experiment: add completed A6000 crossover resultsa6000-x1-results

1 files changed, 4265 insertions, 0 deletions

diff --git a/results/resnet_crossover/r2/resnet-r2-clean_kp-d56-lr0p1.json b/results/resnet_crossover/r2/resnet-r2-clean_kp-d56-lr0p1.json
new file mode 100644
index 0000000..c51dbc4
--- /dev/null
+++ b/results/resnet_crossover/r2/resnet-r2-clean_kp-d56-lr0p1.json
@@ -0,0 +1,4265 @@
+{
+  "architecture": {
+    "adaptive_apical_parameters": 0,
+    "adaptive_feedback_parameters": 848512,
+    "base_width": 16,
+    "blocks_per_stage": 9,
+    "bn_eps": 1e-05,
+    "bn_momentum": 0.1,
+    "depth": 56,
+    "family": "CIFAR 6n+2 ResNet, option-A shortcuts",
+    "fixed_feedback_parameters": 0,
+    "fixed_traffic_coefficients": 0,
+    "forward_parameters": 853018,
+    "hidden_shapes": [
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        16,
+        32,
+        32
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        32,
+        16,
+        16
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ],
+      [
+        64,
+        8,
+        8
+      ]
+    ],
+    "normalization": "batchnorm",
+    "predictor_parameters": 0,
+    "residual_scale": 1.0,
+    "vectorizer_mode": null,
+    "vectorizer_parameters": 0
+  },
+  "args": {
+    "a_scale": 1.0,
+    "a_warmup_steps": 0,
+    "alignment_probe": 32,
+    "apical_calibration_mode": "unit_targets",
+    "apical_seed": null,
+    "augment_train": 1,
+    "batch_size": 128,
+    "bn_eps": 1e-05,
+    "bn_momentum": 0.1,
+    "data_dir": "/orion/u/oscarwan/scratch/sdil-data",
+    "depth": 56,
+    "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-d56-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": 0.9988037016656663,
+    "feedback_forward_cosine": [
+      0.9995574951171875,
+      0.9995169639587402,
+      0.9992508888244629,
+      0.9991180896759033,
+      0.9991408586502075,
+      0.9988906979560852,
+      0.999122142791748,
+      0.9988346099853516,
+      0.9992246627807617,
+      0.9988415837287903,
+      0.9990280866622925,
+      0.9985275864601135,
+      0.9986708164215088,
+      0.9977948069572449,
+      0.9980864524841309,
+      0.9977025985717773,
+      0.9983896017074585,
+      0.9977715611457825,
+      0.9988182783126831,
+      0.9991542100906372,
+      0.9989104866981506,
+      0.9987741708755493,
+      0.9988317489624023,
+      0.9986496567726135,
+      0.9990257620811462,
+      0.9987555146217346,
+      0.9989592432975769,
+      0.9986085891723633,
+      0.9990679621696472,
+      0.998760461807251,
+      0.9990558624267578,
+      0.9986815452575684,
+      0.9991776943206787,
+      0.9988480806350708,
+      0.9991104006767273,
+      0.9987267255783081,
+      0.9992961883544922,
+      0.9995409250259399,
+      0.9993419647216797,
+      0.9993253946304321,
+      0.9993946552276611,
+      0.9992978572845459,
+      0.9994099736213684,
+      0.9992327690124512,
+      0.9994537830352783,
+      0.9991894960403442,
+      0.9993447065353394,
+      0.9989761114120483,
+      0.9992188215255737,
+      0.9986401796340942,
+      0.9989041686058044,
+      0.997963547706604,
+      0.9979575872421265,
+      0.9966350197792053,
+      0.9999642968177795
+    ],
+    "feedback_forward_norm_ratio": [
+      0.9988628625869751,
+      1.00169038772583,
+      0.9988455772399902,
+      1.0018149614334106,
+      0.9983507394790649,
+      1.001532793045044,
+      0.9981478452682495,
+      1.00508713722229,
+      0.9976799488067627,
+      1.002779483795166,
+      0.996472954750061,
+      1.0031712055206299,
+      0.9947890639305115,
+      1.0027819871902466,
+      0.9938843250274658,
+      1.0071712732315063,
+      0.9960893988609314,
+      1.0045490264892578,
+      0.9968646764755249,
+      1.0018043518066406,
+      0.9983800649642944,
+      1.0021274089813232,
+      0.9981029033660889,
+      1.002813696861267,
+      0.998471736907959,
+      1.0024747848510742,
+      0.9984707236289978,
+      1.00377357006073,
+      0.9984996318817139,
+      1.003908634185791,
+      0.9981768727302551,
+      1.00295889377594,
+      0.9980287551879883,
+      1.003975510597229,
+      0.9989507794380188,
+      1.003403663635254,
+      0.9981476068496704,
+      1.0010191202163696,
+      0.9987508058547974,
+      1.0020904541015625,
+      0.9993308186531067,
+      1.0017856359481812,
+      0.9992122650146484,
+      1.0019315481185913,
+      0.9992152452468872,
+      1.0023112297058105,
+      0.9992201924324036,
+      1.0025296211242676,
+      0.9988380670547485,
+      1.002525806427002,
+      0.9991728663444519,
+      1.003301739692688,
+      0.9986709356307983,
+      1.003277063369751,
+      1.0064365863800049
+    ],
+    "feedback_forward_relative_error": [
+      0.029755709692835808,
+      0.03115258552134037,
+      0.03870101645588875,
+      0.042074669152498245,
+      0.04145186021924019,
+      0.04716512933373451,
+      0.041904985904693604,
+      0.04866664856672287,
+      0.03940116614103317,
+      0.04828055575489998,
+      0.04415307566523552,
+      0.054445065557956696,
+      0.05168891325592995,
+      0.06656099855899811,
+      0.061974987387657166,
+      0.0684041976928711,
+      0.05677487328648567,
+      0.0670657828450203,
+      0.04864046350121498,
+      0.041204363107681274,
+      0.04667006433010101,
+      0.04961398243904114,
+      0.04832734540104866,
+      0.05211685970425606,
+      0.04413510486483574,
+      0.05001169815659523,
+      0.04561459273099899,
+      0.05298462510108948,
+      0.043167274445295334,
+      0.050040122121572495,
+      0.04345225915312767,
+      0.05151088535785675,
+      0.04056119546294212,
+      0.048257168382406235,
+      0.04217178001999855,
+      0.05066346749663353,
+      0.03753187507390976,
+      0.030329721048474312,
+      0.03627943620085716,
+      0.03683214634656906,
+      0.034789882600307465,
+      0.03754897788167,
+      0.03434815630316734,
+      0.03925550356507301,
+      0.03304633870720863,
+      0.040373899042606354,
+      0.03619636967778206,
+      0.045381058007478714,
+      0.039520811289548874,
+      0.05227590724825859,
+      0.046804919838905334,
+      0.06400913000106812,
+      0.06388391554355621,
+      0.08223697543144226,
+      0.01064779981970787
+    ],
+    "innovation_negative_gradient_cosine": [
+      0.9990317821502686,
+      0.9989892244338989,
+      0.9989813566207886,
+      0.9987704753875732,
+      0.9989952445030212,
+      0.9988008737564087,
+      0.9990298748016357,
+      0.9987648129463196,
+      0.99904465675354,
+      0.9989824891090393,
+      0.9990545511245728,
+      0.9986403584480286,
+      0.9990695714950562,
+      0.9982811212539673,
+      0.9990877509117126,
+      0.9978867769241333,
+      0.9990891814231873,
+      0.9979665279388428,
+      0.9990986585617065,
+      0.998831033706665,
+      0.9991048574447632,
+      0.9985469579696655,
+      0.9991083145141602,
+      0.9985823035240173,
+      0.9991117715835571,
+      0.9986124634742737,
+      0.9991379380226135,
+      0.9986610412597656,
+      0.9991350173950195,
+      0.9989628195762634,
+      0.9991565942764282,
+      0.9987471699714661,
+      0.9991878867149353,
+      0.9989113807678223,
+      0.9992119073867798,
+      0.9989118576049805,
+      0.9992275238037109,
+      0.9992373585700989,
+      0.9992811679840088,
+      0.999153733253479,
+      0.9993259906768799,
+      0.9991576671600342,
+      0.9993447661399841,
+      0.9992607235908508,
+      0.9994027614593506,
+      0.9993295073509216,
+      0.9994453191757202,
+      0.9994139671325684,
+      0.9994925260543823,
+      0.9995317459106445,
+      0.9995478391647339,
+      0.9996591210365295,
+      0.9996875524520874,
+      0.9998064041137695,
+      0.7163748741149902
+    ],
+    "mean_feedback_forward_cosine": 0.9988813335245306,
+    "mean_feedback_forward_relative_error": 0.04600105993449688,
+    "normalization_state": "training_batch_stats_without_running_update",
+    "raw_negative_gradient_cosine": [
+      0.9990317821502686,
+      0.9989892244338989,
+      0.9989813566207886,
+      0.9987704753875732,
+      0.9989952445030212,
+      0.9988008737564087,
+      0.9990298748016357,
+      0.9987648129463196,
+      0.99904465675354,
+      0.9989824891090393,
+      0.9990545511245728,
+      0.9986403584480286,
+      0.9990695714950562,
+      0.9982811212539673,
+      0.9990877509117126,
+      0.9978867769241333,
+      0.9990891814231873,
+      0.9979665279388428,
+      0.9990986585617065,
+      0.998831033706665,
+      0.9991048574447632,
+      0.9985469579696655,
+      0.9991083145141602,
+      0.9985823035240173,
+      0.9991117715835571,
+      0.9986124634742737,
+      0.9991379380226135,
+      0.9986610412597656,
+      0.9991350173950195,
+      0.9989628195762634,
+      0.9991565942764282,
+      0.9987471699714661,
+      0.9991878867149353,
+      0.9989113807678223,
+      0.9992119073867798,
+      0.9989118576049805,
+      0.9992275238037109,
+      0.9992373585700989,
+      0.9992811679840088,
+      0.999153733253479,
+      0.9993259906768799,
+      0.9991576671600342,
+      0.9993447661399841,
+      0.9992607235908508,
+      0.9994027614593506,
+      0.9993295073509216,
+      0.9994453191757202,
+      0.9994139671325684,
+      0.9994925260543823,
+      0.9995317459106445,
+      0.9995478391647339,
+      0.9996591210365295,
+      0.9996875524520874,
+      0.9998064041137695,
+      0.7163748741149902
+    ],
+    "teaching_negative_gradient_cosine": [
+      0.9990317821502686,
+      0.9989892244338989,
+      0.9989813566207886,
+      0.9987704753875732,
+      0.9989952445030212,
+      0.9988008737564087,
+      0.9990298748016357,
+      0.9987648129463196,
+      0.99904465675354,
+      0.9989824891090393,
+      0.9990545511245728,
+      0.9986403584480286,
+      0.9990695714950562,
+      0.9982811212539673,
+      0.9990877509117126,
+      0.9978867769241333,
+      0.9990891814231873,
+      0.9979665279388428,
+      0.9990986585617065,
+      0.998831033706665,
+      0.9991048574447632,
+      0.9985469579696655,
+      0.9991083145141602,
+      0.9985823035240173,
+      0.9991117715835571,
+      0.9986124634742737,
+      0.9991379380226135,
+      0.9986610412597656,
+      0.9991350173950195,
+      0.9989628195762634,
+      0.9991565942764282,
+      0.9987471699714661,
+      0.9991878867149353,
+      0.9989113807678223,
+      0.9992119073867798,
+      0.9989118576049805,
+      0.9992275238037109,
+      0.9992373585700989,
+      0.9992811679840088,
+      0.999153733253479,
+      0.9993259906768799,
+      0.9991576671600342,
+      0.9993447661399841,
+      0.9992607235908508,
+      0.9994027614593506,
+      0.9993295073509216,
+      0.9994453191757202,
+      0.9994139671325684,
+      0.9994925260543823,
+      0.9995317459106445,
+      0.9995478391647339,
+      0.9996591210365295,
+      0.9996875524520874,
+      0.9998064041137695,
+      0.7163748741149902
+    ],
+    "wall_s": 0.11626768112182617
+  },
+  "epochs": [
+    {
+      "epoch": 1,
+      "eval_accuracy": 0.2936,
+      "eval_loss": 2.060192431640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.4729894399642944,
+        "mean_feedback_forward_cosine": 0.08207923307676207,
+        "mean_feedback_forward_relative_error": 1.3605215722864323,
+        "min_feedback_forward_cosine": 0.011852272786200047
+      },
+      "lr": 0.02,
+      "output_lr": 0.02,
+      "step": 352,
+      "train_examples": 45000,
+      "train_loss": 2.123790234099494
+    },
+    {
+      "epoch": 2,
+      "eval_accuracy": 0.4336,
+      "eval_loss": 1.644146044921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3818473815917969,
+        "mean_feedback_forward_cosine": 0.1517215799878944,
+        "mean_feedback_forward_relative_error": 1.306954277645458,
+        "min_feedback_forward_cosine": 0.04962993040680885
+      },
+      "lr": 0.04,
+      "output_lr": 0.04,
+      "step": 704,
+      "train_examples": 45000,
+      "train_loss": 1.6988532908757528
+    },
+    {
+      "epoch": 3,
+      "eval_accuracy": 0.5284,
+      "eval_loss": 1.41752919921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3391482830047607,
+        "mean_feedback_forward_cosine": 0.2346487126567147,
+        "mean_feedback_forward_relative_error": 1.2409382657571273,
+        "min_feedback_forward_cosine": 0.11064116656780243
+      },
+      "lr": 0.06000000000000001,
+      "output_lr": 0.06000000000000001,
+      "step": 1056,
+      "train_examples": 45000,
+      "train_loss": 1.40682961162991
+    },
+    {
+      "epoch": 4,
+      "eval_accuracy": 0.5884,
+      "eval_loss": 1.1971573974609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.3097155094146729,
+        "mean_feedback_forward_cosine": 0.3149321618405255,
+        "mean_feedback_forward_relative_error": 1.1737277052619242,
+        "min_feedback_forward_cosine": 0.14912492036819458
+      },
+      "lr": 0.08,
+      "output_lr": 0.08,
+      "step": 1408,
+      "train_examples": 45000,
+      "train_loss": 1.1264890681160822
+    },
+    {
+      "epoch": 5,
+      "eval_accuracy": 0.7082,
+      "eval_loss": 0.8434026123046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.2819832563400269,
+        "mean_feedback_forward_cosine": 0.387133172696287,
+        "mean_feedback_forward_relative_error": 1.109520400654186,
+        "min_feedback_forward_cosine": 0.1867155283689499
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 1760,
+      "train_examples": 45000,
+      "train_loss": 0.9187799799389309
+    },
+    {
+      "epoch": 6,
