grt-30m-mor / eval_results.json
davidhouseholter's picture
Add 8-way classical geometric Mixture-of-Rotations optimizer ablation: eval_results.json
279645a verified
Raw History Blame Contribute Delete
11.3 kB
{
"model_id": "MatterWaveAI/grt-30m-mor",
"architecture": "GRTForCausalLM (Mixture-of-Rotations)",
"base_model": "MatterWaveAI/grt-30m",
"parameters": 31634086,
"base_parameters": 29305498,
"ffn_n_experts": 4,
"mor_mode": "direct",
"mor_router_mode": "rotational",
"nn_linear_layers": 0,
"before_upcycle_specialization": {
"val_loss": 5.8467,
"val_ppl": 346.102,
"domain_val_loss": {
"code_and_algorithms": 5.4092,
"reasoning_and_language": 6.172,
"rotational_geometry": 6.6421,
"science_and_physics": 5.3337
}
},
"after_upcycle_specialization": {
"val_loss": 0.5115,
"val_ppl": 1.6678,
"domain_val_loss": {
"code_and_algorithms": 0.4334,
"reasoning_and_language": 0.6442,
"rotational_geometry": 0.4991,
"science_and_physics": 0.5293
},
"router_diagnostics": [
{
"layer": 0,
"mean_expert_probs": [
0.2494,
0.2488,
0.253,
0.2487
],
"marginal_entropy": 1.3863,
"mean_top1_confidence": 0.2545,
"expert_angle_l2_divergence": 2.5326
},
{
"layer": 1,
"mean_expert_probs": [
0.2492,
0.247,
0.2574,
0.2464
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.2599,
"expert_angle_l2_divergence": 2.5353
},
{
"layer": 2,
"mean_expert_probs": [
0.2526,
0.2482,
0.2528,
0.2465
],
"marginal_entropy": 1.3862,
"mean_top1_confidence": 0.2541,
"expert_angle_l2_divergence": 2.5524
},
{
"layer": 3,
"mean_expert_probs": [
0.251,
0.2486,
0.2548,
0.2457
],
"marginal_entropy": 1.3862,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5425
},
{
"layer": 4,
"mean_expert_probs": [
0.2511,
0.2486,
0.2562,
0.2441
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.2586,
"expert_angle_l2_divergence": 2.5487
},
{
"layer": 5,
"mean_expert_probs": [
0.2508,
0.2485,
0.2572,
0.2436
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.26,
"expert_angle_l2_divergence": 2.5342
},
{
"layer": 6,
"mean_expert_probs": [
0.2504,
0.2486,
0.2571,
0.2439
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.2602,
"expert_angle_l2_divergence": 2.5501
},
{
"layer": 7,
"mean_expert_probs": [
0.251,
0.2488,
0.2567,
0.2435
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.2592,
"expert_angle_l2_divergence": 2.5291
},
{
"layer": 8,
"mean_expert_probs": [
0.2506,
0.2461,
0.2605,
0.2427
],
"marginal_entropy": 1.3859,
"mean_top1_confidence": 0.2658,
"expert_angle_l2_divergence": 2.5455
},
{
"layer": 9,
"mean_expert_probs": [
0.2502,
0.247,
0.2582,
0.2445
],
"marginal_entropy": 1.3861,
"mean_top1_confidence": 0.2647,
"expert_angle_l2_divergence": 2.5364
},
{
"layer": 10,
"mean_expert_probs": [
0.2501,
0.2484,
0.2563,
0.2452
],
"marginal_entropy": 1.3862,
"mean_top1_confidence": 0.2636,
"expert_angle_l2_divergence": 2.5368
},
{
"layer": 11,
"mean_expert_probs": [
0.2509,
0.2481,
0.256,
0.245
],
"marginal_entropy": 1.3862,
"mean_top1_confidence": 0.2588,
"expert_angle_l2_divergence": 2.524
}
]
},
"optimizer_ablations_8way": {
"pair_adam": {
"variant": "PairAdam (Factored Toroidal Baseline)",
"val_loss": 0.4829,
"val_ppl": 1.6208,
"domain_val_loss": {
"code_and_algorithms": 0.3669,
"reasoning_and_language": 0.6553,
"rotational_geometry": 0.4772,
"science_and_physics": 0.5155
},
"domain_val_ppl": {
"code_and_algorithms": 1.4432,
"reasoning_and_language": 1.9257,
"rotational_geometry": 1.6115,
"science_and_physics": 1.6745
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5386,
"top4_expert_angle_l2_divergence": 2.5347,
"top4_full_expert_angle_l2_divergence": 3.5883,
"elapsed_sec": 180.15
},
"hessian_pair_adam": {
"variant": "HessianPairAdam (Sophia-Torus Cauchy-Riemann Curvature)",
"val_loss": 0.5403,
"val_ppl": 1.7165,
"domain_val_loss": {
"code_and_algorithms": 0.4082,
"reasoning_and_language": 0.7201,
"rotational_geometry": 0.533,
"science_and_physics": 0.5916
},
"domain_val_ppl": {
"code_and_algorithms": 1.5041,
"reasoning_and_language": 2.0545,
