experiment: add completed A6000 crossover resultsa6000-x1-results

1 files changed, 4001 insertions, 0 deletions

diff --git a/results/resnet_crossover/r2/resnet-r2-clean_kp-d32-lr0p1.json b/results/resnet_crossover/r2/resnet-r2-clean_kp-d32-lr0p1.json
new file mode 100644
index 0000000..7ca9d56
--- /dev/null
+++ b/results/resnet_crossover/r2/resnet-r2-clean_kp-d32-lr0p1.json
@@ -0,0 +1,4001 @@
+{
+  "architecture": {
+    "adaptive_apical_parameters": 0,
+    "adaptive_feedback_parameters": 461440,
+    "base_width": 16,
+    "blocks_per_stage": 5,
+    "bn_eps": 1e-05,
+    "bn_momentum": 0.1,
+    "depth": 32,
+    "family": "CIFAR 6n+2 ResNet, option-A shortcuts",
+    "fixed_feedback_parameters": 0,
+    "fixed_traffic_coefficients": 0,
+    "forward_parameters": 464154,
+    "hidden_shapes": [
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ]
+    ],
+    "normalization": "batchnorm",
+    "predictor_parameters": 0,
+    "residual_scale": 1.0,
+    "vectorizer_mode": null,
+    "vectorizer_parameters": 0
+  },
+  "args": {
+    "a_scale": 1.0,
+    "a_warmup_steps": 0,
+    "alignment_probe": 32,
+    "apical_calibration_mode": "unit_targets",
+    "apical_seed": null,
+    "augment_train": 1,
+    "batch_size": 128,
+    "bn_eps": 1e-05,
+    "bn_momentum": 0.1,
+    "data_dir": "/orion/u/oscarwan/scratch/sdil-data",
+    "depth": 32,
+    "device": "cuda",
+    "epochs": 200,
+    "eta_A": 0.01,
+    "eta_P": 0.01,
+    "eval_every": 1,
+    "eval_split": "validation",
+    "learn_P": 0,
+    "loader_seed": 0,
+    "lr": 0.1,
+    "lr_gamma": 0.1,
+    "lr_milestones": "100,150",
+    "lr_schedule": "step",
+    "max_steps": 0,
+    "mirror_batch_size": 1,
+    "mirror_eta": 0.1,
+    "mirror_every": 16,
+    "mirror_noise_std": 1.0,
+    "mirror_seed": 3000,
+    "mirror_warmup_steps": 0,
+    "mode": "kp",
+    "momentum": 0.9,
+    "neutral_projection": 0,
+    "normalization": "batchnorm",
+    "nuisance_scale": 0.0,
+    "out": "/orion/u/oscarwan/scratch/sdil-work/sdil-a6000-full-rerun/results/resnet_crossover/r2/resnet-r2-clean_kp-d32-lr0p1.json",
+    "output_lr": 0.1,
+    "pert_directions": 1,
+    "pert_every": 4,
+    "pert_sigma": 0.01,
+    "perturb_seed": 1000,
+    "predictor_every": 0,
+    "predictor_mode": "nlms",
+    "predictor_warmup_steps": 0,
+    "residual_scale": null,
+    "seed": 0,
+    "split_seed": 2027,
+    "traffic_calibration_examples": 64,
+    "traffic_ratio": 4.0,
+    "traffic_rule": "innovation",
+    "traffic_seed": 4000,
+    "train_limit": 0,
+    "use_residual": 1,
+    "val_examples": 5000,
+    "vectorizer_mode": "spatial_template",
+    "warmup_epochs": 5,
+    "weight_decay": 0.0001,
+    "weight_scale": 1.0,
+    "width": 16
+  },
+  "calibration_metric_space": "reciprocal_local_activity_products",
+  "counters": {
+    "apical_warmup_examples": 0,
+    "calibration_event_examples": 0,
+    "causal_scalar_observations": 0,
+    "logical_batch_loss_queries": 0,
+    "mirror_conv_examples": 0,
+    "mirror_events": 0,
+    "mirror_readout_examples": 0,
+    "neutral_projection_examples": 0,
+    "ordinary_examples": 9000000,
+    "per_example_loss_terms": 0,
+    "perturbation_events": 0,
+    "perturbation_forward_examples": 0,
+    "predictor_update_examples": 0,
+    "predictor_warmup_examples": 0,
+    "traffic_audit_examples": 0,
+    "traffic_calibration_examples": 0
+  },
+  "diagnostics": {
+    "early_third_mean": 0.9990160942077637,
+    "feedback_forward_cosine": [
+      0.9995843172073364,
+      0.9995485544204712,
+      0.9993047714233398,
+      0.9992356300354004,
+      0.999225914478302,
+      0.9990967512130737,
+      0.9992426037788391,
+      0.9990879893302917,
+      0.9994840621948242,
+      0.9992990493774414,
+      0.999346911907196,
+      0.9995205402374268,
+      0.9992691278457642,
+      0.9992083311080933,
+      0.9994052648544312,
+      0.9992618560791016,
+      0.9994874000549316,
+      0.9992926120758057,
+      0.999449610710144,
+      0.9992502331733704,
+      0.999462366104126,
+      0.9996726512908936,
+      0.9996097087860107,
+      0.9995791912078857,
+      0.9996154308319092,
+      0.9994989633560181,
+      0.9995627403259277,
+      0.999337375164032,
+      0.9994921088218689,
+      0.9988247156143188,
+      0.9999656677246094
+    ],
+    "feedback_forward_norm_ratio": [
+      0.9991628527641296,
+      1.001270055770874,
+      0.9982960224151611,
+      1.0040854215621948,
+      0.9988148212432861,
+      1.0012292861938477,
+      0.9982872605323792,
+      1.0016695261001587,
+      0.9984201192855835,
+      1.0020995140075684,
+      0.9984967708587646,
+      1.000747799873352,
+      0.9983559846878052,
+      1.0024127960205078,
+      0.9985647797584534,
+      1.0024219751358032,
+      0.9993768334388733,
+      1.0025209188461304,
+      0.9989345669746399,
+      1.0021547079086304,
+      0.9993236064910889,
+      1.000596046447754,
+      0.9998040199279785,
+      1.0014662742614746,
+      1.0001678466796875,
+      1.0017902851104736,
+      0.9997692108154297,
+      1.0016337633132935,
+      0.9996170401573181,
+      1.0019867420196533,
+      1.005363941192627
+    ],
+    "feedback_forward_relative_error": [
+      0.028834059834480286,
+      0.03009512461721897,
+      0.037293966859579086,
+      0.03939209133386612,
+      0.03934045135974884,
+      0.04254467412829399,
+      0.038926754146814346,
+      0.04277598485350609,
+      0.03213443234562874,
+      0.03754224628210068,
+      0.036144670099020004,
+      0.030987465754151344,
+      0.038239117711782455,
+      0.039911918342113495,
+      0.03449464589357376,
+      0.038545139133930206,
+      0.03201395645737648,
+      0.037745725363492966,
+      0.03317772224545479,
+      0.03882520645856857,
+      0.03278729319572449,
+      0.02560167759656906,
+      0.027935020625591278,
+      0.02906847931444645,
+      0.027740592136979103,
+      0.03173808380961418,
+      0.0295693501830101,
+      0.03646891936659813,
+      0.03187037259340286,
+      0.04857011139392853,
+      0.009890343062579632
+    ],
+    "innovation_negative_gradient_cosine": [
+      0.999089241027832,
+      0.9990065097808838,
+      0.9990898370742798,
+      0.9988292455673218,
+      0.999106764793396,
+      0.9988035559654236,
+      0.9991393089294434,
+      0.9989110231399536,
+      0.999163031578064,
+      0.9990224242210388,
+      0.9991826415061951,
+      0.9991794228553772,
+      0.9992508888244629,
+      0.9989583492279053,
+      0.9992785453796387,
+      0.9990521669387817,
+      0.9993072748184204,
+      0.9991351366043091,
+      0.9993723630905151,
+      0.9991949796676636,
+      0.9994121789932251,
+      0.999466598033905,
+      0.9995043873786926,
+      0.9994280338287354,
+      0.999567985534668,
+      0.9995042085647583,
+      0.9996387362480164,
+      0.9997044801712036,
+      0.999808669090271,
+      0.9998710751533508,
+      0.7805278897285461
+    ],
+    "mean_feedback_forward_cosine": 0.9993942726042963,
+    "mean_feedback_forward_relative_error": 0.034200180532230486,
+    "normalization_state": "training_batch_stats_without_running_update",
+    "raw_negative_gradient_cosine": [
+      0.999089241027832,
+      0.9990065097808838,
+      0.9990898370742798,
+      0.9988292455673218,
+      0.999106764793396,
+      0.9988035559654236,
+      0.9991393089294434,
+      0.9989110231399536,
+      0.999163031578064,
+      0.9990224242210388,
+      0.9991826415061951,
+      0.9991794228553772,
+      0.9992508888244629,
+      0.9989583492279053,
+      0.9992785453796387,
+      0.9990521669387817,
+      0.9993072748184204,
+      0.9991351366043091,
+      0.9993723630905151,
+      0.9991949796676636,
+      0.9994121789932251,
+      0.999466598033905,
+      0.9995043873786926,
+      0.9994280338287354,
+      0.999567985534668,
+      0.9995042085647583,
+      0.9996387362480164,
+      0.9997044801712036,
+      0.999808669090271,
+      0.9998710751533508,
+      0.7805278897285461
+    ],
+    "teaching_negative_gradient_cosine": [
+      0.999089241027832,
+      0.9990065097808838,
+      0.9990898370742798,
+      0.9988292455673218,
+      0.999106764793396,
+      0.9988035559654236,
+      0.9991393089294434,
+      0.9989110231399536,
+      0.999163031578064,
+      0.9990224242210388,
+      0.9991826415061951,
+      0.9991794228553772,
+      0.9992508888244629,
+      0.9989583492279053,
+      0.9992785453796387,
+      0.9990521669387817,
+      0.9993072748184204,
+      0.9991351366043091,
+      0.9993723630905151,
+      0.9991949796676636,
+      0.9994121789932251,
+      0.999466598033905,
+      0.9995043873786926,
+      0.9994280338287354,
+      0.999567985534668,
+      0.9995042085647583,
+      0.9996387362480164,
+      0.9997044801712036,
+      0.999808669090271,
+      0.9998710751533508,
+      0.7805278897285461
+    ],
+    "wall_s": 0.08741545677185059
+  },
+  "epochs": [
+    {
+      "epoch": 1,
+      "eval_accuracy": 0.3352,
+      "eval_loss": 1.85539453125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3970344066619873,
+        "mean_feedback_forward_cosine": 0.09891654959609432,
+        "mean_feedback_forward_relative_error": 1.3545093651740783,
+        "min_feedback_forward_cosine": 0.030816495418548584
+      },
+      "lr": 0.02,
+      "output_lr": 0.02,
+      "step": 352,
+      "train_examples": 45000,
+      "train_loss": 2.0165474932140772
+    },
+    {
+      "epoch": 2,
+      "eval_accuracy": 0.474,
+      "eval_loss": 1.5613578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3619446754455566,
+        "mean_feedback_forward_cosine": 0.21640455626672314,
+        "mean_feedback_forward_relative_error": 1.2635457823353429,
+        "min_feedback_forward_cosine": 0.08645570278167725
+      },
+      "lr": 0.04,
+      "output_lr": 0.04,
+      "step": 704,
+      "train_examples": 45000,
+      "train_loss": 1.574120481385125
+    },
+    {
+      "epoch": 3,
+      "eval_accuracy": 0.5534,
+      "eval_loss": 1.43058212890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.31855046749115,
+        "mean_feedback_forward_cosine": 0.33052211519210567,
+        "mean_feedback_forward_relative_error": 1.1678868647544616,
+        "min_feedback_forward_cosine": 0.15054291486740112
+      },
+      "lr": 0.06000000000000001,
+      "output_lr": 0.06000000000000001,
+      "step": 1056,
+      "train_examples": 45000,
+      "train_loss": 1.2379753240161473
+    },
+    {
+      "epoch": 4,
+      "eval_accuracy": 0.6548,
+      "eval_loss": 1.0312754150390624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.2712821960449219,
+        "mean_feedback_forward_cosine": 0.42535690578722185,
+        "mean_feedback_forward_relative_error": 1.0820181158281141,
+        "min_feedback_forward_cosine": 0.21556095778942108
+      },
+      "lr": 0.08,
+      "output_lr": 0.08,
+      "step": 1408,
+      "train_examples": 45000,
+      "train_loss": 0.9983540509329902
+    },
+    {
+      "epoch": 5,
+      "eval_accuracy": 0.6956,
+      "eval_loss": 0.9139252319335938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.2214312553405762,
+        "mean_feedback_forward_cosine": 0.5065497131116928,
+        "mean_feedback_forward_relative_error": 1.0019216018338357,
+        "min_feedback_forward_cosine": 0.2794457972049713
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 1760,
+      "train_examples": 45000,
+      "train_loss": 0.8451392686208089
+    },
+    {
+      "epoch": 6,
+      "eval_accuracy": 0.7304,
+      "eval_loss": 0.777873779296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1813551187515259,
+        "mean_feedback_forward_cosine": 0.559566403588941,
+        "mean_feedback_forward_relative_error": 0.9465686207817446,
+        "min_feedback_forward_cosine": 0.32730311155319214
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2112,
+      "train_examples": 45000,
+      "train_loss": 0.7229798585573832
+    },
+    {
+      "epoch": 7,
+      "eval_accuracy": 0.7702,
+      "eval_loss": 0.6828182861328125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1461174488067627,
+        "mean_feedback_forward_cosine": 0.6001128660094354,
+        "mean_feedback_forward_relative_error": 0.9015945369197477,
+        "min_feedback_forward_cosine": 0.36858493089675903
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2464,
+      "train_examples": 45000,
+      "train_loss": 0.6540972160975138
+    },
+    {
+      "epoch": 8,
+      "eval_accuracy": 0.747,
+      "eval_loss": 0.7888611328125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1137171983718872,
+        "mean_feedback_forward_cosine": 0.634972870349884,
+        "mean_feedback_forward_relative_error": 0.8613815144185097,
+        "min_feedback_forward_cosine": 0.4044458270072937
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2816,
+      "train_examples": 45000,
+      "train_loss": 0.6051028917206658
+    },
+    {
+      "epoch": 9,
+      "eval_accuracy": 0.755,
+      "eval_loss": 0.8275123046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0851657390594482,
+        "mean_feedback_forward_cosine": 0.6635892391204834,
+        "mean_feedback_forward_relative_error": 0.8266697622114613,
+        "min_feedback_forward_cosine": 0.4357264041900635
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3168,
+      "train_examples": 45000,
+      "train_loss": 0.5544356402185228
+    },
+    {
+      "epoch": 10,
+      "eval_accuracy": 0.7566,
+      "eval_loss": 0.7798774658203125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0581560134887695,
+        "mean_feedback_forward_cosine": 0.689135649511891,
+        "mean_feedback_forward_relative_error": 0.7944988915997167,