+      "eval_accuracy": 0.728,
+      "eval_loss": 0.8178329467773438,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.260962724685669,
+        "mean_feedback_forward_cosine": 0.43850364685058596,
+        "mean_feedback_forward_relative_error": 1.0616818113760516,
+        "min_feedback_forward_cosine": 0.2145915925502777
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2112,
+      "train_examples": 45000,
+      "train_loss": 0.7764964687347412
+    },
+    {
+      "epoch": 7,
+      "eval_accuracy": 0.7022,
+      "eval_loss": 0.9358474975585938,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.2412327527999878,
+        "mean_feedback_forward_cosine": 0.4796803523193706,
+        "mean_feedback_forward_relative_error": 1.0219877210530368,
+        "min_feedback_forward_cosine": 0.24086108803749084
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2464,
+      "train_examples": 45000,
+      "train_loss": 0.6912625069724189
+    },
+    {
+      "epoch": 8,
+      "eval_accuracy": 0.7508,
+      "eval_loss": 0.762169921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.222935676574707,
+        "mean_feedback_forward_cosine": 0.5153755480592901,
+        "mean_feedback_forward_relative_error": 0.986245432767001,
+        "min_feedback_forward_cosine": 0.26393070816993713
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 2816,
+      "train_examples": 45000,
+      "train_loss": 0.6286246035257975
+    },
+    {
+      "epoch": 9,
+      "eval_accuracy": 0.7888,
+      "eval_loss": 0.6249593627929687,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.2051806449890137,
+        "mean_feedback_forward_cosine": 0.5469873997298154,
+        "mean_feedback_forward_relative_error": 0.9534298192371021,
+        "min_feedback_forward_cosine": 0.2858530282974243
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3168,
+      "train_examples": 45000,
+      "train_loss": 0.5727805726899041
+    },
+    {
+      "epoch": 10,
+      "eval_accuracy": 0.768,
+      "eval_loss": 0.7355963623046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1889394521713257,
+        "mean_feedback_forward_cosine": 0.5755576019937342,
+        "mean_feedback_forward_relative_error": 0.9226999954743819,
+        "min_feedback_forward_cosine": 0.30642619729042053
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3520,
+      "train_examples": 45000,
+      "train_loss": 0.5328205897172292
+    },
+    {
+      "epoch": 11,
+      "eval_accuracy": 0.8112,
+      "eval_loss": 0.5712257080078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1723517179489136,
+        "mean_feedback_forward_cosine": 0.6018412416631526,
+        "mean_feedback_forward_relative_error": 0.8934933689507571,
+        "min_feedback_forward_cosine": 0.3266719579696655
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 3872,
+      "train_examples": 45000,
+      "train_loss": 0.5016598609076606
+    },
+    {
+      "epoch": 12,
+      "eval_accuracy": 0.7896,
+      "eval_loss": 0.6734730712890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1551438570022583,
+        "mean_feedback_forward_cosine": 0.625412240895358,
+        "mean_feedback_forward_relative_error": 0.8661270428787579,
+        "min_feedback_forward_cosine": 0.34698301553726196
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4224,
+      "train_examples": 45000,
+      "train_loss": 0.47733594166437787
+    },
+    {
+      "epoch": 13,
+      "eval_accuracy": 0.799,
+      "eval_loss": 0.6092771240234375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.138596773147583,
+        "mean_feedback_forward_cosine": 0.6471591234207154,
+        "mean_feedback_forward_relative_error": 0.8402352311394431,
+        "min_feedback_forward_cosine": 0.36682766675949097
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4576,
+      "train_examples": 45000,
+      "train_loss": 0.4463163204775916
+    },
+    {
+      "epoch": 14,
+      "eval_accuracy": 0.8226,
+      "eval_loss": 0.5384412719726562,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.122653603553772,
+        "mean_feedback_forward_cosine": 0.6674662720073353,
+        "mean_feedback_forward_relative_error": 0.8154430156404322,
+        "min_feedback_forward_cosine": 0.3852914571762085
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 4928,
+      "train_examples": 45000,
+      "train_loss": 0.4283656439781189
+    },
+    {
+      "epoch": 15,
+      "eval_accuracy": 0.829,
+      "eval_loss": 0.5133809020996094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.1056475639343262,
+        "mean_feedback_forward_cosine": 0.6869502766565843,
+        "mean_feedback_forward_relative_error": 0.7906913573091681,
+        "min_feedback_forward_cosine": 0.4046149253845215
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5280,
+      "train_examples": 45000,
+      "train_loss": 0.4076533457332187
+    },
+    {
+      "epoch": 16,
+      "eval_accuracy": 0.8106,
+      "eval_loss": 0.6083410766601562,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0893713235855103,
+        "mean_feedback_forward_cosine": 0.7045990922234275,
+        "mean_feedback_forward_relative_error": 0.7676530827175487,
+        "min_feedback_forward_cosine": 0.42278337478637695
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5632,
+      "train_examples": 45000,
+      "train_loss": 0.3880259382724762
+    },
+    {
+      "epoch": 17,
+      "eval_accuracy": 0.8292,
+      "eval_loss": 0.503566552734375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0723820924758911,
+        "mean_feedback_forward_cosine": 0.7214057423851706,
+        "mean_feedback_forward_relative_error": 0.7450004333799536,
+        "min_feedback_forward_cosine": 0.44093620777130127
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 5984,
+      "train_examples": 45000,
+      "train_loss": 0.3736611439810859
+    },
+    {
+      "epoch": 18,
+      "eval_accuracy": 0.8352,
+      "eval_loss": 0.5023697204589844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0543376207351685,
+        "mean_feedback_forward_cosine": 0.737669068033045,
+        "mean_feedback_forward_relative_error": 0.7226391394029964,
+        "min_feedback_forward_cosine": 0.4603174924850464
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6336,
+      "train_examples": 45000,
+      "train_loss": 0.364715851465861
+    },
+    {
+      "epoch": 19,
+      "eval_accuracy": 0.8164,
+      "eval_loss": 0.603892041015625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.035783290863037,
+        "mean_feedback_forward_cosine": 0.7531207366423174,
+        "mean_feedback_forward_relative_error": 0.70054920993068,
+        "min_feedback_forward_cosine": 0.4795728921890259
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 6688,
+      "train_examples": 45000,
+      "train_loss": 0.3488689161459605
+    },
+    {
+      "epoch": 20,
+      "eval_accuracy": 0.8404,
+      "eval_loss": 0.5018754943847656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 1.0169116258621216,
+        "mean_feedback_forward_cosine": 0.7675304158167405,
+        "mean_feedback_forward_relative_error": 0.6793056907978925,
+        "min_feedback_forward_cosine": 0.4986239969730377
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7040,
+      "train_examples": 45000,
+      "train_loss": 0.33735975733862983
+    },
+    {
+      "epoch": 21,
+      "eval_accuracy": 0.8446,
+      "eval_loss": 0.4680168029785156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9975670576095581,
+        "mean_feedback_forward_cosine": 0.7813035347244957,
+        "mean_feedback_forward_relative_error": 0.658438462560827,
+        "min_feedback_forward_cosine": 0.5182281732559204
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7392,
+      "train_examples": 45000,
+      "train_loss": 0.3261894694434272
+    },
+    {
+      "epoch": 22,
+      "eval_accuracy": 0.8524,
+      "eval_loss": 0.4533489501953125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9778302311897278,
+        "mean_feedback_forward_cosine": 0.7944718913598494,
+        "mean_feedback_forward_relative_error": 0.6378612987019799,
+        "min_feedback_forward_cosine": 0.5373923778533936
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 7744,
+      "train_examples": 45000,
+      "train_loss": 0.3175746521313985
+    },
+    {
+      "epoch": 23,
+      "eval_accuracy": 0.8608,
+      "eval_loss": 0.4396246276855469,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9584059715270996,
+        "mean_feedback_forward_cosine": 0.8066396474838257,
+        "mean_feedback_forward_relative_error": 0.6182341255924918,
+        "min_feedback_forward_cosine": 0.5559208393096924
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8096,
+      "train_examples": 45000,
+      "train_loss": 0.30371766968833075
+    },
+    {
+      "epoch": 24,
+      "eval_accuracy": 0.8596,
+      "eval_loss": 0.4563602478027344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9386764168739319,
+        "mean_feedback_forward_cosine": 0.8177748517556624,
+        "mean_feedback_forward_relative_error": 0.5997207224369049,
+        "min_feedback_forward_cosine": 0.5744795799255371
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8448,
+      "train_examples": 45000,
+      "train_loss": 0.2900086401462555
+    },
+    {
+      "epoch": 25,
+      "eval_accuracy": 0.8586,
+      "eval_loss": 0.4243275390625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.9187594652175903,
+        "mean_feedback_forward_cosine": 0.8285161896185441,
+        "mean_feedback_forward_relative_error": 0.5813728538426486,
+        "min_feedback_forward_cosine": 0.5922850370407104
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 8800,
+      "train_examples": 45000,
+      "train_loss": 0.2870960140122308
+    },
+    {
+      "epoch": 26,
+      "eval_accuracy": 0.866,
+      "eval_loss": 0.41880641479492187,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8995776772499084,
+        "mean_feedback_forward_cosine": 0.8383215871724216,
+        "mean_feedback_forward_relative_error": 0.5641180030324242,
+        "min_feedback_forward_cosine": 0.6091452240943909
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9152,
+      "train_examples": 45000,
+      "train_loss": 0.2801684870455
+    },
+    {
+      "epoch": 27,
+      "eval_accuracy": 0.8608,
+      "eval_loss": 0.45782495727539063,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8790687918663025,
+        "mean_feedback_forward_cosine": 0.8478951313278892,
+        "mean_feedback_forward_relative_error": 0.5467334628105164,
+        "min_feedback_forward_cosine": 0.6270616054534912
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9504,
+      "train_examples": 45000,
+      "train_loss": 0.276506783774164
+    },
+    {
+      "epoch": 28,
+      "eval_accuracy": 0.8738,
+      "eval_loss": 0.41229105224609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8599552512168884,
+        "mean_feedback_forward_cosine": 0.8569087667898698,
+        "mean_feedback_forward_relative_error": 0.5298725553534248,
+        "min_feedback_forward_cosine": 0.6432315111160278
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 9856,
+      "train_examples": 45000,
+      "train_loss": 0.25949633451567755
+    },
+    {
+      "epoch": 29,
+      "eval_accuracy": 0.8548,
+      "eval_loss": 0.47992633666992185,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8409947752952576,
+        "mean_feedback_forward_cosine": 0.8654873620380055,
+        "mean_feedback_forward_relative_error": 0.5133874578909441,
+        "min_feedback_forward_cosine": 0.658990740776062
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10208,
+      "train_examples": 45000,
+      "train_loss": 0.2583834177679486
+    },
+    {
+      "epoch": 30,
+      "eval_accuracy": 0.8498,
+      "eval_loss": 0.49892302856445314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8211984038352966,
+        "mean_feedback_forward_cosine": 0.8737081343477423,
+        "mean_feedback_forward_relative_error": 0.4970983491702513,
+        "min_feedback_forward_cosine": 0.6749225854873657
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10560,
+      "train_examples": 45000,
+      "train_loss": 0.25416444894737666
+    },
+    {
+      "epoch": 31,
+      "eval_accuracy": 0.8446,
+      "eval_loss": 0.5110629577636718,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.8005052804946899,
+        "mean_feedback_forward_cosine": 0.8813880790363658,
+        "mean_feedback_forward_relative_error": 0.4814133836464448,
+        "min_feedback_forward_cosine": 0.6910892724990845
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 10912,
+      "train_examples": 45000,
+      "train_loss": 0.2507732811000612
+    },
+    {
+      "epoch": 32,
+      "eval_accuracy": 0.8732,
+      "eval_loss": 0.40844691162109376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7807812094688416,
+        "mean_feedback_forward_cosine": 0.8887567043304443,
+        "mean_feedback_forward_relative_error": 0.4659483096816323,
+        "min_feedback_forward_cosine": 0.7062845230102539
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11264,
+      "train_examples": 45000,
+      "train_loss": 0.24385885693762038
+    },
+    {
+      "epoch": 33,
+      "eval_accuracy": 0.8492,
+      "eval_loss": 0.49456326904296877,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7614753842353821,
+        "mean_feedback_forward_cosine": 0.8955482537096197,
+        "mean_feedback_forward_relative_error": 0.4511532629078085,
+        "min_feedback_forward_cosine": 0.720395565032959
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11616,
+      "train_examples": 45000,
