From 3c7c221ae6cfcdf6840da1d5262d94c75d56288b Mon Sep 17 00:00:00 2001
From: Oscar Wan <oscarwan@orionp2.stanford.edu>
Date: Wed, 26 Aug 2026 08:50:07 -0700
Subject: experiment: add completed A6000 crossover results

---
 .../r2/resnet-r2-clean_kp-d20-lr0p1.json           | 3869 ++++++++++++++++++++
 1 file changed, 3869 insertions(+)
 create mode 100644 results/resnet_crossover/r2/resnet-r2-clean_kp-d20-lr0p1.json

(limited to 'results/resnet_crossover/r2/resnet-r2-clean_kp-d20-lr0p1.json')

diff --git a/results/resnet_crossover/r2/resnet-r2-clean_kp-d20-lr0p1.json b/results/resnet_crossover/r2/resnet-r2-clean_kp-d20-lr0p1.json
new file mode 100644
index 0000000..b440202
--- /dev/null
+++ b/results/resnet_crossover/r2/resnet-r2-clean_kp-d20-lr0p1.json
@@ -0,0 +1,3869 @@
+{
+  "architecture": {
+    "adaptive_apical_parameters": 0,
+    "adaptive_feedback_parameters": 267904,
+    "base_width": 16,
+    "blocks_per_stage": 3,
+    "bn_eps": 1e-05,
+    "bn_momentum": 0.1,
+    "depth": 20,
+    "family": "CIFAR 6n+2 ResNet, option-A shortcuts",
+    "fixed_feedback_parameters": 0,
+    "fixed_traffic_coefficients": 0,
+    "forward_parameters": 269722,
+    "hidden_shapes": [
+      [
+        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
+      ],
+      [
+        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": 20,
+    "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-d20-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": 0.9992254873116811,
+    "feedback_forward_cosine": [
+      0.9996554255485535,
+      0.9996746182441711,
+      0.9996549487113953,
+      0.9996075630187988,
+      0.9996635913848877,
+      0.9995270371437073,
+      0.9995354413986206,
+      0.9996858835220337,
+      0.9996363520622253,
+      0.9995448589324951,
+      0.999627411365509,
+      0.9995063543319702,
+      0.9995788335800171,
+      0.999750554561615,
+      0.9997305870056152,
+      0.999675989151001,
+      0.9997004270553589,
+      0.9992839097976685,
+      0.9999722242355347
+    ],
+    "feedback_forward_norm_ratio": [
+      0.9991434216499329,
+      1.0004491806030273,
+      0.9987601041793823,
+      1.0013868808746338,
+      0.999297022819519,
+      1.0023059844970703,
+      0.9987834692001343,
+      1.0002436637878418,
+      0.9993070363998413,
+      1.0014671087265015,
+      0.9994328618049622,
+      1.0015302896499634,
+      0.9993026256561279,
+      1.0005580186843872,
+      1.0002413988113403,
+      1.0010871887207031,
+      1.0002120733261108,
+      1.0017412900924683,
+      1.0046420097351074
+    ],
+    "feedback_forward_relative_error": [
+      0.02625316195189953,
+      0.025520561262965202,
+      0.026284009218215942,
+      0.028065064921975136,
+      0.025940094143152237,
+      0.03087770752608776,
+      0.03048705868422985,
+      0.02506886050105095,
+      0.026966994628310204,
+      0.030226128175854683,
+      0.02730029635131359,
+      0.03148436173796654,
+      0.029020819813013077,
+      0.022351281717419624,
+      0.02321099117398262,
+      0.02549651823937893,
+      0.024476630613207817,
+      0.03791650012135506,
+      0.008795272558927536
+    ],
+    "innovation_negative_gradient_cosine": [
+      0.9992389678955078,
+      0.9991669058799744,
+      0.999206006526947,
+      0.9992231726646423,
+      0.9992378950119019,
+      0.9992799758911133,
+      0.9992815256118774,
+      0.9993095397949219,
+      0.999367356300354,
+      0.9992690086364746,
+      0.9994315505027771,
+      0.999388575553894,
+      0.9994919300079346,
+      0.9995896816253662,
+      0.9996410608291626,
+      0.999697208404541,
+      0.9998255968093872,
+      0.9999359846115112,
+      0.8563728332519531
+    ],
+    "mean_feedback_forward_cosine": 0.9996322111079567,
+    "mean_feedback_forward_relative_error": 0.026618016491595068,
+    "normalization_state": "training_batch_stats_without_running_update",
+    "raw_negative_gradient_cosine": [
+      0.9992389678955078,
+      0.9991669058799744,
+      0.999206006526947,
+      0.9992231726646423,
+      0.9992378950119019,
+      0.9992799758911133,
+      0.9992815256118774,
+      0.9993095397949219,
+      0.999367356300354,
+      0.9992690086364746,
+      0.9994315505027771,
+      0.999388575553894,
+      0.9994919300079346,
+      0.9995896816253662,
+      0.9996410608291626,
+      0.999697208404541,
+      0.9998255968093872,
+      0.9999359846115112,
+      0.8563728332519531
+    ],
+    "teaching_negative_gradient_cosine": [
+      0.9992389678955078,
+      0.9991669058799744,
+      0.999206006526947,
+      0.9992231726646423,
+      0.9992378950119019,
+      0.9992799758911133,
+      0.9992815256118774,
+      0.9993095397949219,
+      0.999367356300354,
+      0.9992690086364746,
+      0.9994315505027771,
+      0.999388575553894,
+      0.9994919300079346,
+      0.9995896816253662,
+      0.9996410608291626,
+      0.999697208404541,
+      0.9998255968093872,
+      0.9999359846115112,
+      0.8563728332519531
+    ],
+    "wall_s": 0.06803202629089355
+  },
+  "epochs": [
+    {
+      "epoch": 1,
+      "eval_accuracy": 0.344,
+      "eval_loss": 1.81885517578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.396217942237854,
+        "mean_feedback_forward_cosine": 0.12113288064536296,
+        "mean_feedback_forward_relative_error": 1.3444790902890658,
+        "min_feedback_forward_cosine": 0.037865087389945984
+      },
+      "lr": 0.02,
+      "output_lr": 0.02,
+      "step": 352,
+      "train_examples": 45000,
+      "train_loss": 1.9119543294694687
+    },
+    {
+      "epoch": 2,
+      "eval_accuracy": 0.4864,
+      "eval_loss": 1.565498681640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3504620790481567,
+        "mean_feedback_forward_cosine": 0.27226419001817703,
+        "mean_feedback_forward_relative_error": 1.222043868742491,
+        "min_feedback_forward_cosine": 0.11271074414253235
+      },
+      "lr": 0.04,
+      "output_lr": 0.04,
+      "step": 704,
+      "train_examples": 45000,
+      "train_loss": 1.4822090652677749
+    },
+    {
+      "epoch": 3,
+      "eval_accuracy": 0.523,
+      "eval_loss": 1.6564058837890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3010425567626953,
+        "mean_feedback_forward_cosine": 0.39278248579878555,
+        "mean_feedback_forward_relative_error": 1.11243883873287,
+        "min_feedback_forward_cosine": 0.18469145894050598
+      },
+      "lr": 0.06000000000000001,
+      "output_lr": 0.06000000000000001,
+      "step": 1056,
+      "train_examples": 45000,
+      "train_loss": 1.19166419444614
+    },
+    {
+      "epoch": 4,
+      "eval_accuracy": 0.6216,
+      "eval_loss": 1.09284404296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.246029257774353,
+        "mean_feedback_forward_cosine": 0.49751683755924825,
+        "mean_feedback_forward_relative_error": 1.0094205486147028,
+        "min_feedback_forward_cosine": 0.25789177417755127
+      },
+      "lr": 0.08,
+      "output_lr": 0.08,
+      "step": 1408,
+      "train_examples": 45000,
+      "train_loss": 1.0127930777443779
+    },
+    {
+      "epoch": 5,
+      "eval_accuracy": 0.7116,
+      "eval_loss": 0.8747228881835938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1837794780731201,
+        "mean_feedback_forward_cosine": 0.5858596798620725,
+        "mean_feedback_forward_relative_error": 0.9147494780389887,
+        "min_feedback_forward_cosine": 0.3337680697441101
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 1760,
+      "train_examples": 45000,
+      "train_loss": 0.884297638130188
+    },
+    {
+      "epoch": 6,
+      "eval_accuracy": 0.6672,
+      "eval_loss": 0.985279248046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1347962617874146,
+        "mean_feedback_forward_cosine": 0.6410800830314034,
+        "mean_feedback_forward_relative_error": 0.8509495085791537,
+        "min_feedback_forward_cosine": 0.38860657811164856
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2112,
+      "train_examples": 45000,
+      "train_loss": 0.7689341519885593
+    },
+    {
+      "epoch": 7,
+      "eval_accuracy": 0.745,
+      "eval_loss": 0.7586026489257812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0899189710617065,
+        "mean_feedback_forward_cosine": 0.6840777710864419,
+        "mean_feedback_forward_relative_error": 0.798096848161597,
+        "min_feedback_forward_cosine": 0.4376838803291321
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2464,
+      "train_examples": 45000,
+      "train_loss": 0.7034475085046556
+    },
+    {
+      "epoch": 8,
+      "eval_accuracy": 0.702,
+      "eval_loss": 1.04657080078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0519425868988037,
+        "mean_feedback_forward_cosine": 0.7168333953932712,
+        "mean_feedback_forward_relative_error": 0.7551552151378832,
+        "min_feedback_forward_cosine": 0.4767589867115021
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2816,
+      "train_examples": 45000,
+      "train_loss": 0.6440844105932447
+    },
+    {
+      "epoch": 9,
+      "eval_accuracy": 0.7628,
+      "eval_loss": 0.7188642578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0166592597961426,
+        "mean_feedback_forward_cosine": 0.744070269559559,
+        "mean_feedback_forward_relative_error": 0.7177636921405792,
+        "min_feedback_forward_cosine": 0.5122615694999695
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3168,
+      "train_examples": 45000,
+      "train_loss": 0.5978653068330553
+    },
+    {
+      "epoch": 10,
+      "eval_accuracy": 0.7684,
+      "eval_loss": 0.7278686767578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9829434752464294,
+        "mean_feedback_forward_cosine": 0.7671844739662973,
+        "mean_feedback_forward_relative_error": 0.6843707843830711,
+        "min_feedback_forward_cosine": 0.5441762208938599
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3520,
+      "train_examples": 45000,
+      "train_loss": 0.5641649359703064
+    },
+    {
+      "epoch": 11,
+      "eval_accuracy": 0.7844,
+      "eval_loss": 0.6429546142578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9501974582672119,
+        "mean_feedback_forward_cosine": 0.787681868201808,
+        "mean_feedback_forward_relative_error": 0.6532509734756068,
+        "min_feedback_forward_cosine": 0.574102520942688
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3872,
+      "train_examples": 45000,
+      "train_loss": 0.5336678455458747
+    },
+    {
+      "epoch": 12,
+      "eval_accuracy": 0.7888,
+      "eval_loss": 0.6300875366210937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.918483316898346,
+        "mean_feedback_forward_cosine": 0.8054801445258292,
+        "mean_feedback_forward_relative_error": 0.6250106943281073,
+        "min_feedback_forward_cosine": 0.6022139191627502
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4224,
+      "train_examples": 45000,
+      "train_loss": 0.5122270748244392
+    },
+    {
+      "epoch": 13,
+      "eval_accuracy": 0.6694,
+      "eval_loss": 1.1285370361328124,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8890088796615601,
+        "mean_feedback_forward_cosine": 0.8212189956715232,
+        "mean_feedback_forward_relative_error": 0.5988754441863612,
+        "min_feedback_forward_cosine": 0.6274921894073486
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4576,
+      "train_examples": 45000,
+      "train_loss": 0.4805156542142232
+    },
+    {
+      "epoch": 14,
+      "eval_accuracy": 0.8094,
+      "eval_loss": 0.5873873657226563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.859317421913147,
+        "mean_feedback_forward_cosine": 0.8354297970470629,
+        "mean_feedback_forward_relative_error": 0.5742914959004051,
+        "min_feedback_forward_cosine": 0.651719868183136
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4928,
+      "train_examples": 45000,
+      "train_loss": 0.4709179120169746
+    },
+    {
+      "epoch": 15,
+      "eval_accuracy": 0.8018,
+      "eval_loss": 0.631718359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8302932977676392,
+        "mean_feedback_forward_cosine": 0.8483938072857103,
+        "mean_feedback_forward_relative_error": 0.550973863978135,
+        "min_feedback_forward_cosine": 0.674503743648529
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5280,
+      "train_examples": 45000,
+      "train_loss": 0.44848755186398825
+    },
+    {
+      "epoch": 16,
+      "eval_accuracy": 0.8008,
+      "eval_loss": 0.6399617309570312,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8032397627830505,
+        "mean_feedback_forward_cosine": 0.8598326538738451,
+        "mean_feedback_forward_relative_error": 0.529568424350337,
+        "min_feedback_forward_cosine": 0.6950962543487549
