diff --git "a/trainer_state.json" "b/trainer_state.json"
new file mode 100644--- /dev/null
+++ "b/trainer_state.json"
@@ -0,0 +1,24033 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.5,
+  "eval_steps": 2000,
+  "global_step": 20000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.00025,
+      "grad_norm": 32.5,
+      "learning_rate": 0.0001,
+      "loss": 7.9852,
+      "loss/crossentropy": 2.2558943748474123,
+      "loss/hidden": 3.53671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22032691352069378,
+      "step": 10
+    },
+    {
+      "epoch": 0.0005,
+      "grad_norm": 39.0,
+      "grad_norm_var": 6.1306640625,
+      "learning_rate": 0.0001,
+      "loss": 8.0827,
+      "loss/crossentropy": 2.219619666039944,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20763051956892015,
+      "step": 20
+    },
+    {
+      "epoch": 0.00075,
+      "grad_norm": 30.375,
+      "grad_norm_var": 7.383072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.8479,
+      "loss/crossentropy": 2.185364603996277,
+      "loss/hidden": 3.59296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23391152992844583,
+      "step": 30
+    },
+    {
+      "epoch": 0.001,
+      "grad_norm": 32.75,
+      "grad_norm_var": 187.0322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.8341,
+      "loss/crossentropy": 2.083733668923378,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2040597340092063,
+      "step": 40
+    },
+    {
+      "epoch": 0.00125,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.111393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6815,
+      "loss/crossentropy": 2.182037726044655,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20423058047890663,
+      "step": 50
+    },
+    {
+      "epoch": 0.0015,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.3853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6919,
+      "loss/crossentropy": 2.1419573068618774,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19463330563157796,
+      "step": 60
+    },
+    {
+      "epoch": 0.00175,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.228125,
+      "learning_rate": 0.0001,
+      "loss": 7.838,
+      "loss/crossentropy": 2.242894622683525,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22062199115753173,
+      "step": 70
+    },
+    {
+      "epoch": 0.002,
+      "grad_norm": 32.5,
+      "grad_norm_var": 8.371809895833334,
+      "learning_rate": 0.0001,
+      "loss": 8.018,
+      "loss/crossentropy": 2.0408543169498445,
+      "loss/hidden": 3.506640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20080858804285526,
+      "step": 80
+    },
+    {
+      "epoch": 0.00225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.322916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.8807,
+      "loss/crossentropy": 2.0654082030057905,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2109438929706812,
+      "step": 90
+    },
+    {
+      "epoch": 0.0025,
+      "grad_norm": 34.5,
+      "grad_norm_var": 2805.3603515625,
+      "learning_rate": 0.0001,
+      "loss": 8.0497,
+      "loss/crossentropy": 2.0930048365145923,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2103888330049813,
+      "step": 100
+    },
+    {
+      "epoch": 0.00275,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2798.430143229167,
+      "learning_rate": 0.0001,
+      "loss": 7.8583,
+      "loss/crossentropy": 2.3308374524116515,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2045755073428154,
+      "step": 110
+    },
+    {
+      "epoch": 0.003,
+      "grad_norm": 34.0,
+      "grad_norm_var": 1.6999348958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.7088,
+      "loss/crossentropy": 2.120428466796875,
+      "loss/hidden": 3.56796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21561774536967276,
+      "step": 120
+    },
+    {
+      "epoch": 0.00325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 40.73899739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.762,
+      "loss/crossentropy": 2.1464689180254934,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2084670951589942,
+      "step": 130
+    },
+    {
+      "epoch": 0.0035,
+      "grad_norm": 33.0,
+      "grad_norm_var": 42.67265625,
+      "learning_rate": 0.0001,
+      "loss": 7.711,
+      "loss/crossentropy": 2.1301105961203577,
+      "loss/hidden": 3.512890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21644905991852284,
+      "step": 140
+    },
+    {
+      "epoch": 0.00375,
+      "grad_norm": 32.75,
+      "grad_norm_var": 8.066080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.7295,
+      "loss/crossentropy": 2.115240353345871,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2064087452366948,
+      "step": 150
+    },
+    {
+      "epoch": 0.004,
+      "grad_norm": 32.5,
+      "grad_norm_var": 422.72057291666664,
+      "learning_rate": 0.0001,
+      "loss": 7.9367,
+      "loss/crossentropy": 2.141779786348343,
+      "loss/hidden": 3.457421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21747791785746812,
+      "step": 160
+    },
+    {
+      "epoch": 0.00425,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.705208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.756,
+      "loss/crossentropy": 2.103468084335327,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2027540819719434,