+      "train_loss": 0.23739461857477823
+    },
+    {
+      "epoch": 34,
+      "eval_accuracy": 0.8498,
+      "eval_loss": 0.5223807678222656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.742235541343689,
+        "mean_feedback_forward_cosine": 0.9018027283928611,
+        "mean_feedback_forward_relative_error": 0.43716591732068494,
+        "min_feedback_forward_cosine": 0.7343444228172302
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 11968,
+      "train_examples": 45000,
+      "train_loss": 0.23336360233359868
+    },
+    {
+      "epoch": 35,
+      "eval_accuracy": 0.8744,
+      "eval_loss": 0.41874742431640627,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7225374579429626,
+        "mean_feedback_forward_cosine": 0.9077971696853637,
+        "mean_feedback_forward_relative_error": 0.4234206673773852,
+        "min_feedback_forward_cosine": 0.7480750679969788
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12320,
+      "train_examples": 45000,
+      "train_loss": 0.22778456059561836
+    },
+    {
+      "epoch": 36,
+      "eval_accuracy": 0.8776,
+      "eval_loss": 0.3977263732910156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.7031345367431641,
+        "mean_feedback_forward_cosine": 0.9135233012112705,
+        "mean_feedback_forward_relative_error": 0.4098250017924742,
+        "min_feedback_forward_cosine": 0.7613703012466431
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 12672,
+      "train_examples": 45000,
+      "train_loss": 0.2255130520661672
+    },
+    {
+      "epoch": 37,
+      "eval_accuracy": 0.867,
+      "eval_loss": 0.4421707946777344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6846021413803101,
+        "mean_feedback_forward_cosine": 0.9187919194048101,
+        "mean_feedback_forward_relative_error": 0.39689874730326913,
+        "min_feedback_forward_cosine": 0.7737734913825989
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13024,
+      "train_examples": 45000,
+      "train_loss": 0.2240035187959671
+    },
+    {
+      "epoch": 38,
+      "eval_accuracy": 0.851,
+      "eval_loss": 0.4847253173828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6660941243171692,
+        "mean_feedback_forward_cosine": 0.9238915584304116,
+        "mean_feedback_forward_relative_error": 0.3839693303812634,
+        "min_feedback_forward_cosine": 0.7856288552284241
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13376,
+      "train_examples": 45000,
+      "train_loss": 0.21737773218684725
+    },
+    {
+      "epoch": 39,
+      "eval_accuracy": 0.8788,
+      "eval_loss": 0.41233887939453123,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6483564972877502,
+        "mean_feedback_forward_cosine": 0.9285368594256315,
+        "mean_feedback_forward_relative_error": 0.3718270332975821,
+        "min_feedback_forward_cosine": 0.7968446016311646
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 13728,
+      "train_examples": 45000,
+      "train_loss": 0.2110462031231986
+    },
+    {
+      "epoch": 40,
+      "eval_accuracy": 0.8744,
+      "eval_loss": 0.42741173706054686,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6304129362106323,
+        "mean_feedback_forward_cosine": 0.9328895178708163,
+        "mean_feedback_forward_relative_error": 0.3601139377463948,
+        "min_feedback_forward_cosine": 0.8078715801239014
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14080,
+      "train_examples": 45000,
+      "train_loss": 0.20277180547714233
+    },
+    {
+      "epoch": 41,
+      "eval_accuracy": 0.881,
+      "eval_loss": 0.40434840698242186,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.6114498376846313,
+        "mean_feedback_forward_cosine": 0.9371648885987022,
+        "mean_feedback_forward_relative_error": 0.34830014421181243,
+        "min_feedback_forward_cosine": 0.8190717101097107
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14432,
+      "train_examples": 45000,
+      "train_loss": 0.20585658853848776
+    },
+    {
+      "epoch": 42,
+      "eval_accuracy": 0.878,
+      "eval_loss": 0.4233556945800781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5938428640365601,
+        "mean_feedback_forward_cosine": 0.9411213528026234,
+        "mean_feedback_forward_relative_error": 0.33697369437326086,
+        "min_feedback_forward_cosine": 0.8294274210929871
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 14784,
+      "train_examples": 45000,
+      "train_loss": 0.2012337659464942
+    },
+    {
+      "epoch": 43,
+      "eval_accuracy": 0.8768,
+      "eval_loss": 0.41400225830078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5771043300628662,
+        "mean_feedback_forward_cosine": 0.944767753644423,
+        "mean_feedback_forward_relative_error": 0.32624045339497654,
+        "min_feedback_forward_cosine": 0.8388245105743408
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15136,
+      "train_examples": 45000,
+      "train_loss": 0.19511909804344177
+    },
+    {
+      "epoch": 44,
+      "eval_accuracy": 0.8104,
+      "eval_loss": 0.7237696166992188,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5596101880073547,
+        "mean_feedback_forward_cosine": 0.948338532447815,
+        "mean_feedback_forward_relative_error": 0.31540666994723404,
+        "min_feedback_forward_cosine": 0.8483365774154663
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15488,
+      "train_examples": 45000,
+      "train_loss": 0.1980742193804847
+    },
+    {
+      "epoch": 45,
+      "eval_accuracy": 0.8706,
+      "eval_loss": 0.46084964599609374,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5431535243988037,
+        "mean_feedback_forward_cosine": 0.9515498302199624,
+        "mean_feedback_forward_relative_error": 0.3053073207085783,
+        "min_feedback_forward_cosine": 0.8570070266723633
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 15840,
+      "train_examples": 45000,
+      "train_loss": 0.19275552099015977
+    },
+    {
+      "epoch": 46,
+      "eval_accuracy": 0.8814,
+      "eval_loss": 0.40947244873046873,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5276136994361877,
+        "mean_feedback_forward_cosine": 0.9546002951535312,
+        "mean_feedback_forward_relative_error": 0.2954358386722478,
+        "min_feedback_forward_cosine": 0.8650082349777222
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16192,
+      "train_examples": 45000,
+      "train_loss": 0.18924892360104456
+    },
+    {
+      "epoch": 47,
+      "eval_accuracy": 0.8824,
+      "eval_loss": 0.41092626342773436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.5122539401054382,
+        "mean_feedback_forward_cosine": 0.9575513818047263,
+        "mean_feedback_forward_relative_error": 0.2855187581344084,
+        "min_feedback_forward_cosine": 0.8727079629898071
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16544,
+      "train_examples": 45000,
+      "train_loss": 0.18873042728106182
+    },
+    {
+      "epoch": 48,
+      "eval_accuracy": 0.873,
+      "eval_loss": 0.4387502258300781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.49630874395370483,
+        "mean_feedback_forward_cosine": 0.9602571855891835,
+        "mean_feedback_forward_relative_error": 0.2761423014781692,
+        "min_feedback_forward_cosine": 0.8804479241371155
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 16896,
+      "train_examples": 45000,
+      "train_loss": 0.18430980693764157
+    },
+    {
+      "epoch": 49,
+      "eval_accuracy": 0.8618,
+      "eval_loss": 0.4662376098632813,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.48133349418640137,
+        "mean_feedback_forward_cosine": 0.9627048893408342,
+        "mean_feedback_forward_relative_error": 0.26738952113823455,
+        "min_feedback_forward_cosine": 0.8874017000198364
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17248,
+      "train_examples": 45000,
+      "train_loss": 0.1800568481153912
+    },
+    {
+      "epoch": 50,
+      "eval_accuracy": 0.8826,
+      "eval_loss": 0.4051923889160156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.466120183467865,
+        "mean_feedback_forward_cosine": 0.9650715307755904,
+        "mean_feedback_forward_relative_error": 0.2586525908925317,
+        "min_feedback_forward_cosine": 0.8943054676055908
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17600,
+      "train_examples": 45000,
+      "train_loss": 0.1789180386662483
+    },
+    {
+      "epoch": 51,
+      "eval_accuracy": 0.8702,
+      "eval_loss": 0.4498435302734375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4508410096168518,
+        "mean_feedback_forward_cosine": 0.9673665068366311,
+        "mean_feedback_forward_relative_error": 0.24993357875130393,
+        "min_feedback_forward_cosine": 0.9010016918182373
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 17952,
+      "train_examples": 45000,
+      "train_loss": 0.18010127037101323
+    },
+    {
+      "epoch": 52,
+      "eval_accuracy": 0.87,
+      "eval_loss": 0.477175048828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4368517994880676,
+        "mean_feedback_forward_cosine": 0.9694177584214644,
+        "mean_feedback_forward_relative_error": 0.24190838066014378,
+        "min_feedback_forward_cosine": 0.9070504307746887
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18304,
+      "train_examples": 45000,
+      "train_loss": 0.16981975546148087
+    },
+    {
+      "epoch": 53,
+      "eval_accuracy": 0.8716,
+      "eval_loss": 0.4523770324707031,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.4235183298587799,
+        "mean_feedback_forward_cosine": 0.9713853868571195,
+        "mean_feedback_forward_relative_error": 0.23388134376569228,
+        "min_feedback_forward_cosine": 0.9125978946685791
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 18656,
+      "train_examples": 45000,
+      "train_loss": 0.17002208354208204
+    },
+    {
+      "epoch": 54,
+      "eval_accuracy": 0.8596,
+      "eval_loss": 0.5152298583984375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.40936705470085144,
+        "mean_feedback_forward_cosine": 0.9733334248716181,
+        "mean_feedback_forward_relative_error": 0.2257189244031906,
+        "min_feedback_forward_cosine": 0.9182888865470886
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19008,
+      "train_examples": 45000,
+      "train_loss": 0.17592860356701745
+    },
+    {
+      "epoch": 55,
+      "eval_accuracy": 0.8734,
+      "eval_loss": 0.4423846618652344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3969172537326813,
+        "mean_feedback_forward_cosine": 0.9750462586229498,
+        "mean_feedback_forward_relative_error": 0.21827290789647535,
+        "min_feedback_forward_cosine": 0.9230918884277344
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19360,
+      "train_examples": 45000,
+      "train_loss": 0.17371374536090428
+    },
+    {
+      "epoch": 56,
+      "eval_accuracy": 0.8784,
+      "eval_loss": 0.4110608825683594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.38447803258895874,
+        "mean_feedback_forward_cosine": 0.976641688563607,
+        "mean_feedback_forward_relative_error": 0.2111172491176562,
+        "min_feedback_forward_cosine": 0.9277975559234619
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 19712,
+      "train_examples": 45000,
+      "train_loss": 0.16767991562949286
+    },
+    {
+      "epoch": 57,
+      "eval_accuracy": 0.8784,
+      "eval_loss": 0.4192649353027344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3719860017299652,
+        "mean_feedback_forward_cosine": 0.97813284072009,
+        "mean_feedback_forward_relative_error": 0.2041995076293295,
+        "min_feedback_forward_cosine": 0.9323537349700928
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20064,
+      "train_examples": 45000,
+      "train_loss": 0.16788513456980386
+    },
+    {
+      "epoch": 58,
+      "eval_accuracy": 0.886,
+      "eval_loss": 0.37670152587890626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3604535162448883,
+        "mean_feedback_forward_cosine": 0.9794968269088051,
+        "mean_feedback_forward_relative_error": 0.1976644179360433,
+        "min_feedback_forward_cosine": 0.9364805221557617
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20416,
+      "train_examples": 45000,
+      "train_loss": 0.16070309614075554
+    },
+    {
+      "epoch": 59,
+      "eval_accuracy": 0.8636,
+      "eval_loss": 0.4858129577636719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.3487560451030731,
+        "mean_feedback_forward_cosine": 0.9808387929742987,
+        "mean_feedback_forward_relative_error": 0.1910235660997304,
+        "min_feedback_forward_cosine": 0.9404898881912231
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 20768,
+      "train_examples": 45000,
+      "train_loss": 0.1631975841826863
+    },
+    {
+      "epoch": 60,
+      "eval_accuracy": 0.8766,
+      "eval_loss": 0.4174827453613281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.33727237582206726,
+        "mean_feedback_forward_cosine": 0.982091445272619,
+        "mean_feedback_forward_relative_error": 0.18464546623555098,
+        "min_feedback_forward_cosine": 0.9443246722221375
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21120,
+      "train_examples": 45000,
+      "train_loss": 0.1633899968147278
+    },
+    {
+      "epoch": 61,
+      "eval_accuracy": 0.873,
+      "eval_loss": 0.452418115234375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.32584843039512634,
+        "mean_feedback_forward_cosine": 0.9832355997779153,
+        "mean_feedback_forward_relative_error": 0.17862626761198044,