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5632,
+      "train_examples": 45000,
+      "train_loss": 0.4364324880705939
+    },
+    {
+      "epoch": 17,
+      "eval_accuracy": 0.8222,
+      "eval_loss": 0.5439112243652344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7772430777549744,
+        "mean_feedback_forward_cosine": 0.8702934415716874,
+        "mean_feedback_forward_relative_error": 0.5092110178972545,
+        "min_feedback_forward_cosine": 0.7146843671798706
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5984,
+      "train_examples": 45000,
+      "train_loss": 0.4180992658085293
+    },
+    {
+      "epoch": 18,
+      "eval_accuracy": 0.8316,
+      "eval_loss": 0.5281030395507812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7511368989944458,
+        "mean_feedback_forward_cosine": 0.8800302555686549,
+        "mean_feedback_forward_relative_error": 0.48952555499578776,
+        "min_feedback_forward_cosine": 0.7333462238311768
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6336,
+      "train_examples": 45000,
+      "train_loss": 0.4113643139521281
+    },
+    {
+      "epoch": 19,
+      "eval_accuracy": 0.834,
+      "eval_loss": 0.5129178466796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7260403633117676,
+        "mean_feedback_forward_cosine": 0.8888662520207857,
+        "mean_feedback_forward_relative_error": 0.47090583176989304,
+        "min_feedback_forward_cosine": 0.7505145072937012
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6688,
+      "train_examples": 45000,
+      "train_loss": 0.398461009144783
+    },
+    {
+      "epoch": 20,
+      "eval_accuracy": 0.8144,
+      "eval_loss": 0.5892133605957032,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7021915912628174,
+        "mean_feedback_forward_cosine": 0.8966989234874123,
+        "mean_feedback_forward_relative_error": 0.45369475923086466,
+        "min_feedback_forward_cosine": 0.7663941979408264
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7040,
+      "train_examples": 45000,
+      "train_loss": 0.3833774073653751
+    },
+    {
+      "epoch": 21,
+      "eval_accuracy": 0.8162,
+      "eval_loss": 0.6017692626953125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6776108145713806,
+        "mean_feedback_forward_cosine": 0.9041856401845029,
+        "mean_feedback_forward_relative_error": 0.4366687630352221,
+        "min_feedback_forward_cosine": 0.7823076248168945
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7392,
+      "train_examples": 45000,
+      "train_loss": 0.373022927708096
+    },
+    {
+      "epoch": 22,
+      "eval_accuracy": 0.7892,
+      "eval_loss": 0.7000044067382812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6551166772842407,
+        "mean_feedback_forward_cosine": 0.911015090189482,
+        "mean_feedback_forward_relative_error": 0.4206081620956722,
+        "min_feedback_forward_cosine": 0.796286940574646
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7744,
+      "train_examples": 45000,
+      "train_loss": 0.368713469309277
+    },
+    {
+      "epoch": 23,
+      "eval_accuracy": 0.8424,
+      "eval_loss": 0.4894224792480469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6320359110832214,
+        "mean_feedback_forward_cosine": 0.9173811485892848,
+        "mean_feedback_forward_relative_error": 0.4050847472328889,
+        "min_feedback_forward_cosine": 0.8102049827575684
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8096,
+      "train_examples": 45000,
+      "train_loss": 0.36200492996639677
+    },
+    {
+      "epoch": 24,
+      "eval_accuracy": 0.8508,
+      "eval_loss": 0.4847625244140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6104835271835327,
+        "mean_feedback_forward_cosine": 0.923206090927124,
+        "mean_feedback_forward_relative_error": 0.39038152600589554,
+        "min_feedback_forward_cosine": 0.8226820230484009
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8448,
+      "train_examples": 45000,
+      "train_loss": 0.35031071446206835
+    },
+    {
+      "epoch": 25,
+      "eval_accuracy": 0.8472,
+      "eval_loss": 0.48265894165039064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5895683169364929,
+        "mean_feedback_forward_cosine": 0.9285419740174946,
+        "mean_feedback_forward_relative_error": 0.37639613763282176,
+        "min_feedback_forward_cosine": 0.8343791961669922
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8800,
+      "train_examples": 45000,
+      "train_loss": 0.3469265750249227
+    },
+    {
+      "epoch": 26,
+      "eval_accuracy": 0.8544,
+      "eval_loss": 0.44391314697265627,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5687019228935242,
+        "mean_feedback_forward_cosine": 0.9335442655964902,
+        "mean_feedback_forward_relative_error": 0.3628646017689454,
+        "min_feedback_forward_cosine": 0.8457885384559631
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9152,
+      "train_examples": 45000,
+      "train_loss": 0.34203060767385696
+    },
+    {
+      "epoch": 27,
+      "eval_accuracy": 0.7932,
+      "eval_loss": 0.6590688598632812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.549074113368988,
+        "mean_feedback_forward_cosine": 0.9380777917410198,
+        "mean_feedback_forward_relative_error": 0.35010523309833125,
+        "min_feedback_forward_cosine": 0.8560115098953247
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9504,
+      "train_examples": 45000,
+      "train_loss": 0.3337548480987549
+    },
+    {
+      "epoch": 28,
+      "eval_accuracy": 0.8348,
+      "eval_loss": 0.5370311096191406,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5305789709091187,
+        "mean_feedback_forward_cosine": 0.9421965573963366,
+        "mean_feedback_forward_relative_error": 0.338118544534633,
+        "min_feedback_forward_cosine": 0.8653731942176819
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9856,
+      "train_examples": 45000,
+      "train_loss": 0.3222589003774855
+    },
+    {
+      "epoch": 29,
+      "eval_accuracy": 0.8502,
+      "eval_loss": 0.4641535888671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5120760798454285,
+        "mean_feedback_forward_cosine": 0.9460670446094713,
+        "mean_feedback_forward_relative_error": 0.32643518400819677,
+        "min_feedback_forward_cosine": 0.8744317293167114
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10208,
+      "train_examples": 45000,
+      "train_loss": 0.31773917949464586
+    },
+    {
+      "epoch": 30,
+      "eval_accuracy": 0.8528,
+      "eval_loss": 0.4488963806152344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.49475499987602234,
+        "mean_feedback_forward_cosine": 0.9497027867718747,
+        "mean_feedback_forward_relative_error": 0.3151285601289649,
+        "min_feedback_forward_cosine": 0.882604718208313
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10560,
+      "train_examples": 45000,
+      "train_loss": 0.3195970536020067
+    },
+    {
+      "epoch": 31,
+      "eval_accuracy": 0.83,
+      "eval_loss": 0.5759987670898438,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.47704029083251953,
+        "mean_feedback_forward_cosine": 0.9531683921813965,
+        "mean_feedback_forward_relative_error": 0.3039606629233611,
+        "min_feedback_forward_cosine": 0.8906716704368591
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10912,
+      "train_examples": 45000,
+      "train_loss": 0.312859878677792
+    },
+    {
+      "epoch": 32,
+      "eval_accuracy": 0.8274,
+      "eval_loss": 0.5443523193359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.46091774106025696,
+        "mean_feedback_forward_cosine": 0.9562506518865886,
+        "mean_feedback_forward_relative_error": 0.2936018642626311,
+        "min_feedback_forward_cosine": 0.8978977203369141
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11264,
+      "train_examples": 45000,
+      "train_loss": 0.3004345349841648
+    },
+    {
+      "epoch": 33,
+      "eval_accuracy": 0.8412,
+      "eval_loss": 0.5110523986816407,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.44484663009643555,
+        "mean_feedback_forward_cosine": 0.9592594598468981,
+        "mean_feedback_forward_relative_error": 0.28327554934903193,
+        "min_feedback_forward_cosine": 0.904775857925415
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11616,
+      "train_examples": 45000,
+      "train_loss": 0.3016047252337138
+    },
+    {
+      "epoch": 34,
+      "eval_accuracy": 0.8304,
+      "eval_loss": 0.5683611450195313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.42912110686302185,
+        "mean_feedback_forward_cosine": 0.9620242558027569,
+        "mean_feedback_forward_relative_error": 0.2734151631593704,
+        "min_feedback_forward_cosine": 0.9112752079963684
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11968,
+      "train_examples": 45000,
+      "train_loss": 0.29844820080068374
+    },
+    {
+      "epoch": 35,
+      "eval_accuracy": 0.8534,
+      "eval_loss": 0.47788355712890623,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4136193096637726,
+        "mean_feedback_forward_cosine": 0.9646216285856146,
+        "mean_feedback_forward_relative_error": 0.26376486295147944,
+        "min_feedback_forward_cosine": 0.9175251126289368
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12320,
+      "train_examples": 45000,
+      "train_loss": 0.29182118734253776
+    },
+    {
+      "epoch": 36,
+      "eval_accuracy": 0.829,
+      "eval_loss": 0.5739090393066406,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3991897404193878,
+        "mean_feedback_forward_cosine": 0.9669801373230783,
+        "mean_feedback_forward_relative_error": 0.2547991220888339,
+        "min_feedback_forward_cosine": 0.9230731725692749
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12672,
+      "train_examples": 45000,
+      "train_loss": 0.2934884912861718
+    },
+    {
+      "epoch": 37,
+      "eval_accuracy": 0.8344,
+      "eval_loss": 0.5700015991210937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3849577009677887,
+        "mean_feedback_forward_cosine": 0.9692228216873972,
+        "mean_feedback_forward_relative_error": 0.24591049551963806,
+        "min_feedback_forward_cosine": 0.9283632636070251
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13024,
+      "train_examples": 45000,
+      "train_loss": 0.28581399149894715
+    },
+    {
+      "epoch": 38,
+      "eval_accuracy": 0.8496,
+      "eval_loss": 0.46519459228515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3715720474720001,
+        "mean_feedback_forward_cosine": 0.9713154152819985,
+        "mean_feedback_forward_relative_error": 0.2373458535263413,
+        "min_feedback_forward_cosine": 0.9331613779067993
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13376,
+      "train_examples": 45000,
+      "train_loss": 0.2865924285676744
+    },
+    {
+      "epoch": 39,
+      "eval_accuracy": 0.8558,
+      "eval_loss": 0.46593084716796873,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3588463366031647,
+        "mean_feedback_forward_cosine": 0.9731747727645071,
+        "mean_feedback_forward_relative_error": 0.22942296886130384,
+        "min_feedback_forward_cosine": 0.937605619430542
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13728,
+      "train_examples": 45000,
+      "train_loss": 0.2737637662622664
+    },
+    {
+      "epoch": 40,
+      "eval_accuracy": 0.8636,
+      "eval_loss": 0.44457278442382814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3455275595188141,
+        "mean_feedback_forward_cosine": 0.9750205466621801,
+        "mean_feedback_forward_relative_error": 0.2213271828858476,
+        "min_feedback_forward_cosine": 0.9420970678329468
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14080,
+      "train_examples": 45000,
+      "train_loss": 0.27786194910473294
+    },
+    {
+      "epoch": 41,
+      "eval_accuracy": 0.846,
+      "eval_loss": 0.5155881774902343,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3331117630004883,
+        "mean_feedback_forward_cosine": 0.9767144510620519,
+        "mean_feedback_forward_relative_error": 0.21364634640906988,
+        "min_feedback_forward_cosine": 0.946147084236145
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14432,
+      "train_examples": 45000,
+      "train_loss": 0.2707260728518168
+    },
+    {
+      "epoch": 42,
+      "eval_accuracy": 0.8526,
+      "eval_loss": 0.49671691284179686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3209340572357178,
+        "mean_feedback_forward_cosine": 0.9782712177226418,
+        "mean_feedback_forward_relative_error": 0.20631324695913414,
+        "min_feedback_forward_cosine": 0.9499820470809937
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14784,
