grt-30m-qa / eval_results.json
davidhouseholter's picture
Release MatterWaveAI/grt-30m-qa (100% rotational head + 8-way geometric optimizer ablation)
92454d2 verified
Raw History Blame Contribute Delete
3.97 kB
{
"model_id": "MatterWaveAI/grt-30m-qa",
"architecture": "GRTForQuestionAnswering",
"base_model": "MatterWaveAI/grt-30m",
"parameters": 29306020,
"trainable_parameters": 812430,
"nn_linear_layers": 0,
"nn_conv1d_layers": 0,
"num_labels": 2,
"best_optimizer": "polar_natural_adam",
"benchmarks": {
"squad_extractive_qa_val": {
"val_span_loss": 0.6512,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25
},
"step0_untrained_head": {
"val_span_loss": 8.9793,
"exact_match": 0.0,
"span_token_f1": 0.45,
"start_acc": 3.75,
"end_acc": 0.0
}
},
"optimizer_ablation_8way": {
"pair_adam": {
"optimizer_type": "pair_adam",
"variant": "PairAdam (Factored Toroidal Baseline)",
"final_train_loss": 0.1383,
"last_batch_train_loss": 0.5309,
"val_span_loss": 0.6532,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 207.23
},
"hessian_pair_adam": {
"optimizer_type": "hessian_pair_adam",
"variant": "HessianPairAdam (Sophia-Torus Cauchy-Riemann Curvature)",
"final_train_loss": 0.1387,
"last_batch_train_loss": 0.5307,
"val_span_loss": 0.6506,
"exact_match": 73.75,
"span_token_f1": 83.39,
"start_acc": 83.75,
"end_acc": 75.0,
"wall_time_s": 232.96
},
"polar_natural_adam": {
"optimizer_type": "polar_natural_adam",
"variant": "PolarNaturalAdam (Riemannian Polar Metric + Cauchy-Riemann Hessian)",
"final_train_loss": 0.1385,
"last_batch_train_loss": 0.5308,
"val_span_loss": 0.6512,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 219.08
},
"cyclide_adam": {
"optimizer_type": "cyclide_adam",
"variant": "CyclideAdam (Dupin Cyclide Conic Metric + Deviatoric Twist)",
"final_train_loss": 0.1382,
"last_batch_train_loss": 0.5305,
"val_span_loss": 0.652,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 225.19
},
"symplectic_adam": {
"optimizer_type": "symplectic_adam",
"variant": "SymplecticAdam (Gromov-Liouville Non-Squeezing Floor + Casimir Cavity)",
"final_train_loss": 0.1391,
"last_batch_train_loss": 0.5317,
"val_span_loss": 0.6537,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 261.38
},
"gyric_adam": {
"optimizer_type": "gyric_adam",
"variant": "GyricAdam (Poinsot-Lande Gyroscopic Projection + Thomas Precession)",
"final_train_loss": 0.1393,
"last_batch_train_loss": 0.532,
"val_span_loss": 0.6541,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 221.44
},
"huygens_adam": {
"optimizer_type": "huygens_adam",
"variant": "HuygensAdam (Circular Order Parameter R_p Splay Torque)",
"final_train_loss": 0.1378,
"last_batch_train_loss": 0.532,
"val_span_loss": 0.6557,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 152.14
},
"huygens_symplectic_adam": {
"optimizer_type": "huygens_symplectic_adam",
"variant": "HuygensSymplecticAdam (Two-Stage Huygens -> Symplectic Curriculum)",
"final_train_loss": 0.1391,
"last_batch_train_loss": 0.5317,
"val_span_loss": 0.6537,
"exact_match": 75.0,
"span_token_f1": 83.57,
"start_acc": 83.75,
"end_acc": 76.25,
"wall_time_s": 146.7
}
}
}