+        "min_feedback_forward_cosine": 0.9479833245277405
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21472,
+      "train_examples": 45000,
+      "train_loss": 0.159811972117424
+    },
+    {
+      "epoch": 62,
+      "eval_accuracy": 0.8828,
+      "eval_loss": 0.4084390075683594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.31567618250846863,
+        "mean_feedback_forward_cosine": 0.9843033476309343,
+        "mean_feedback_forward_relative_error": 0.17278472164815123,
+        "min_feedback_forward_cosine": 0.9511356353759766
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 21824,
+      "train_examples": 45000,
+      "train_loss": 0.15469107734097376
+    },
+    {
+      "epoch": 63,
+      "eval_accuracy": 0.878,
+      "eval_loss": 0.44356014404296873,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.30494773387908936,
+        "mean_feedback_forward_cosine": 0.9853211641311646,
+        "mean_feedback_forward_relative_error": 0.16705804108218714,
+        "min_feedback_forward_cosine": 0.9543746709823608
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22176,
+      "train_examples": 45000,
+      "train_loss": 0.1573526954491933
+    },
+    {
+      "epoch": 64,
+      "eval_accuracy": 0.8818,
+      "eval_loss": 0.4164936279296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.29509037733078003,
+        "mean_feedback_forward_cosine": 0.9862557031891562,
+        "mean_feedback_forward_relative_error": 0.16163663694804364,
+        "min_feedback_forward_cosine": 0.9572689533233643
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22528,
+      "train_examples": 45000,
+      "train_loss": 0.15185034220880933
+    },
+    {
+      "epoch": 65,
+      "eval_accuracy": 0.8884,
+      "eval_loss": 0.429320068359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.28480038046836853,
+        "mean_feedback_forward_cosine": 0.9871627710082315,
+        "mean_feedback_forward_relative_error": 0.15618729252706876,
+        "min_feedback_forward_cosine": 0.9601784944534302
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 22880,
+      "train_examples": 45000,
+      "train_loss": 0.15663046832614475
+    },
+    {
+      "epoch": 66,
+      "eval_accuracy": 0.8818,
+      "eval_loss": 0.3846602600097656,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.27500370144844055,
+        "mean_feedback_forward_cosine": 0.9880119280381636,
+        "mean_feedback_forward_relative_error": 0.15090371072292327,
+        "min_feedback_forward_cosine": 0.9628419280052185
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23232,
+      "train_examples": 45000,
+      "train_loss": 0.152567342710495
+    },
+    {
+      "epoch": 67,
+      "eval_accuracy": 0.8742,
+      "eval_loss": 0.4532151306152344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2653149366378784,
+        "mean_feedback_forward_cosine": 0.9888372052799571,
+        "mean_feedback_forward_relative_error": 0.14560815332965418,
+        "min_feedback_forward_cosine": 0.9653981924057007
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23584,
+      "train_examples": 45000,
+      "train_loss": 0.15156706712577078
+    },
+    {
+      "epoch": 68,
+      "eval_accuracy": 0.8854,
+      "eval_loss": 0.413903759765625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2560727298259735,
+        "mean_feedback_forward_cosine": 0.9895687645131891,
+        "mean_feedback_forward_relative_error": 0.1407341560179537,
+        "min_feedback_forward_cosine": 0.9677327871322632
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 23936,
+      "train_examples": 45000,
+      "train_loss": 0.15645641708903843
+    },
+    {
+      "epoch": 69,
+      "eval_accuracy": 0.8872,
+      "eval_loss": 0.40043238525390623,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.24709388613700867,
+        "mean_feedback_forward_cosine": 0.9902782396836715,
+        "mean_feedback_forward_relative_error": 0.13584166609428144,
+        "min_feedback_forward_cosine": 0.9699423909187317
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24288,
+      "train_examples": 45000,
+      "train_loss": 0.15490477414660983
+    },
+    {
+      "epoch": 70,
+      "eval_accuracy": 0.885,
+      "eval_loss": 0.397725732421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.23907920718193054,
+        "mean_feedback_forward_cosine": 0.9909099470485341,
+        "mean_feedback_forward_relative_error": 0.13134042986414649,
+        "min_feedback_forward_cosine": 0.9718361496925354
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24640,
+      "train_examples": 45000,
+      "train_loss": 0.1508726365831163
+    },
+    {
+      "epoch": 71,
+      "eval_accuracy": 0.8558,
+      "eval_loss": 0.5760824829101563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.2312377542257309,
+        "mean_feedback_forward_cosine": 0.9915095892819491,
+        "mean_feedback_forward_relative_error": 0.12692420821298253,
+        "min_feedback_forward_cosine": 0.9736427068710327
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 24992,
+      "train_examples": 45000,
+      "train_loss": 0.14491609041425918
+    },
+    {
+      "epoch": 72,
+      "eval_accuracy": 0.8696,
+      "eval_loss": 0.458574853515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.22312340140342712,
+        "mean_feedback_forward_cosine": 0.9920836979692632,
+        "mean_feedback_forward_relative_error": 0.12254367403008722,
+        "min_feedback_forward_cosine": 0.9754534363746643
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25344,
+      "train_examples": 45000,
+      "train_loss": 0.1479260725710127
+    },
+    {
+      "epoch": 73,
+      "eval_accuracy": 0.8892,
+      "eval_loss": 0.40083897705078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.21591386198997498,
+        "mean_feedback_forward_cosine": 0.9926146171309731,
+        "mean_feedback_forward_relative_error": 0.11832766709002582,
+        "min_feedback_forward_cosine": 0.976998507976532
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 25696,
+      "train_examples": 45000,
+      "train_loss": 0.14577782358063593
+    },
+    {
+      "epoch": 74,
+      "eval_accuracy": 0.8836,
+      "eval_loss": 0.4026210205078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.20882253348827362,
+        "mean_feedback_forward_cosine": 0.9930870424617421,
+        "mean_feedback_forward_relative_error": 0.11447676802900704,
+        "min_feedback_forward_cosine": 0.9784666299819946
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26048,
+      "train_examples": 45000,
+      "train_loss": 0.14426382974121305
+    },
+    {
+      "epoch": 75,
+      "eval_accuracy": 0.883,
+      "eval_loss": 0.42882922973632814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.20159201323986053,
+        "mean_feedback_forward_cosine": 0.9935504837469621,
+        "mean_feedback_forward_relative_error": 0.11056213744662025,
+        "min_feedback_forward_cosine": 0.9799209237098694
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26400,
+      "train_examples": 45000,
+      "train_loss": 0.14628605412244797
+    },
+    {
+      "epoch": 76,
+      "eval_accuracy": 0.878,
+      "eval_loss": 0.41888806762695313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1942160427570343,
+        "mean_feedback_forward_cosine": 0.9940022219311108,
+        "mean_feedback_forward_relative_error": 0.10662056024778974,
+        "min_feedback_forward_cosine": 0.9813616275787354
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 26752,
+      "train_examples": 45000,
+      "train_loss": 0.15107803481419882
+    },
+    {
+      "epoch": 77,
+      "eval_accuracy": 0.892,
+      "eval_loss": 0.3812136474609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.18739287555217743,
+        "mean_feedback_forward_cosine": 0.9944140856916254,
+        "mean_feedback_forward_relative_error": 0.10289184376597404,
+        "min_feedback_forward_cosine": 0.982636034488678
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27104,
+      "train_examples": 45000,
+      "train_loss": 0.145390626361635
+    },
+    {
+      "epoch": 78,
+      "eval_accuracy": 0.8826,
+      "eval_loss": 0.4390092895507812,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1810525357723236,
+        "mean_feedback_forward_cosine": 0.9947815277359703,
+        "mean_feedback_forward_relative_error": 0.0994456759230657,
+        "min_feedback_forward_cosine": 0.9837870597839355
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27456,
+      "train_examples": 45000,
+      "train_loss": 0.14267062601778244
+    },
+    {
+      "epoch": 79,
+      "eval_accuracy": 0.8756,
+      "eval_loss": 0.41855242919921876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1744450032711029,
+        "mean_feedback_forward_cosine": 0.9951367963444103,
+        "mean_feedback_forward_relative_error": 0.09600104195150462,
+        "min_feedback_forward_cosine": 0.9849445819854736
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 27808,
+      "train_examples": 45000,
+      "train_loss": 0.14812842531204223
+    },
+    {
+      "epoch": 80,
+      "eval_accuracy": 0.8682,
+      "eval_loss": 0.4982807067871094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.16858580708503723,
+        "mean_feedback_forward_cosine": 0.9954568223519759,
+        "mean_feedback_forward_relative_error": 0.0927825734696605,
+        "min_feedback_forward_cosine": 0.985930323600769
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28160,
+      "train_examples": 45000,
+      "train_loss": 0.14163471700880262
+    },
+    {
+      "epoch": 81,
+      "eval_accuracy": 0.8788,
+      "eval_loss": 0.4318050842285156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1628764420747757,
+        "mean_feedback_forward_cosine": 0.9957529859109359,
+        "mean_feedback_forward_relative_error": 0.08970520404929465,
+        "min_feedback_forward_cosine": 0.9868632555007935
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28512,
+      "train_examples": 45000,
+      "train_loss": 0.14164452673594158
+    },
+    {
+      "epoch": 82,
+      "eval_accuracy": 0.8824,
+      "eval_loss": 0.4356259216308594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.15716975927352905,
+        "mean_feedback_forward_cosine": 0.9960444970564408,
+        "mean_feedback_forward_relative_error": 0.08655340471728282,
+        "min_feedback_forward_cosine": 0.9877631068229675
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 28864,
+      "train_examples": 45000,
+      "train_loss": 0.1429812172730764
+    },
+    {
+      "epoch": 83,
+      "eval_accuracy": 0.8834,
+      "eval_loss": 0.4211099670410156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1516208052635193,
+        "mean_feedback_forward_cosine": 0.996317475492304,
+        "mean_feedback_forward_relative_error": 0.08351170965893702,
+        "min_feedback_forward_cosine": 0.9886099696159363
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29216,
+      "train_examples": 45000,
+      "train_loss": 0.14329515885247124
+    },
+    {
+      "epoch": 84,
+      "eval_accuracy": 0.9018,
+      "eval_loss": 0.35779424438476565,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14630314707756042,
+        "mean_feedback_forward_cosine": 0.9965671257539229,
+        "mean_feedback_forward_relative_error": 0.08063399812037295,
+        "min_feedback_forward_cosine": 0.9893922805786133
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29568,
+      "train_examples": 45000,
+      "train_loss": 0.14221931150092018
+    },
+    {
+      "epoch": 85,
+      "eval_accuracy": 0.8702,
+      "eval_loss": 0.4682876647949219,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.14148655533790588,
+        "mean_feedback_forward_cosine": 0.9967920650135387,
+        "mean_feedback_forward_relative_error": 0.0779483347792517,
+        "min_feedback_forward_cosine": 0.9900739192962646
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 29920,
+      "train_examples": 45000,
+      "train_loss": 0.137186794895596
+    },
+    {
+      "epoch": 86,
+      "eval_accuracy": 0.8932,
+      "eval_loss": 0.3757369873046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1365615576505661,
+        "mean_feedback_forward_cosine": 0.9970066233114763,
+        "mean_feedback_forward_relative_error": 0.07529233558611437,
+        "min_feedback_forward_cosine": 0.9907506704330444
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30272,
+      "train_examples": 45000,
+      "train_loss": 0.1364652813633283
+    },
+    {
+      "epoch": 87,
+      "eval_accuracy": 0.8844,
+      "eval_loss": 0.4202875732421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.13171401619911194,
+        "mean_feedback_forward_cosine": 0.9972147313031283,
+        "mean_feedback_forward_relative_error": 0.07261722778732127,
+        "min_feedback_forward_cosine": 0.9913944602012634
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30624,
+      "train_examples": 45000,
+      "train_loss": 0.14023481232060325
+    },
+    {
+      "epoch": 88,
+      "eval_accuracy": 0.8998,
+      "eval_loss": 0.3637839233398438,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1271563023328781,
+        "mean_feedback_forward_cosine": 0.997396829995242,
+        "mean_feedback_forward_relative_error": 0.07020163962786848,
+        "min_feedback_forward_cosine": 0.9919781684875488
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 30976,
+      "train_examples": 45000,
+      "train_loss": 0.13764463415410783
+    },
+    {
+      "epoch": 89,
+      "eval_accuracy": 0.8942,
+      "eval_loss": 0.37466764526367186,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.12278088927268982,