+      "train_examples": 45000,
+      "train_loss": 0.2681165763537089
+    },
+    {
+      "epoch": 43,
+      "eval_accuracy": 0.8602,
+      "eval_loss": 0.4529365417480469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.30942007899284363,
+        "mean_feedback_forward_cosine": 0.9797280869985882,
+        "mean_feedback_forward_relative_error": 0.19923599417272367,
+        "min_feedback_forward_cosine": 0.9534589648246765
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15136,
+      "train_examples": 45000,
+      "train_loss": 0.26589359577231936
+    },
+    {
+      "epoch": 44,
+      "eval_accuracy": 0.816,
+      "eval_loss": 0.6069492797851562,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.29825714230537415,
+        "mean_feedback_forward_cosine": 0.9811234944745114,
+        "mean_feedback_forward_relative_error": 0.1922379583120346,
+        "min_feedback_forward_cosine": 0.9567015171051025
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15488,
+      "train_examples": 45000,
+      "train_loss": 0.2647001281181971
+    },
+    {
+      "epoch": 45,
+      "eval_accuracy": 0.8246,
+      "eval_loss": 0.6435483520507812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.28758206963539124,
+        "mean_feedback_forward_cosine": 0.9823637510600843,
+        "mean_feedback_forward_relative_error": 0.18576494329854062,
+        "min_feedback_forward_cosine": 0.9596829414367676
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15840,
+      "train_examples": 45000,
+      "train_loss": 0.25985294319258795
+    },
+    {
+      "epoch": 46,
+      "eval_accuracy": 0.845,
+      "eval_loss": 0.5288216491699219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2769462466239929,
+        "mean_feedback_forward_cosine": 0.9835391013245833,
+        "mean_feedback_forward_relative_error": 0.17942392904507487,
+        "min_feedback_forward_cosine": 0.96257483959198
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16192,
+      "train_examples": 45000,
+      "train_loss": 0.2615274556848738
+    },
+    {
+      "epoch": 47,
+      "eval_accuracy": 0.8418,
+      "eval_loss": 0.54138017578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2668825387954712,
+        "mean_feedback_forward_cosine": 0.984656035900116,
+        "mean_feedback_forward_relative_error": 0.17317641526460648,
+        "min_feedback_forward_cosine": 0.9652228355407715
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16544,
+      "train_examples": 45000,
+      "train_loss": 0.2564132101482815
+    },
+    {
+      "epoch": 48,
+      "eval_accuracy": 0.8372,
+      "eval_loss": 0.5357763732910156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.25704485177993774,
+        "mean_feedback_forward_cosine": 0.9856965698693928,
+        "mean_feedback_forward_relative_error": 0.1671780539970649,
+        "min_feedback_forward_cosine": 0.9677011966705322
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16896,
+      "train_examples": 45000,
+      "train_loss": 0.25630424637264676
+    },
+    {
+      "epoch": 49,
+      "eval_accuracy": 0.8638,
+      "eval_loss": 0.44562494506835937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.24754087626934052,
+        "mean_feedback_forward_cosine": 0.9866711967869809,
+        "mean_feedback_forward_relative_error": 0.1613692139324389,
+        "min_feedback_forward_cosine": 0.9700177311897278
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17248,
+      "train_examples": 45000,
+      "train_loss": 0.25752347220314875
+    },
+    {
+      "epoch": 50,
+      "eval_accuracy": 0.865,
+      "eval_loss": 0.4701431762695312,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.23877203464508057,
+        "mean_feedback_forward_cosine": 0.9875494618164865,
+        "mean_feedback_forward_relative_error": 0.1559329197594994,
+        "min_feedback_forward_cosine": 0.9720914363861084
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17600,
+      "train_examples": 45000,
+      "train_loss": 0.2516157629966736
+    },
+    {
+      "epoch": 51,
+      "eval_accuracy": 0.854,
+      "eval_loss": 0.4998943542480469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.22992095351219177,
+        "mean_feedback_forward_cosine": 0.9883927734274613,
+        "mean_feedback_forward_relative_error": 0.1505319970218759,
+        "min_feedback_forward_cosine": 0.9741135239601135
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17952,
+      "train_examples": 45000,
+      "train_loss": 0.24805901011890835
+    },
+    {
+      "epoch": 52,
+      "eval_accuracy": 0.8616,
+      "eval_loss": 0.47119757690429687,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.22150832414627075,
+        "mean_feedback_forward_cosine": 0.9891759470889443,
+        "mean_feedback_forward_relative_error": 0.1453473183669542,
+        "min_feedback_forward_cosine": 0.9759514927864075
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18304,
+      "train_examples": 45000,
+      "train_loss": 0.24537688566843668
+    },
+    {
+      "epoch": 53,
+      "eval_accuracy": 0.8556,
+      "eval_loss": 0.48224800415039065,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.21376891434192657,
+        "mean_feedback_forward_cosine": 0.989898261271025,
+        "mean_feedback_forward_relative_error": 0.14038500679950966,
+        "min_feedback_forward_cosine": 0.9775823354721069
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18656,
+      "train_examples": 45000,
+      "train_loss": 0.2405990086025662
+    },
+    {
+      "epoch": 54,
+      "eval_accuracy": 0.8436,
+      "eval_loss": 0.5470617431640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.20610255002975464,
+        "mean_feedback_forward_cosine": 0.9905684119776675,
+        "mean_feedback_forward_relative_error": 0.1356346622893685,
+        "min_feedback_forward_cosine": 0.9791518449783325
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19008,
+      "train_examples": 45000,
+      "train_loss": 0.2398143388748169
+    },
+    {
+      "epoch": 55,
+      "eval_accuracy": 0.8622,
+      "eval_loss": 0.45401356811523436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.19825775921344757,
+        "mean_feedback_forward_cosine": 0.9912248034226266,
+        "mean_feedback_forward_relative_error": 0.13083300131716227,
+        "min_feedback_forward_cosine": 0.9806845784187317
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19360,
+      "train_examples": 45000,
+      "train_loss": 0.2417209722412957
+    },
+    {
+      "epoch": 56,
+      "eval_accuracy": 0.8272,
+      "eval_loss": 0.6230890991210938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.19106264412403107,
+        "mean_feedback_forward_cosine": 0.9918148360754314,
+        "mean_feedback_forward_relative_error": 0.12633993947192243,
+        "min_feedback_forward_cosine": 0.9820552468299866
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19712,
+      "train_examples": 45000,
+      "train_loss": 0.23602942765553792
+    },
+    {
+      "epoch": 57,
+      "eval_accuracy": 0.8342,
+      "eval_loss": 0.5783703979492187,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.18371108174324036,
+        "mean_feedback_forward_cosine": 0.9923666746992814,
+        "mean_feedback_forward_relative_error": 0.12198558371318013,
+        "min_feedback_forward_cosine": 0.9833942651748657
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20064,
+      "train_examples": 45000,
+      "train_loss": 0.2385865336894989
+    },
+    {
+      "epoch": 58,
+      "eval_accuracy": 0.8152,
+      "eval_loss": 0.6801926147460937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.17712336778640747,
+        "mean_feedback_forward_cosine": 0.992893824451848,
+        "mean_feedback_forward_relative_error": 0.11769728735089302,
+        "min_feedback_forward_cosine": 0.9845579266548157
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20416,
+      "train_examples": 45000,
+      "train_loss": 0.23616963873969185
+    },
+    {
+      "epoch": 59,
+      "eval_accuracy": 0.8444,
+      "eval_loss": 0.5382642761230468,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1707431674003601,
+        "mean_feedback_forward_cosine": 0.9933741876953527,
+        "mean_feedback_forward_relative_error": 0.113632299594189,
+        "min_feedback_forward_cosine": 0.9856466054916382
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20768,
+      "train_examples": 45000,
+      "train_loss": 0.2340568670378791
+    },
+    {
+      "epoch": 60,
+      "eval_accuracy": 0.8684,
+      "eval_loss": 0.42916826171875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1647421419620514,
+        "mean_feedback_forward_cosine": 0.9938058602182489,
+        "mean_feedback_forward_relative_error": 0.10984297096729279,
+        "min_feedback_forward_cosine": 0.9866265654563904
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21120,
+      "train_examples": 45000,
+      "train_loss": 0.22967808141708373
+    },
+    {
+      "epoch": 61,
+      "eval_accuracy": 0.8644,
+      "eval_loss": 0.4622958190917969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1589381992816925,
+        "mean_feedback_forward_cosine": 0.9942190270674857,
+        "mean_feedback_forward_relative_error": 0.1061004608085281,
+        "min_feedback_forward_cosine": 0.9875530004501343
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21472,
+      "train_examples": 45000,
+      "train_loss": 0.22695105500353707
+    },
+    {
+      "epoch": 62,
+      "eval_accuracy": 0.8694,
+      "eval_loss": 0.4372173278808594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.15320304036140442,
+        "mean_feedback_forward_cosine": 0.9946047820542988,
+        "mean_feedback_forward_relative_error": 0.1024947821309692,
+        "min_feedback_forward_cosine": 0.9884278774261475
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21824,
+      "train_examples": 45000,
+      "train_loss": 0.22514011753400168
+    },
+    {
+      "epoch": 63,
+      "eval_accuracy": 0.8506,
+      "eval_loss": 0.534511572265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14787736535072327,
+        "mean_feedback_forward_cosine": 0.9949603143491244,
+        "mean_feedback_forward_relative_error": 0.0990475590683912,
+        "min_feedback_forward_cosine": 0.9892134666442871
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22176,
+      "train_examples": 45000,
+      "train_loss": 0.22366815836164686
+    },
+    {
+      "epoch": 64,
+      "eval_accuracy": 0.8492,
+      "eval_loss": 0.4916988586425781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14245589077472687,
+        "mean_feedback_forward_cosine": 0.9952998600508037,
+        "mean_feedback_forward_relative_error": 0.09564386621901863,
+        "min_feedback_forward_cosine": 0.9899836778640747
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22528,
+      "train_examples": 45000,
+      "train_loss": 0.22651206641462115
+    },
+    {
+      "epoch": 65,
+      "eval_accuracy": 0.8582,
+      "eval_loss": 0.4878775207519531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1371895670890808,
+        "mean_feedback_forward_cosine": 0.9956255743378087,
+        "mean_feedback_forward_relative_error": 0.09226729152233977,
+        "min_feedback_forward_cosine": 0.9907054901123047
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22880,
+      "train_examples": 45000,
+      "train_loss": 0.23046863102383083
+    },
+    {
+      "epoch": 66,
+      "eval_accuracy": 0.8532,
+      "eval_loss": 0.5311714599609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.13216394186019897,
+        "mean_feedback_forward_cosine": 0.9959230548457095,
+        "mean_feedback_forward_relative_error": 0.08907792697611608,
+        "min_feedback_forward_cosine": 0.991371750831604
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23232,
+      "train_examples": 45000,
+      "train_loss": 0.22570315517319572
+    },
+    {
+      "epoch": 67,
+      "eval_accuracy": 0.8616,
+      "eval_loss": 0.47444708251953127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.12750616669654846,
+        "mean_feedback_forward_cosine": 0.9961913918194018,
+        "mean_feedback_forward_relative_error": 0.08607606923109606,
+        "min_feedback_forward_cosine": 0.9919623136520386
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23584,
+      "train_examples": 45000,
+      "train_loss": 0.2197730148103502
+    },
+    {
+      "epoch": 68,
+      "eval_accuracy": 0.8686,
+      "eval_loss": 0.4310839172363281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.12284544110298157,
+        "mean_feedback_forward_cosine": 0.9964495489471837,
+        "mean_feedback_forward_relative_error": 0.08310531922861149,
+        "min_feedback_forward_cosine": 0.9925363063812256
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23936,
+      "train_examples": 45000,