+        "min_feedback_forward_cosine": 0.4646669626235962
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3520,
+      "train_examples": 45000,
+      "train_loss": 0.5214713230662875
+    },
+    {
+      "epoch": 11,
+      "eval_accuracy": 0.8146,
+      "eval_loss": 0.5679540466308594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0313493013381958,
+        "mean_feedback_forward_cosine": 0.7127979820774447,
+        "mean_feedback_forward_relative_error": 0.76344284222972,
+        "min_feedback_forward_cosine": 0.4922131299972534
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3872,
+      "train_examples": 45000,
+      "train_loss": 0.4940733931541443
+    },
+    {
+      "epoch": 12,
+      "eval_accuracy": 0.8034,
+      "eval_loss": 0.5874515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0037760734558105,
+        "mean_feedback_forward_cosine": 0.7339891156842632,
+        "mean_feedback_forward_relative_error": 0.7344710971078565,
+        "min_feedback_forward_cosine": 0.5191805362701416
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4224,
+      "train_examples": 45000,
+      "train_loss": 0.4745852570427789
+    },
+    {
+      "epoch": 13,
+      "eval_accuracy": 0.772,
+      "eval_loss": 0.727918994140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9788193106651306,
+        "mean_feedback_forward_cosine": 0.7526846374234846,
+        "mean_feedback_forward_relative_error": 0.7078268095370261,
+        "min_feedback_forward_cosine": 0.5433098077774048
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4576,
+      "train_examples": 45000,
+      "train_loss": 0.447769571240743
+    },
+    {
+      "epoch": 14,
+      "eval_accuracy": 0.8288,
+      "eval_loss": 0.51325146484375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.953144371509552,
+        "mean_feedback_forward_cosine": 0.7700060529093589,
+        "mean_feedback_forward_relative_error": 0.6821638220740903,
+        "min_feedback_forward_cosine": 0.5672977566719055
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4928,
+      "train_examples": 45000,
+      "train_loss": 0.43158293512132434
+    },
+    {
+      "epoch": 15,
+      "eval_accuracy": 0.817,
+      "eval_loss": 0.5514772705078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9272802472114563,
+        "mean_feedback_forward_cosine": 0.7858748801292912,
+        "mean_feedback_forward_relative_error": 0.6579482872639933,
+        "min_feedback_forward_cosine": 0.590796709060669
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5280,
+      "train_examples": 45000,
+      "train_loss": 0.41741799857351514
+    },
+    {
+      "epoch": 16,
+      "eval_accuracy": 0.8046,
+      "eval_loss": 0.6012989379882813,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9033938050270081,
+        "mean_feedback_forward_cosine": 0.8003424771370427,
+        "mean_feedback_forward_relative_error": 0.6350040791496154,
+        "min_feedback_forward_cosine": 0.6120290756225586
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5632,
+      "train_examples": 45000,
+      "train_loss": 0.39520628892580667
+    },
+    {
+      "epoch": 17,
+      "eval_accuracy": 0.8054,
+      "eval_loss": 0.6249626220703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8786085247993469,
+        "mean_feedback_forward_cosine": 0.8139687603519808,
+        "mean_feedback_forward_relative_error": 0.6127420500401528,
+        "min_feedback_forward_cosine": 0.6332476139068604
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5984,
+      "train_examples": 45000,
+      "train_loss": 0.38424887357287935
+    },
+    {
+      "epoch": 18,
+      "eval_accuracy": 0.826,
+      "eval_loss": 0.5176603759765624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8530976176261902,
+        "mean_feedback_forward_cosine": 0.8264068634279312,
+        "mean_feedback_forward_relative_error": 0.5914711466720027,
+        "min_feedback_forward_cosine": 0.654157280921936
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6336,
+      "train_examples": 45000,
+      "train_loss": 0.3791499035358429
+    },
+    {
+      "epoch": 19,
+      "eval_accuracy": 0.844,
+      "eval_loss": 0.4981465942382812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8279763460159302,
+        "mean_feedback_forward_cosine": 0.8380863012806061,
+        "mean_feedback_forward_relative_error": 0.5709285841834161,
+        "min_feedback_forward_cosine": 0.674217939376831
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6688,
+      "train_examples": 45000,
+      "train_loss": 0.3649848249647352
+    },
+    {
+      "epoch": 20,
+      "eval_accuracy": 0.841,
+      "eval_loss": 0.4900862609863281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8037641644477844,
+        "mean_feedback_forward_cosine": 0.8487852657994917,
+        "mean_feedback_forward_relative_error": 0.5513923168182373,
+        "min_feedback_forward_cosine": 0.6925684809684753
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7040,
+      "train_examples": 45000,
+      "train_loss": 0.3498300515757667
+    },
+    {
+      "epoch": 21,
+      "eval_accuracy": 0.8254,
+      "eval_loss": 0.5741294128417969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7798997163772583,
+        "mean_feedback_forward_cosine": 0.8588206902627022,
+        "mean_feedback_forward_relative_error": 0.5324669249596135,
+        "min_feedback_forward_cosine": 0.7107298374176025
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7392,
+      "train_examples": 45000,
+      "train_loss": 0.3389130860964457
+    },
+    {
+      "epoch": 22,
+      "eval_accuracy": 0.8452,
+      "eval_loss": 0.48783974609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7567204236984253,
+        "mean_feedback_forward_cosine": 0.8682854021749189,
+        "mean_feedback_forward_relative_error": 0.5141304625618842,
+        "min_feedback_forward_cosine": 0.7275015115737915
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7744,
+      "train_examples": 45000,
+      "train_loss": 0.3313982831213209
+    },
+    {
+      "epoch": 23,
+      "eval_accuracy": 0.8432,
+      "eval_loss": 0.503389306640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7334077954292297,
+        "mean_feedback_forward_cosine": 0.8772400752190621,
+        "mean_feedback_forward_relative_error": 0.4961547774653281,
+        "min_feedback_forward_cosine": 0.7437876462936401
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8096,
+      "train_examples": 45000,
+      "train_loss": 0.32460021476745604
+    },
+    {
+      "epoch": 24,
+      "eval_accuracy": 0.8662,
+      "eval_loss": 0.4224889404296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.710907518863678,
+        "mean_feedback_forward_cosine": 0.8852154189540494,
+        "mean_feedback_forward_relative_error": 0.4794098816571697,
+        "min_feedback_forward_cosine": 0.7589513063430786
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8448,
+      "train_examples": 45000,
+      "train_loss": 0.3124096696138382
+    },
+    {
+      "epoch": 25,
+      "eval_accuracy": 0.8464,
+      "eval_loss": 0.5034165222167969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6883212327957153,
+        "mean_feedback_forward_cosine": 0.8927982334167727,
+        "mean_feedback_forward_relative_error": 0.46311643960014465,
+        "min_feedback_forward_cosine": 0.7737209796905518
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8800,
+      "train_examples": 45000,
+      "train_loss": 0.30748651559087964
+    },
+    {
+      "epoch": 26,
+      "eval_accuracy": 0.8578,
+      "eval_loss": 0.4695193481445312,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6669498682022095,
+        "mean_feedback_forward_cosine": 0.8997289615292703,
+        "mean_feedback_forward_relative_error": 0.4476435371944981,
+        "min_feedback_forward_cosine": 0.7872694730758667
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9152,
+      "train_examples": 45000,
+      "train_loss": 0.29803875002331204
+    },
+    {
+      "epoch": 27,
+      "eval_accuracy": 0.8472,
+      "eval_loss": 0.49544317626953127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6458977460861206,
+        "mean_feedback_forward_cosine": 0.9063345193862915,
+        "mean_feedback_forward_relative_error": 0.43247731029987335,
+        "min_feedback_forward_cosine": 0.80039381980896
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9504,
+      "train_examples": 45000,
+      "train_loss": 0.29557929752137924
+    },
+    {
+      "epoch": 28,
+      "eval_accuracy": 0.8522,
+      "eval_loss": 0.4892141296386719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.626682698726654,
+        "mean_feedback_forward_cosine": 0.9123510795254861,
+        "mean_feedback_forward_relative_error": 0.4181007968802606,
+        "min_feedback_forward_cosine": 0.8119982481002808
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9856,
+      "train_examples": 45000,
+      "train_loss": 0.2796997773885727
+    },
+    {
+      "epoch": 29,
+      "eval_accuracy": 0.8618,
+      "eval_loss": 0.4524398193359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6067200303077698,
+        "mean_feedback_forward_cosine": 0.9181783026264559,
+        "mean_feedback_forward_relative_error": 0.4038001053756283,
+        "min_feedback_forward_cosine": 0.823555588722229
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10208,
+      "train_examples": 45000,
+      "train_loss": 0.28144939953486126
+    },
+    {
+      "epoch": 30,
+      "eval_accuracy": 0.8616,
+      "eval_loss": 0.431435400390625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5873559713363647,
+        "mean_feedback_forward_cosine": 0.9236067668084176,
+        "mean_feedback_forward_relative_error": 0.3899670858536997,
+        "min_feedback_forward_cosine": 0.8343836069107056
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10560,
+      "train_examples": 45000,
+      "train_loss": 0.27674690407911934
+    },
+    {
+      "epoch": 31,
+      "eval_accuracy": 0.8578,
+      "eval_loss": 0.4860141845703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5684483051300049,
+        "mean_feedback_forward_cosine": 0.9285307911134535,
+        "mean_feedback_forward_relative_error": 0.3769292422840672,
+        "min_feedback_forward_cosine": 0.8449437618255615
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10912,
+      "train_examples": 45000,
+      "train_loss": 0.2692386422687107
+    },
+    {
+      "epoch": 32,
+      "eval_accuracy": 0.84,
+      "eval_loss": 0.5316449401855469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5495582222938538,
+        "mean_feedback_forward_cosine": 0.9332671357739356,
+        "mean_feedback_forward_relative_error": 0.3640456214066475,
+        "min_feedback_forward_cosine": 0.8548435568809509
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11264,
+      "train_examples": 45000,
+      "train_loss": 0.26496250779893665
+    },
+    {
+      "epoch": 33,
+      "eval_accuracy": 0.8594,
+      "eval_loss": 0.46718753662109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5314036011695862,
+        "mean_feedback_forward_cosine": 0.9376626629983226,
+        "mean_feedback_forward_relative_error": 0.351678614174166,
+        "min_feedback_forward_cosine": 0.864128053188324
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11616,
+      "train_examples": 45000,
+      "train_loss": 0.2612344945536719
+    },
+    {
+      "epoch": 34,
+      "eval_accuracy": 0.8486,
+      "eval_loss": 0.5073669921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5140032768249512,
+        "mean_feedback_forward_cosine": 0.9416962304422932,
+        "mean_feedback_forward_relative_error": 0.3399790626379751,
+        "min_feedback_forward_cosine": 0.8728002309799194
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11968,
+      "train_examples": 45000,
+      "train_loss": 0.25332926568190256
+    },
+    {
+      "epoch": 35,
+      "eval_accuracy": 0.8634,
+      "eval_loss": 0.4601386779785156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.49668073654174805,
+        "mean_feedback_forward_cosine": 0.9455604130221952,
+        "mean_feedback_forward_relative_error": 0.3283394209800228,
+        "min_feedback_forward_cosine": 0.8810791969299316
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12320,
+      "train_examples": 45000,
+      "train_loss": 0.24765972827275595
+    },
+    {
+      "epoch": 36,
+      "eval_accuracy": 0.8416,
+      "eval_loss": 0.49482080688476565,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.48018109798431396,
+        "mean_feedback_forward_cosine": 0.949139166262842,
+        "mean_feedback_forward_relative_error": 0.3172376600484694,
+        "min_feedback_forward_cosine": 0.8887827396392822
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12672,
+      "train_examples": 45000,
+      "train_loss": 0.24774828658103942
+    },
+    {
+      "epoch": 37,
+      "eval_accuracy": 0.858,
+      "eval_loss": 0.45331826782226564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4640062749385834,
+        "mean_feedback_forward_cosine": 0.9525774159739094,
+        "mean_feedback_forward_relative_error": 0.30619221156643284,
+        "min_feedback_forward_cosine": 0.8959776163101196
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13024,
+      "train_examples": 45000,
+      "train_loss": 0.24434037954939736
+    },
+    {
+      "epoch": 38,
+      "eval_accuracy": 0.858,
+      "eval_loss": 0.45240281982421876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4489477574825287,
+        "mean_feedback_forward_cosine": 0.9556983863153765,
+        "mean_feedback_forward_relative_error": 0.29581394238818076,
+        "min_feedback_forward_cosine": 0.9025242328643799