+        "mean_feedback_forward_cosine": 0.9975727265531367,
+        "mean_feedback_forward_relative_error": 0.06778270504014058,
+        "min_feedback_forward_cosine": 0.9925185441970825
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31328,
+      "train_examples": 45000,
+      "train_loss": 0.13379682774146398
+    },
+    {
+      "epoch": 90,
+      "eval_accuracy": 0.8878,
+      "eval_loss": 0.41289153442382814,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11860892176628113,
+        "mean_feedback_forward_cosine": 0.9977324615825306,
+        "mean_feedback_forward_relative_error": 0.06551586030559106,
+        "min_feedback_forward_cosine": 0.9930169582366943
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 31680,
+      "train_examples": 45000,
+      "train_loss": 0.13367654723591274
+    },
+    {
+      "epoch": 91,
+      "eval_accuracy": 0.8796,
+      "eval_loss": 0.4534392883300781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11397954821586609,
+        "mean_feedback_forward_cosine": 0.9978942480954257,
+        "mean_feedback_forward_relative_error": 0.06314405470409176,
+        "min_feedback_forward_cosine": 0.9935492277145386
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32032,
+      "train_examples": 45000,
+      "train_loss": 0.1437890195131302
+    },
+    {
+      "epoch": 92,
+      "eval_accuracy": 0.877,
+      "eval_loss": 0.4637198852539062,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.11020797491073608,
+        "mean_feedback_forward_cosine": 0.9980307849970731,
+        "mean_feedback_forward_relative_error": 0.061057206615805625,
+        "min_feedback_forward_cosine": 0.9939678311347961
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32384,
+      "train_examples": 45000,
+      "train_loss": 0.1297376984834671
+    },
+    {
+      "epoch": 93,
+      "eval_accuracy": 0.8974,
+      "eval_loss": 0.38976353759765625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.10636366158723831,
+        "mean_feedback_forward_cosine": 0.9981618870388378,
+        "mean_feedback_forward_relative_error": 0.05898658186197281,
+        "min_feedback_forward_cosine": 0.9943795204162598
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 32736,
+      "train_examples": 45000,
+      "train_loss": 0.13253259558942582
+    },
+    {
+      "epoch": 94,
+      "eval_accuracy": 0.881,
+      "eval_loss": 0.45050140991210935,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.1024697870016098,
+        "mean_feedback_forward_cosine": 0.9982858592813665,
+        "mean_feedback_forward_relative_error": 0.05696909615600651,
+        "min_feedback_forward_cosine": 0.9947818517684937
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33088,
+      "train_examples": 45000,
+      "train_loss": 0.13139092605246439
+    },
+    {
+      "epoch": 95,
+      "eval_accuracy": 0.884,
+      "eval_loss": 0.42125306396484374,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09880731254816055,
+        "mean_feedback_forward_cosine": 0.9984017632224343,
+        "mean_feedback_forward_relative_error": 0.05500872118229216,
+        "min_feedback_forward_cosine": 0.9951469302177429
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33440,
+      "train_examples": 45000,
+      "train_loss": 0.13053015309174856
+    },
+    {
+      "epoch": 96,
+      "eval_accuracy": 0.8994,
+      "eval_loss": 0.37246392211914064,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09534408897161484,
+        "mean_feedback_forward_cosine": 0.9985089713876898,
+        "mean_feedback_forward_relative_error": 0.053137221051888034,
+        "min_feedback_forward_cosine": 0.9954792261123657
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 33792,
+      "train_examples": 45000,
+      "train_loss": 0.13170791505177815
+    },
+    {
+      "epoch": 97,
+      "eval_accuracy": 0.8916,
+      "eval_loss": 0.37829818115234376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.09184065461158752,
+        "mean_feedback_forward_cosine": 0.9986103144558993,
+        "mean_feedback_forward_relative_error": 0.051301561533050107,
+        "min_feedback_forward_cosine": 0.9958046674728394
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34144,
+      "train_examples": 45000,
+      "train_loss": 0.1322530888001124
+    },
+    {
+      "epoch": 98,
+      "eval_accuracy": 0.8822,
+      "eval_loss": 0.44099720458984376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08854187279939651,
+        "mean_feedback_forward_cosine": 0.9987025206739252,
+        "mean_feedback_forward_relative_error": 0.04957084210420197,
+        "min_feedback_forward_cosine": 0.9960996508598328
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34496,
+      "train_examples": 45000,
+      "train_loss": 0.1289476733631558
+    },
+    {
+      "epoch": 99,
+      "eval_accuracy": 0.898,
+      "eval_loss": 0.3525207153320313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08554445952177048,
+        "mean_feedback_forward_cosine": 0.9987886905670166,
+        "mean_feedback_forward_relative_error": 0.047894603725184094,
+        "min_feedback_forward_cosine": 0.9963589906692505
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 34848,
+      "train_examples": 45000,
+      "train_loss": 0.1272515431880951
+    },
+    {
+      "epoch": 100,
+      "eval_accuracy": 0.8954,
+      "eval_loss": 0.400292919921875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08254595100879669,
+        "mean_feedback_forward_cosine": 0.9988718823953109,
+        "mean_feedback_forward_relative_error": 0.04622269711372527,
+        "min_feedback_forward_cosine": 0.9966089129447937
+      },
+      "lr": 0.1,
+      "output_lr": 0.1,
+      "step": 35200,
+      "train_examples": 45000,
+      "train_loss": 0.1293600611527761
+    },
+    {
+      "epoch": 101,
+      "eval_accuracy": 0.9198,
+      "eval_loss": 0.28196166381835935,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08252088725566864,
+        "mean_feedback_forward_cosine": 0.9988725673068654,
+        "mean_feedback_forward_relative_error": 0.04620732959698547,
+        "min_feedback_forward_cosine": 0.9966108798980713
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35552,
+      "train_examples": 45000,
+      "train_loss": 0.06943065629336569
+    },
+    {
+      "epoch": 102,
+      "eval_accuracy": 0.9222,
+      "eval_loss": 0.28367977294921876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08251013606786728,
+        "mean_feedback_forward_cosine": 0.9988728794184598,
+        "mean_feedback_forward_relative_error": 0.04619971408762715,
+        "min_feedback_forward_cosine": 0.9966118931770325
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 35904,
+      "train_examples": 45000,
+      "train_loss": 0.0442628238108423
+    },
+    {
+      "epoch": 103,
+      "eval_accuracy": 0.9236,
+      "eval_loss": 0.27802401123046877,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08250072598457336,
+        "mean_feedback_forward_cosine": 0.9988731460137801,
+        "mean_feedback_forward_relative_error": 0.046192759241570125,
+        "min_feedback_forward_cosine": 0.9966127872467041
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36256,
+      "train_examples": 45000,
+      "train_loss": 0.03693747533584634
+    },
+    {
+      "epoch": 104,
+      "eval_accuracy": 0.925,
+      "eval_loss": 0.2826873077392578,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0824924111366272,
+        "mean_feedback_forward_cosine": 0.9988733931021256,
+        "mean_feedback_forward_relative_error": 0.04618666246533394,
+        "min_feedback_forward_cosine": 0.9966134428977966
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36608,
+      "train_examples": 45000,
+      "train_loss": 0.030649015731364488
+    },
+    {
+      "epoch": 105,
+      "eval_accuracy": 0.9262,
+      "eval_loss": 0.2913733154296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.082485131919384,
+        "mean_feedback_forward_cosine": 0.9988736239346591,
+        "mean_feedback_forward_relative_error": 0.04618093506856398,
+        "min_feedback_forward_cosine": 0.9966140389442444
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 36960,
+      "train_examples": 45000,
+      "train_loss": 0.02735501567953163
+    },
+    {
+      "epoch": 106,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.2939286346435547,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08247771114110947,
+        "mean_feedback_forward_cosine": 0.9988738623532382,
+        "mean_feedback_forward_relative_error": 0.04617522387680682,
+        "min_feedback_forward_cosine": 0.9966146349906921
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37312,
+      "train_examples": 45000,
+      "train_loss": 0.024195579751663737
+    },
+    {
+      "epoch": 107,
+      "eval_accuracy": 0.9272,
+      "eval_loss": 0.29316375732421873,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08247034251689911,
+        "mean_feedback_forward_cosine": 0.9988740758462386,
+        "mean_feedback_forward_relative_error": 0.046169685081324796,
+        "min_feedback_forward_cosine": 0.9966151714324951
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 37664,
+      "train_examples": 45000,
+      "train_loss": 0.022561528134014873
+    },
+    {
+      "epoch": 108,
+      "eval_accuracy": 0.9244,
+      "eval_loss": 0.30675077514648436,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08246370404958725,
+        "mean_feedback_forward_cosine": 0.9988743045113303,
+        "mean_feedback_forward_relative_error": 0.04616451940753243,
+        "min_feedback_forward_cosine": 0.9966158270835876
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38016,
+      "train_examples": 45000,
+      "train_loss": 0.02048250459664398
+    },
+    {
+      "epoch": 109,
+      "eval_accuracy": 0.9252,
+      "eval_loss": 0.3081547302246094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08245643228292465,
+        "mean_feedback_forward_cosine": 0.9988745136694475,
+        "mean_feedback_forward_relative_error": 0.04615932416848161,
+        "min_feedback_forward_cosine": 0.9966166019439697
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38368,
+      "train_examples": 45000,
+      "train_loss": 0.019332495719856686
+    },
+    {
+      "epoch": 110,
+      "eval_accuracy": 0.9258,
+      "eval_loss": 0.3063327392578125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08244861662387848,
+        "mean_feedback_forward_cosine": 0.9988747369159352,
+        "mean_feedback_forward_relative_error": 0.04615418922833421,
+        "min_feedback_forward_cosine": 0.9966172575950623
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 38720,
+      "train_examples": 45000,
+      "train_loss": 0.018840931750006146
+    },
+    {
+      "epoch": 111,
+      "eval_accuracy": 0.926,
+      "eval_loss": 0.30996431884765624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08244211226701736,
+        "mean_feedback_forward_cosine": 0.9988749330694026,
+        "mean_feedback_forward_relative_error": 0.04614953860979189,
+        "min_feedback_forward_cosine": 0.9966177344322205
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39072,
+      "train_examples": 45000,
+      "train_loss": 0.016484283983376292
+    },
+    {
+      "epoch": 112,
+      "eval_accuracy": 0.9262,
+      "eval_loss": 0.31271492309570315,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0824364498257637,
+        "mean_feedback_forward_cosine": 0.9988750945438039,
+        "mean_feedback_forward_relative_error": 0.04614522804252126,
+        "min_feedback_forward_cosine": 0.9966182708740234
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39424,
+      "train_examples": 45000,
+      "train_loss": 0.014451664542489582
+    },
+    {
+      "epoch": 113,
+      "eval_accuracy": 0.926,
+      "eval_loss": 0.322545263671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08243008702993393,
+        "mean_feedback_forward_cosine": 0.9988752787763422,
+        "mean_feedback_forward_relative_error": 0.04614073788239197,
+        "min_feedback_forward_cosine": 0.9966187477111816
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 39776,
+      "train_examples": 45000,
+      "train_loss": 0.01427282026078966
+    },
+    {
+      "epoch": 114,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.32204202270507815,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08242438733577728,
+        "mean_feedback_forward_cosine": 0.9988754662600431,
+        "mean_feedback_forward_relative_error": 0.04613645566119389,
+        "min_feedback_forward_cosine": 0.9966192245483398
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40128,
+      "train_examples": 45000,
+      "train_loss": 0.01288464905437496
+    },
+    {
+      "epoch": 115,
+      "eval_accuracy": 0.9238,
+      "eval_loss": 0.32415628662109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08241870999336243,
+        "mean_feedback_forward_cosine": 0.9988756277344444,
+        "mean_feedback_forward_relative_error": 0.046132276952266696,
+        "min_feedback_forward_cosine": 0.996619701385498
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40480,
+      "train_examples": 45000,
+      "train_loss": 0.012184910022715727
+    },
+    {
+      "epoch": 116,
+      "eval_accuracy": 0.927,
+      "eval_loss": 0.32800870971679685,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0824129655957222,
+        "mean_feedback_forward_cosine": 0.9988758065483787,
+        "mean_feedback_forward_relative_error": 0.0461282127621499,
+        "min_feedback_forward_cosine": 0.9966202974319458
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 40832,
