davidhouseholter commited on
Commit
279645a
·
verified ·
1 Parent(s): 3a42e51

Add 8-way classical geometric Mixture-of-Rotations optimizer ablation: eval_results.json

Browse files
Files changed (1) hide show
  1. eval_results.json +194 -0
eval_results.json CHANGED
@@ -173,5 +173,199 @@
173
  "expert_angle_l2_divergence": 2.524
174
  }
175
  ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
176
  }
177
  }
 
173
  "expert_angle_l2_divergence": 2.524
174
  }
175
  ]
176
+ },
177
+ "optimizer_ablations_8way": {
178
+ "pair_adam": {
179
+ "variant": "PairAdam (Factored Toroidal Baseline)",
180
+ "val_loss": 0.4829,
181
+ "val_ppl": 1.6208,
182
+ "domain_val_loss": {
183
+ "code_and_algorithms": 0.3669,
184
+ "reasoning_and_language": 0.6553,
185
+ "rotational_geometry": 0.4772,
186
+ "science_and_physics": 0.5155
187
+ },
188
+ "domain_val_ppl": {
189
+ "code_and_algorithms": 1.4432,
190
+ "reasoning_and_language": 1.9257,
191
+ "rotational_geometry": 1.6115,
192
+ "science_and_physics": 1.6745
193
+ },
194
+ "mean_routing_entropy_nats": 1.3862,
195
+ "mean_routing_entropy_bits": 1.9999,
196
+ "mean_top1_confidence": 0.2559,
197
+ "expert_angle_l2_divergence": 2.5386,
198
+ "top4_expert_angle_l2_divergence": 2.5347,
199
+ "top4_full_expert_angle_l2_divergence": 3.5883,
200
+ "elapsed_sec": 180.15
201
+ },
202
+ "hessian_pair_adam": {
203
+ "variant": "HessianPairAdam (Sophia-Torus Cauchy-Riemann Curvature)",
204
+ "val_loss": 0.5403,
205
+ "val_ppl": 1.7165,
206
+ "domain_val_loss": {
207
+ "code_and_algorithms": 0.4082,
208
+ "reasoning_and_language": 0.7201,
209
+ "rotational_geometry": 0.533,
210
+ "science_and_physics": 0.5916
211
+ },
212
+ "domain_val_ppl": {
213
+ "code_and_algorithms": 1.5041,
214
+ "reasoning_and_language": 2.0545,
215
+ "rotational_geometry": 1.7041,
216
+ "science_and_physics": 1.8068
217
+ },
218
+ "mean_routing_entropy_nats": 1.3862,
219
+ "mean_routing_entropy_bits": 1.9999,
220
+ "mean_top1_confidence": 0.2559,
221
+ "expert_angle_l2_divergence": 2.5387,
222
+ "top4_expert_angle_l2_divergence": 2.5349,
223
+ "top4_full_expert_angle_l2_divergence": 3.5888,
224
+ "elapsed_sec": 180.09
225
+ },
226
+ "polar_natural_adam": {
227
+ "variant": "PolarNaturalAdam (Riemannian Polar Metric + Cauchy-Riemann Hessian)",
228
+ "val_loss": 0.4848,
229
+ "val_ppl": 1.6239,
230
+ "domain_val_loss": {
231
+ "code_and_algorithms": 0.3744,
232
+ "reasoning_and_language": 0.6466,
233
+ "rotational_geometry": 0.4747,
234
+ "science_and_physics": 0.5226
235
+ },
236
+ "domain_val_ppl": {
237
+ "code_and_algorithms": 1.4542,
238
+ "reasoning_and_language": 1.9091,
239
+ "rotational_geometry": 1.6076,
240
+ "science_and_physics": 1.6865
241
+ },
242
+ "mean_routing_entropy_nats": 1.3862,
243
+ "mean_routing_entropy_bits": 1.9999,
244
+ "mean_top1_confidence": 0.2559,
245
+ "expert_angle_l2_divergence": 2.5387,
246
+ "top4_expert_angle_l2_divergence": 2.5349,
247
+ "top4_full_expert_angle_l2_divergence": 3.5887,
248
+ "elapsed_sec": 180.75
249
+ },
250
+ "cyclide_adam": {
251
+ "variant": "CyclideAdam (Dupin Cyclide Conic Metric + Deviatoric Twist)",
252
+ "val_loss": 0.5006,
253
+ "val_ppl": 1.6498,
254
+ "domain_val_loss": {
255
+ "code_and_algorithms": 0.3758,
256
+ "reasoning_and_language": 0.6718,
257
+ "rotational_geometry": 0.4986,
258
+ "science_and_physics": 0.5426
259
+ },
260
+ "domain_val_ppl": {
261
+ "code_and_algorithms": 1.4561,
262
+ "reasoning_and_language": 1.9578,
263
+ "rotational_geometry": 1.6464,
264
+ "science_and_physics": 1.7205
265
+ },
266
+ "mean_routing_entropy_nats": 1.3862,
267
+ "mean_routing_entropy_bits": 1.9999,
