control_trained.json

{
 "episodes": 100,
 "seed": 0,
 "model": "google/gemma-4-12B-it + AUBIN-12B-Control LoRA (Colab T4, 4-bit)",
 "train": "control_train.py --episodes 3000 --train 1200 --bs 4 (15161 examples, seeds 1000+), best dev step 1200 (0.875)",
 "rows": {
  "random": {
   "episodes": 100,
   "decisions": 668,
   "step_accuracy": 0.3533,
   "success": 0.06,
   "lava_deaths": 40,
   "ms_per_decision_median": 0.0
  },
  "random + safety shield": {
   "episodes": 100,
   "decisions": 858,
   "step_accuracy": 0.5338,
   "success": 0.13,
   "lava_deaths": 0,
   "ms_per_decision_median": 0.0
  },
  "greedy (ignores obstacles)": {
   "episodes": 100,
   "decisions": 456,
   "step_accuracy": 0.568,
   "success": 0.58,
   "lava_deaths": 26,
   "ms_per_decision_median": 0.0
  },
  "greedy + safety shield": {
   "episodes": 100,
   "decisions": 540,
   "step_accuracy": 0.8759,
   "success": 0.89,
   "lava_deaths": 0,
   "ms_per_decision_median": 0.0
  },
  "AUBIN-12B-Control (one pass per move)": {
   "episodes": 100,
   "decisions": 404,
   "step_accuracy": 0.8639,
   "success": 0.74,
   "lava_deaths": 24,
   "ms_per_decision_median": 853.7,
   "acc_when_conf_ge_0.8": 0.8773,
   "coverage_conf_ge_0.8": 0.6658,
   "acc_when_conf_ge_0.9": 0.8972,
   "coverage_conf_ge_0.9": 0.5297
  },
  "AUBIN-12B-Control + safety shield (never steps on visible lava/walls)": {
   "episodes": 100,
   "decisions": 536,
   "step_accuracy": 0.9179,
   "success": 0.92,
   "lava_deaths": 0,
   "ms_per_decision_median": 856.1,
   "acc_when_conf_ge_0.8": 0.9815,
   "coverage_conf_ge_0.8": 0.6045,
   "acc_when_conf_ge_0.9": 0.985,
   "coverage_conf_ge_0.9": 0.4981
  }
 }
}