+      "train_examples": 45000,
+      "train_loss": 0.011643429762952857
+    },
+    {
+      "epoch": 117,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.3297094970703125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08240707963705063,
+        "mean_feedback_forward_cosine": 0.9988759918646379,
+        "mean_feedback_forward_relative_error": 0.04612408006055788,
+        "min_feedback_forward_cosine": 0.996620774269104
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41184,
+      "train_examples": 45000,
+      "train_loss": 0.011566649926371044
+    },
+    {
+      "epoch": 118,
+      "eval_accuracy": 0.9274,
+      "eval_loss": 0.33520316162109376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0824013203382492,
+        "mean_feedback_forward_cosine": 0.9988761804320595,
+        "mean_feedback_forward_relative_error": 0.04611968616531654,
+        "min_feedback_forward_cosine": 0.996621310710907
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41536,
+      "train_examples": 45000,
+      "train_loss": 0.01146308294420855
+    },
+    {
+      "epoch": 119,
+      "eval_accuracy": 0.9252,
+      "eval_loss": 0.3379811279296875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08239556103944778,
+        "mean_feedback_forward_cosine": 0.9988763310692527,
+        "mean_feedback_forward_relative_error": 0.04611563034016978,
+        "min_feedback_forward_cosine": 0.9966216087341309
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 41888,
+      "train_examples": 45000,
+      "train_loss": 0.010218534137308598
+    },
+    {
+      "epoch": 120,
+      "eval_accuracy": 0.9264,
+      "eval_loss": 0.33973890380859373,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08238988369703293,
+        "mean_feedback_forward_cosine": 0.9988765250552785,
+        "mean_feedback_forward_relative_error": 0.04611164289103313,
+        "min_feedback_forward_cosine": 0.9966222047805786
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42240,
+      "train_examples": 45000,
+      "train_loss": 0.00960727816954669
+    },
+    {
+      "epoch": 121,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.34343662719726564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08238489925861359,
+        "mean_feedback_forward_cosine": 0.998876690864563,
+        "mean_feedback_forward_relative_error": 0.046107920445501806,
+        "min_feedback_forward_cosine": 0.9966225624084473
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42592,
+      "train_examples": 45000,
+      "train_loss": 0.009085430883988738
+    },
+    {
+      "epoch": 122,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.34254222412109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08237940073013306,
+        "mean_feedback_forward_cosine": 0.9988768653436141,
+        "mean_feedback_forward_relative_error": 0.04610400257462805,
+        "min_feedback_forward_cosine": 0.996623158454895
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 42944,
+      "train_examples": 45000,
+      "train_loss": 0.008975493807718157
+    },
+    {
+      "epoch": 123,
+      "eval_accuracy": 0.9276,
+      "eval_loss": 0.34295497436523437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0823746845126152,
+        "mean_feedback_forward_cosine": 0.9988770127296448,
+        "mean_feedback_forward_relative_error": 0.04610043037682772,
+        "min_feedback_forward_cosine": 0.9966235756874084
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43296,
+      "train_examples": 45000,
+      "train_loss": 0.008150759787774749
+    },
+    {
+      "epoch": 124,
+      "eval_accuracy": 0.9246,
+      "eval_loss": 0.3488459045410156,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08236976712942123,
+        "mean_feedback_forward_cosine": 0.998877163366838,
+        "mean_feedback_forward_relative_error": 0.04609683183106509,
+        "min_feedback_forward_cosine": 0.9966238737106323
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 43648,
+      "train_examples": 45000,
+      "train_loss": 0.008191198907544215
+    },
+    {
+      "epoch": 125,
+      "eval_accuracy": 0.9276,
+      "eval_loss": 0.34850970458984376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08236398547887802,
+        "mean_feedback_forward_cosine": 0.9988773215900768,
+        "mean_feedback_forward_relative_error": 0.0460928878154267,
+        "min_feedback_forward_cosine": 0.9966244101524353
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44000,
+      "train_examples": 45000,
+      "train_loss": 0.008717926258511014
+    },
+    {
+      "epoch": 126,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.3463415405273437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08235813677310944,
+        "mean_feedback_forward_cosine": 0.9988774852319198,
+        "mean_feedback_forward_relative_error": 0.04608909797600724,
+        "min_feedback_forward_cosine": 0.9966248273849487
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44352,
+      "train_examples": 45000,
+      "train_loss": 0.008424199743734465
+    },
+    {
+      "epoch": 127,
+      "eval_accuracy": 0.9242,
+      "eval_loss": 0.3489424072265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08235276490449905,
+        "mean_feedback_forward_cosine": 0.9988776391202753,
+        "mean_feedback_forward_relative_error": 0.046085428514263845,
+        "min_feedback_forward_cosine": 0.9966253042221069
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 44704,
+      "train_examples": 45000,
+      "train_loss": 0.00735124247815046
+    },
+    {
+      "epoch": 128,
+      "eval_accuracy": 0.9268,
+      "eval_loss": 0.3469889953613281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08234743773937225,
+        "mean_feedback_forward_cosine": 0.9988778103481639,
+        "mean_feedback_forward_relative_error": 0.04608147801323371,
+        "min_feedback_forward_cosine": 0.9966257810592651
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45056,
+      "train_examples": 45000,
+      "train_loss": 0.0076482207431561415
+    },
+    {
+      "epoch": 129,
+      "eval_accuracy": 0.9268,
+      "eval_loss": 0.3487642822265625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08234293758869171,
+        "mean_feedback_forward_cosine": 0.9988779599016363,
+        "mean_feedback_forward_relative_error": 0.04607815569774671,
+        "min_feedback_forward_cosine": 0.9966261386871338
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45408,
+      "train_examples": 45000,
+      "train_loss": 0.006547223501776655
+    },
+    {
+      "epoch": 130,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.35033135986328123,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08233887702226639,
+        "mean_feedback_forward_cosine": 0.9988781018690629,
+        "mean_feedback_forward_relative_error": 0.04607508634640412,
+        "min_feedback_forward_cosine": 0.9966265559196472
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 45760,
+      "train_examples": 45000,
+      "train_loss": 0.006076416883534855
+    },
+    {
+      "epoch": 131,
+      "eval_accuracy": 0.9282,
+      "eval_loss": 0.35761566162109376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08233317732810974,
+        "mean_feedback_forward_cosine": 0.998878276348114,
+        "mean_feedback_forward_relative_error": 0.04607112404297699,
+        "min_feedback_forward_cosine": 0.9966269731521606
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46112,
+      "train_examples": 45000,
+      "train_loss": 0.007004095608658261
+    },
+    {
+      "epoch": 132,
+      "eval_accuracy": 0.9256,
+      "eval_loss": 0.35786424560546876,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08232785016298294,
+        "mean_feedback_forward_cosine": 0.9988784291527488,
+        "mean_feedback_forward_relative_error": 0.046067386340688576,
+        "min_feedback_forward_cosine": 0.9966274499893188
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46464,
+      "train_examples": 45000,
+      "train_loss": 0.00657963326341576
+    },
+    {
+      "epoch": 133,
+      "eval_accuracy": 0.927,
+      "eval_loss": 0.3533356994628906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08232413232326508,
+        "mean_feedback_forward_cosine": 0.9988785591992465,
+        "mean_feedback_forward_relative_error": 0.04606463924388994,
+        "min_feedback_forward_cosine": 0.996627688407898
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 46816,
+      "train_examples": 45000,
+      "train_loss": 0.005323932779717466
+    },
+    {
+      "epoch": 134,
+      "eval_accuracy": 0.9276,
+      "eval_loss": 0.35352208251953127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08231943100690842,
+        "mean_feedback_forward_cosine": 0.9988786946643483,
+        "mean_feedback_forward_relative_error": 0.04606130861423232,
+        "min_feedback_forward_cosine": 0.9966281056404114
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47168,
+      "train_examples": 45000,
+      "train_loss": 0.005774464671355155
+    },
+    {
+      "epoch": 135,
+      "eval_accuracy": 0.927,
+      "eval_loss": 0.3579586608886719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08231444656848907,
+        "mean_feedback_forward_cosine": 0.9988788355480541,
+        "mean_feedback_forward_relative_error": 0.046057646366005595,
+        "min_feedback_forward_cosine": 0.9966285228729248
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47520,
+      "train_examples": 45000,
+      "train_loss": 0.006122449112104045
+    },
+    {
+      "epoch": 136,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.3627355712890625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08231008797883987,
+        "mean_feedback_forward_cosine": 0.9988789666782726,
+        "mean_feedback_forward_relative_error": 0.046054657260802656,
+        "min_feedback_forward_cosine": 0.9966288805007935
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 47872,
+      "train_examples": 45000,
+      "train_loss": 0.005008096188720729
+    },
+    {
+      "epoch": 137,
+      "eval_accuracy": 0.9254,
+      "eval_loss": 0.36479898681640627,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08230576664209366,
+        "mean_feedback_forward_cosine": 0.9988791205666282,
+        "mean_feedback_forward_relative_error": 0.04605141476812688,
+        "min_feedback_forward_cosine": 0.9966292381286621
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48224,
+      "train_examples": 45000,
+      "train_loss": 0.00519936425000843
+    },
+    {
+      "epoch": 138,
+      "eval_accuracy": 0.9272,
+      "eval_loss": 0.36759619140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08230125904083252,
+        "mean_feedback_forward_cosine": 0.9988792657852172,
+        "mean_feedback_forward_relative_error": 0.04604820312762802,
+        "min_feedback_forward_cosine": 0.9966295957565308
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48576,
+      "train_examples": 45000,
+      "train_loss": 0.005097926148772239
+    },
+    {
+      "epoch": 139,
+      "eval_accuracy": 0.928,
+      "eval_loss": 0.367848046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08229579031467438,
+        "mean_feedback_forward_cosine": 0.9988794348456643,
+        "mean_feedback_forward_relative_error": 0.04604436362331564,
+        "min_feedback_forward_cosine": 0.9966301321983337
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 48928,
+      "train_examples": 45000,
+      "train_loss": 0.00578090578098264
+    },
+    {
+      "epoch": 140,
+      "eval_accuracy": 0.9258,
+      "eval_loss": 0.3670384765625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08228998631238937,
+        "mean_feedback_forward_cosine": 0.9988796158270402,
+        "mean_feedback_forward_relative_error": 0.04604042116552591,
+        "min_feedback_forward_cosine": 0.9966304898262024
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49280,
+      "train_examples": 45000,
+      "train_loss": 0.005641720619507962
+    },
+    {
+      "epoch": 141,
+      "eval_accuracy": 0.9266,
+      "eval_loss": 0.36997833251953127,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08228462934494019,
+        "mean_feedback_forward_cosine": 0.998879792473533,
+        "mean_feedback_forward_relative_error": 0.04603636854074218,
+        "min_feedback_forward_cosine": 0.9966309070587158
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49632,
+      "train_examples": 45000,
+      "train_loss": 0.0057345138311800025
+    },
+    {
+      "epoch": 142,
+      "eval_accuracy": 0.9278,
+      "eval_loss": 0.36403873291015626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08227932453155518,
+        "mean_feedback_forward_cosine": 0.998879955031655,
+        "mean_feedback_forward_relative_error": 0.046032756346870554,
+        "min_feedback_forward_cosine": 0.9966314435005188
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 49984,
+      "train_examples": 45000,
+      "train_loss": 0.004907455729775958
+    },
+    {
+      "epoch": 143,
+      "eval_accuracy": 0.928,
+      "eval_loss": 0.36518594360351564,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08227457106113434,
+        "mean_feedback_forward_cosine": 0.9988800861618735,
+        "mean_feedback_forward_relative_error": 0.046029586328024216,
+        "min_feedback_forward_cosine": 0.9966318607330322
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50336,
+      "train_examples": 45000,
+      "train_loss": 0.004633074706503087
+    },
+    {
+      "epoch": 144,
+      "eval_accuracy": 0.9288,
+      "eval_loss": 0.36308513793945313,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822695642709732,