"rotational_geometry": 1.7041,
"science_and_physics": 1.8068
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5387,
"top4_expert_angle_l2_divergence": 2.5349,
"top4_full_expert_angle_l2_divergence": 3.5888,
"elapsed_sec": 180.09
},
"polar_natural_adam": {
"variant": "PolarNaturalAdam (Riemannian Polar Metric + Cauchy-Riemann Hessian)",
"val_loss": 0.4848,
"val_ppl": 1.6239,
"domain_val_loss": {
"code_and_algorithms": 0.3744,
"reasoning_and_language": 0.6466,
"rotational_geometry": 0.4747,
"science_and_physics": 0.5226
},
"domain_val_ppl": {
"code_and_algorithms": 1.4542,
"reasoning_and_language": 1.9091,
"rotational_geometry": 1.6076,
"science_and_physics": 1.6865
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5387,
"top4_expert_angle_l2_divergence": 2.5349,
"top4_full_expert_angle_l2_divergence": 3.5887,
"elapsed_sec": 180.75
},
"cyclide_adam": {
"variant": "CyclideAdam (Dupin Cyclide Conic Metric + Deviatoric Twist)",
"val_loss": 0.5006,
"val_ppl": 1.6498,
"domain_val_loss": {
"code_and_algorithms": 0.3758,
"reasoning_and_language": 0.6718,
"rotational_geometry": 0.4986,
"science_and_physics": 0.5426
},
"domain_val_ppl": {
"code_and_algorithms": 1.4561,
"reasoning_and_language": 1.9578,
"rotational_geometry": 1.6464,
"science_and_physics": 1.7205
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5534,
"top4_expert_angle_l2_divergence": 2.5789,
"top4_full_expert_angle_l2_divergence": 3.6659,
"elapsed_sec": 180.09
},
"symplectic_adam": {
"variant": "SymplecticAdam (Gromov-Liouville Non-Squeezing Floor + Casimir Cavity)",
"val_loss": 0.5567,
"val_ppl": 1.7448,
"domain_val_loss": {
"code_and_algorithms": 0.4185,
"reasoning_and_language": 0.7417,
"rotational_geometry": 0.553,
"science_and_physics": 0.6082
},
"domain_val_ppl": {
"code_and_algorithms": 1.5197,
"reasoning_and_language": 2.0996,
"rotational_geometry": 1.7385,
"science_and_physics": 1.8372
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5386,
"top4_expert_angle_l2_divergence": 2.5345,
"top4_full_expert_angle_l2_divergence": 3.588,
"elapsed_sec": 180.2
},
"gyric_adam": {
"variant": "GyricAdam (Poinsot-Lande Gyroscopic Projection + Thomas Precession)",
"val_loss": 0.5745,
"val_ppl": 1.7763,
"domain_val_loss": {
"code_and_algorithms": 0.4279,
"reasoning_and_language": 0.7609,
"rotational_geometry": 0.5672,
"science_and_physics": 0.6412
},
"domain_val_ppl": {
"code_and_algorithms": 1.5341,
"reasoning_and_language": 2.1401,
"rotational_geometry": 1.7633,
"science_and_physics": 1.8987
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5387,
"top4_expert_angle_l2_divergence": 2.535,
"top4_full_expert_angle_l2_divergence": 3.5888,
"elapsed_sec": 177.29
},
"huygens_adam": {
"variant": "HuygensAdam (Circular Order Parameter R_p Splay Torque)",
"val_loss": 0.5225,
"val_ppl": 1.6862,
"domain_val_loss": {
"code_and_algorithms": 0.3957,
"reasoning_and_language": 0.7041,
"rotational_geometry": 0.5107,
"science_and_physics": 0.5696
},
"domain_val_ppl": {
"code_and_algorithms": 1.4854,
"reasoning_and_language": 2.022,
"rotational_geometry": 1.6665,
"science_and_physics": 1.7676
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5676,
"top4_expert_angle_l2_divergence": 2.6216,
"top4_full_expert_angle_l2_divergence": 3.7258,
"elapsed_sec": 176.07
},
"huygens_symplectic_adam": {
"variant": "HuygensSymplecticAdam (Two-Stage Huygens -> Symplectic Curriculum)",
"val_loss": 0.523,
"val_ppl": 1.6871,
"domain_val_loss": {
"code_and_algorithms": 0.4091,
"reasoning_and_language": 0.6859,
"rotational_geometry": 0.5202,
"science_and_physics": 0.557
},
"domain_val_ppl": {
"code_and_algorithms": 1.5055,
"reasoning_and_language": 1.9856,
"rotational_geometry": 1.6823,
"science_and_physics": 1.7454
},
"mean_routing_entropy_nats": 1.3862,
"mean_routing_entropy_bits": 1.9999,
"mean_top1_confidence": 0.2559,
"expert_angle_l2_divergence": 2.5388,
"top4_expert_angle_l2_divergence": 2.5351,
"top4_full_expert_angle_l2_divergence": 3.5896,
"elapsed_sec": 176.01
}
}
}