AUBIN-E4B-Control / control_trained.json
emrevrg's picture
AUBIN release card, calibration, code (measured numbers only)
5ef0246 verified
Raw History Blame Contribute Delete
2.16 kB
{
"episodes": 100,
"seed": 0,
"model": "google/gemma-4-E4B-it + AUBIN-E4B-Control LoRA (Colab T4, 4-bit)",
"train": "control_train.py --episodes 3000 --train 1200 --bs 4 (15161 examples, seeds 1000+), best dev step 1000 (0.852)",
"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
},
"Gemma-4-E4B base, no control training (one pass per move)": {
"episodes": 100,
"decisions": 683,
"step_accuracy": 0.3924,
"success": 0.08,
"lava_deaths": 26,
"ms_per_decision_median": 328.6
},
"Gemma-4-E4B base, no control training + safety shield": {
"episodes": 100,
"decisions": 858,
"step_accuracy": 0.5536,
"success": 0.11,
"lava_deaths": 0,
"ms_per_decision_median": 330.1
},
"AUBIN-E4B-Control (one pass per move)": {
"episodes": 100,
"decisions": 430,
"step_accuracy": 0.6,
"success": 0.58,
"lava_deaths": 28,
"ms_per_decision_median": 384.1,
"acc_when_conf_ge_0.8": 0.6364,
"coverage_conf_ge_0.8": 0.5372,
"acc_when_conf_ge_0.9": 0.6528,
"coverage_conf_ge_0.9": 0.4488
},
"AUBIN-E4B-Control + safety shield (never steps on visible lava/walls)": {
"episodes": 100,
"decisions": 530,
"step_accuracy": 0.9321,
"success": 0.95,
"lava_deaths": 0,
"ms_per_decision_median": 385.2,
"acc_when_conf_ge_0.8": 0.9891,
"coverage_conf_ge_0.8": 0.5208,
"acc_when_conf_ge_0.9": 0.9872,
"coverage_conf_ge_0.9": 0.4415
}
}
}