+        "mean_feedback_forward_cosine": 0.9988802649758078,
+        "mean_feedback_forward_relative_error": 0.04602613601494919,
+        "min_feedback_forward_cosine": 0.9966320991516113
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 50688,
+      "train_examples": 45000,
+      "train_loss": 0.004616090032748051
+    },
+    {
+      "epoch": 145,
+      "eval_accuracy": 0.928,
+      "eval_loss": 0.3633923767089844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08226427435874939,
+        "mean_feedback_forward_cosine": 0.998880432952534,
+        "mean_feedback_forward_relative_error": 0.04602236039936543,
+        "min_feedback_forward_cosine": 0.9966326951980591
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51040,
+      "train_examples": 45000,
+      "train_loss": 0.005115178668167856
+    },
+    {
+      "epoch": 146,
+      "eval_accuracy": 0.9282,
+      "eval_loss": 0.3682091247558594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822596624493599,
+        "mean_feedback_forward_cosine": 0.9988805814222856,
+        "mean_feedback_forward_relative_error": 0.0460187761282379,
+        "min_feedback_forward_cosine": 0.9966330528259277
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51392,
+      "train_examples": 45000,
+      "train_loss": 0.004550867236633268
+    },
+    {
+      "epoch": 147,
+      "eval_accuracy": 0.927,
+      "eval_loss": 0.3651136474609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08225411176681519,
+        "mean_feedback_forward_cosine": 0.99888076023622,
+        "mean_feedback_forward_relative_error": 0.04601478727365082,
+        "min_feedback_forward_cosine": 0.9966334700584412
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 51744,
+      "train_examples": 45000,
+      "train_loss": 0.004593415929294295
+    },
+    {
+      "epoch": 148,
+      "eval_accuracy": 0.9262,
+      "eval_loss": 0.3646269836425781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08224842697381973,
+        "mean_feedback_forward_cosine": 0.9988809282129462,
+        "mean_feedback_forward_relative_error": 0.046011200902814214,
+        "min_feedback_forward_cosine": 0.9966338872909546
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52096,
+      "train_examples": 45000,
+      "train_loss": 0.004782757617574599
+    },
+    {
+      "epoch": 149,
+      "eval_accuracy": 0.9272,
+      "eval_loss": 0.365397119140625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822431892156601,
+        "mean_feedback_forward_cosine": 0.9988811276175759,
+        "mean_feedback_forward_relative_error": 0.04600701702928001,
+        "min_feedback_forward_cosine": 0.9966343641281128
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52448,
+      "train_examples": 45000,
+      "train_loss": 0.004701021136467655
+    },
+    {
+      "epoch": 150,
+      "eval_accuracy": 0.9304,
+      "eval_loss": 0.3664459228515625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223860710859299,
+        "mean_feedback_forward_cosine": 0.9988812652501193,
+        "mean_feedback_forward_relative_error": 0.0460035984150388,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.010000000000000002,
+      "output_lr": 0.010000000000000002,
+      "step": 52800,
+      "train_examples": 45000,
+      "train_loss": 0.004233766572032538
+    },
+    {
+      "epoch": 151,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.36865162353515624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223855495452881,
+        "mean_feedback_forward_cosine": 0.9988812836733731,
+        "mean_feedback_forward_relative_error": 0.046003527346659794,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53152,
+      "train_examples": 45000,
+      "train_loss": 0.0036866345710224574
+    },
+    {
+      "epoch": 152,
+      "eval_accuracy": 0.93,
+      "eval_loss": 0.36565570068359377,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223850280046463,
+        "mean_feedback_forward_cosine": 0.9988812869245356,
+        "mean_feedback_forward_relative_error": 0.046003469282930545,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53504,
+      "train_examples": 45000,
+      "train_loss": 0.003957245495936109
+    },
+    {
+      "epoch": 153,
+      "eval_accuracy": 0.929,
+      "eval_loss": 0.3695786437988281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223845064640045,
+        "mean_feedback_forward_cosine": 0.9988812869245356,
+        "mean_feedback_forward_relative_error": 0.04600340781564062,
+        "min_feedback_forward_cosine": 0.9966347813606262
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 53856,
+      "train_examples": 45000,
+      "train_loss": 0.003925301420285056
+    },
+    {
+      "epoch": 154,
+      "eval_accuracy": 0.9302,
+      "eval_loss": 0.36479189453125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223842084407806,
+        "mean_feedback_forward_cosine": 0.998881278254769,
+        "mean_feedback_forward_relative_error": 0.046003351631489664,
+        "min_feedback_forward_cosine": 0.9966347813606262
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54208,
+      "train_examples": 45000,
+      "train_loss": 0.0035732608032723267
+    },
+    {
+      "epoch": 155,
+      "eval_accuracy": 0.9302,
+      "eval_loss": 0.3620485961914062,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223836869001389,
+        "mean_feedback_forward_cosine": 0.9988812825896524,
+        "mean_feedback_forward_relative_error": 0.04600329211151058,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54560,
+      "train_examples": 45000,
+      "train_loss": 0.0037292424756826628
+    },
+    {
+      "epoch": 156,
+      "eval_accuracy": 0.9308,
+      "eval_loss": 0.3636377868652344,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822383388876915,
+        "mean_feedback_forward_cosine": 0.9988812825896524,
+        "mean_feedback_forward_relative_error": 0.046003235639496286,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 54912,
+      "train_examples": 45000,
+      "train_loss": 0.0035863490228851638
+    },
+    {
+      "epoch": 157,
+      "eval_accuracy": 0.93,
+      "eval_loss": 0.36617817993164065,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223830908536911,
+        "mean_feedback_forward_cosine": 0.9988812869245356,
+        "mean_feedback_forward_relative_error": 0.04600318036973476,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55264,
+      "train_examples": 45000,
+      "train_loss": 0.0034266716464733085
+    },
+    {
+      "epoch": 158,
+      "eval_accuracy": 0.9284,
+      "eval_loss": 0.36658130493164065,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223827183246613,
+        "mean_feedback_forward_cosine": 0.9988812825896524,
+        "mean_feedback_forward_relative_error": 0.04600312345745888,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55616,
+      "train_examples": 45000,
+      "train_loss": 0.0035902588734713693
+    },
+    {
+      "epoch": 159,
+      "eval_accuracy": 0.93,
+      "eval_loss": 0.3643918701171875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223822712898254,
+        "mean_feedback_forward_cosine": 0.998881290175698,
+        "mean_feedback_forward_relative_error": 0.04600307231938297,
+        "min_feedback_forward_cosine": 0.9966347813606262
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 55968,
+      "train_examples": 45000,
+      "train_loss": 0.003073369382456359
+    },
+    {
+      "epoch": 160,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.3648558898925781,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223821222782135,
+        "mean_feedback_forward_cosine": 0.998881290175698,
+        "mean_feedback_forward_relative_error": 0.04600302441553636,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56320,
+      "train_examples": 45000,
+      "train_loss": 0.0029175158710322445
+    },
+    {
+      "epoch": 161,
+      "eval_accuracy": 0.9304,
+      "eval_loss": 0.367481298828125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223816752433777,
+        "mean_feedback_forward_cosine": 0.9988812869245356,
+        "mean_feedback_forward_relative_error": 0.04600296841764992,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 56672,
+      "train_examples": 45000,
+      "train_loss": 0.0034142314993465942
+    },
+    {
+      "epoch": 162,
+      "eval_accuracy": 0.9284,
+      "eval_loss": 0.3675195251464844,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223813772201538,
+        "mean_feedback_forward_cosine": 0.9988812934268605,
+        "mean_feedback_forward_relative_error": 0.04600291731344028,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57024,
+      "train_examples": 45000,
+      "train_loss": 0.0030324965446359583
+    },
+    {
+      "epoch": 163,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.36677179565429685,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822380930185318,
+        "mean_feedback_forward_cosine": 0.998881302096627,
+        "mean_feedback_forward_relative_error": 0.04600286566737023,
+        "min_feedback_forward_cosine": 0.9966347217559814
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57376,
+      "train_examples": 45000,
+      "train_loss": 0.0034874971714284686
+    },
+    {
+      "epoch": 164,
+      "eval_accuracy": 0.9286,
+      "eval_loss": 0.36896765747070315,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223806321620941,
+        "mean_feedback_forward_cosine": 0.998881302096627,
+        "mean_feedback_forward_relative_error": 0.046002815003422175,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 57728,
+      "train_examples": 45000,
+      "train_loss": 0.0029989460522101986
+    },
+    {
+      "epoch": 165,
+      "eval_accuracy": 0.9306,
+      "eval_loss": 0.366157421875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223803341388702,
+        "mean_feedback_forward_cosine": 0.9988813031803477,
+        "mean_feedback_forward_relative_error": 0.04600276467813687,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58080,
+      "train_examples": 45000,
+      "train_loss": 0.0031791578263044356
+    },
+    {
+      "epoch": 166,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.3687070068359375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223800361156464,
+        "mean_feedback_forward_cosine": 0.9988812880082564,
+        "mean_feedback_forward_relative_error": 0.0460027188570662,
+        "min_feedback_forward_cosine": 0.9966347813606262
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58432,
+      "train_examples": 45000,
+      "train_loss": 0.002584952381066978
+    },
+    {
+      "epoch": 167,
+      "eval_accuracy": 0.93,
+      "eval_loss": 0.3658307983398438,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223797380924225,
+        "mean_feedback_forward_cosine": 0.9988812923431396,
+        "mean_feedback_forward_relative_error": 0.04600266805765304,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 58784,
+      "train_examples": 45000,
+      "train_loss": 0.0029853353516198695
+    },
+    {
+      "epoch": 168,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.36633271484375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223794400691986,
+        "mean_feedback_forward_cosine": 0.9988813064315102,
+        "mean_feedback_forward_relative_error": 0.04600261669944633,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59136,
+      "train_examples": 45000,
+      "train_loss": 0.0029499868230480285
+    },
+    {
+      "epoch": 169,
+      "eval_accuracy": 0.9298,
+      "eval_loss": 0.36940147705078125,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223791420459747,
+        "mean_feedback_forward_cosine": 0.9988812934268605,
+        "mean_feedback_forward_relative_error": 0.046002566916021434,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59488,
+      "train_examples": 45000,
+      "train_loss": 0.002881501671630475
+    },
+    {
+      "epoch": 170,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.369974560546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822378620505333,
+        "mean_feedback_forward_cosine": 0.9988812999291854,
+        "mean_feedback_forward_relative_error": 0.04600251288237897,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 59840,
+      "train_examples": 45000,
+      "train_loss": 0.0030441269177322585
+    },
+    {
+      "epoch": 171,
+      "eval_accuracy": 0.9282,
+      "eval_loss": 0.36931307983398437,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223782479763031,
+        "mean_feedback_forward_cosine": 0.9988813096826726,
+        "mean_feedback_forward_relative_error": 0.04600245668129487,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60192,
+      "train_examples": 45000,
+      "train_loss": 0.0033884187957892815
+    },
+    {
+      "epoch": 172,
+      "eval_accuracy": 0.9286,
+      "eval_loss": 0.36912783813476563,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223780989646912,
+        "mean_feedback_forward_cosine": 0.9988813031803477,
+        "mean_feedback_forward_relative_error": 0.04600240898064591,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60544,
+      "train_examples": 45000,
+      "train_loss": 0.0030959877320668764
+    },
+    {
+      "epoch": 173,
+      "eval_accuracy": 0.929,
+      "eval_loss": 0.36873092041015626,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223778009414673,
+        "mean_feedback_forward_cosine": 0.9988813042640686,
+        "mean_feedback_forward_relative_error": 0.04600236126306382,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 60896,
+      "train_examples": 45000,
+      "train_loss": 0.0028156025466819605
+    },
+    {
+      "epoch": 174,
+      "eval_accuracy": 0.9298,
