Kurtosis-calibrated gradient clipping / bench_report.json

✓✓ Beats tuned baseline

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 25.60164165496826
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 13.860755920410156
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 10.442761182785034
 33      }
 34    ],
 35    "full": {
 36      "mean": 10.930935978889465,
 37      "std": 1.3983305692789296,
 38      "per_seed": [
 39        10.067484855651855,
 40        12.066851615905762,
 41        9.622394561767578,
 42        10.014313697814941,
 43        8.787652015686035,
 44        11.810267448425293,
 45        12.032318115234375,
 46        13.046205520629883
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "cfg": {
 53      "lr": 0.006,
 54      "weight_decay": 0.0
 55    },
 56    "mean": 10.922250509262085,
 57    "std": 1.396535179145413,
 58    "per_seed": [
 59      10.052755355834961,
 60      12.058212280273438,
 61      9.612079620361328,
 62      10.001825332641602,
 63      8.791247367858887,
 64      11.804107666015625,
 65      12.02346420288086,
 66      13.03431224822998
 67    ],
 68    "n": 8,
 69    "behavior_all_seeds": [
 70      {
 71        "raw_spike": 97,
 72        "clipped_entries": 462,
 73        "entries": 354888,
 74        "tail_sum": 0.4687500017462298,
 75        "k_sum": 4870.490382072394,
 76        "tau_sum": 3035.111967606231,
 77        "steps": 72,
 78        "mean_tail_above_tau": 0.0010850694484866431,
 79        "mean_kurtosis_bound": 11.274283291834244,
 80        "mean_tau": 7.025722147236645,
 81        "clip_fraction": 0.0013018191654831947
 82      },
 83      {
 84        "raw_spike": 102,
 85        "clipped_entries": 623,
 86        "entries": 354888,
 87        "tail_sum": 0.6460449252044782,
 88        "k_sum": 4870.597906613239,
 89        "tau_sum": 3113.7841693589585,
 90        "steps": 72,
 91        "mean_tail_above_tau": 0.001495474363899255,
 92        "mean_kurtosis_bound": 11.274532191234348,
 93        "mean_tau": 7.207833725367959,
 94        "clip_fraction": 0.0017554834201212778
 95      },
 96      {
 97        "raw_spike": 90,
 98        "clipped_entries": 539,
 99        "entries": 354888,
100        "tail_sum": 0.4866699246922508,
101        "k_sum": 5089.335932393671,
102        "tau_sum": 2659.756519132028,
103        "steps": 72,
104        "mean_tail_above_tau": 0.0011265507516024325,
105        "mean_kurtosis_bound": 11.78087021387424,
106        "mean_tau": 6.156843794287102,
107        "clip_fraction": 0.0015187890263970605
108      },
109      {
110        "raw_spike": 100,
111        "clipped_entries": 822,
112        "entries": 354888,
113        "tail_sum": 0.7038574230391532,
114        "k_sum": 4833.979456970323,
115        "tau_sum": 2974.376847470511,
116        "steps": 72,
117        "mean_tail_above_tau": 0.0016292995903684103,
118        "mean_kurtosis_bound": 11.189767261505377,
119        "mean_tau": 6.885131591366924,
120        "clip_fraction": 0.0023162237100155542
121      },
122      {
123        "raw_spike": 81,
124        "clipped_entries": 358,
125        "entries": 354888,
126        "tail_sum": 0.29570312541909516,
127        "k_sum": 5140.149971492376,
128        "tau_sum": 3046.6057521423736,
129        "steps": 72,
130        "mean_tail_above_tau": 0.0006844979755071647,
131        "mean_kurtosis_bound": 11.8984953043805,
132        "mean_tau": 7.052328129959198,
133        "clip_fraction": 0.001008768963729402
134      },
135      {
136        "raw_spike": 106,
137        "clipped_entries": 853,
138        "entries": 354888,
139        "tail_sum": 0.7641601584618911,
140        "k_sum": 5438.268085443231,
141        "tau_sum": 3097.6346601389573,
142        "steps": 72,
143        "mean_tail_above_tau": 0.001768889255698822,
144        "mean_kurtosis_bound": 12.58858353111859,
145        "mean_tau": 7.170450602173513,
146        "clip_fraction": 0.0024035752124613964
147      },
148      {
149        "raw_spike": 94,
150        "clipped_entries": 469,
151        "entries": 354888,
152        "tail_sum": 0.5587890637107193,
153        "k_sum": 5913.583425274048,
154        "tau_sum": 3219.2745450983753,
155        "steps": 72,
156        "mean_tail_above_tau": 0.0012934932030340726,
157        "mean_kurtosis_bound": 13.688850521467703,
158        "mean_tau": 7.452024409949943,
159        "clip_fraction": 0.0013215436982935461
160      },
161      {
162        "raw_spike": 97,
163        "clipped_entries": 813,
164        "entries": 354888,
165        "tail_sum": 0.8062500024680048,
166        "k_sum": 5628.396702211232,
167        "tau_sum": 3650.706699383914,
168        "steps": 72,
169        "mean_tail_above_tau": 0.0018663194501574185,
170        "mean_kurtosis_bound": 13.028696069933407,
171        "mean_tau": 8.45070995227758,
172        "clip_fraction": 0.002290863596402245
173      }
174    ]
175  },
176  "comparison": {
177    "delta_mean": -0.008685469627380371,
178    "idea_wins": 7,
179    "n_pairs": 8,
180    "per_seed_diffs": [
181      -0.014729499816894531,
182      -0.008639335632324219,
183      -0.01031494140625,
184      -0.012488365173339844,
185      0.0035953521728515625,
186      -0.006159782409667969,
187      -0.008853912353515625,
188      -0.011893272399902344
189    ],
190    "p_value": 0.01475,
191    "mde": 0.004699979815996483,
192    "mde_rel_pct": 0.04299704824063914,
193    "verdict": "idea better (significant)",
194    "system_worked": true
195  },
196  "math_check": {
197    "max_inverse_error": 8.881784197001252e-16,
198    "max_transition_polynomial_error": 1.4210854715202004e-14,
199    "min_projected_tail_margin": 0.0,
200    "passed": true
201  },
202  "idea_sweep": [
203    {
204      "cfg": {
205        "lr": 0.001,
206        "weight_decay": 0.0
207      },
208      "mean": 25.558531761169434
209    },
210    {
211      "cfg": {
212        "lr": 0.003,
213        "weight_decay": 0.0
214      },
215      "mean": 13.852189540863037
216    },
217    {
218      "cfg": {
219        "lr": 0.006,
220        "weight_decay": 0.0
221      },
222      "mean": 10.431218147277832
223    }
224  ],
225  "protocol_note": "8 paired seeds; baseline sweep and idea sweep share the full learning-rate union; Adam and architecture are otherwise identical.",
226  "mechanism_signature": {
227    "prediction": "trained gradient-coordinate tail above calibrated tau is controlled near target delta=.02",
228    "target_delta": 0.02,
229    "observed_mean_tail_above_tau": 0.0013686992548442773,
230    "mean_clip_fraction": 0.0017396333491129594,
231    "mean_kurtosis_bound": 12.09050979816855,
232    "confirmed": true
233  }
234}