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13376,
+      "train_examples": 45000,
+      "train_loss": 0.23976409662034776
+    },
+    {
+      "epoch": 39,
+      "eval_accuracy": 0.8412,
+      "eval_loss": 0.5876475219726562,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4343760013580322,
+        "mean_feedback_forward_cosine": 0.9584913234556874,
+        "mean_feedback_forward_relative_error": 0.2862166019697343,
+        "min_feedback_forward_cosine": 0.9086928963661194
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13728,
+      "train_examples": 45000,
+      "train_loss": 0.23123540512985655
+    },
+    {
+      "epoch": 40,
+      "eval_accuracy": 0.8564,
+      "eval_loss": 0.4765188659667969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.41979628801345825,
+        "mean_feedback_forward_cosine": 0.9612213392411509,
+        "mean_feedback_forward_relative_error": 0.2765878001528402,
+        "min_feedback_forward_cosine": 0.9146870970726013
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14080,
+      "train_examples": 45000,
+      "train_loss": 0.2318395030286577
+    },
+    {
+      "epoch": 41,
+      "eval_accuracy": 0.8616,
+      "eval_loss": 0.44323438720703123,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.40531429648399353,
+        "mean_feedback_forward_cosine": 0.9637255207184823,
+        "mean_feedback_forward_relative_error": 0.26743238106850653,
+        "min_feedback_forward_cosine": 0.9204114675521851
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14432,
+      "train_examples": 45000,
+      "train_loss": 0.22619207657972973
+    },
+    {
+      "epoch": 42,
+      "eval_accuracy": 0.8578,
+      "eval_loss": 0.4867693237304688,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.391878217458725,
+        "mean_feedback_forward_cosine": 0.966075841457613,
+        "mean_feedback_forward_relative_error": 0.2584934775387087,
+        "min_feedback_forward_cosine": 0.9255408644676208
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14784,
+      "train_examples": 45000,
+      "train_loss": 0.22141475805441538
+    },
+    {
+      "epoch": 43,
+      "eval_accuracy": 0.8746,
+      "eval_loss": 0.41213989868164064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3782900869846344,
+        "mean_feedback_forward_cosine": 0.9682946012866113,
+        "mean_feedback_forward_relative_error": 0.2497686491377892,
+        "min_feedback_forward_cosine": 0.9305530786514282
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15136,
+      "train_examples": 45000,
+      "train_loss": 0.2203050850868225
+    },
+    {
+      "epoch": 44,
+      "eval_accuracy": 0.8488,
+      "eval_loss": 0.5225678588867188,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.36502012610435486,
+        "mean_feedback_forward_cosine": 0.9704237772572425,
+        "mean_feedback_forward_relative_error": 0.2411658537964667,
+        "min_feedback_forward_cosine": 0.9352737665176392
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15488,
+      "train_examples": 45000,
+      "train_loss": 0.21957209897041322
+    },
+    {
+      "epoch": 45,
+      "eval_accuracy": 0.849,
+      "eval_loss": 0.5496326416015626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.35209617018699646,
+        "mean_feedback_forward_cosine": 0.9723464500519537,
+        "mean_feedback_forward_relative_error": 0.23312005520828308,
+        "min_feedback_forward_cosine": 0.9396935701370239
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15840,
+      "train_examples": 45000,
+      "train_loss": 0.21895495686531066
+    },
+    {
+      "epoch": 46,
+      "eval_accuracy": 0.86,
+      "eval_loss": 0.4849064758300781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.33995580673217773,
+        "mean_feedback_forward_cosine": 0.9741548203652904,
+        "mean_feedback_forward_relative_error": 0.2253057812490771,
+        "min_feedback_forward_cosine": 0.9437055587768555
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16192,
+      "train_examples": 45000,
+      "train_loss": 0.21140181116792892
+    },
+    {
+      "epoch": 47,
+      "eval_accuracy": 0.8596,
+      "eval_loss": 0.48711629028320313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.32750967144966125,
+        "mean_feedback_forward_cosine": 0.9759039148207633,
+        "mean_feedback_forward_relative_error": 0.21750406536363787,
+        "min_feedback_forward_cosine": 0.9477157592773438
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16544,
+      "train_examples": 45000,
+      "train_loss": 0.21739240936703152
+    },
+    {
+      "epoch": 48,
+      "eval_accuracy": 0.8674,
+      "eval_loss": 0.49024164428710937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3152298629283905,
+        "mean_feedback_forward_cosine": 0.9775446115001556,
+        "mean_feedback_forward_relative_error": 0.20989031272549782,
+        "min_feedback_forward_cosine": 0.9515104293823242
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16896,
+      "train_examples": 45000,
+      "train_loss": 0.21653300149705676
+    },
+    {
+      "epoch": 49,
+      "eval_accuracy": 0.8672,
+      "eval_loss": 0.4632137329101563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.30373454093933105,
+        "mean_feedback_forward_cosine": 0.9790571889569682,
+        "mean_feedback_forward_relative_error": 0.20261672836157582,
+        "min_feedback_forward_cosine": 0.954938530921936
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17248,
+      "train_examples": 45000,
+      "train_loss": 0.20910456528663635
+    },
+    {
+      "epoch": 50,
+      "eval_accuracy": 0.8686,
+      "eval_loss": 0.44734768676757813,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2930586338043213,
+        "mean_feedback_forward_cosine": 0.980450362928452,
+        "mean_feedback_forward_relative_error": 0.19571300307589193,
+        "min_feedback_forward_cosine": 0.958027184009552
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17600,
+      "train_examples": 45000,
+      "train_loss": 0.20682951056427426
+    },
+    {
+      "epoch": 51,
+      "eval_accuracy": 0.865,
+      "eval_loss": 0.4583529357910156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.28304439783096313,
+        "mean_feedback_forward_cosine": 0.9817609517805038,
+        "mean_feedback_forward_relative_error": 0.18900507039600803,
+        "min_feedback_forward_cosine": 0.9608170986175537
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17952,
+      "train_examples": 45000,
+      "train_loss": 0.20303286356396144
+    },
+    {
+      "epoch": 52,
+      "eval_accuracy": 0.8758,
+      "eval_loss": 0.42540660400390623,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.27335530519485474,
+        "mean_feedback_forward_cosine": 0.9829263110314647,
+        "mean_feedback_forward_relative_error": 0.18280399518628274,
+        "min_feedback_forward_cosine": 0.9634319543838501
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18304,
+      "train_examples": 45000,
+      "train_loss": 0.1975233178668552
+    },
+    {
+      "epoch": 53,
+      "eval_accuracy": 0.8726,
+      "eval_loss": 0.45839552612304685,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2639857232570648,
+        "mean_feedback_forward_cosine": 0.9840461400247389,
+        "mean_feedback_forward_relative_error": 0.1766628747986209,
+        "min_feedback_forward_cosine": 0.9658682346343994
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18656,
+      "train_examples": 45000,
+      "train_loss": 0.19599860397444832
+    },
+    {
+      "epoch": 54,
+      "eval_accuracy": 0.854,
+      "eval_loss": 0.5225734069824218,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.25491610169410706,
+        "mean_feedback_forward_cosine": 0.9850916497169002,
+        "mean_feedback_forward_relative_error": 0.17073152082100992,
+        "min_feedback_forward_cosine": 0.9681482911109924
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19008,
+      "train_examples": 45000,
+      "train_loss": 0.19324108230802747
+    },
+    {
+      "epoch": 55,
+      "eval_accuracy": 0.8656,
+      "eval_loss": 0.4879412841796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.24540641903877258,
+        "mean_feedback_forward_cosine": 0.9861299530152352,
+        "mean_feedback_forward_relative_error": 0.1646571411721168,
+        "min_feedback_forward_cosine": 0.9704469442367554
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19360,
+      "train_examples": 45000,
+      "train_loss": 0.20186779841846889
+    },
+    {
+      "epoch": 56,
+      "eval_accuracy": 0.8788,
+      "eval_loss": 0.4124189697265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.23672057688236237,
+        "mean_feedback_forward_cosine": 0.9870482760090982,
+        "mean_feedback_forward_relative_error": 0.15906442100963286,
+        "min_feedback_forward_cosine": 0.9724932312965393
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19712,
+      "train_examples": 45000,
+      "train_loss": 0.1911707753446367
+    },
+    {
+      "epoch": 57,
+      "eval_accuracy": 0.8544,
+      "eval_loss": 0.4946718139648438,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.22838672995567322,
+        "mean_feedback_forward_cosine": 0.9879102610772655,
+        "mean_feedback_forward_relative_error": 0.1536721505465046,
+        "min_feedback_forward_cosine": 0.9743702411651611
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20064,
+      "train_examples": 45000,
+      "train_loss": 0.19003599887424044
+    },
+    {
+      "epoch": 58,
+      "eval_accuracy": 0.8834,
+      "eval_loss": 0.4138064086914062,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.22036190330982208,
+        "mean_feedback_forward_cosine": 0.9887177309682292,
+        "mean_feedback_forward_relative_error": 0.14841678906832972,
+        "min_feedback_forward_cosine": 0.9761249423027039
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20416,
+      "train_examples": 45000,
+      "train_loss": 0.19302229064305623
+    },
+    {
+      "epoch": 59,
+      "eval_accuracy": 0.862,
+      "eval_loss": 0.49976304931640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.21271266043186188,
+        "mean_feedback_forward_cosine": 0.9894644348852096,
+        "mean_feedback_forward_relative_error": 0.14339836915173837,
+        "min_feedback_forward_cosine": 0.9777436852455139
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20768,
+      "train_examples": 45000,
+      "train_loss": 0.1886476122485267
+    },
+    {
+      "epoch": 60,
+      "eval_accuracy": 0.886,
+      "eval_loss": 0.3878099365234375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.20517779886722565,
+        "mean_feedback_forward_cosine": 0.9901608478638434,
+        "mean_feedback_forward_relative_error": 0.1385545569562143,
+        "min_feedback_forward_cosine": 0.9792742133140564
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21120,
+      "train_examples": 45000,
+      "train_loss": 0.184096456334326
+    },
+    {
+      "epoch": 61,
+      "eval_accuracy": 0.87,
+      "eval_loss": 0.45004430541992185,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1976340413093567,
+        "mean_feedback_forward_cosine": 0.9908320807641552,
+        "mean_feedback_forward_relative_error": 0.13373789431587343,
+        "min_feedback_forward_cosine": 0.9807661771774292
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21472,
+      "train_examples": 45000,
+      "train_loss": 0.1952857186794281
+    },
+    {
+      "epoch": 62,
+      "eval_accuracy": 0.8708,
+      "eval_loss": 0.42280288696289064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.19097411632537842,
+        "mean_feedback_forward_cosine": 0.9914130280094762,
+        "mean_feedback_forward_relative_error": 0.12940076714561832,
+        "min_feedback_forward_cosine": 0.9820346832275391
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21824,
+      "train_examples": 45000,
+      "train_loss": 0.1781928380449613
+    },
+    {
+      "epoch": 63,
+      "eval_accuracy": 0.8804,
+      "eval_loss": 0.40265328369140624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.18434743583202362,
+        "mean_feedback_forward_cosine": 0.9919711351394653,
+        "mean_feedback_forward_relative_error": 0.125099069769344,
+        "min_feedback_forward_cosine": 0.9832502007484436
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22176,
+      "train_examples": 45000,
+      "train_loss": 0.18024166654480828
+    },
+    {
+      "epoch": 64,
+      "eval_accuracy": 0.8768,
+      "eval_loss": 0.4114610107421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.17806272208690643,
+        "mean_feedback_forward_cosine": 0.9924912568061582,
+        "mean_feedback_forward_relative_error": 0.12095758090576818,
+        "min_feedback_forward_cosine": 0.9843647480010986
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22528,
+      "train_examples": 45000,
+      "train_loss": 0.17777254914310242
+    },
+    {
+      "epoch": 65,
+      "eval_accuracy": 0.8712,
+      "eval_loss": 0.4435559143066406,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1717073768377304,
+        "mean_feedback_forward_cosine": 0.9929938431709043,
+        "mean_feedback_forward_relative_error": 0.1168181098276569,
+        "min_feedback_forward_cosine": 0.9854501485824585
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22880,
+      "train_examples": 45000,
+      "train_loss": 0.18547025267018213
+    },
+    {
+      "epoch": 66,
+      "eval_accuracy": 0.859,
+      "eval_loss": 0.5121481506347656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.16578994691371918,
+        "mean_feedback_forward_cosine": 0.9934470845806983,