+      "train_loss": 0.22210470549265543
+    },
+    {
+      "epoch": 69,
+      "eval_accuracy": 0.875,
+      "eval_loss": 0.39954692993164065,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11851232498884201,
+        "mean_feedback_forward_cosine": 0.9966876569547152,
+        "mean_feedback_forward_relative_error": 0.08026892731064245,
+        "min_feedback_forward_cosine": 0.9930498600006104
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24288,
+      "train_examples": 45000,
+      "train_loss": 0.22326679899957444
+    },
+    {
+      "epoch": 70,
+      "eval_accuracy": 0.8728,
+      "eval_loss": 0.431579052734375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1143125668168068,
+        "mean_feedback_forward_cosine": 0.9969074663363005,
+        "mean_feedback_forward_relative_error": 0.07755606699930995,
+        "min_feedback_forward_cosine": 0.9935322403907776
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24640,
+      "train_examples": 45000,
+      "train_loss": 0.21717988074090747
+    },
+    {
+      "epoch": 71,
+      "eval_accuracy": 0.8372,
+      "eval_loss": 0.6019211547851563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11009687930345535,
+        "mean_feedback_forward_cosine": 0.9971224697012651,
+        "mean_feedback_forward_relative_error": 0.07480305041137494,
+        "min_feedback_forward_cosine": 0.9939974546432495
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24992,
+      "train_examples": 45000,
+      "train_loss": 0.21987548260688783
+    },
+    {
+      "epoch": 72,
+      "eval_accuracy": 0.844,
+      "eval_loss": 0.5271776062011718,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1060536578297615,
+        "mean_feedback_forward_cosine": 0.9973184591845462,
+        "mean_feedback_forward_relative_error": 0.07221450186089466,
+        "min_feedback_forward_cosine": 0.9944281578063965
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25344,
+      "train_examples": 45000,
+      "train_loss": 0.21815802181826696
+    },
+    {
+      "epoch": 73,
+      "eval_accuracy": 0.8644,
+      "eval_loss": 0.4716430725097656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.10231104493141174,
+        "mean_feedback_forward_cosine": 0.9974972291996604,
+        "mean_feedback_forward_relative_error": 0.06975691196949858,
+        "min_feedback_forward_cosine": 0.994812548160553
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25696,
+      "train_examples": 45000,
+      "train_loss": 0.21477413632074993
+    },
+    {
+      "epoch": 74,
+      "eval_accuracy": 0.8756,
+      "eval_loss": 0.39738220825195314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09869157522916794,
+        "mean_feedback_forward_cosine": 0.9976618415430972,
+        "mean_feedback_forward_relative_error": 0.06742280917732339,
+        "min_feedback_forward_cosine": 0.9951719045639038
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26048,
+      "train_examples": 45000,
+      "train_loss": 0.2134038785510593
+    },
+    {
+      "epoch": 75,
+      "eval_accuracy": 0.8396,
+      "eval_loss": 0.5778799682617187,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09520690888166428,
+        "mean_feedback_forward_cosine": 0.9978216541440863,
+        "mean_feedback_forward_relative_error": 0.0650710579203932,
+        "min_feedback_forward_cosine": 0.9955058097839355
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26400,
+      "train_examples": 45000,
+      "train_loss": 0.21306624372800193
+    },
+    {
+      "epoch": 76,
+      "eval_accuracy": 0.8564,
+      "eval_loss": 0.50355625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09168129414319992,
+        "mean_feedback_forward_cosine": 0.9979705936030338,
+        "mean_feedback_forward_relative_error": 0.06280808003717347,
+        "min_feedback_forward_cosine": 0.9958318471908569
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26752,
+      "train_examples": 45000,
+      "train_loss": 0.21620365244812437
+    },
+    {
+      "epoch": 77,
+      "eval_accuracy": 0.8696,
+      "eval_loss": 0.44153466796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08841861039400101,
+        "mean_feedback_forward_cosine": 0.9981047730696829,
+        "mean_feedback_forward_relative_error": 0.0606959780776187,
+        "min_feedback_forward_cosine": 0.9961221218109131
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27104,
+      "train_examples": 45000,
+      "train_loss": 0.2079311021433936
+    },
+    {
+      "epoch": 78,
+      "eval_accuracy": 0.8704,
+      "eval_loss": 0.4515508117675781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08516950905323029,
+        "mean_feedback_forward_cosine": 0.9982353982172514,
+        "mean_feedback_forward_relative_error": 0.058566737037740256,
+        "min_feedback_forward_cosine": 0.9964010715484619
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27456,
+      "train_examples": 45000,
+      "train_loss": 0.21112659882969326
+    },
+    {
+      "epoch": 79,
+      "eval_accuracy": 0.8624,
+      "eval_loss": 0.47880274658203126,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08209028840065002,
+        "mean_feedback_forward_cosine": 0.9983568599349574,
+        "mean_feedback_forward_relative_error": 0.05651370122244483,
+        "min_feedback_forward_cosine": 0.9966554045677185
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27808,
+      "train_examples": 45000,
+      "train_loss": 0.21050622888141207
+    },
+    {
+      "epoch": 80,
+      "eval_accuracy": 0.8636,
+      "eval_loss": 0.4666690185546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.07915076613426208,
+        "mean_feedback_forward_cosine": 0.9984677810417978,
+        "mean_feedback_forward_relative_error": 0.05457279988025364,
+        "min_feedback_forward_cosine": 0.9968898296356201
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28160,
+      "train_examples": 45000,
+      "train_loss": 0.21067021816306644
+    },
+    {
+      "epoch": 81,
+      "eval_accuracy": 0.8598,
+      "eval_loss": 0.5046466125488281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.07637186348438263,
+        "mean_feedback_forward_cosine": 0.9985681176185608,
+        "mean_feedback_forward_relative_error": 0.05274927233786959,
+        "min_feedback_forward_cosine": 0.9971036911010742
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28512,
+      "train_examples": 45000,
+      "train_loss": 0.20502160912619696
+    },
+    {
+      "epoch": 82,
+      "eval_accuracy": 0.8576,
+      "eval_loss": 0.4760834167480469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0736217275261879,
+        "mean_feedback_forward_cosine": 0.9986653202458432,
+        "mean_feedback_forward_relative_error": 0.050925137965302715,
+        "min_feedback_forward_cosine": 0.997308611869812
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28864,
+      "train_examples": 45000,
+      "train_loss": 0.20797349837091234
+    },
+    {
+      "epoch": 83,
+      "eval_accuracy": 0.8376,
+      "eval_loss": 0.611455859375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.07095649093389511,
+        "mean_feedback_forward_cosine": 0.9987575662763495,
+        "mean_feedback_forward_relative_error": 0.049138400703668594,
+        "min_feedback_forward_cosine": 0.997499406337738
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29216,
+      "train_examples": 45000,
+      "train_loss": 0.20943366742134095
+    },
+    {
+      "epoch": 84,
+      "eval_accuracy": 0.8716,
+      "eval_loss": 0.4457356201171875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06834620982408524,
+        "mean_feedback_forward_cosine": 0.9988417311718589,
+        "mean_feedback_forward_relative_error": 0.04744118040329531,
+        "min_feedback_forward_cosine": 0.9976792335510254
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29568,
+      "train_examples": 45000,
+      "train_loss": 0.2061102368089888
+    },
+    {
+      "epoch": 85,
+      "eval_accuracy": 0.8524,
+      "eval_loss": 0.5395504333496094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06592147052288055,
+        "mean_feedback_forward_cosine": 0.9989195466041565,
+        "mean_feedback_forward_relative_error": 0.04582055912990319,
+        "min_feedback_forward_cosine": 0.9978403449058533
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29920,
+      "train_examples": 45000,
+      "train_loss": 0.20714529931810166
+    },
+    {
+      "epoch": 86,
+      "eval_accuracy": 0.875,
+      "eval_loss": 0.4334523864746094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06359504163265228,
+        "mean_feedback_forward_cosine": 0.9989917152806332,
+        "mean_feedback_forward_relative_error": 0.04426226617866441,
+        "min_feedback_forward_cosine": 0.9979900121688843
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30272,
+      "train_examples": 45000,
+      "train_loss": 0.20295997999509174
+    },
+    {
+      "epoch": 87,
+      "eval_accuracy": 0.873,
+      "eval_loss": 0.43320928955078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.06130767986178398,
+        "mean_feedback_forward_cosine": 0.9990615468276175,
+        "mean_feedback_forward_relative_error": 0.04270692159862895,
+        "min_feedback_forward_cosine": 0.9981316328048706
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30624,
+      "train_examples": 45000,
+      "train_loss": 0.20736424657503763
+    },
+    {
+      "epoch": 88,
+      "eval_accuracy": 0.8666,
+      "eval_loss": 0.45640607299804686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.059117674827575684,
+        "mean_feedback_forward_cosine": 0.9991248218636763,
+        "mean_feedback_forward_relative_error": 0.04123614924518686,
+        "min_feedback_forward_cosine": 0.9982626438140869
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30976,
+      "train_examples": 45000,
+      "train_loss": 0.20548082500298817
+    },
+    {
+      "epoch": 89,
+      "eval_accuracy": 0.874,
+      "eval_loss": 0.41520469360351564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.056990038603544235,
+        "mean_feedback_forward_cosine": 0.9991847339429354,
+        "mean_feedback_forward_relative_error": 0.039800861948414853,
+        "min_feedback_forward_cosine": 0.9983851909637451
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31328,
+      "train_examples": 45000,
+      "train_loss": 0.20585563192102643
+    },
+    {
+      "epoch": 90,
+      "eval_accuracy": 0.839,
+      "eval_loss": 0.569453662109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.055011142045259476,
+        "mean_feedback_forward_cosine": 0.9992381836238661,
+        "mean_feedback_forward_relative_error": 0.038467583197512124,
+        "min_feedback_forward_cosine": 0.9984948635101318
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31680,
+      "train_examples": 45000,
+      "train_loss": 0.20076531094180214
+    },
+    {
+      "epoch": 91,
+      "eval_accuracy": 0.8722,
+      "eval_loss": 0.4262968078613281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.052972324192523956,
+        "mean_feedback_forward_cosine": 0.9992925712936803,
+        "mean_feedback_forward_relative_error": 0.03706920274386281,
+        "min_feedback_forward_cosine": 0.9986041784286499
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32032,
+      "train_examples": 45000,
+      "train_loss": 0.20624251328044468
+    },
+    {
+      "epoch": 92,
+      "eval_accuracy": 0.871,
+      "eval_loss": 0.4491199462890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.05109149217605591,
+        "mean_feedback_forward_cosine": 0.9993400699213931,
+        "mean_feedback_forward_relative_error": 0.03580162468317308,
+        "min_feedback_forward_cosine": 0.9987013339996338
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32384,
+      "train_examples": 45000,
+      "train_loss": 0.19799577930238513
+    },
+    {
+      "epoch": 93,
+      "eval_accuracy": 0.8736,
+      "eval_loss": 0.4416592041015625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0493144765496254,
+        "mean_feedback_forward_cosine": 0.9993836095458583,
+        "mean_feedback_forward_relative_error": 0.03459886961469525,
+        "min_feedback_forward_cosine": 0.9987897276878357
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32736,
+      "train_examples": 45000,
+      "train_loss": 0.19725176190800137
+    },
+    {
+      "epoch": 94,
+      "eval_accuracy": 0.8626,
+      "eval_loss": 0.47947452392578127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.047581423074007034,
+        "mean_feedback_forward_cosine": 0.9994248214520907,
+        "mean_feedback_forward_relative_error": 0.03341968107576433,
+        "min_feedback_forward_cosine": 0.9988730549812317
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33088,