268
+ "mean_top1_confidence": 0.2559,
269
+ "expert_angle_l2_divergence": 2.5534,
270
+ "top4_expert_angle_l2_divergence": 2.5789,
271
+ "top4_full_expert_angle_l2_divergence": 3.6659,
272
+ "elapsed_sec": 180.09
273
+ },
274
+ "symplectic_adam": {
275
+ "variant": "SymplecticAdam (Gromov-Liouville Non-Squeezing Floor + Casimir Cavity)",
276
+ "val_loss": 0.5567,
277
+ "val_ppl": 1.7448,
278
+ "domain_val_loss": {
279
+ "code_and_algorithms": 0.4185,
280
+ "reasoning_and_language": 0.7417,
281
+ "rotational_geometry": 0.553,
282
+ "science_and_physics": 0.6082
283
+ },
284
+ "domain_val_ppl": {
285
+ "code_and_algorithms": 1.5197,
286
+ "reasoning_and_language": 2.0996,
287
+ "rotational_geometry": 1.7385,
288
+ "science_and_physics": 1.8372
289
+ },
290
+ "mean_routing_entropy_nats": 1.3862,
291
+ "mean_routing_entropy_bits": 1.9999,
292
+ "mean_top1_confidence": 0.2559,
293
+ "expert_angle_l2_divergence": 2.5386,
294
+ "top4_expert_angle_l2_divergence": 2.5345,
295
+ "top4_full_expert_angle_l2_divergence": 3.588,
296
+ "elapsed_sec": 180.2
297
+ },
298
+ "gyric_adam": {
299
+ "variant": "GyricAdam (Poinsot-Lande Gyroscopic Projection + Thomas Precession)",
300
+ "val_loss": 0.5745,
301
+ "val_ppl": 1.7763,
302
+ "domain_val_loss": {
303
+ "code_and_algorithms": 0.4279,
304
+ "reasoning_and_language": 0.7609,
305
+ "rotational_geometry": 0.5672,
306
+ "science_and_physics": 0.6412
307
+ },
308
+ "domain_val_ppl": {
309
+ "code_and_algorithms": 1.5341,
310
+ "reasoning_and_language": 2.1401,
311
+ "rotational_geometry": 1.7633,
312
+ "science_and_physics": 1.8987
313
+ },
314
+ "mean_routing_entropy_nats": 1.3862,
315
+ "mean_routing_entropy_bits": 1.9999,
316
+ "mean_top1_confidence": 0.2559,
317
+ "expert_angle_l2_divergence": 2.5387,
318
+ "top4_expert_angle_l2_divergence": 2.535,
319
+ "top4_full_expert_angle_l2_divergence": 3.5888,
320
+ "elapsed_sec": 177.29
321
+ },
322
+ "huygens_adam": {
323
+ "variant": "HuygensAdam (Circular Order Parameter R_p Splay Torque)",
324
+ "val_loss": 0.5225,
325
+ "val_ppl": 1.6862,
326
+ "domain_val_loss": {
327
+ "code_and_algorithms": 0.3957,
328
+ "reasoning_and_language": 0.7041,
329
+ "rotational_geometry": 0.5107,
330
+ "science_and_physics": 0.5696
331
+ },
332
+ "domain_val_ppl": {
333
+ "code_and_algorithms": 1.4854,
334
+ "reasoning_and_language": 2.022,
335
+ "rotational_geometry": 1.6665,
336
+ "science_and_physics": 1.7676
337
+ },
338
+ "mean_routing_entropy_nats": 1.3862,
339
+ "mean_routing_entropy_bits": 1.9999,
340
+ "mean_top1_confidence": 0.2559,
341
+ "expert_angle_l2_divergence": 2.5676,
342
+ "top4_expert_angle_l2_divergence": 2.6216,
343
+ "top4_full_expert_angle_l2_divergence": 3.7258,
344
+ "elapsed_sec": 176.07
345
+ },
346
+ "huygens_symplectic_adam": {
347
+ "variant": "HuygensSymplecticAdam (Two-Stage Huygens -> Symplectic Curriculum)",
348
+ "val_loss": 0.523,
349
+ "val_ppl": 1.6871,
350
+ "domain_val_loss": {
351
+ "code_and_algorithms": 0.4091,
352
+ "reasoning_and_language": 0.6859,
353
+ "rotational_geometry": 0.5202,
354
+ "science_and_physics": 0.557
355
+ },
356
+ "domain_val_ppl": {
357
+ "code_and_algorithms": 1.5055,
358
+ "reasoning_and_language": 1.9856,
359
+ "rotational_geometry": 1.6823,
360
+ "science_and_physics": 1.7454
361
+ },
362
+ "mean_routing_entropy_nats": 1.3862,
363
+ "mean_routing_entropy_bits": 1.9999,
364
+ "mean_top1_confidence": 0.2559,
365
+ "expert_angle_l2_divergence": 2.5388,
366
+ "top4_expert_angle_l2_divergence": 2.5351,
367
+ "top4_full_expert_angle_l2_divergence": 3.5896,
368
+ "elapsed_sec": 176.01
369
+ }
370
  }
371
  }