+        "mean_feedback_forward_relative_error": 0.11296291853631696,
+        "min_feedback_forward_cosine": 0.986427903175354
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23232,
+      "train_examples": 45000,
+      "train_loss": 0.1755339354938931
+    },
+    {
+      "epoch": 67,
+      "eval_accuracy": 0.855,
+      "eval_loss": 0.5351258056640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.15990713238716125,
+        "mean_feedback_forward_cosine": 0.993884357713884,
+        "mean_feedback_forward_relative_error": 0.10910652541825848,
+        "min_feedback_forward_cosine": 0.9873668551445007
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23584,
+      "train_examples": 45000,
+      "train_loss": 0.1767852502849367
+    },
+    {
+      "epoch": 68,
+      "eval_accuracy": 0.8736,
+      "eval_loss": 0.44542548828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.15426933765411377,
+        "mean_feedback_forward_cosine": 0.9942979985667814,
+        "mean_feedback_forward_relative_error": 0.10534040029010465,
+        "min_feedback_forward_cosine": 0.9882389307022095
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23936,
+      "train_examples": 45000,
+      "train_loss": 0.17708395219379
+    },
+    {
+      "epoch": 69,
+      "eval_accuracy": 0.8866,
+      "eval_loss": 0.37270454711914064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14849302172660828,
+        "mean_feedback_forward_cosine": 0.9946968151677039,
+        "mean_feedback_forward_relative_error": 0.10158244051760243,
+        "min_feedback_forward_cosine": 0.9890974164009094
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24288,
+      "train_examples": 45000,
+      "train_loss": 0.18407729478412205
+    },
+    {
+      "epoch": 70,
+      "eval_accuracy": 0.8848,
+      "eval_loss": 0.3816723022460938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14344274997711182,
+        "mean_feedback_forward_cosine": 0.9950423971299203,
+        "mean_feedback_forward_relative_error": 0.09819682723572178,
+        "min_feedback_forward_cosine": 0.989820122718811
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24640,
+      "train_examples": 45000,
+      "train_loss": 0.17256281439993115
+    },
+    {
+      "epoch": 71,
+      "eval_accuracy": 0.871,
+      "eval_loss": 0.47614798583984375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.13826242089271545,
+        "mean_feedback_forward_cosine": 0.9953846220047243,
+        "mean_feedback_forward_relative_error": 0.09473993376858773,
+        "min_feedback_forward_cosine": 0.9905393719673157
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24992,
+      "train_examples": 45000,
+      "train_loss": 0.17987764293617672
+    },
+    {
+      "epoch": 72,
+      "eval_accuracy": 0.8598,
+      "eval_loss": 0.46955314331054687,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.13329215347766876,
+        "mean_feedback_forward_cosine": 0.9956990288149926,
+        "mean_feedback_forward_relative_error": 0.09145044511364352,
+        "min_feedback_forward_cosine": 0.991203784942627
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25344,
+      "train_examples": 45000,
+      "train_loss": 0.17530702959431543
+    },
+    {
+      "epoch": 73,
+      "eval_accuracy": 0.8896,
+      "eval_loss": 0.37162421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.12890847027301788,
+        "mean_feedback_forward_cosine": 0.99597418115985,
+        "mean_feedback_forward_relative_error": 0.08846384130658642,
+        "min_feedback_forward_cosine": 0.9917711019515991
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25696,
+      "train_examples": 45000,
+      "train_loss": 0.16974435313542685
+    },
+    {
+      "epoch": 74,
+      "eval_accuracy": 0.8748,
+      "eval_loss": 0.4273910888671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.12445683777332306,
+        "mean_feedback_forward_cosine": 0.9962359916779303,
+        "mean_feedback_forward_relative_error": 0.0855239714465795,
+        "min_feedback_forward_cosine": 0.9923257827758789
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26048,
+      "train_examples": 45000,
+      "train_loss": 0.16758987526098887
+    },
+    {
+      "epoch": 75,
+      "eval_accuracy": 0.8744,
+      "eval_loss": 0.44162723388671876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1202189177274704,
+        "mean_feedback_forward_cosine": 0.996487186801049,
+        "mean_feedback_forward_relative_error": 0.08261592692184833,
+        "min_feedback_forward_cosine": 0.9928371906280518
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26400,
+      "train_examples": 45000,
+      "train_loss": 0.16726809644434187
+    },
+    {
+      "epoch": 76,
+      "eval_accuracy": 0.8748,
+      "eval_loss": 0.4214301574707031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11572326719760895,
+        "mean_feedback_forward_cosine": 0.9967352132643422,
+        "mean_feedback_forward_relative_error": 0.07963655528522307,
+        "min_feedback_forward_cosine": 0.993360161781311
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26752,
+      "train_examples": 45000,
+      "train_loss": 0.1751393345170551
+    },
+    {
+      "epoch": 77,
+      "eval_accuracy": 0.8956,
+      "eval_loss": 0.34109723510742185,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11186333745718002,
+        "mean_feedback_forward_cosine": 0.9969496880808184,
+        "mean_feedback_forward_relative_error": 0.07697323192992518,
+        "min_feedback_forward_cosine": 0.9937939643859863
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27104,
+      "train_examples": 45000,
+      "train_loss": 0.16733670739597745
+    },
+    {
+      "epoch": 78,
+      "eval_accuracy": 0.872,
+      "eval_loss": 0.44417415771484375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.10790038108825684,
+        "mean_feedback_forward_cosine": 0.9971546011586343,
+        "mean_feedback_forward_relative_error": 0.07434488015790139,
+        "min_feedback_forward_cosine": 0.9942234754562378
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27456,
+      "train_examples": 45000,
+      "train_loss": 0.17388162036736807
+    },
+    {
+      "epoch": 79,
+      "eval_accuracy": 0.8782,
+      "eval_loss": 0.438529736328125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.10394710302352905,
+        "mean_feedback_forward_cosine": 0.997349654474566,
+        "mean_feedback_forward_relative_error": 0.07174537811548479,
+        "min_feedback_forward_cosine": 0.9946376085281372
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27808,
+      "train_examples": 45000,
+      "train_loss": 0.1652422828303443
+    },
+    {
+      "epoch": 80,
+      "eval_accuracy": 0.8816,
+      "eval_loss": 0.44652198486328126,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.10049350559711456,
+        "mean_feedback_forward_cosine": 0.9975251978443515,
+        "mean_feedback_forward_relative_error": 0.06931953246314679,
+        "min_feedback_forward_cosine": 0.9949864745140076
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28160,
+      "train_examples": 45000,
+      "train_loss": 0.16283351455264622
+    },
+    {
+      "epoch": 81,
+      "eval_accuracy": 0.8752,
+      "eval_loss": 0.4451551513671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09698614478111267,
+        "mean_feedback_forward_cosine": 0.9976857150754621,
+        "mean_feedback_forward_relative_error": 0.06702608723313577,
+        "min_feedback_forward_cosine": 0.9953293800354004
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28512,
+      "train_examples": 45000,
+      "train_loss": 0.1627018078883489
+    },
+    {
+      "epoch": 82,
+      "eval_accuracy": 0.8712,
+      "eval_loss": 0.4183915771484375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09360911697149277,
+        "mean_feedback_forward_cosine": 0.9978386798212605,
+        "mean_feedback_forward_relative_error": 0.06476938940825001,
+        "min_feedback_forward_cosine": 0.9956478476524353
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28864,
+      "train_examples": 45000,
+      "train_loss": 0.1625231981145011
+    },
+    {
+      "epoch": 83,
+      "eval_accuracy": 0.864,
+      "eval_loss": 0.479559423828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09036646038293839,
+        "mean_feedback_forward_cosine": 0.9979804735029897,
+        "mean_feedback_forward_relative_error": 0.06260339933778009,
+        "min_feedback_forward_cosine": 0.9959428310394287
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29216,
+      "train_examples": 45000,
+      "train_loss": 0.15952994990878636
+    },
+    {
+      "epoch": 84,
+      "eval_accuracy": 0.8664,
+      "eval_loss": 0.5053081115722656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08696158975362778,
+        "mean_feedback_forward_cosine": 0.9981216576791578,
+        "mean_feedback_forward_relative_error": 0.06037566810846329,
+        "min_feedback_forward_cosine": 0.9962424039840698
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29568,
+      "train_examples": 45000,
+      "train_loss": 0.16951513698101045
+    },
+    {
+      "epoch": 85,
+      "eval_accuracy": 0.878,
+      "eval_loss": 0.46674307861328124,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08394859731197357,
+        "mean_feedback_forward_cosine": 0.9982455045946182,
+        "mean_feedback_forward_relative_error": 0.05834392362063931,
+        "min_feedback_forward_cosine": 0.996497392654419
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29920,
+      "train_examples": 45000,
+      "train_loss": 0.16335849794281854
+    },
+    {
+      "epoch": 86,
+      "eval_accuracy": 0.8914,
+      "eval_loss": 0.3704887084960938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08105872571468353,
+        "mean_feedback_forward_cosine": 0.9983624296803628,
+        "mean_feedback_forward_relative_error": 0.05636267388059247,
+        "min_feedback_forward_cosine": 0.9967339634895325
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30272,
+      "train_examples": 45000,
+      "train_loss": 0.15780939542982314
+    },
+    {
+      "epoch": 87,
+      "eval_accuracy": 0.8826,
+      "eval_loss": 0.3811180114746094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0780579075217247,
+        "mean_feedback_forward_cosine": 0.9984782568870052,
+        "mean_feedback_forward_relative_error": 0.054333017117554144,
+        "min_feedback_forward_cosine": 0.9969696998596191
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30624,
+      "train_examples": 45000,
+      "train_loss": 0.16681171383327908
+    },
+    {
+      "epoch": 88,
+      "eval_accuracy": 0.8872,
+      "eval_loss": 0.4038883117675781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.07522793114185333,
+        "mean_feedback_forward_cosine": 0.998582543865327,
+        "mean_feedback_forward_relative_error": 0.05243379346305324,
+        "min_feedback_forward_cosine": 0.9971845746040344
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30976,
+      "train_examples": 45000,
+      "train_loss": 0.16635543229844835
+    },
+    {
+      "epoch": 89,
+      "eval_accuracy": 0.8768,
+      "eval_loss": 0.4656008239746094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.07262639701366425,
+        "mean_feedback_forward_cosine": 0.9986755482612117,
+        "mean_feedback_forward_relative_error": 0.05068007844590371,
+        "min_feedback_forward_cosine": 0.9973758459091187
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31328,
+      "train_examples": 45000,
+      "train_loss": 0.15748041656017303
+    },
+    {
+      "epoch": 90,
+      "eval_accuracy": 0.8846,
+      "eval_loss": 0.4177832824707031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06999342143535614,
+        "mean_feedback_forward_cosine": 0.9987672355867201,
+        "mean_feedback_forward_relative_error": 0.048894857086481586,
+        "min_feedback_forward_cosine": 0.9975628852844238
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31680,
+      "train_examples": 45000,
+      "train_loss": 0.16211600266562567
+    },
+    {
+      "epoch": 91,
+      "eval_accuracy": 0.8716,
+      "eval_loss": 0.4431572265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06744895130395889,
+        "mean_feedback_forward_cosine": 0.9988512204539391,
+        "mean_feedback_forward_relative_error": 0.04720152908515546,
+        "min_feedback_forward_cosine": 0.997735857963562
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32032,
+      "train_examples": 45000,
+      "train_loss": 0.16203772835466596
+    },
+    {
+      "epoch": 92,
+      "eval_accuracy": 0.8826,
+      "eval_loss": 0.40832467041015624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06500688195228577,
+        "mean_feedback_forward_cosine": 0.9989292698521768,
+        "mean_feedback_forward_relative_error": 0.04557065719798688,
+        "min_feedback_forward_cosine": 0.9978965520858765
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32384,
+      "train_examples": 45000,
+      "train_loss": 0.15830527657402887
+    },
+    {
+      "epoch": 93,
+      "eval_accuracy": 0.8902,
+      "eval_loss": 0.37456282958984377,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06275323033332825,
+        "mean_feedback_forward_cosine": 0.9990006685256958,
+        "mean_feedback_forward_relative_error": 0.04402178369702831,
+        "min_feedback_forward_cosine": 0.9980393648147583
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32736,
+      "train_examples": 45000,
+      "train_loss": 0.1551544521861606
+    },
+    {
+      "epoch": 94,
+      "eval_accuracy": 0.8852,
+      "eval_loss": 0.43639714965820314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0605853907763958,