+      "train_examples": 45000,
+      "train_loss": 0.19693422346115114
+    },
+    {
+      "epoch": 95,
+      "eval_accuracy": 0.8508,
+      "eval_loss": 0.5477066467285157,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04589957743883133,
+        "mean_feedback_forward_cosine": 0.9994630186181319,
+        "mean_feedback_forward_relative_error": 0.032289010129476846,
+        "min_feedback_forward_cosine": 0.9989511966705322
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33440,
+      "train_examples": 45000,
+      "train_loss": 0.19458590910699633
+    },
+    {
+      "epoch": 96,
+      "eval_accuracy": 0.8694,
+      "eval_loss": 0.4502195617675781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04428688809275627,
+        "mean_feedback_forward_cosine": 0.9994986276877554,
+        "mean_feedback_forward_relative_error": 0.031198943376933273,
+        "min_feedback_forward_cosine": 0.9990235567092896
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33792,
+      "train_examples": 45000,
+      "train_loss": 0.19458247496816847
+    },
+    {
+      "epoch": 97,
+      "eval_accuracy": 0.8682,
+      "eval_loss": 0.46479819946289064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.04265602305531502,
+        "mean_feedback_forward_cosine": 0.9995334713082564,
+        "mean_feedback_forward_relative_error": 0.030095937563792655,
+        "min_feedback_forward_cosine": 0.9990940690040588
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34144,
+      "train_examples": 45000,
+      "train_loss": 0.19885163509315915
+    },
+    {
+      "epoch": 98,
+      "eval_accuracy": 0.8678,
+      "eval_loss": 0.4972033447265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.041172705590724945,
+        "mean_feedback_forward_cosine": 0.9995646257149545,
+        "mean_feedback_forward_relative_error": 0.029072351851745656,
+        "min_feedback_forward_cosine": 0.9991556406021118
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34496,
+      "train_examples": 45000,
+      "train_loss": 0.19531077443758646
+    },
+    {
+      "epoch": 99,
+      "eval_accuracy": 0.8752,
+      "eval_loss": 0.40619945068359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03969261795282364,
+        "mean_feedback_forward_cosine": 0.9995948766407213,
+        "mean_feedback_forward_relative_error": 0.028044214913327443,
+        "min_feedback_forward_cosine": 0.9992151260375977
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34848,
+      "train_examples": 45000,
+      "train_loss": 0.1985523914443122
+    },
+    {
+      "epoch": 100,
+      "eval_accuracy": 0.8546,
+      "eval_loss": 0.5547154479980468,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03826097771525383,
+        "mean_feedback_forward_cosine": 0.9996227025985718,
+        "mean_feedback_forward_relative_error": 0.027064508856519273,
+        "min_feedback_forward_cosine": 0.9992707967758179
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 35200,
+      "train_examples": 45000,
+      "train_loss": 0.19719866987334358
+    },
+    {
+      "epoch": 101,
+      "eval_accuracy": 0.9096,
+      "eval_loss": 0.31504560546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03824806213378906,
+        "mean_feedback_forward_cosine": 0.9996229880734494,
+        "mean_feedback_forward_relative_error": 0.02705035456701329,
+        "min_feedback_forward_cosine": 0.9992711544036865
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35552,
+      "train_examples": 45000,
+      "train_loss": 0.12498004048400455
+    },
+    {
+      "epoch": 102,
+      "eval_accuracy": 0.9116,
+      "eval_loss": 0.31542617797851563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03824153542518616,
+        "mean_feedback_forward_cosine": 0.9996231637502971,
+        "mean_feedback_forward_relative_error": 0.02704000105395129,
+        "min_feedback_forward_cosine": 0.9992715716362
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35904,
+      "train_examples": 45000,
+      "train_loss": 0.09414871436489954
+    },
+    {
+      "epoch": 103,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.3152183624267578,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03823477402329445,
+        "mean_feedback_forward_cosine": 0.9996233425642315,
+        "mean_feedback_forward_relative_error": 0.02702999928672063,
+        "min_feedback_forward_cosine": 0.999271810054779
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36256,
+      "train_examples": 45000,
+      "train_loss": 0.08886900408930248
+    },
+    {
+      "epoch": 104,
+      "eval_accuracy": 0.9156,
+      "eval_loss": 0.3212423614501953,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038228511810302734,
+        "mean_feedback_forward_cosine": 0.999623533926512,
+        "mean_feedback_forward_relative_error": 0.027020629740467195,
+        "min_feedback_forward_cosine": 0.9992719888687134
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36608,
+      "train_examples": 45000,
+      "train_loss": 0.07943104676802953
+    },
+    {
+      "epoch": 105,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.3224820953369141,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03822202607989311,
+        "mean_feedback_forward_cosine": 0.9996236876437539,
+        "mean_feedback_forward_relative_error": 0.02701137523706022,
+        "min_feedback_forward_cosine": 0.9992722272872925
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36960,
+      "train_examples": 45000,
+      "train_loss": 0.07692426940202712
+    },
+    {
+      "epoch": 106,
+      "eval_accuracy": 0.9132,
+      "eval_loss": 0.3243196716308594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03821607679128647,
+        "mean_feedback_forward_cosine": 0.9996238539093419,
+        "mean_feedback_forward_relative_error": 0.027002569288015366,
+        "min_feedback_forward_cosine": 0.9992725253105164
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37312,
+      "train_examples": 45000,
+      "train_loss": 0.0701384553776847
+    },
+    {
+      "epoch": 107,
+      "eval_accuracy": 0.9132,
+      "eval_loss": 0.3256701171875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038209863007068634,
+        "mean_feedback_forward_cosine": 0.9996240139007568,
+        "mean_feedback_forward_relative_error": 0.02699356859451846,
+        "min_feedback_forward_cosine": 0.9992727041244507
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37664,
+      "train_examples": 45000,
+      "train_loss": 0.06748027348253462
+    },
+    {
+      "epoch": 108,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.333216845703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038203537464141846,
+        "mean_feedback_forward_cosine": 0.999624208400124,
+        "mean_feedback_forward_relative_error": 0.026984702501642078,
+        "min_feedback_forward_cosine": 0.9992730021476746
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38016,
+      "train_examples": 45000,
+      "train_loss": 0.06455075730615192
+    },
+    {
+      "epoch": 109,
+      "eval_accuracy": 0.9104,
+      "eval_loss": 0.34080203857421876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03819728270173073,
+        "mean_feedback_forward_cosine": 0.9996243401577598,
+        "mean_feedback_forward_relative_error": 0.02697615119579591,
+        "min_feedback_forward_cosine": 0.9992731809616089
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38368,
+      "train_examples": 45000,
+      "train_loss": 0.06166491029262543
+    },
+    {
+      "epoch": 110,
+      "eval_accuracy": 0.914,
+      "eval_loss": 0.3373171447753906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038191262632608414,
+        "mean_feedback_forward_cosine": 0.9996245283829538,
+        "mean_feedback_forward_relative_error": 0.02696770125705945,
+        "min_feedback_forward_cosine": 0.999273419380188
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38720,
+      "train_examples": 45000,
+      "train_loss": 0.05793297163115607
+    },
+    {
+      "epoch": 111,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.3412690979003906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0381854884326458,
+        "mean_feedback_forward_cosine": 0.9996246883743688,
+        "mean_feedback_forward_relative_error": 0.026959390379488468,
+        "min_feedback_forward_cosine": 0.9992736577987671
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39072,
+      "train_examples": 45000,
+      "train_loss": 0.05538705672224363
+    },
+    {
+      "epoch": 112,
+      "eval_accuracy": 0.9112,
+      "eval_loss": 0.34866596069335937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03817934915423393,
+        "mean_feedback_forward_cosine": 0.9996248546399569,
+        "mean_feedback_forward_relative_error": 0.026950864709521596,
+        "min_feedback_forward_cosine": 0.9992738962173462
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39424,
+      "train_examples": 45000,
+      "train_loss": 0.05465268326600393
+    },
+    {
+      "epoch": 113,
+      "eval_accuracy": 0.9114,
+      "eval_loss": 0.3563314697265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03817310184240341,
+        "mean_feedback_forward_cosine": 0.9996250209055448,
+        "mean_feedback_forward_relative_error": 0.026942566282262926,
+        "min_feedback_forward_cosine": 0.9992742538452148
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39776,
+      "train_examples": 45000,
+      "train_loss": 0.05348444962501526
+    },
+    {
+      "epoch": 114,
+      "eval_accuracy": 0.9146,
+      "eval_loss": 0.3597590148925781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03816705942153931,
+        "mean_feedback_forward_cosine": 0.9996251683486136,
+        "mean_feedback_forward_relative_error": 0.026934282364029633,
+        "min_feedback_forward_cosine": 0.9992744326591492
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40128,
+      "train_examples": 45000,
+      "train_loss": 0.05038708394269149
+    },
+    {
+      "epoch": 115,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.3545485961914063,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038161080330610275,
+        "mean_feedback_forward_cosine": 0.9996253597108942,
+        "mean_feedback_forward_relative_error": 0.02692613059556798,
+        "min_feedback_forward_cosine": 0.9992746710777283
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40480,
+      "train_examples": 45000,
+      "train_loss": 0.047704472191300654
+    },
+    {
+      "epoch": 116,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.3614710632324219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038155052810907364,
+        "mean_feedback_forward_cosine": 0.9996255008797896,
+        "mean_feedback_forward_relative_error": 0.026917944368171066,
+        "min_feedback_forward_cosine": 0.999274730682373
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40832,
+      "train_examples": 45000,
+      "train_loss": 0.046668650693363616
+    },
+    {
+      "epoch": 117,
+      "eval_accuracy": 0.9118,
+      "eval_loss": 0.36147152099609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038148727267980576,
+        "mean_feedback_forward_cosine": 0.9996256859678971,
+        "mean_feedback_forward_relative_error": 0.026909909701268924,
+        "min_feedback_forward_cosine": 0.9992750883102417
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41184,
+      "train_examples": 45000,
+      "train_loss": 0.0468660827504264
+    },
+    {
+      "epoch": 118,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.36134048461914064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03814224898815155,
+        "mean_feedback_forward_cosine": 0.9996258585076583,
+        "mean_feedback_forward_relative_error": 0.026901894298038985,
+        "min_feedback_forward_cosine": 0.9992753863334656
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41536,
+      "train_examples": 45000,
+      "train_loss": 0.04667721270521482
+    },
+    {
+      "epoch": 119,
+      "eval_accuracy": 0.9118,
+      "eval_loss": 0.3710879089355469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03813629969954491,
+        "mean_feedback_forward_cosine": 0.9996260247732464,
+        "mean_feedback_forward_relative_error": 0.026893928549007365,
+        "min_feedback_forward_cosine": 0.9992755651473999
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41888,
+      "train_examples": 45000,
+      "train_loss": 0.04263128900461727
+    },
+    {
+      "epoch": 120,
+      "eval_accuracy": 0.9124,
+      "eval_loss": 0.37179076538085937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038130130618810654,
+        "mean_feedback_forward_cosine": 0.9996262004500941,
+        "mean_feedback_forward_relative_error": 0.026885935301451308,
+        "min_feedback_forward_cosine": 0.9992758631706238