+      "eval_loss": 0.36716961669921877,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223774284124374,
+        "mean_feedback_forward_cosine": 0.9988812934268605,
+        "mean_feedback_forward_relative_error": 0.04600231291895563,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61248,
+      "train_examples": 45000,
+      "train_loss": 0.0027567072859654824
+    },
+    {
+      "epoch": 175,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3688115600585937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223771303892136,
+        "mean_feedback_forward_cosine": 0.9988813085989519,
+        "mean_feedback_forward_relative_error": 0.04600225873291493,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61600,
+      "train_examples": 45000,
+      "train_loss": 0.003222914586795701
+    },
+    {
+      "epoch": 176,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.36736928100585936,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223768323659897,
+        "mean_feedback_forward_cosine": 0.9988813031803477,
+        "mean_feedback_forward_relative_error": 0.04600221047347242,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 61952,
+      "train_examples": 45000,
+      "train_loss": 0.0028732660255498354
+    },
+    {
+      "epoch": 177,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.372470849609375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223765343427658,
+        "mean_feedback_forward_cosine": 0.9988813107663935,
+        "mean_feedback_forward_relative_error": 0.04600216378881173,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62304,
+      "train_examples": 45000,
+      "train_loss": 0.0025538987044348483
+    },
+    {
+      "epoch": 178,
+      "eval_accuracy": 0.9284,
+      "eval_loss": 0.36910101928710937,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822376236319542,
+        "mean_feedback_forward_cosine": 0.9988813064315102,
+        "mean_feedback_forward_relative_error": 0.04600211475044489,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 62656,
+      "train_examples": 45000,
+      "train_loss": 0.002970595129744874
+    },
+    {
+      "epoch": 179,
+      "eval_accuracy": 0.9298,
+      "eval_loss": 0.37037894287109374,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822376161813736,
+        "mean_feedback_forward_cosine": 0.998881314017556,
+        "mean_feedback_forward_relative_error": 0.046002069691365416,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63008,
+      "train_examples": 45000,
+      "train_loss": 0.0025246295870385237
+    },
+    {
+      "epoch": 180,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.36845943603515624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223757892847061,
+        "mean_feedback_forward_cosine": 0.9988813075152311,
+        "mean_feedback_forward_relative_error": 0.0460020206191323,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63360,
+      "train_examples": 45000,
+      "train_loss": 0.002713063888189693
+    },
+    {
+      "epoch": 181,
+      "eval_accuracy": 0.9286,
+      "eval_loss": 0.3703472412109375,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223755657672882,
+        "mean_feedback_forward_cosine": 0.9988813118501143,
+        "mean_feedback_forward_relative_error": 0.04600197120823644,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 63712,
+      "train_examples": 45000,
+      "train_loss": 0.0025600299073590175
+    },
+    {
+      "epoch": 182,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3755940856933594,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223753422498703,
+        "mean_feedback_forward_cosine": 0.9988812988454645,
+        "mean_feedback_forward_relative_error": 0.0460019274699417,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64064,
+      "train_examples": 45000,
+      "train_loss": 0.0027186543648027714
+    },
+    {
+      "epoch": 183,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3694640441894531,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223750442266464,
+        "mean_feedback_forward_cosine": 0.9988813085989519,
+        "mean_feedback_forward_relative_error": 0.04600187953222882,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64416,
+      "train_examples": 45000,
+      "train_loss": 0.0028076297329221334
+    },
+    {
+      "epoch": 184,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3719008056640625,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223746716976166,
+        "mean_feedback_forward_cosine": 0.9988813118501143,
+        "mean_feedback_forward_relative_error": 0.046001831357452005,
+        "min_feedback_forward_cosine": 0.9966347813606262
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 64768,
+      "train_examples": 45000,
+      "train_loss": 0.00280507246243457
+    },
+    {
+      "epoch": 185,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3697983947753906,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223744481801987,
+        "mean_feedback_forward_cosine": 0.9988813129338351,
+        "mean_feedback_forward_relative_error": 0.046001782928678125,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65120,
+      "train_examples": 45000,
+      "train_loss": 0.002665952594392002
+    },
+    {
+      "epoch": 186,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.3683447998046875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223741501569748,
+        "mean_feedback_forward_cosine": 0.9988813216036017,
+        "mean_feedback_forward_relative_error": 0.04600173328071833,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65472,
+      "train_examples": 45000,
+      "train_loss": 0.0028820886498110163
+    },
+    {
+      "epoch": 187,
+      "eval_accuracy": 0.9282,
+      "eval_loss": 0.37149351806640624,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223740011453629,
+        "mean_feedback_forward_cosine": 0.9988813172687184,
+        "mean_feedback_forward_relative_error": 0.04600168620659546,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 65824,
+      "train_examples": 45000,
+      "train_loss": 0.00259548195157614
+    },
+    {
+      "epoch": 188,
+      "eval_accuracy": 0.9298,
+      "eval_loss": 0.37394226684570314,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822373703122139,
+        "mean_feedback_forward_cosine": 0.9988813085989519,
+        "mean_feedback_forward_relative_error": 0.0460016364401037,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66176,
+      "train_examples": 45000,
+      "train_loss": 0.0025710957238243687
+    },
+    {
+      "epoch": 189,
+      "eval_accuracy": 0.9294,
+      "eval_loss": 0.3728668823242188,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223734050989151,
+        "mean_feedback_forward_cosine": 0.9988813161849975,
+        "mean_feedback_forward_relative_error": 0.04600158809599551,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66528,
+      "train_examples": 45000,
+      "train_loss": 0.002441268061970671
+    },
+    {
+      "epoch": 190,
+      "eval_accuracy": 0.9278,
+      "eval_loss": 0.37325433959960935,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223730325698853,
+        "mean_feedback_forward_cosine": 0.9988813237710432,
+        "mean_feedback_forward_relative_error": 0.04600153495980935,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 66880,
+      "train_examples": 45000,
+      "train_loss": 0.0030400358228633803
+    },
+    {
+      "epoch": 191,
+      "eval_accuracy": 0.929,
+      "eval_loss": 0.3715138000488281,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223726600408554,
+        "mean_feedback_forward_cosine": 0.9988813443617387,
+        "mean_feedback_forward_relative_error": 0.04600149122151462,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67232,
+      "train_examples": 45000,
+      "train_loss": 0.0025101503582723025
+    },
+    {
+      "epoch": 192,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3713754638671875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223722875118256,
+        "mean_feedback_forward_cosine": 0.9988813129338351,
+        "mean_feedback_forward_relative_error": 0.04600143927064809,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67584,
+      "train_examples": 45000,
+      "train_loss": 0.0031009860323948993
+    },
+    {
+      "epoch": 193,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3713826843261719,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223719894886017,
+        "mean_feedback_forward_cosine": 0.9988813270222057,
+        "mean_feedback_forward_relative_error": 0.046001392823051324,
+        "min_feedback_forward_cosine": 0.996634840965271
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 67936,
+      "train_examples": 45000,
+      "train_loss": 0.0025039921342084807
+    },
+    {
+      "epoch": 194,
+      "eval_accuracy": 0.9298,
+      "eval_loss": 0.3682757446289063,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223717659711838,
+        "mean_feedback_forward_cosine": 0.9988813324408098,
+        "mean_feedback_forward_relative_error": 0.046001344089480964,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68288,
+      "train_examples": 45000,
+      "train_loss": 0.002788148091774848
+    },
+    {
+      "epoch": 195,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.37077115478515627,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.0822371393442154,
+        "mean_feedback_forward_cosine": 0.9988813226873224,
+        "mean_feedback_forward_relative_error": 0.046001296236433765,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68640,
+      "train_examples": 45000,
+      "train_loss": 0.002563142611583074
+    },
+    {
+      "epoch": 196,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.37052415771484376,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223710209131241,
+        "mean_feedback_forward_cosine": 0.9988813378594138,
+        "mean_feedback_forward_relative_error": 0.04600124660540711,
+        "min_feedback_forward_cosine": 0.9966350197792053
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 68992,
+      "train_examples": 45000,
+      "train_loss": 0.0027777536426153446
+    },
+    {
+      "epoch": 197,
+      "eval_accuracy": 0.929,
+      "eval_loss": 0.36994708862304687,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223707228899002,
+        "mean_feedback_forward_cosine": 0.9988813497803428,
+        "mean_feedback_forward_relative_error": 0.04600120046260682,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69344,
+      "train_examples": 45000,
+      "train_loss": 0.0028230392361473706
+    },
+    {
+      "epoch": 198,
+      "eval_accuracy": 0.9292,
+      "eval_loss": 0.3723174560546875,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223704993724823,
+        "mean_feedback_forward_cosine": 0.9988813291896473,
+        "mean_feedback_forward_relative_error": 0.04600115440447222,
+        "min_feedback_forward_cosine": 0.9966349005699158
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 69696,
+      "train_examples": 45000,
+      "train_loss": 0.0022861307460400795
+    },
+    {
+      "epoch": 199,
+      "eval_accuracy": 0.9296,
+      "eval_loss": 0.36884978637695315,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223701268434525,
+        "mean_feedback_forward_cosine": 0.9988813216036017,
+        "mean_feedback_forward_relative_error": 0.046001107211817396,
+        "min_feedback_forward_cosine": 0.9966349601745605
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70048,
+      "train_examples": 45000,
+      "train_loss": 0.0030935465381791193
+    },
+    {
+      "epoch": 200,
+      "eval_accuracy": 0.9274,
+      "eval_loss": 0.3740764099121094,
+      "eval_split": "validation",
+      "feedback_tracking": {
+        "max_feedback_forward_relative_error": 0.08223697543144226,
+        "mean_feedback_forward_cosine": 0.9988813335245306,
+        "mean_feedback_forward_relative_error": 0.04600105993449688,
+        "min_feedback_forward_cosine": 0.9966350197792053
+      },
+      "lr": 0.0010000000000000002,
+      "output_lr": 0.0010000000000000002,
+      "step": 70400,
+      "train_examples": 45000,
+      "train_loss": 0.002820015812975665
+    }
+  ],
+  "evaluation_protocol": {
+    "test_evaluations": 0,
+    "test_used_for_selection": false,
+    "validation_evaluations": 201
+  },
+  "final": {
+    "accuracy": 0.9274,
+    "epoch": 200,
+    "evaluation_split": "validation",
+    "finite": true,
+    "loss": 0.3740764099121094,
+    "step": 70400
+  },
+  "hardware": {
+    "cuda_device_name": "NVIDIA RTX A6000",
+    "cuda_visible_devices": "0",
+    "device": "cuda",
+    "device_total_memory_bytes": 50896568320,
+    "peak_memory_allocated_bytes": 3256025088,
+    "peak_memory_reserved_bytes": 4517265408,
+    "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": 60.171127796173096,
+    "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": 5054.333840370178,
+    "train_wall_s": 4991.6432955265045
+  },
+  "work": {
+    "apical_macs_per_example": 125043328,
+    "causal_scalar_observations": 0,
+    "components": {
+      "apical_projection_macs": 1125389952000000,
+      "apical_regression_macs": 0,
+      "bp_reverse_macs_estimate": 0,
+      "kp_reciprocal_correlation_macs": 1125389952000000,
+      "local_weight_correlation_macs": 1129371264000000,
+      "mirror_feedback_prediction_macs": 0,
+      "mirror_local_correlation_macs": 0,
+      "mirror_response_macs": 0,
+      "ordinary_forward_macs": 1129371264000000,
+      "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": 125485696,
+    "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": 4509522432000000
+  }
+}