+        "mean_feedback_forward_cosine": 0.9990669758089127,
+        "mean_feedback_forward_relative_error": 0.04253085158885487,
+        "min_feedback_forward_cosine": 0.998172402381897
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33088,
+      "train_examples": 45000,
+      "train_loss": 0.15768385762373607
+    },
+    {
+      "epoch": 95,
+      "eval_accuracy": 0.8828,
+      "eval_loss": 0.42318448486328125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.05844554677605629,
+        "mean_feedback_forward_cosine": 0.9991313084479301,
+        "mean_feedback_forward_relative_error": 0.04103709913549885,
+        "min_feedback_forward_cosine": 0.9982991218566895
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33440,
+      "train_examples": 45000,
+      "train_loss": 0.1587442197839419
+    },
+    {
+      "epoch": 96,
+      "eval_accuracy": 0.8878,
+      "eval_loss": 0.4036041809082031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.05644100531935692,
+        "mean_feedback_forward_cosine": 0.9991878405694039,
+        "mean_feedback_forward_relative_error": 0.03967912615306916,
+        "min_feedback_forward_cosine": 0.9984134435653687
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33792,
+      "train_examples": 45000,
+      "train_loss": 0.15544112352265252
+    },
+    {
+      "epoch": 97,
+      "eval_accuracy": 0.8914,
+      "eval_loss": 0.37009312133789063,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.054409269243478775,
+        "mean_feedback_forward_cosine": 0.9992446053412652,
+        "mean_feedback_forward_relative_error": 0.038267940523162965,
+        "min_feedback_forward_cosine": 0.9985255599021912
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34144,
+      "train_examples": 45000,
+      "train_loss": 0.15759313933319516
+    },
+    {
+      "epoch": 98,
+      "eval_accuracy": 0.88,
+      "eval_loss": 0.4166911987304687,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.05243669077754021,
+        "mean_feedback_forward_cosine": 0.9992959845450616,
+        "mean_feedback_forward_relative_error": 0.03694337375101543,
+        "min_feedback_forward_cosine": 0.9986302256584167
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34496,
+      "train_examples": 45000,
+      "train_loss": 0.1612001913335588
+    },
+    {
+      "epoch": 99,
+      "eval_accuracy": 0.8714,
+      "eval_loss": 0.4708991271972656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.05059374123811722,
+        "mean_feedback_forward_cosine": 0.9993423665723493,
+        "mean_feedback_forward_relative_error": 0.035702559405997876,
+        "min_feedback_forward_cosine": 0.9987248778343201
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34848,
+      "train_examples": 45000,
+      "train_loss": 0.15767538166840872
+    },
+    {
+      "epoch": 100,
+      "eval_accuracy": 0.8808,
+      "eval_loss": 0.450357470703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04881930351257324,
+        "mean_feedback_forward_cosine": 0.9993862932728182,
+        "mean_feedback_forward_relative_error": 0.03448866133487994,
+        "min_feedback_forward_cosine": 0.9988125562667847
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 35200,
+      "train_examples": 45000,
+      "train_loss": 0.15321411666737664
+    },
+    {
+      "epoch": 101,
+      "eval_accuracy": 0.9136,
+      "eval_loss": 0.29533384399414064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048803918063640594,
+        "mean_feedback_forward_cosine": 0.9993866758961831,
+        "mean_feedback_forward_relative_error": 0.03447543967875742,
+        "min_feedback_forward_cosine": 0.998813271522522
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35552,
+      "train_examples": 45000,
+      "train_loss": 0.09057808789942
+    },
+    {
+      "epoch": 102,
+      "eval_accuracy": 0.9196,
+      "eval_loss": 0.28288860778808594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048796601593494415,
+        "mean_feedback_forward_cosine": 0.9993869008556489,
+        "mean_feedback_forward_relative_error": 0.03446682533549686,
+        "min_feedback_forward_cosine": 0.9988136291503906
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35904,
+      "train_examples": 45000,
+      "train_loss": 0.06322400026718776
+    },
+    {
+      "epoch": 103,
+      "eval_accuracy": 0.9212,
+      "eval_loss": 0.2858207183837891,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04879026859998703,
+        "mean_feedback_forward_cosine": 0.9993870719786613,
+        "mean_feedback_forward_relative_error": 0.03445892669861355,
+        "min_feedback_forward_cosine": 0.9988139867782593
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36256,
+      "train_examples": 45000,
+      "train_loss": 0.0535766043484211
+    },
+    {
+      "epoch": 104,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.2870035583496094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04878412187099457,
+        "mean_feedback_forward_cosine": 0.9993872354107518,
+        "mean_feedback_forward_relative_error": 0.03445159391530098,
+        "min_feedback_forward_cosine": 0.9988143444061279
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36608,
+      "train_examples": 45000,
+      "train_loss": 0.04680160608887672
+    },
+    {
+      "epoch": 105,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.29164626159667967,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048778191208839417,
+        "mean_feedback_forward_cosine": 0.9993874142246861,
+        "mean_feedback_forward_relative_error": 0.034444511237163696,
+        "min_feedback_forward_cosine": 0.9988146424293518
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36960,
+      "train_examples": 45000,
+      "train_loss": 0.043014900404877134
+    },
+    {
+      "epoch": 106,
+      "eval_accuracy": 0.9218,
+      "eval_loss": 0.29402776794433594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04877280443906784,
+        "mean_feedback_forward_cosine": 0.9993875834249681,
+        "mean_feedback_forward_relative_error": 0.03443780937983144,
+        "min_feedback_forward_cosine": 0.9988148212432861
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37312,
+      "train_examples": 45000,
+      "train_loss": 0.03942505518330468
+    },
+    {
+      "epoch": 107,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.30112089538574216,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04876742511987686,
+        "mean_feedback_forward_cosine": 0.9993877507025196,
+        "mean_feedback_forward_relative_error": 0.03443125629376981,
+        "min_feedback_forward_cosine": 0.99881511926651
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37664,
+      "train_examples": 45000,
+      "train_loss": 0.035907337534758776
+    },
+    {
+      "epoch": 108,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.3071636474609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04876227304339409,
+        "mean_feedback_forward_cosine": 0.999387917980071,
+        "mean_feedback_forward_relative_error": 0.03442491798271095,
+        "min_feedback_forward_cosine": 0.9988153576850891
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38016,
+      "train_examples": 45000,
+      "train_loss": 0.03364239944087134
+    },
+    {
+      "epoch": 109,
+      "eval_accuracy": 0.922,
+      "eval_loss": 0.3126777282714844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048756614327430725,
+        "mean_feedback_forward_cosine": 0.9993880660303177,
+        "mean_feedback_forward_relative_error": 0.03441854629425272,
+        "min_feedback_forward_cosine": 0.998815655708313
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38368,
+      "train_examples": 45000,
+      "train_loss": 0.03245734071003066
+    },
+    {
+      "epoch": 110,
+      "eval_accuracy": 0.9204,
+      "eval_loss": 0.30980113525390623,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04875122755765915,
+        "mean_feedback_forward_cosine": 0.9993882583033654,
+        "mean_feedback_forward_relative_error": 0.034412265725193486,
+        "min_feedback_forward_cosine": 0.9988158941268921
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38720,
+      "train_examples": 45000,
+      "train_loss": 0.031768651305966907
+    },
+    {
+      "epoch": 111,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.31072572021484374,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04874620586633682,
+        "mean_feedback_forward_cosine": 0.9993884063536121,
+        "mean_feedback_forward_relative_error": 0.0344061431805453,
+        "min_feedback_forward_cosine": 0.9988161325454712
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39072,
+      "train_examples": 45000,
+      "train_loss": 0.029296634758843317
+    },
+    {
+      "epoch": 112,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.31651325073242187,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0487414114177227,
+        "mean_feedback_forward_cosine": 0.9993885736311635,
+        "mean_feedback_forward_relative_error": 0.03440025157385295,
+        "min_feedback_forward_cosine": 0.9988164901733398
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39424,
+      "train_examples": 45000,
+      "train_loss": 0.026635085297955406
+    },
+    {
+      "epoch": 113,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.3184968017578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04873606190085411,
+        "mean_feedback_forward_cosine": 0.9993887293723321,
+        "mean_feedback_forward_relative_error": 0.03439420197279223,
+        "min_feedback_forward_cosine": 0.9988166093826294
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39776,
+      "train_examples": 45000,
+      "train_loss": 0.02689870360626115
+    },
+    {
+      "epoch": 114,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.3188640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04873174428939819,
+        "mean_feedback_forward_cosine": 0.9993888601180045,
+        "mean_feedback_forward_relative_error": 0.03438880129326736,
+        "min_feedback_forward_cosine": 0.9988168478012085
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40128,
+      "train_examples": 45000,
+      "train_loss": 0.02336271217018366
+    },
+    {
+      "epoch": 115,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.3224325378417969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04872721806168556,
+        "mean_feedback_forward_cosine": 0.9993890216273646,
+        "mean_feedback_forward_relative_error": 0.0343832528038371,
+        "min_feedback_forward_cosine": 0.9988170862197876
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40480,
+      "train_examples": 45000,
+      "train_loss": 0.022886741083198124
+    },
+    {
+      "epoch": 116,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.3264565856933594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048722509294748306,
+        "mean_feedback_forward_cosine": 0.9993891754458027,
+        "mean_feedback_forward_relative_error": 0.034377753614417965,
+        "min_feedback_forward_cosine": 0.9988173246383667
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40832,
+      "train_examples": 45000,
+      "train_loss": 0.02161789578249057
+    },
+    {
+      "epoch": 117,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.3350888854980469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048718128353357315,
+        "mean_feedback_forward_cosine": 0.9993893273415104,
+        "mean_feedback_forward_relative_error": 0.034372437234607436,
+        "min_feedback_forward_cosine": 0.9988175630569458
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41184,
+      "train_examples": 45000,
+      "train_loss": 0.02008457115491231
+    },
+    {
+      "epoch": 118,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.3312977966308594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04871320724487305,
+        "mean_feedback_forward_cosine": 0.9993894561644523,
+        "mean_feedback_forward_relative_error": 0.034366853925729954,
+        "min_feedback_forward_cosine": 0.9988176822662354
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41536,
+      "train_examples": 45000,
+      "train_loss": 0.021647869668073125
+    },
+    {
+      "epoch": 119,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.33949063720703127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0487087108194828,
+        "mean_feedback_forward_cosine": 0.9993896061374296,
+        "mean_feedback_forward_relative_error": 0.034361511408802003,
+        "min_feedback_forward_cosine": 0.998818039894104
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41888,
+      "train_examples": 45000,
+      "train_loss": 0.02027899324049552
+    },
+    {
+      "epoch": 120,
+      "eval_accuracy": 0.921,
+      "eval_loss": 0.34436965942382813,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04870382696390152,
+        "mean_feedback_forward_cosine": 0.9993897695695201,
+        "mean_feedback_forward_relative_error": 0.03435609096120442,
+        "min_feedback_forward_cosine": 0.9988182187080383
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42240,
+      "train_examples": 45000,
+      "train_loss": 0.01882550497353077
+    },
+    {
+      "epoch": 121,
+      "eval_accuracy": 0.9208,
+      "eval_loss": 0.342112255859375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04869933798909187,
+        "mean_feedback_forward_cosine": 0.9993899022379229,
+        "mean_feedback_forward_relative_error": 0.034350986412215616,
+        "min_feedback_forward_cosine": 0.9988184571266174
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42592,
+      "train_examples": 45000,
+      "train_loss": 0.01746830699750119
+    },