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42240,
+      "train_examples": 45000,
+      "train_loss": 0.04206890949805578
+    },
+    {
+      "epoch": 121,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.3723926452636719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038124118000268936,
+        "mean_feedback_forward_cosine": 0.9996263635785956,
+        "mean_feedback_forward_relative_error": 0.02687794715166092,
+        "min_feedback_forward_cosine": 0.9992761015892029
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42592,
+      "train_examples": 45000,
+      "train_loss": 0.04026431170701981
+    },
+    {
+      "epoch": 122,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.3758395263671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03811798617243767,
+        "mean_feedback_forward_cosine": 0.9996265361183568,
+        "mean_feedback_forward_relative_error": 0.02686990219119348,
+        "min_feedback_forward_cosine": 0.9992762804031372
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42944,
+      "train_examples": 45000,
+      "train_loss": 0.03920797461072604
+    },
+    {
+      "epoch": 123,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.3701650634765625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03811146691441536,
+        "mean_feedback_forward_cosine": 0.9996267274806374,
+        "mean_feedback_forward_relative_error": 0.02686190257143033,
+        "min_feedback_forward_cosine": 0.9992765188217163
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43296,
+      "train_examples": 45000,
+      "train_loss": 0.040231052759620876
+    },
+    {
+      "epoch": 124,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.3734594177246094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038105301558971405,
+        "mean_feedback_forward_cosine": 0.9996268906091389,
+        "mean_feedback_forward_relative_error": 0.026853734676383044,
+        "min_feedback_forward_cosine": 0.9992768168449402
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43648,
+      "train_examples": 45000,
+      "train_loss": 0.038030407783057954
+    },
+    {
+      "epoch": 125,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.37734255981445314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038098517805337906,
+        "mean_feedback_forward_cosine": 0.9996270788343329,
+        "mean_feedback_forward_relative_error": 0.02684540046673072,
+        "min_feedback_forward_cosine": 0.9992770552635193
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44000,
+      "train_examples": 45000,
+      "train_loss": 0.03932609410782655
+    },
+    {
+      "epoch": 126,
+      "eval_accuracy": 0.9162,
+      "eval_loss": 0.37866737670898437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0380915105342865,
+        "mean_feedback_forward_cosine": 0.999627267059527,
+        "mean_feedback_forward_relative_error": 0.026837026700377464,
+        "min_feedback_forward_cosine": 0.9992772936820984
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44352,
+      "train_examples": 45000,
+      "train_loss": 0.037950586869650416
+    },
+    {
+      "epoch": 127,
+      "eval_accuracy": 0.9132,
+      "eval_loss": 0.39278297119140626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038084693253040314,
+        "mean_feedback_forward_cosine": 0.9996274395992881,
+        "mean_feedback_forward_relative_error": 0.02682868611851805,
+        "min_feedback_forward_cosine": 0.9992775321006775
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44704,
+      "train_examples": 45000,
+      "train_loss": 0.03678174229264259
+    },
+    {
+      "epoch": 128,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.38518326416015625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03807844966650009,
+        "mean_feedback_forward_cosine": 0.9996276246873956,
+        "mean_feedback_forward_relative_error": 0.026820657823823,
+        "min_feedback_forward_cosine": 0.9992778301239014
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45056,
+      "train_examples": 45000,
+      "train_loss": 0.03376755590836207
+    },
+    {
+      "epoch": 129,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.3975419860839844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03807184472680092,
+        "mean_feedback_forward_cosine": 0.9996278191867628,
+        "mean_feedback_forward_relative_error": 0.026812335966449035,
+        "min_feedback_forward_cosine": 0.9992780685424805
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45408,
+      "train_examples": 45000,
+      "train_loss": 0.03459894592629539
+    },
+    {
+      "epoch": 130,
+      "eval_accuracy": 0.9134,
+      "eval_loss": 0.3980976745605469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03806491941213608,
+        "mean_feedback_forward_cosine": 0.9996280074119568,
+        "mean_feedback_forward_relative_error": 0.02680385995068048,
+        "min_feedback_forward_cosine": 0.9992783665657043
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45760,
+      "train_examples": 45000,
+      "train_loss": 0.034467040161291755
+    },
+    {
+      "epoch": 131,
+      "eval_accuracy": 0.9106,
+      "eval_loss": 0.401936376953125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03805799409747124,
+        "mean_feedback_forward_cosine": 0.999628157992112,
+        "mean_feedback_forward_relative_error": 0.026795381287995138,
+        "min_feedback_forward_cosine": 0.9992786049842834
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46112,
+      "train_examples": 45000,
+      "train_loss": 0.03290436211890645
+    },
+    {
+      "epoch": 132,
+      "eval_accuracy": 0.9158,
+      "eval_loss": 0.39828818969726565,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03805088624358177,
+        "mean_feedback_forward_cosine": 0.9996284026848642,
+        "mean_feedback_forward_relative_error": 0.02678674944725476,
+        "min_feedback_forward_cosine": 0.9992789030075073
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46464,
+      "train_examples": 45000,
+      "train_loss": 0.03265204352868928
+    },
+    {
+      "epoch": 133,
+      "eval_accuracy": 0.9134,
+      "eval_loss": 0.4009542297363281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0380437932908535,
+        "mean_feedback_forward_cosine": 0.999628600321318,
+        "mean_feedback_forward_relative_error": 0.02677821461111307,
+        "min_feedback_forward_cosine": 0.9992791414260864
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46816,
+      "train_examples": 45000,
+      "train_loss": 0.03201477077437772
+    },
+    {
+      "epoch": 134,
+      "eval_accuracy": 0.9138,
+      "eval_loss": 0.40583228759765627,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038036979734897614,
+        "mean_feedback_forward_cosine": 0.9996287540385598,
+        "mean_feedback_forward_relative_error": 0.026769806189756645,
+        "min_feedback_forward_cosine": 0.999279260635376
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47168,
+      "train_examples": 45000,
+      "train_loss": 0.031194838628504012
+    },
+    {
+      "epoch": 135,
+      "eval_accuracy": 0.9122,
+      "eval_loss": 0.3971060607910156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038029663264751434,
+        "mean_feedback_forward_cosine": 0.999628976771706,
+        "mean_feedback_forward_relative_error": 0.026760906618284553,
+        "min_feedback_forward_cosine": 0.9992796182632446
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47520,
+      "train_examples": 45000,
+      "train_loss": 0.03187097754743364
+    },
+    {
+      "epoch": 136,
+      "eval_accuracy": 0.9136,
+      "eval_loss": 0.40547701416015625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038022592663764954,
+        "mean_feedback_forward_cosine": 0.9996291806823329,
+        "mean_feedback_forward_relative_error": 0.026752426828208723,
+        "min_feedback_forward_cosine": 0.999279797077179
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47872,
+      "train_examples": 45000,
+      "train_loss": 0.029618267251054445
+    },
+    {
+      "epoch": 137,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.40041107177734375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.038015320897102356,
+        "mean_feedback_forward_cosine": 0.9996294034154791,
+        "mean_feedback_forward_relative_error": 0.02674361695780566,
+        "min_feedback_forward_cosine": 0.9992802143096924
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48224,
+      "train_examples": 45000,
+      "train_loss": 0.03065378427306811
+    },
+    {
+      "epoch": 138,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.4129467590332031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03800799697637558,
+        "mean_feedback_forward_cosine": 0.9996295885035866,
+        "mean_feedback_forward_relative_error": 0.026734913454244013,
+        "min_feedback_forward_cosine": 0.9992804527282715
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48576,
+      "train_examples": 45000,
+      "train_loss": 0.02922247618569268
+    },
+    {
+      "epoch": 139,
+      "eval_accuracy": 0.916,
+      "eval_loss": 0.40856820678710937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03800143301486969,
+        "mean_feedback_forward_cosine": 0.999629773591694,
+        "mean_feedback_forward_relative_error": 0.02672667232783217,
+        "min_feedback_forward_cosine": 0.9992806911468506
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48928,
+      "train_examples": 45000,
+      "train_loss": 0.02787504948510064
+    },
+    {
+      "epoch": 140,
+      "eval_accuracy": 0.912,
+      "eval_loss": 0.43249401245117186,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03799450770020485,
+        "mean_feedback_forward_cosine": 0.999629977502321,
+        "mean_feedback_forward_relative_error": 0.026718208664341977,
+        "min_feedback_forward_cosine": 0.9992810487747192
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49280,
+      "train_examples": 45000,
+      "train_loss": 0.027838887613349492
+    },
+    {
+      "epoch": 141,
+      "eval_accuracy": 0.9126,
+      "eval_loss": 0.4213609069824219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037987120449543,
+        "mean_feedback_forward_cosine": 0.9996301782758612,
+        "mean_feedback_forward_relative_error": 0.02670954457043033,
+        "min_feedback_forward_cosine": 0.9992812275886536
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49632,
+      "train_examples": 45000,
+      "train_loss": 0.02806180555969477
+    },
+    {
+      "epoch": 142,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.42442941284179686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037980616092681885,
+        "mean_feedback_forward_cosine": 0.9996303570897955,
+        "mean_feedback_forward_relative_error": 0.02670134569665319,
+        "min_feedback_forward_cosine": 0.9992814064025879
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49984,
+      "train_examples": 45000,
+      "train_loss": 0.025535029594269065
+    },
+    {
+      "epoch": 143,
+      "eval_accuracy": 0.914,
+      "eval_loss": 0.42216495971679685,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03797249123454094,
+        "mean_feedback_forward_cosine": 0.9996305798229418,
+        "mean_feedback_forward_relative_error": 0.026691938505360956,
+        "min_feedback_forward_cosine": 0.9992818236351013
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50336,
+      "train_examples": 45000,
+      "train_loss": 0.028624223914245763
+    },
+    {
+      "epoch": 144,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.4278658020019531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037965077906847,
+        "mean_feedback_forward_cosine": 0.9996307805964821,
+        "mean_feedback_forward_relative_error": 0.026683067559803788,
+        "min_feedback_forward_cosine": 0.9992820620536804
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50688,
+      "train_examples": 45000,
+      "train_loss": 0.026180000118662915
+    },
+    {
+      "epoch": 145,
+      "eval_accuracy": 0.9138,
+      "eval_loss": 0.427544677734375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03795746713876724,
+        "mean_feedback_forward_cosine": 0.9996310190150612,
+        "mean_feedback_forward_relative_error": 0.026674046377210242,
+        "min_feedback_forward_cosine": 0.9992823600769043
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51040,
+      "train_examples": 45000,
+      "train_loss": 0.026351813669668305
+    },
+    {
+      "epoch": 146,
+      "eval_accuracy": 0.9158,
+      "eval_loss": 0.42938318481445314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03794943168759346,
+        "mean_feedback_forward_cosine": 0.9996312386111209,
+        "mean_feedback_forward_relative_error": 0.02666486049757192,
