Lattice Error-Feedback Residual Blocks / bench_report.json

✓✓ Beats tuned baseline

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "quantized_residual_rnn",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "h": 0.05
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "h": 0.05
 17        },
 18        "mean": 0.16015676967799664
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "h": 0.05
 24        },
 25        "mean": 0.016941760201007128
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "h": 0.05
 31        },
 32        "mean": 0.0486269072862342
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.017625866690650582,
 37      "std": 0.006777111362893624,
 38      "per_seed": [
 39        0.013591268099844456,
 40        0.009187710471451283,
 41        0.02248837985098362,
 42        0.022499682381749153,
 43        0.010225045494735241,
 44        0.029711289331316948,
 45        0.012789911590516567,
 46        0.02051364630460739
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.002115632814820856,
 53    "std": 0.00044668106939017545,
 54    "per_seed": [
 55      0.0019963281229138374,
 56      0.0019065372180193663,
 57      0.001684049959294498,
 58      0.001756328041665256,
 59      0.002363975392654538,
 60      0.0028063422068953514,
 61      0.0016278764232993126,
 62      0.002783625153824687
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.015510233875829726,
 68    "idea_wins": 8,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.011594939976930618,
 72      -0.007281173253431916,
 73      -0.02080432989168912,
 74      -0.020743354340083897,
 75      -0.007861070102080703,
 76      -0.026904947124421597,
 77      -0.011162035167217255,
 78      -0.017730021150782704
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.0059179323712754825,
 82    "mde_rel_pct": 33.575270227219946,
 83    "verdict": "idea better (significant)",
 84    "system_worked": true
 85  },
 86  "mechanism_signature": {
 87    "prediction": "increment feedback telescopes quantization error; unsaturated final carry is bounded by h/2",
 88    "observed_conservation_residual_mean": 8.455314948196957e-09,
 89    "observed_conservation_residual_max": 1.0588774301822923e-08,
 90    "observed_max_carry_over_h": 0.49999260902404785,
 91    "observed_saturation_total": 58,
 92    "per_seed": [
 93      {
 94        "seed": 0,
 95        "conservation_residual": 8.429561049183576e-09,
 96        "max_carry_over_h": 0.49997812509536743,
 97        "saturation": 0
 98      },
 99      {
100        "seed": 1,
101        "conservation_residual": 5.565331751000713e-09,
102        "max_carry_over_h": 0.49962854385375977,
103        "saturation": 0
104      },
105      {
106        "seed": 2,
107        "conservation_residual": 1.0588774301822923e-08,
108        "max_carry_over_h": 0.499947190284729,
109        "saturation": 35
110      },
111      {
112        "seed": 3,
113        "conservation_residual": 1.0041367737301243e-08,
114        "max_carry_over_h": 0.49995261430740356,
115        "saturation": 16
116      },
117      {
118        "seed": 4,
119        "conservation_residual": 8.129112494259516e-09,
120        "max_carry_over_h": 0.4999336004257202,
121        "saturation": 0
122      },
123      {
124        "seed": 5,
125        "conservation_residual": 9.632834974127036e-09,
126        "max_carry_over_h": 0.49989578127861023,
127        "saturation": 7
128      },
129      {
130        "seed": 6,
131        "conservation_residual": 7.62265717213495e-09,
132        "max_carry_over_h": 0.49999260902404785,
133        "saturation": 0
134      },
135      {
136        "seed": 7,
137        "conservation_residual": 7.632880105745699e-09,
138        "max_carry_over_h": 0.4997541308403015,
139        "saturation": 0
140      }
141    ],
142    "confirmed": false,
143    "idea_sweep": {
144      "{'lr': 0.001, 'h': 0.05}": {
145        "mean": 0.006243410549359396,
146        "std": 0.0022887590986672723,
147        "per_seed": [
148          0.005430769640952349,
149          0.0037102538626641035,
150          0.007385811768472195,
151          0.011516324244439602,
152          0.006037216633558273,
153          0.00639760959893465,
154          0.003992232959717512,
155          0.005477065686136484
156        ],
157        "n": 8
158      },
159      "{'lr': 0.003, 'h': 0.05}": {
160        "mean": 0.004214856220642105,
161        "std": 0.0010927308726895986,
162        "per_seed": [
163          0.004151166416704655,
164          0.0027393524069339037,
165          0.004888014867901802,
166          0.006238944828510284,
167          0.0037882074248045683,
168          0.004983744118362665,
169          0.0027855250518769026,
170          0.004143894650042057
171        ],
172        "n": 8
173      },
174      "{'lr': 0.01, 'h': 0.05}": {
175        "mean": 0.002115632814820856,
176        "std": 0.00044668106939017545,
177        "per_seed": [
178          0.0019963281229138374,
179          0.0019065372180193663,
180          0.001684049959294498,
181          0.001756328041665256,
182          0.002363975392654538,
183          0.0028063422068953514,
184          0.0016278764232993126,
185          0.002783625153824687
186        ],
187        "n": 8
188      }
189    },
190    "idea_best_cfg": {
191      "lr": 0.01,
192      "h": 0.05
193    },
194    "baseline_best_cfg": {
195      "lr": 0.003,
196      "h": 0.05
197    }
198  }
199}