+    {
+      "epoch": 122,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.34617767333984373,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04869527742266655,
+        "mean_feedback_forward_cosine": 0.999390056056361,
+        "mean_feedback_forward_relative_error": 0.03434607020068553,
+        "min_feedback_forward_cosine": 0.9988187551498413
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42944,
+      "train_examples": 45000,
+      "train_loss": 0.016621154091341628
+    },
+    {
+      "epoch": 123,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.3514703369140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04869047924876213,
+        "mean_feedback_forward_cosine": 0.9993902041066077,
+        "mean_feedback_forward_relative_error": 0.03434080309084346,
+        "min_feedback_forward_cosine": 0.9988188743591309
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43296,
+      "train_examples": 45000,
+      "train_loss": 0.01763997759670019
+    },
+    {
+      "epoch": 124,
+      "eval_accuracy": 0.9196,
+      "eval_loss": 0.3572915588378906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04868617281317711,
+        "mean_feedback_forward_cosine": 0.9993903194704363,
+        "mean_feedback_forward_relative_error": 0.034335700164158496,
+        "min_feedback_forward_cosine": 0.9988191723823547
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43648,
+      "train_examples": 45000,
+      "train_loss": 0.01719067797842953
+    },
+    {
+      "epoch": 125,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.35623099365234373,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04868151620030403,
+        "mean_feedback_forward_cosine": 0.9993904636752221,
+        "mean_feedback_forward_relative_error": 0.034330394599706895,
+        "min_feedback_forward_cosine": 0.9988192915916443
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44000,
+      "train_examples": 45000,
+      "train_loss": 0.017087505733304552
+    },
+    {
+      "epoch": 126,
+      "eval_accuracy": 0.9236,
+      "eval_loss": 0.35596282958984377,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04867681860923767,
+        "mean_feedback_forward_cosine": 0.9993906117254688,
+        "mean_feedback_forward_relative_error": 0.03432521590542409,
+        "min_feedback_forward_cosine": 0.9988195300102234
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44352,
+      "train_examples": 45000,
+      "train_loss": 0.01596804515255822
+    },
+    {
+      "epoch": 127,
+      "eval_accuracy": 0.9226,
+      "eval_loss": 0.356564501953125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04867197573184967,
+        "mean_feedback_forward_cosine": 0.9993907713120983,
+        "mean_feedback_forward_relative_error": 0.034319929628362576,
+        "min_feedback_forward_cosine": 0.9988197088241577
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44704,
+      "train_examples": 45000,
+      "train_loss": 0.01586283933967352
+    },
+    {
+      "epoch": 128,
+      "eval_accuracy": 0.9218,
+      "eval_loss": 0.3588898681640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0486680269241333,
+        "mean_feedback_forward_cosine": 0.9993909020577708,
+        "mean_feedback_forward_relative_error": 0.034315337787472436,
+        "min_feedback_forward_cosine": 0.9988199472427368
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45056,
+      "train_examples": 45000,
+      "train_loss": 0.013547314140200615
+    },
+    {
+      "epoch": 129,
+      "eval_accuracy": 0.9224,
+      "eval_loss": 0.36866251831054686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04866375774145126,
+        "mean_feedback_forward_cosine": 0.9993910404943651,
+        "mean_feedback_forward_relative_error": 0.034310508098813794,
+        "min_feedback_forward_cosine": 0.9988202452659607
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45408,
+      "train_examples": 45000,
+      "train_loss": 0.014398164713051585
+    },
+    {
+      "epoch": 130,
+      "eval_accuracy": 0.9218,
+      "eval_loss": 0.365063037109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04865895211696625,
+        "mean_feedback_forward_cosine": 0.9993912231537604,
+        "mean_feedback_forward_relative_error": 0.03430531098837814,
+        "min_feedback_forward_cosine": 0.9988205432891846
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45760,
+      "train_examples": 45000,
+      "train_loss": 0.01393503595052494
+    },
+    {
+      "epoch": 131,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.3608143493652344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04865493252873421,
+        "mean_feedback_forward_cosine": 0.999391338517589,
+        "mean_feedback_forward_relative_error": 0.03430056653075641,
+        "min_feedback_forward_cosine": 0.9988206624984741
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46112,
+      "train_examples": 45000,
+      "train_loss": 0.012824879109197192
+    },
+    {
+      "epoch": 132,
+      "eval_accuracy": 0.9224,
+      "eval_loss": 0.3637261047363281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04865088686347008,
+        "mean_feedback_forward_cosine": 0.9993914654178004,
+        "mean_feedback_forward_relative_error": 0.03429574687634745,
+        "min_feedback_forward_cosine": 0.9988207817077637
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46464,
+      "train_examples": 45000,
+      "train_loss": 0.012399559729629092
+    },
+    {
+      "epoch": 133,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.36333014526367186,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048646971583366394,
+        "mean_feedback_forward_cosine": 0.9993916269271604,
+        "mean_feedback_forward_relative_error": 0.03429104835395851,
+        "min_feedback_forward_cosine": 0.998820960521698
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46816,
+      "train_examples": 45000,
+      "train_loss": 0.012052169627075395
+    },
+    {
+      "epoch": 134,
+      "eval_accuracy": 0.9244,
+      "eval_loss": 0.3689380126953125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048642974346876144,
+        "mean_feedback_forward_cosine": 0.9993917595955634,
+        "mean_feedback_forward_relative_error": 0.03428640291695633,
+        "min_feedback_forward_cosine": 0.9988212585449219
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47168,
+      "train_examples": 45000,
+      "train_loss": 0.01183781708114677
+    },
+    {
+      "epoch": 135,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.36527835083007815,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04863850027322769,
+        "mean_feedback_forward_cosine": 0.9993918884185052,
+        "mean_feedback_forward_relative_error": 0.034281394594619354,
+        "min_feedback_forward_cosine": 0.9988214373588562
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47520,
+      "train_examples": 45000,
+      "train_loss": 0.012423885153399573
+    },
+    {
+      "epoch": 136,
+      "eval_accuracy": 0.9228,
+      "eval_loss": 0.37191195068359373,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048634033650159836,
+        "mean_feedback_forward_cosine": 0.9993920499278653,
+        "mean_feedback_forward_relative_error": 0.03427618378472905,
+        "min_feedback_forward_cosine": 0.9988217353820801
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47872,
+      "train_examples": 45000,
+      "train_loss": 0.011888691175646253
+    },
+    {
+      "epoch": 137,
+      "eval_accuracy": 0.9252,
+      "eval_loss": 0.3752788452148437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04862968251109123,
+        "mean_feedback_forward_cosine": 0.999392197978112,
+        "mean_feedback_forward_relative_error": 0.03427136731484244,
+        "min_feedback_forward_cosine": 0.9988219141960144
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48224,
+      "train_examples": 45000,
+      "train_loss": 0.011469858308798738
+    },
+    {
+      "epoch": 138,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.3746169250488281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04862533137202263,
+        "mean_feedback_forward_cosine": 0.9993923421828977,
+        "mean_feedback_forward_relative_error": 0.03426646318046316,
+        "min_feedback_forward_cosine": 0.9988220930099487
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48576,
+      "train_examples": 45000,
+      "train_loss": 0.011192755055841472
+    },
+    {
+      "epoch": 139,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.3725655700683594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04862077161669731,
+        "mean_feedback_forward_cosine": 0.9993924863876835,
+        "mean_feedback_forward_relative_error": 0.034261413399250276,
+        "min_feedback_forward_cosine": 0.9988223910331726
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48928,
+      "train_examples": 45000,
+      "train_loss": 0.011524058209856351
+    },
+    {
+      "epoch": 140,
+      "eval_accuracy": 0.9226,
+      "eval_loss": 0.37775662841796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04861658811569214,
+        "mean_feedback_forward_cosine": 0.9993926363606607,
+        "mean_feedback_forward_relative_error": 0.03425657896385077,
+        "min_feedback_forward_cosine": 0.9988225698471069
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49280,
+      "train_examples": 45000,
+      "train_loss": 0.010894056089719136
+    },
+    {
+      "epoch": 141,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.3685088928222656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048611849546432495,
+        "mean_feedback_forward_cosine": 0.9993928036382121,
+        "mean_feedback_forward_relative_error": 0.03425116128017826,
+        "min_feedback_forward_cosine": 0.9988226294517517
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49632,
+      "train_examples": 45000,
+      "train_loss": 0.011159929553253784
+    },
+    {
+      "epoch": 142,
+      "eval_accuracy": 0.9258,
+      "eval_loss": 0.3666699462890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048607733100652695,
+        "mean_feedback_forward_cosine": 0.9993929363066151,
+        "mean_feedback_forward_relative_error": 0.03424643322585091,
+        "min_feedback_forward_cosine": 0.9988230466842651
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49984,
+      "train_examples": 45000,
+      "train_loss": 0.010106286029020944
+    },
+    {
+      "epoch": 143,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.3864923645019531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0486031174659729,
+        "mean_feedback_forward_cosine": 0.9993930862795922,
+        "mean_feedback_forward_relative_error": 0.034241410603206006,
+        "min_feedback_forward_cosine": 0.9988232254981995
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50336,
+      "train_examples": 45000,
+      "train_loss": 0.010455552863805659
+    },
+    {
+      "epoch": 144,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.37877412109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04859888553619385,
+        "mean_feedback_forward_cosine": 0.9993932766299094,
+        "mean_feedback_forward_relative_error": 0.034236317200045434,
+        "min_feedback_forward_cosine": 0.998823344707489
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50688,
+      "train_examples": 45000,
+      "train_loss": 0.010130042814215024
+    },
+    {
+      "epoch": 145,
+      "eval_accuracy": 0.9224,
+      "eval_loss": 0.38209491577148436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048594359308481216,
+        "mean_feedback_forward_cosine": 0.9993934035301208,
+        "mean_feedback_forward_relative_error": 0.0342312341916465,
+        "min_feedback_forward_cosine": 0.9988236427307129
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51040,
+      "train_examples": 45000,
+      "train_loss": 0.009922520908858213
+    },
+    {
+      "epoch": 146,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.3881881774902344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04858988896012306,
+        "mean_feedback_forward_cosine": 0.9993935381212542,
+        "mean_feedback_forward_relative_error": 0.034226064540205464,
+        "min_feedback_forward_cosine": 0.9988237023353577
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51392,
+      "train_examples": 45000,
+      "train_loss": 0.009807878665212127
+    },
+    {
+      "epoch": 147,
+      "eval_accuracy": 0.9248,
+      "eval_loss": 0.38153494262695314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048584867268800735,
+        "mean_feedback_forward_cosine": 0.9993937188579191,
+        "mean_feedback_forward_relative_error": 0.03422067278335171,
+        "min_feedback_forward_cosine": 0.998823881149292
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51744,
+      "train_examples": 45000,
+      "train_loss": 0.010108151466813353
+    },
+    {
+      "epoch": 148,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.39089484252929685,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04858054965734482,
+        "mean_feedback_forward_cosine": 0.9993938476808609,
+        "mean_feedback_forward_relative_error": 0.034215778472923467,
+        "min_feedback_forward_cosine": 0.9988242387771606
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52096,
+      "train_examples": 45000,
+      "train_loss": 0.00944744374089771
+    },
+    {
+      "epoch": 149,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.3983838195800781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857631400227547,
+        "mean_feedback_forward_cosine": 0.9993940303402562,
+        "mean_feedback_forward_relative_error": 0.03421078793584339,
+        "min_feedback_forward_cosine": 0.9988244771957397
