IQC-Certified Training Dynamics / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.0014904916170053184
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0010051239951280877
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.00037887565849814564
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0003780504302994814,
 33      "std": 7.314250470278514e-05,
 34      "per_seed": [
 35        0.0003051091334782541,
 36        0.00044008868280798197,
 37        0.0002926669840235263,
 38        0.0004776378336828202,
 39        0.00027556720306165516,
 40        0.00042108612251468003,
 41        0.00037083675852045417,
 42        0.0004414107243064791
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.00030132598112686537,
 49    "std": 6.586114249212311e-05,
 50    "per_seed": [
 51      0.00032067738357000053,
 52      0.00030712690204381943,
 53      0.00022303542937152088,
 54      0.0003347629099152982,
 55      0.000343360414262861,
 56      0.00025801677838899195,
 57      0.00020414931350387633,
 58      0.0004194787179585546
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -7.6724449172616e-05,
 64    "idea_wins": 6,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      1.556825009174645e-05,
 68      -0.00013296178076416254,
 69      -6.963155465200543e-05,
 70      -0.00014287492376752198,
 71      6.779321120120585e-05,
 72      -0.00016306934412568808,
 73      -0.00016668744501657784,
 74      -2.193200634792447e-05
 75    ],
 76    "p_value": 0.0522,
 77    "mde": 7.448216191901723e-05,
 78    "mde_rel_pct": 19.701647174429734,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "track_match": "dynamics/control",
 84    "intervention": "IQC-inspired certified gradient clipping",
 85    "idea_sweep": [
 86      {
 87        "cfg": {
 88          "lr": 0.001,
 89          "clip": 1.0
 90        },
 91        "mean": 0.001604030971066095
 92      },
 93      {
 94        "cfg": {
 95          "lr": 0.003,
 96          "clip": 1.0
 97        },
 98        "mean": 0.0008869413213687949
 99      },
100      {
101        "cfg": {
102          "lr": 0.01,
103          "clip": 1.0
104        },
105        "mean": 0.00030132598112686537
106      }
107    ],
108    "best_idea_cfg": {
109      "lr": 0.01,
110      "clip": 1.0
111    },
112    "signature": {
113      "predicted_gamma_proxy": 12.570822464322077,
114      "observed_output_gain": 7.226121114322568e-05,
115      "disturbance_epsilon": 5.01467227935791,
116      "output_shift": 0.000362366292392835,
117      "parameter_shift": 0.03903782367706299,
118      "estimated_incremental_gain_L": 0.08642427995792329,
119      "rho": 0.9204507101394397,
120      "clip_events": 6,
121      "confirmed": true
122    }
123  },
124  "protocol_notes": {
125    "epochs": 12,
126    "batch": 64,
127    "n_train": 400,
128    "n_test": 200,
129    "baseline_grid": [
130      {
131        "lr": 0.001
132      },
133      {
134        "lr": 0.003
135      },
136      {
137        "lr": 0.01
138      }
139    ],
140    "idea_grid": [
141      {
142        "lr": 0.001,
143        "clip": 1.0
144      },
145      {
146        "lr": 0.003,
147        "clip": 1.0
148      },
149      {
150        "lr": 0.01,
151        "clip": 1.0
152      }
153    ],
154    "paired_seeds": [
155      0,
156      1,
157      2,
158      3,
159      4,
160      5,
161      6,
162      7
163    ]
164  }
165}