+        "min_feedback_forward_cosine": 0.999282717704773
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51392,
+      "train_examples": 45000,
+      "train_loss": 0.026147830514444245
+    },
+    {
+      "epoch": 147,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.43108264770507815,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037941813468933105,
+        "mean_feedback_forward_cosine": 0.9996314613442672,
+        "mean_feedback_forward_relative_error": 0.026655893233653746,
+        "min_feedback_forward_cosine": 0.999282956123352
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51744,
+      "train_examples": 45000,
+      "train_loss": 0.025178877196378176
+    },
+    {
+      "epoch": 148,
+      "eval_accuracy": 0.912,
+      "eval_loss": 0.43701239624023436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03793426975607872,
+        "mean_feedback_forward_cosine": 0.9996316495694613,
+        "mean_feedback_forward_relative_error": 0.026646792251420647,
+        "min_feedback_forward_cosine": 0.9992832541465759
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52096,
+      "train_examples": 45000,
+      "train_loss": 0.024480403775639006
+    },
+    {
+      "epoch": 149,
+      "eval_accuracy": 0.9118,
+      "eval_loss": 0.4395541748046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037926580756902695,
+        "mean_feedback_forward_cosine": 0.9996318691655209,
+        "mean_feedback_forward_relative_error": 0.02663764043858177,
+        "min_feedback_forward_cosine": 0.9992836117744446
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52448,
+      "train_examples": 45000,
+      "train_loss": 0.024357059074110454
+    },
+    {
+      "epoch": 150,
+      "eval_accuracy": 0.914,
+      "eval_loss": 0.4387383911132812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791917487978935,
+        "mean_feedback_forward_cosine": 0.9996321075841,
+        "mean_feedback_forward_relative_error": 0.02662865788136658,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52800,
+      "train_examples": 45000,
+      "train_loss": 0.023285648666487798
+    },
+    {
+      "epoch": 151,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.43026748046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0379190668463707,
+        "mean_feedback_forward_cosine": 0.9996320981728403,
+        "mean_feedback_forward_relative_error": 0.026628399218775724,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53152,
+      "train_examples": 45000,
+      "train_loss": 0.0211612816479471
+    },
+    {
+      "epoch": 152,
+      "eval_accuracy": 0.914,
+      "eval_loss": 0.42774334716796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918996065855026,
+        "mean_feedback_forward_cosine": 0.9996321013099269,
+        "mean_feedback_forward_relative_error": 0.026628176142510614,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53504,
+      "train_examples": 45000,
+      "train_loss": 0.02011557665401035
+    },
+    {
+      "epoch": 153,
+      "eval_accuracy": 0.9134,
+      "eval_loss": 0.4272701232910156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791893273591995,
+        "mean_feedback_forward_cosine": 0.9996320981728403,
+        "mean_feedback_forward_relative_error": 0.0266279557131623,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53856,
+      "train_examples": 45000,
+      "train_loss": 0.019754009400142564
+    },
+    {
+      "epoch": 154,
+      "eval_accuracy": 0.9158,
+      "eval_loss": 0.423570166015625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791886940598488,
+        "mean_feedback_forward_cosine": 0.9996320950357538,
+        "mean_feedback_forward_relative_error": 0.026627739205172186,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54208,
+      "train_examples": 45000,
+      "train_loss": 0.019484071310361225
+    },
+    {
+      "epoch": 155,
+      "eval_accuracy": 0.9162,
+      "eval_loss": 0.42812459716796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918820977211,
+        "mean_feedback_forward_cosine": 0.9996320981728403,
+        "mean_feedback_forward_relative_error": 0.02662751343297331,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54560,
+      "train_examples": 45000,
+      "train_loss": 0.01772839131222831
+    },
+    {
+      "epoch": 156,
+      "eval_accuracy": 0.9168,
+      "eval_loss": 0.4284971862792969,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791876137256622,
+        "mean_feedback_forward_cosine": 0.9996321044470134,
+        "mean_feedback_forward_relative_error": 0.026627286141248124,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54912,
+      "train_examples": 45000,
+      "train_loss": 0.01800470546417766
+    },
+    {
+      "epoch": 157,
+      "eval_accuracy": 0.9156,
+      "eval_loss": 0.425212060546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791870176792145,
+        "mean_feedback_forward_cosine": 0.9996321138582731,
+        "mean_feedback_forward_relative_error": 0.026627064094339545,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55264,
+      "train_examples": 45000,
+      "train_loss": 0.019227959014640913
+    },
+    {
+      "epoch": 158,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.4282466613769531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791864588856697,
+        "mean_feedback_forward_cosine": 0.9996321326807925,
+        "mean_feedback_forward_relative_error": 0.026626851164588804,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55616,
+      "train_examples": 45000,
+      "train_loss": 0.017664594351748625
+    },
+    {
+      "epoch": 159,
+      "eval_accuracy": 0.9174,
+      "eval_loss": 0.42084263916015624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791859373450279,
+        "mean_feedback_forward_cosine": 0.9996321138582731,
+        "mean_feedback_forward_relative_error": 0.026626633970361007,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55968,
+      "train_examples": 45000,
+      "train_loss": 0.017749026487767695
+    },
+    {
+      "epoch": 160,
+      "eval_accuracy": 0.9164,
+      "eval_loss": 0.42557537231445314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918537855148315,
+        "mean_feedback_forward_cosine": 0.9996321201324463,
+        "mean_feedback_forward_relative_error": 0.02662641520758993,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56320,
+      "train_examples": 45000,
+      "train_loss": 0.01739938084317578
+    },
+    {
+      "epoch": 161,
+      "eval_accuracy": 0.9164,
+      "eval_loss": 0.42110528564453126,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918489426374435,
+        "mean_feedback_forward_cosine": 0.9996321389549657,
+        "mean_feedback_forward_relative_error": 0.026626198503531907,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56672,
+      "train_examples": 45000,
+      "train_loss": 0.01759473463681837
+    },
+    {
+      "epoch": 162,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.42479563598632814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918440997600555,
+        "mean_feedback_forward_cosine": 0.999632151503312,
+        "mean_feedback_forward_relative_error": 0.026625978564353364,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57024,
+      "train_examples": 45000,
+      "train_loss": 0.01624706518434816
+    },
+    {
+      "epoch": 163,
+      "eval_accuracy": 0.9128,
+      "eval_loss": 0.4251840393066406,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791838884353638,
+        "mean_feedback_forward_cosine": 0.9996321075841,
+        "mean_feedback_forward_relative_error": 0.02662576034076904,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57376,
+      "train_examples": 45000,
+      "train_loss": 0.01809051398370001
+    },
+    {
+      "epoch": 164,
+      "eval_accuracy": 0.9162,
+      "eval_loss": 0.42689862060546874,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0379183292388916,
+        "mean_feedback_forward_cosine": 0.9996321232695329,
+        "mean_feedback_forward_relative_error": 0.02662554500918639,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57728,
+      "train_examples": 45000,
+      "train_loss": 0.017154754579729505
+    },
+    {
+      "epoch": 165,
+      "eval_accuracy": 0.9156,
+      "eval_loss": 0.42907059326171876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791828826069832,
+        "mean_feedback_forward_cosine": 0.9996321420920523,
+        "mean_feedback_forward_relative_error": 0.026625329530552813,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58080,
+      "train_examples": 45000,
+      "train_loss": 0.015940935181909137
+    },
+    {
+      "epoch": 166,
+      "eval_accuracy": 0.9146,
+      "eval_loss": 0.42847799072265624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791823983192444,
+        "mean_feedback_forward_cosine": 0.9996321483662254,
+        "mean_feedback_forward_relative_error": 0.026625118463447218,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58432,
+      "train_examples": 45000,
+      "train_loss": 0.01682934757007493
+    },
+    {
+      "epoch": 167,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.42574083251953126,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791818395256996,
+        "mean_feedback_forward_cosine": 0.9996321169953597,
+        "mean_feedback_forward_relative_error": 0.026624903916136214,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58784,
+      "train_examples": 45000,
+      "train_loss": 0.016822115646633835
+    },
+    {
+      "epoch": 168,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.4300515563964844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918128073215485,
+        "mean_feedback_forward_cosine": 0.9996321326807925,
+        "mean_feedback_forward_relative_error": 0.026624686182721666,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59136,
+      "train_examples": 45000,
+      "train_loss": 0.016840671036309665
+    },
+    {
+      "epoch": 169,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.42753214111328125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037918079644441605,
+        "mean_feedback_forward_cosine": 0.999632157777485,
+        "mean_feedback_forward_relative_error": 0.026624476488091443,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59488,
+      "train_examples": 45000,
+      "train_loss": 0.01653002626688944
+    },
+    {
+      "epoch": 170,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.4329725341796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791803494095802,
+        "mean_feedback_forward_cosine": 0.9996321452291388,
+        "mean_feedback_forward_relative_error": 0.026624266499359357,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59840,
+      "train_examples": 45000,
+      "train_loss": 0.01631236160033279
+    },
+    {
+      "epoch": 171,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.43096278686523437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791797533631325,
+        "mean_feedback_forward_cosine": 0.9996321483662254,
+        "mean_feedback_forward_relative_error": 0.026624056167508427,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60192,
+      "train_examples": 45000,
+      "train_loss": 0.017680182982815637
+    },
+    {
+      "epoch": 172,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.4310341735839844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791791573166847,
+        "mean_feedback_forward_cosine": 0.9996321640516582,
+        "mean_feedback_forward_relative_error": 0.0266238445121991,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60544,
+      "train_examples": 45000,
+      "train_loss": 0.01705649436513583
+    },
+    {
+      "epoch": 173,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.4284998229980469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037917863577604294,
+        "mean_feedback_forward_cosine": 0.999632157777485,
+        "mean_feedback_forward_relative_error": 0.02662362472007149,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60896,
+      "train_examples": 45000,
+      "train_loss": 0.016325897006193798
+    },
+    {
+      "epoch": 174,
+      "eval_accuracy": 0.9156,
+      "eval_loss": 0.42368416137695314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791780769824982,
+        "mean_feedback_forward_cosine": 0.9996321671887448,
+        "mean_feedback_forward_relative_error": 0.026623414927407316,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61248,
+      "train_examples": 45000,
+      "train_loss": 0.016306531962255635
+    },
+    {
+      "epoch": 175,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.42613728637695314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791775926947594,
+        "mean_feedback_forward_cosine": 0.9996321546403986,
+        "mean_feedback_forward_relative_error": 0.026623194743143886,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61600,