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52448,
+      "train_examples": 45000,
+      "train_loss": 0.009071883625868294
+    },
+    {
+      "epoch": 150,
+      "eval_accuracy": 0.922,
+      "eval_loss": 0.3920445617675781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571620136499405,
+        "mean_feedback_forward_cosine": 0.9993941918496163,
+        "mean_feedback_forward_relative_error": 0.03420544635023801,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52800,
+      "train_examples": 45000,
+      "train_loss": 0.009368490175157786
+    },
+    {
+      "epoch": 151,
+      "eval_accuracy": 0.9202,
+      "eval_loss": 0.3945726989746094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857156425714493,
+        "mean_feedback_forward_cosine": 0.9993941880041554,
+        "mean_feedback_forward_relative_error": 0.03420531115825138,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53152,
+      "train_examples": 45000,
+      "train_loss": 0.008574614812102583
+    },
+    {
+      "epoch": 152,
+      "eval_accuracy": 0.9208,
+      "eval_loss": 0.3895898193359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857151582837105,
+        "mean_feedback_forward_cosine": 0.9993941841586944,
+        "mean_feedback_forward_relative_error": 0.034205187893202225,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53504,
+      "train_examples": 45000,
+      "train_loss": 0.00854987682186895
+    },
+    {
+      "epoch": 153,
+      "eval_accuracy": 0.9218,
+      "eval_loss": 0.3936317932128906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571471124887466,
+        "mean_feedback_forward_cosine": 0.9993941918496163,
+        "mean_feedback_forward_relative_error": 0.03420507057660049,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53856,
+      "train_examples": 45000,
+      "train_loss": 0.007438057490438223
+    },
+    {
+      "epoch": 154,
+      "eval_accuracy": 0.9226,
+      "eval_loss": 0.38970969848632814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857144504785538,
+        "mean_feedback_forward_cosine": 0.9993941976178077,
+        "mean_feedback_forward_relative_error": 0.034204961972371224,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54208,
+      "train_examples": 45000,
+      "train_loss": 0.007093484061956406
+    },
+    {
+      "epoch": 155,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.3903849304199219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857141524553299,
+        "mean_feedback_forward_cosine": 0.9993941899268858,
+        "mean_feedback_forward_relative_error": 0.03420485216643541,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54560,
+      "train_examples": 45000,
+      "train_loss": 0.007574405139767462
+    },
+    {
+      "epoch": 156,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.3892175231933594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0485713854432106,
+        "mean_feedback_forward_cosine": 0.9993941918496163,
+        "mean_feedback_forward_relative_error": 0.034204739085849256,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54912,
+      "train_examples": 45000,
+      "train_loss": 0.00694707155831986
+    },
+    {
+      "epoch": 157,
+      "eval_accuracy": 0.9224,
+      "eval_loss": 0.38980198974609376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571355640888214,
+        "mean_feedback_forward_cosine": 0.9993941880041554,
+        "mean_feedback_forward_relative_error": 0.034204619486005075,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55264,
+      "train_examples": 45000,
+      "train_loss": 0.007672812810457415
+    },
+    {
+      "epoch": 158,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.3905103515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857131466269493,
+        "mean_feedback_forward_cosine": 0.9993942110769211,
+        "mean_feedback_forward_relative_error": 0.03420450631529093,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55616,
+      "train_examples": 45000,
+      "train_loss": 0.0080210231008629
+    },
+    {
+      "epoch": 159,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.38340865478515623,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571277409791946,
+        "mean_feedback_forward_cosine": 0.9993941880041554,
+        "mean_feedback_forward_relative_error": 0.034204392543723504,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55968,
+      "train_examples": 45000,
+      "train_loss": 0.0068947837022650576
+    },
+    {
+      "epoch": 160,
+      "eval_accuracy": 0.9236,
+      "eval_loss": 0.3892705810546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571255058050156,
+        "mean_feedback_forward_cosine": 0.9993941995405382,
+        "mean_feedback_forward_relative_error": 0.03420428601243804,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56320,
+      "train_examples": 45000,
+      "train_loss": 0.006723156801238656
+    },
+    {
+      "epoch": 161,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.387517919921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857123643159866,
+        "mean_feedback_forward_cosine": 0.9993942110769211,
+        "mean_feedback_forward_relative_error": 0.03420418536951465,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56672,
+      "train_examples": 45000,
+      "train_loss": 0.006679223413185941
+    },
+    {
+      "epoch": 162,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.3882124084472656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857119545340538,
+        "mean_feedback_forward_cosine": 0.9993941956950773,
+        "mean_feedback_forward_relative_error": 0.0342040729498671,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57024,
+      "train_examples": 45000,
+      "train_loss": 0.007147442474319703
+    },
+    {
+      "epoch": 163,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.3867006103515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571158200502396,
+        "mean_feedback_forward_cosine": 0.9993942129996515,
+        "mean_feedback_forward_relative_error": 0.034203966688965595,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57376,
+      "train_examples": 45000,
+      "train_loss": 0.006905054523878627
+    },
+    {
+      "epoch": 164,
+      "eval_accuracy": 0.9226,
+      "eval_loss": 0.38806248779296876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857112839818001,
+        "mean_feedback_forward_cosine": 0.9993942091541905,
+        "mean_feedback_forward_relative_error": 0.03420385676285913,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57728,
+      "train_examples": 45000,
+      "train_loss": 0.006833887228535281
+    },
+    {
+      "epoch": 165,
+      "eval_accuracy": 0.9228,
+      "eval_loss": 0.389813525390625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857109487056732,
+        "mean_feedback_forward_cosine": 0.999394218767843,
+        "mean_feedback_forward_relative_error": 0.03420375596972242,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58080,
+      "train_examples": 45000,
+      "train_loss": 0.006642061247221298
+    },
+    {
+      "epoch": 166,
+      "eval_accuracy": 0.9226,
+      "eval_loss": 0.390858056640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048571065068244934,
+        "mean_feedback_forward_cosine": 0.9993942206905734,
+        "mean_feedback_forward_relative_error": 0.03420365085044215,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58432,
+      "train_examples": 45000,
+      "train_loss": 0.006464248421953784
+    },
+    {
+      "epoch": 167,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.38848478393554686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857103154063225,
+        "mean_feedback_forward_cosine": 0.9993942283814953,
+        "mean_feedback_forward_relative_error": 0.03420353828058127,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58784,
+      "train_examples": 45000,
+      "train_loss": 0.006532345111005836
+    },
+    {
+      "epoch": 168,
+      "eval_accuracy": 0.9248,
+      "eval_loss": 0.38907427978515624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857099801301956,
+        "mean_feedback_forward_cosine": 0.999394214922382,
+        "mean_feedback_forward_relative_error": 0.03420343000682131,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59136,
+      "train_examples": 45000,
+      "train_loss": 0.006440749948720137
+    },
+    {
+      "epoch": 169,
+      "eval_accuracy": 0.9236,
+      "eval_loss": 0.38989791870117185,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570968210697174,
+        "mean_feedback_forward_cosine": 0.9993942206905734,
+        "mean_feedback_forward_relative_error": 0.034203322544213266,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59488,
+      "train_examples": 45000,
+      "train_loss": 0.006686322028934956
+    },
+    {
+      "epoch": 170,
+      "eval_accuracy": 0.9236,
+      "eval_loss": 0.38908255004882814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570938408374786,
+        "mean_feedback_forward_cosine": 0.9993942091541905,
+        "mean_feedback_forward_relative_error": 0.03420321913736482,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59840,
+      "train_examples": 45000,
+      "train_loss": 0.006311853679766257
+    },
+    {
+      "epoch": 171,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.38905315551757813,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570916056632996,
+        "mean_feedback_forward_cosine": 0.9993942341496868,
+        "mean_feedback_forward_relative_error": 0.034203112335695375,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60192,
+      "train_examples": 45000,
+      "train_loss": 0.006446439210573832
+    },
+    {
+      "epoch": 172,
+      "eval_accuracy": 0.9258,
+      "eval_loss": 0.3909088073730469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857088252902031,
+        "mean_feedback_forward_cosine": 0.9993942168451124,
+        "mean_feedback_forward_relative_error": 0.0342030058945379,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60544,
+      "train_examples": 45000,
+      "train_loss": 0.006494547173836165
+    },
+    {
+      "epoch": 173,
+      "eval_accuracy": 0.9252,
+      "eval_loss": 0.3877605590820313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570845276117325,
+        "mean_feedback_forward_cosine": 0.9993942264587649,
+        "mean_feedback_forward_relative_error": 0.03420290344905469,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60896,
+      "train_examples": 45000,
+      "train_loss": 0.006705886676493618
+    },
+    {
+      "epoch": 174,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.3882752746582031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857081547379494,
+        "mean_feedback_forward_cosine": 0.9993942303042258,
+        "mean_feedback_forward_relative_error": 0.034202803166643266,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61248,
+      "train_examples": 45000,
+      "train_loss": 0.00622351783282227
+    },
+    {
+      "epoch": 175,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.38828839111328123,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570796847343445,
+        "mean_feedback_forward_cosine": 0.9993942533769915,
+        "mean_feedback_forward_relative_error": 0.034202703304829135,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61600,
+      "train_examples": 45000,
+      "train_loss": 0.005534759585931897
+    },
+    {
+      "epoch": 176,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.3911351684570312,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570770770311356,
+        "mean_feedback_forward_cosine": 0.9993942437633392,
+        "mean_feedback_forward_relative_error": 0.03420259406970393,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61952,
+      "train_examples": 45000,
+      "train_loss": 0.005964737568009231
+    },
+    {
+      "epoch": 177,
+      "eval_accuracy": 0.9248,
+      "eval_loss": 0.3912963134765625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570748418569565,
+        "mean_feedback_forward_cosine": 0.9993942360724172,
+        "mean_feedback_forward_relative_error": 0.03420249706194285,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62304,
+      "train_examples": 45000,
+      "train_loss": 0.005666052417498496
+    },
+    {
+      "epoch": 178,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.39074390869140624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857071861624718,
+        "mean_feedback_forward_cosine": 0.9993942322269562,
+        "mean_feedback_forward_relative_error": 0.03420238794698831,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62656,
+      "train_examples": 45000,
+      "train_loss": 0.00602857864888178
+    },
+    {
+      "epoch": 179,
+      "eval_accuracy": 0.9262,
+      "eval_loss": 0.38614667358398436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570699989795685,
+        "mean_feedback_forward_cosine": 0.9993942303042258,
+        "mean_feedback_forward_relative_error": 0.03420228733410759,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63008,
+      "train_examples": 45000,
+      "train_loss": 0.005615825949195358
+    },
+    {
+      "epoch": 180,
+      "eval_accuracy": 0.9244,
+      "eval_loss": 0.38893945922851564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570673912763596,
+        "mean_feedback_forward_cosine": 0.9993942572224525,
+        "mean_feedback_forward_relative_error": 0.03420218726199481,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63360,
+      "train_examples": 45000,
+      "train_loss": 0.006089276160837875
+    },
+    {
+      "epoch": 181,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.3898399291992187,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857064411044121,
