracy": 0.7345844504021448,
   "selection_accuracy": 0.71
  }
 }
}