+      "train_examples": 45000,
+      "train_loss": 0.016608378148078917
+    },
+    {
+      "epoch": 176,
+      "eval_accuracy": 0.915,
+      "eval_loss": 0.42469633178710936,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791771084070206,
+        "mean_feedback_forward_cosine": 0.9996321483662254,
+        "mean_feedback_forward_relative_error": 0.026622989509058625,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61952,
+      "train_examples": 45000,
+      "train_loss": 0.01599796301147176
+    },
+    {
+      "epoch": 177,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.42996865234375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791766241192818,
+        "mean_feedback_forward_cosine": 0.9996321766000045,
+        "mean_feedback_forward_relative_error": 0.026622781431988665,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62304,
+      "train_examples": 45000,
+      "train_loss": 0.0145107354823086
+    },
+    {
+      "epoch": 178,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.42992142333984373,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0379176065325737,
+        "mean_feedback_forward_cosine": 0.9996321452291388,
+        "mean_feedback_forward_relative_error": 0.026622566737626727,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62656,
+      "train_examples": 45000,
+      "train_loss": 0.016593861082527372
+    },
+    {
+      "epoch": 179,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.429712548828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791755437850952,
+        "mean_feedback_forward_cosine": 0.9996321671887448,
+        "mean_feedback_forward_relative_error": 0.026622350278653596,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63008,
+      "train_examples": 45000,
+      "train_loss": 0.015633646452095775
+    },
+    {
+      "epoch": 180,
+      "eval_accuracy": 0.9142,
+      "eval_loss": 0.4295874938964844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037917498499155045,
+        "mean_feedback_forward_cosine": 0.9996321671887448,
+        "mean_feedback_forward_relative_error": 0.026622139946802667,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63360,
+      "train_examples": 45000,
+      "train_loss": 0.01656670587816172
+    },
+    {
+      "epoch": 181,
+      "eval_accuracy": 0.9166,
+      "eval_loss": 0.4243903991699219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037917450070381165,
+        "mean_feedback_forward_cosine": 0.9996321797370911,
+        "mean_feedback_forward_relative_error": 0.02662193775177002,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63712,
+      "train_examples": 45000,
+      "train_loss": 0.01582161951859792
+    },
+    {
+      "epoch": 182,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.43365278930664064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791740536689758,
+        "mean_feedback_forward_cosine": 0.999632157777485,
+        "mean_feedback_forward_relative_error": 0.02662172746893607,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64064,
+      "train_examples": 45000,
+      "train_loss": 0.014813974743419223
+    },
+    {
+      "epoch": 183,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.42709647216796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791734576225281,
+        "mean_feedback_forward_cosine": 0.9996321891483507,
+        "mean_feedback_forward_relative_error": 0.026621517872339802,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64416,
+      "train_examples": 45000,
+      "train_loss": 0.01592662673095862
+    },
+    {
+      "epoch": 184,
+      "eval_accuracy": 0.9162,
+      "eval_loss": 0.4307603088378906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791729733347893,
+        "mean_feedback_forward_cosine": 0.9996321671887448,
+        "mean_feedback_forward_relative_error": 0.026621317441918348,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64768,
+      "train_examples": 45000,
+      "train_loss": 0.016338840604159566
+    },
+    {
+      "epoch": 185,
+      "eval_accuracy": 0.9146,
+      "eval_loss": 0.43272603149414063,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037917252629995346,
+        "mean_feedback_forward_cosine": 0.9996321860112642,
+        "mean_feedback_forward_relative_error": 0.02662111627624223,
+        "min_feedback_forward_cosine": 0.999284029006958
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65120,
+      "train_examples": 45000,
+      "train_loss": 0.015533369680245717
+    },
+    {
+      "epoch": 186,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.43155269775390626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791719302535057,
+        "mean_feedback_forward_cosine": 0.9996321640516582,
+        "mean_feedback_forward_relative_error": 0.02662091192446257,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65472,
+      "train_examples": 45000,
+      "train_loss": 0.015310422006249427
+    },
+    {
+      "epoch": 187,
+      "eval_accuracy": 0.9148,
+      "eval_loss": 0.431560595703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791714087128639,
+        "mean_feedback_forward_cosine": 0.9996321954225239,
+        "mean_feedback_forward_relative_error": 0.02662071384685604,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65824,
+      "train_examples": 45000,
+      "train_loss": 0.015171491112642817
+    },
+    {
+      "epoch": 188,
+      "eval_accuracy": 0.9154,
+      "eval_loss": 0.4299284484863281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791709616780281,
+        "mean_feedback_forward_cosine": 0.9996321703258314,
+        "mean_feedback_forward_relative_error": 0.026620496799679177,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66176,
+      "train_examples": 45000,
+      "train_loss": 0.015132849889993668
+    },
+    {
+      "epoch": 189,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.43136605224609376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037917036563158035,
+        "mean_feedback_forward_cosine": 0.9996321922854373,
+        "mean_feedback_forward_relative_error": 0.02662028823243944,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66528,
+      "train_examples": 45000,
+      "train_loss": 0.01564434177271194
+    },
+    {
+      "epoch": 190,
+      "eval_accuracy": 0.9138,
+      "eval_loss": 0.42769647216796874,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916988134384155,
+        "mean_feedback_forward_cosine": 0.9996322142450433,
+        "mean_feedback_forward_relative_error": 0.02662007956716575,
+        "min_feedback_forward_cosine": 0.9992838501930237
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66880,
+      "train_examples": 45000,
+      "train_loss": 0.015331490024593141
+    },
+    {
+      "epoch": 191,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.43386570434570315,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791693598031998,
+        "mean_feedback_forward_cosine": 0.9996321891483507,
+        "mean_feedback_forward_relative_error": 0.026619868353009224,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67232,
+      "train_examples": 45000,
+      "train_loss": 0.015315960738725133
+    },
+    {
+      "epoch": 192,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.43133213500976564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916891276836395,
+        "mean_feedback_forward_cosine": 0.9996322111079567,
+        "mean_feedback_forward_relative_error": 0.026619663118923966,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67584,
+      "train_examples": 45000,
+      "train_loss": 0.015171922739346822
+    },
+    {
+      "epoch": 193,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.431591796875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916846573352814,
+        "mean_feedback_forward_cosine": 0.9996322205192164,
+        "mean_feedback_forward_relative_error": 0.02661946067880643,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67936,
+      "train_examples": 45000,
+      "train_loss": 0.014588061490986083
+    },
+    {
+      "epoch": 194,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.4299183837890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916794419288635,
+        "mean_feedback_forward_cosine": 0.9996322079708702,
+        "mean_feedback_forward_relative_error": 0.0266192503469555,
+        "min_feedback_forward_cosine": 0.9992837905883789
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68288,
+      "train_examples": 45000,
+      "train_loss": 0.015228129545847575
+    },
+    {
+      "epoch": 195,
+      "eval_accuracy": 0.916,
+      "eval_loss": 0.4288531494140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791675344109535,
+        "mean_feedback_forward_cosine": 0.9996321985596105,
+        "mean_feedback_forward_relative_error": 0.026619042858089272,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68640,
+      "train_examples": 45000,
+      "train_loss": 0.014976719170146518
+    },
+    {
+      "epoch": 196,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.4313197021484375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916697561740875,
+        "mean_feedback_forward_cosine": 0.9996321891483507,
+        "mean_feedback_forward_relative_error": 0.026618830761627146,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68992,
+      "train_examples": 45000,
+      "train_loss": 0.01579895223768221
+    },
+    {
+      "epoch": 197,
+      "eval_accuracy": 0.9152,
+      "eval_loss": 0.4294109436035156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916649132966995,
+        "mean_feedback_forward_cosine": 0.9996321891483507,
+        "mean_feedback_forward_relative_error": 0.026618630282188718,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69344,
+      "train_examples": 45000,
+      "train_loss": 0.014463566927115123
+    },
+    {
+      "epoch": 198,
+      "eval_accuracy": 0.9168,
+      "eval_loss": 0.42698069458007815,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916600704193115,
+        "mean_feedback_forward_cosine": 0.9996321985596105,
+        "mean_feedback_forward_relative_error": 0.02661842259725458,
+        "min_feedback_forward_cosine": 0.9992839694023132
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69696,
+      "train_examples": 45000,
+      "train_loss": 0.015557571692268053
+    },
+    {
+      "epoch": 199,
+      "eval_accuracy": 0.916,
+      "eval_loss": 0.43350626220703126,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.037916556000709534,
+        "mean_feedback_forward_cosine": 0.9996322205192164,
+        "mean_feedback_forward_relative_error": 0.0266182226570029,
+        "min_feedback_forward_cosine": 0.999284029006958
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70048,
+      "train_examples": 45000,
+      "train_loss": 0.014629167970021565
+    },
+    {
+      "epoch": 200,
+      "eval_accuracy": 0.9144,
+      "eval_loss": 0.4333634704589844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.03791650012135506,
+        "mean_feedback_forward_cosine": 0.9996322111079567,
+        "mean_feedback_forward_relative_error": 0.026618016491595068,
+        "min_feedback_forward_cosine": 0.9992839097976685
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70400,
+      "train_examples": 45000,
+      "train_loss": 0.016335480781065094
+    }
+  ],
+  "evaluation_protocol": {
+    "test_evaluations": 0,
+    "test_used_for_selection": false,
+    "validation_evaluations": 201
+  },
+  "final": {
+    "accuracy": 0.9144,
+    "epoch": 200,
+    "evaluation_split": "validation",
+    "finite": true,
+    "loss": 0.4333634704589844,
+    "step": 70400
+  },
+  "hardware": {
+    "cuda_device_name": "NVIDIA RTX A6000",
+    "cuda_visible_devices": "0",
+    "device": "cuda",
+    "device_total_memory_bytes": 50896568320,
+    "peak_memory_allocated_bytes": 2010747392,
+    "peak_memory_reserved_bytes": 2552233984,
+    "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": 21.03406047821045,
+    "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": 1780.5374817848206,
+    "train_wall_s": 1758.6518173217773
+  },
+  "work": {
+    "apical_macs_per_example": 40108672,
+    "causal_scalar_observations": 0,
+    "components": {
+      "apical_projection_macs": 360978048000000,
+      "apical_regression_macs": 0,
+      "bp_reverse_macs_estimate": 0,
+      "kp_reciprocal_correlation_macs": 360978048000000,
+      "local_weight_correlation_macs": 364959360000000,
+      "mirror_feedback_prediction_macs": 0,
+      "mirror_local_correlation_macs": 0,
+      "mirror_response_macs": 0,
+      "ordinary_forward_macs": 364959360000000,
+      "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": 40551040,
+    "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": 1451874816000000
+  }
+}
-- 
cgit v1.2.3