+        "mean_feedback_forward_cosine": 0.9993942437633392,
+        "mean_feedback_forward_relative_error": 0.03420209064478836,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63712,
+      "train_examples": 45000,
+      "train_loss": 0.006377839122795396
+    },
+    {
+      "epoch": 182,
+      "eval_accuracy": 0.923,
+      "eval_loss": 0.395673095703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857062175869942,
+        "mean_feedback_forward_cosine": 0.9993942245360343,
+        "mean_feedback_forward_relative_error": 0.03420198555555074,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64064,
+      "train_examples": 45000,
+      "train_loss": 0.005886330561753777
+    },
+    {
+      "epoch": 183,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.39157060546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857058823108673,
+        "mean_feedback_forward_cosine": 0.9993942379951477,
+        "mean_feedback_forward_relative_error": 0.034201881517806364,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64416,
+      "train_examples": 45000,
+      "train_loss": 0.00615320463238491
+    },
+    {
+      "epoch": 184,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.3938642578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857056587934494,
+        "mean_feedback_forward_cosine": 0.9993942399178782,
+        "mean_feedback_forward_relative_error": 0.0342017823469735,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64768,
+      "train_examples": 45000,
+      "train_loss": 0.006089492196589708
+    },
+    {
+      "epoch": 185,
+      "eval_accuracy": 0.9244,
+      "eval_loss": 0.39137650146484376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570528626441956,
+        "mean_feedback_forward_cosine": 0.9993942341496868,
+        "mean_feedback_forward_relative_error": 0.03420168437784718,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65120,
+      "train_examples": 45000,
+      "train_loss": 0.005729687671032217
+    },
+    {
+      "epoch": 186,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.3917858642578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857049137353897,
+        "mean_feedback_forward_cosine": 0.9993942418406087,
+        "mean_feedback_forward_relative_error": 0.034201580460273454,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65472,
+      "train_examples": 45000,
+      "train_loss": 0.0060106748471657435
+    },
+    {
+      "epoch": 187,
+      "eval_accuracy": 0.922,
+      "eval_loss": 0.3935861572265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857047274708748,
+        "mean_feedback_forward_cosine": 0.9993942360724172,
+        "mean_feedback_forward_relative_error": 0.034201479006198146,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65824,
+      "train_examples": 45000,
+      "train_loss": 0.006099106899317768
+    },
+    {
+      "epoch": 188,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.3924464660644531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570454120635986,
+        "mean_feedback_forward_cosine": 0.9993942629906439,
+        "mean_feedback_forward_relative_error": 0.03420138022591991,
+        "min_feedback_forward_cosine": 0.9988248944282532
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66176,
+      "train_examples": 45000,
+      "train_loss": 0.005720732238267859
+    },
+    {
+      "epoch": 189,
+      "eval_accuracy": 0.9258,
+      "eval_loss": 0.38947210693359374,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0485704243183136,
+        "mean_feedback_forward_cosine": 0.999394255299722,
+        "mean_feedback_forward_relative_error": 0.03420127672894347,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66528,
+      "train_examples": 45000,
+      "train_loss": 0.0063833029953142
+    },
+    {
+      "epoch": 190,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.3864773376464844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857039451599121,
+        "mean_feedback_forward_cosine": 0.999394255299722,
+        "mean_feedback_forward_relative_error": 0.03420117773836659,
+        "min_feedback_forward_cosine": 0.9988246560096741
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66880,
+      "train_examples": 45000,
+      "train_loss": 0.005768505500422584
+    },
+    {
+      "epoch": 191,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.38983641967773436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857035353779793,
+        "mean_feedback_forward_cosine": 0.9993942476088001,
+        "mean_feedback_forward_relative_error": 0.03420107397100618,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67232,
+      "train_examples": 45000,
+      "train_loss": 0.006166870318022039
+    },
+    {
+      "epoch": 192,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.3892133483886719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570334911346436,
+        "mean_feedback_forward_cosine": 0.9993942668361049,
+        "mean_feedback_forward_relative_error": 0.03420097753405571,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67584,
+      "train_examples": 45000,
+      "train_loss": 0.005772096807261308
+    },
+    {
+      "epoch": 193,
+      "eval_accuracy": 0.926,
+      "eval_loss": 0.3912098205566406,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570312559604645,
+        "mean_feedback_forward_cosine": 0.999394259145183,
+        "mean_feedback_forward_relative_error": 0.03420088317004904,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67936,
+      "train_examples": 45000,
+      "train_loss": 0.005735846981571781
+    },
+    {
+      "epoch": 194,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.38991021728515624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570286482572556,
+        "mean_feedback_forward_cosine": 0.9993942456860696,
+        "mean_feedback_forward_relative_error": 0.03420078535113604,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68288,
+      "train_examples": 45000,
+      "train_loss": 0.005394473344584306
+    },
+    {
+      "epoch": 195,
+      "eval_accuracy": 0.9234,
+      "eval_loss": 0.39238130493164064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857025295495987,
+        "mean_feedback_forward_cosine": 0.9993942456860696,
+        "mean_feedback_forward_relative_error": 0.03420068224471423,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68640,
+      "train_examples": 45000,
+      "train_loss": 0.005845885767539342
+    },
+    {
+      "epoch": 196,
+      "eval_accuracy": 0.924,
+      "eval_loss": 0.388215380859375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857022315263748,
+        "mean_feedback_forward_cosine": 0.9993942649133744,
+        "mean_feedback_forward_relative_error": 0.034200580430126956,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68992,
+      "train_examples": 45000,
+      "train_loss": 0.005960262824677758
+    },
+    {
+      "epoch": 197,
+      "eval_accuracy": 0.925,
+      "eval_loss": 0.3863689880371094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857020080089569,
+        "mean_feedback_forward_cosine": 0.9993942783724877,
+        "mean_feedback_forward_relative_error": 0.034200483362280557,
+        "min_feedback_forward_cosine": 0.9988247752189636
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69344,
+      "train_examples": 45000,
+      "train_loss": 0.005644922489176194
+    },
+    {
+      "epoch": 198,
+      "eval_accuracy": 0.9244,
+      "eval_loss": 0.3886321044921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570167273283005,
+        "mean_feedback_forward_cosine": 0.9993942649133744,
+        "mean_feedback_forward_relative_error": 0.034200379955432104,
+        "min_feedback_forward_cosine": 0.9988248944282532
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69696,
+      "train_examples": 45000,
+      "train_loss": 0.005755332447257307
+    },
+    {
+      "epoch": 199,
+      "eval_accuracy": 0.926,
+      "eval_loss": 0.3910266357421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.048570141196250916,
+        "mean_feedback_forward_cosine": 0.999394255299722,
+        "mean_feedback_forward_relative_error": 0.03420028096485522,
+        "min_feedback_forward_cosine": 0.9988248348236084
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70048,
+      "train_examples": 45000,
+      "train_loss": 0.006012655192530817
+    },
+    {
+      "epoch": 200,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.39594262084960935,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04857011139392853,
+        "mean_feedback_forward_cosine": 0.9993942726042963,
+        "mean_feedback_forward_relative_error": 0.034200180532230486,
+        "min_feedback_forward_cosine": 0.9988247156143188
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70400,
+      "train_examples": 45000,
+      "train_loss": 0.0059044962678311596
+    }
+  ],
+  "evaluation_protocol": {
+    "test_evaluations": 0,
+    "test_used_for_selection": false,
+    "validation_evaluations": 201
+  },
+  "final": {
+    "accuracy": 0.9238,
+    "epoch": 200,
+    "evaluation_split": "validation",
+    "finite": true,
+    "loss": 0.39594262084960935,
+    "step": 70400
+  },
+  "hardware": {
+    "cuda_device_name": "NVIDIA RTX A6000",
+    "cuda_visible_devices": "0",
+    "device": "cuda",
+    "device_total_memory_bytes": 50896568320,
+    "peak_memory_allocated_bytes": 2325996544,
+    "peak_memory_reserved_bytes": 3166699520,
+    "torch_version": "2.3.1+cu118"
+  },
+  "protocol_family": "oral_a_cifar_local_resnet_development",
+  "provenance": {
+    "git_commit": "f55872beaee658ff76ac15fd421b73017820c320",
+    "git_tracked_dirty": false
+  },
+  "schema_version": 1,
+  "split": {
+    "cifar_source_files": [
+      {
+        "bytes": 31035704,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/data_batch_1",
+        "sha256": "54636561a3ce25bd3e19253c6b0d8538147b0ae398331ac4a2d86c6d987368cd"
+      },
+      {
+        "bytes": 31035320,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/data_batch_2",
+        "sha256": "766b2cef9fbc745cf056b3152224f7cf77163b330ea9a15f9392beb8b89bc5a8"
+      },
+      {
+        "bytes": 31035999,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/data_batch_3",
+        "sha256": "0f00d98ebfb30b3ec0ad19f9756dc2630b89003e10525f5e148445e82aa6a1f9"
+      },
+      {
+        "bytes": 31035696,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/data_batch_4",
+        "sha256": "3f7bb240661948b8f4d53e36ec720d8306f5668bd0071dcb4e6c947f78e9682b"
+      },
+      {
+        "bytes": 31035623,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/data_batch_5",
+        "sha256": "d91802434d8376bbaeeadf58a737e3a1b12ac839077e931237e0dcd43adcb154"
+      },
+      {
+        "bytes": 31035526,
+        "path": "/orion/u/oscarwan/scratch/sdil-data/cifar-10-batches-py/test_batch",
+        "sha256": "f53d8d457504f7cff4ea9e021afcf0e0ad8e24a91f3fc42091b8adef61157831"
+      }
+    ],
+    "dataset": "cifar10",
+    "evaluation_split": "validation",
+    "input_layout": "NCHW",
+    "input_shape": [
+      3,
+      32,
+      32
+    ],
+    "loader_seed": 0,
+    "normalization_mean": [
+      0.49140000343322754,
+      0.4821999967098236,
+      0.4465000033378601
+    ],
+    "normalization_std": [
+      0.24699999392032623,
+      0.2434999942779541,
+      0.26159998774528503
+    ],
+    "split_from_training_only": true,
+    "split_seed": 2027,
+    "test_examples": 10000,
+    "train_examples": 45000,
+    "training_augmentation": "random_crop_32_padding_4_zero_then_horizontal_flip_p0.5",
+    "validation_class_counts": {
+      "0": 500,
+      "1": 500,
+      "2": 500,
+      "3": 500,
+      "4": 500,
+      "5": 500,
+      "6": 500,
+      "7": 500,
+      "8": 500,
+      "9": 500
+    },
+    "validation_examples": 5000,
+    "validation_index_sha256": "8328b206a97c420e49e54e3eca4abe3274c4756b084355784ea3fb8059e4515b"
+  },
+  "timing": {
+    "apical_warmup_wall_s": 0.0,
+    "evaluation_wall_s": 34.060356855392456,
+    "mirror_warmup_wall_s": 0.0,
+    "predictor_warmup_wall_s": 0.0,
+    "timing_excludes_data_loading_hashing_and_model_construction": true,
+    "total_timed_wall_s": 2877.0025794506073,
+    "train_wall_s": 2841.6028168201447
+  },
+  "work": {
+    "apical_macs_per_example": 68420224,
+    "causal_scalar_observations": 0,
+    "components": {
+      "apical_projection_macs": 615782016000000,
+      "apical_regression_macs": 0,
+      "bp_reverse_macs_estimate": 0,
+      "kp_reciprocal_correlation_macs": 615782016000000,
+      "local_weight_correlation_macs": 619763328000000,
+      "mirror_feedback_prediction_macs": 0,
+      "mirror_local_correlation_macs": 0,
+      "mirror_response_macs": 0,
+      "ordinary_forward_macs": 619763328000000,
+      "perturbation_forward_macs": 0,
+      "warmup_clean_forward_macs": 0
+    },
+    "definition": "multiply-accumulates in conv/linear maps; one local weight correlation equals one forward-weight MAC count; BP reverse is estimated as one weight-gradient plus one activation-gradient convolution per forward convolution; mixed-traffic/predictor elementwise arithmetic is reported as a conservative operation estimate separately and is not folded into MACs",
+    "elementwise_operations_estimate": 0,
+    "forward_macs_per_example": 68862592,
+    "logical_batch_loss_queries": 0,
+    "mirror_probe_examples": 0,
+    "mirror_readout_probe_examples": 0,
+    "neutral_projection_observations": 0,
+    "per_example_cross_entropy_terms": 0,
+    "total_clean_forward_examples": 9000000,
+    "total_forward_equivalent_examples": 9000000,
+    "total_macs_estimate": 2471090688000000
+  }
+}