Sharp JL Hidden-State Bottleneck / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001,
 10      "epochs": 12
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 12
 17        },
 18        "mean": 0.45439932495355606,
 19        "std": 0.04081052029431072,
 20        "per_seed": [
 21          0.4278562068939209,
 22          0.3956301808357239,
 23          0.46248915791511536,
 24          0.49452289938926697,
 25          0.44014036655426025,
 26          0.4744090139865875,
 27          0.5266208648681641,
 28          0.41352590918540955
 29        ],
 30        "n": 8
 31      },
 32      {
 33        "cfg": {
 34          "lr": 0.003,
 35          "epochs": 12
 36        },
 37        "mean": 0.5644181594252586,
 38        "std": 0.14529394471112436,
 39        "per_seed": [
 40          0.34822821617126465,
 41          0.40346693992614746,
 42          0.5700339674949646,
 43          0.6492372751235962,
 44          0.5649089813232422,
 45          0.8078203201293945,
 46          0.7059332728385925,
 47          0.46571630239486694
 48        ],
 49        "n": 8
 50      },
 51      {
 52        "cfg": {
 53          "lr": 0.01,
 54          "epochs": 12
 55        },
 56        "mean": 1.4830118492245674,
 57        "std": 0.6545389639796526,
 58        "per_seed": [
 59          1.843930959701538,
 60          0.8628994822502136,
 61          2.8634300231933594,
 62          1.0775489807128906,
 63          0.9952502250671387,
 64          2.02944016456604,
 65          1.1516919136047363,
 66          1.0399030447006226
 67        ],
 68        "n": 8
 69      }
 70    ],
 71    "harness_tuning": {
 72      "best_cfg": {
 73        "lr": 0.001,
 74        "epochs": 12
 75      },
 76      "sweep": [
 77        {
 78          "cfg": {
 79            "lr": 0.001,
 80            "epochs": 12
 81          },
 82          "mean": 0.4451246112585068
 83        },
 84        {
 85          "cfg": {
 86            "lr": 0.003,
 87            "epochs": 12
 88          },
 89          "mean": 0.4927415996789932
 90        },
 91        {
 92          "cfg": {
 93            "lr": 0.01,
 94            "epochs": 12
 95          },
 96          "mean": 1.6619523614645004
 97        }
 98      ],
 99      "full": {
100        "mean": 0.45439932495355606,
101        "std": 0.04081052029431072,
102        "per_seed": [
103          0.4278562068939209,
104          0.3956301808357239,
105          0.46248915791511536,
106          0.49452289938926697,
107          0.44014036655426025,
108          0.4744090139865875,
109          0.5266208648681641,
110          0.41352590918540955
111        ],
112        "n": 8
113      }
114    },
115    "full": {
116      "mean": 0.45439932495355606,
117      "std": 0.04081052029431072,
118      "per_seed": [
119        0.4278562068939209,
120        0.3956301808357239,
121        0.46248915791511536,
122        0.49452289938926697,
123        0.44014036655426025,
124        0.4744090139865875,
125        0.5266208648681641,
126        0.41352590918540955
127      ],
128      "n": 8
129    }
130  },
131  "idea": {
132    "per_seed": [
133      0.42581823468208313,
134      0.38251447677612305,
135      0.3930129110813141,
136      0.535940408706665,
137      0.4686683118343353,
138      0.5054774880409241,
139      0.45862287282943726,
140      0.4133540391921997
141    ],
142    "mean": 0.4479260928928852,
143    "std": 0.0507216238042377,
144    "n": 8
145  },
146  "comparison": {
147    "delta_mean": -0.006473232060670853,
148    "idea_wins": 5,
149    "n_pairs": 8,
150    "per_seed_diffs": [
151      -0.0020379722118377686,
152      -0.01311570405960083,
153      -0.06947624683380127,
154      0.04141750931739807,
155      0.028527945280075073,
156      0.031068474054336548,
157      -0.0679979920387268,
158      -0.00017186999320983887
159    ],
160    "p_value": 0.66785,
161    "mde": 0.03573610686518605,
162    "mde_rel_pct": 7.864471820867829,
163    "verdict": "no significant win",
164    "system_worked": false
165  },
166  "mechanism_signature": {
167    "idea_sweep": [
168      {
169        "cfg": {
170          "lr": 0.001,
171          "epochs": 12
172        },
173        "mean": 0.5554435513913631,
174        "std": 0.030751056568709256,
175        "per_seed": [
176          0.5679726004600525,
177          0.48871538043022156,
178          0.5542747378349304,
179          0.5856391191482544,
180          0.56807941198349,
181          0.5934925079345703,
182          0.550084114074707,
183          0.535290539264679
184        ],
185        "n": 8
186      },
187      {
188        "cfg": {
189          "lr": 0.003,
190          "epochs": 12
191        },
192        "mean": 0.4479260928928852,
193        "std": 0.0507216238042377,
194        "per_seed": [
195          0.42581823468208313,
196          0.38251447677612305,
197          0.3930129110813141,
198          0.535940408706665,
199          0.4686683118343353,
200          0.5054774880409241,
201          0.45862287282943726,
202          0.4133540391921997
203        ],
204        "n": 8
205      },
206      {
207        "cfg": {
208          "lr": 0.01,
209          "epochs": 12
210        },
211        "mean": 0.5642099156975746,
212        "std": 0.06657952336519016,
213        "per_seed": [
214          0.571461021900177,
215          0.4690275490283966,
216          0.6746896505355835,
217          0.6299232840538025,
218          0.5943968296051025,
219          0.5776129961013794,
220          0.5029228329658508,
221          0.49364516139030457
222        ],
223        "n": 8
224      }
225    ],
226    "mechanism_signature": {
227      "prediction": "Gaussian JL fixed-pair distance ratio has mean near one and SD near 1/sqrt(2r)",
228      "epsilon": 0.3,
229      "constant": 1.0,
230      "original_width": 64,
231      "projected_width": 18,
232      "raw_width_rule": 17.61272466516729,
233      "trained_test_metric": 0.42581823468208313,
234      "observed_ratio_mean": 1.1171250343322754,
235      "observed_ratio_sd": 5.5652035371167585e-06,
236      "predicted_ratio_mean": 1.0,
237      "predicted_ratio_sd": 0.16666666666666666,
238      "ratio_sd_observed_over_predicted": 3.339122122270055e-05,
239      "p95_absolute_distortion": 0.11712539196014404,
240      "confirmed": false
241    },
242    "matched_structure": "sequence-level jointly processed windows/tokens"
243  },
244  "custom_track": null,
245  "idea_hyperparameters": {
246    "epsilon": 0.3,
247    "constant": 1.0,
248    "width_rule": "ceil(C*min(d,n-1,log(2+epsilon^2*n)/epsilon^2))"
249  }
250}