+      "step": 170
+    },
+    {
+      "epoch": 0.0045,
+      "grad_norm": 32.25,
+      "grad_norm_var": 17.8556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.7008,
+      "loss/crossentropy": 2.1950977832078933,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2186649737879634,
+      "step": 180
+    },
+    {
+      "epoch": 0.00475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 66.61608072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.6552,
+      "loss/crossentropy": 2.029752139747143,
+      "loss/hidden": 3.585546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22110425475984813,
+      "step": 190
+    },
+    {
+      "epoch": 0.005,
+      "grad_norm": 36.0,
+      "grad_norm_var": 13.160872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6964,
+      "loss/crossentropy": 2.169684535264969,
+      "loss/hidden": 3.52265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22122678495943546,
+      "step": 200
+    },
+    {
+      "epoch": 0.00525,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.442643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.7135,
+      "loss/crossentropy": 2.1366169154644012,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19148335698992014,
+      "step": 210
+    },
+    {
+      "epoch": 0.0055,
+      "grad_norm": 35.25,
+      "grad_norm_var": 4.151822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.654,
+      "loss/crossentropy": 2.196457767486572,
+      "loss/hidden": 3.573828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22240160517394542,
+      "step": 220
+    },
+    {
+      "epoch": 0.00575,
+      "grad_norm": 34.5,
+      "grad_norm_var": 61.07057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7286,
+      "loss/crossentropy": 2.099469523131847,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19720815271139144,
+      "step": 230
+    },
+    {
+      "epoch": 0.006,
+      "grad_norm": 33.75,
+      "grad_norm_var": 5.479166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6841,
+      "loss/crossentropy": 2.0730464071035386,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20290055498480797,
+      "step": 240
+    },
+    {
+      "epoch": 0.00625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.9197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6999,
+      "loss/crossentropy": 2.1514100462198256,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013280361890793,
+      "step": 250
+    },
+    {
+      "epoch": 0.0065,
+      "grad_norm": 33.25,
+      "grad_norm_var": 150.86712239583332,
+      "learning_rate": 0.0001,
+      "loss": 7.7617,
+      "loss/crossentropy": 2.0071112543344496,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1923872010782361,
+      "step": 260
+    },
+    {
+      "epoch": 0.00675,
+      "grad_norm": 32.25,
+      "grad_norm_var": 41.373958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.7968,
+      "loss/crossentropy": 2.254330241680145,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20844143405556678,
+      "step": 270
+    },
+    {
+      "epoch": 0.007,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.9625,
+      "learning_rate": 0.0001,
+      "loss": 7.7437,
+      "loss/crossentropy": 2.2058747708797455,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19603454861789943,
+      "step": 280
+    },
+    {
+      "epoch": 0.00725,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.4306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.7308,
+      "loss/crossentropy": 2.158563455939293,
+      "loss/hidden": 3.485546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20696840062737465,
+      "step": 290
+    },
+    {
+      "epoch": 0.0075,
+      "grad_norm": 35.25,
+      "grad_norm_var": 4.163997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.7781,
+      "loss/crossentropy": 2.1728759124875068,
+      "loss/hidden": 3.6984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2232737574726343,
+      "step": 300
+    },
+    {
+      "epoch": 0.00775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.7039,
+      "loss/crossentropy": 2.1304744452238085,
+      "loss/hidden": 3.510546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21460633352398872,
+      "step": 310
+    },
+    {
+      "epoch": 0.008,
+      "grad_norm": 34.25,
+      "grad_norm_var": 102.21979166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7386,
+      "loss/crossentropy": 2.2150494754314423,
+      "loss/hidden": 3.470703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2204894032329321,
+      "step": 320
+    },
+    {
+      "epoch": 0.00825,
+      "grad_norm": 34.5,
+      "grad_norm_var": 103.0806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6843,
+      "loss/crossentropy": 2.1483607694506643,
+      "loss/hidden": 3.525,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2168185070157051,
+      "step": 330
+    },
+    {
+      "epoch": 0.0085,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.245768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6824,
+      "loss/crossentropy": 2.2668254554271696,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978676740080118,
+      "step": 340
+    },
+    {
+      "epoch": 0.00875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 7.785416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6476,
+      "loss/crossentropy": 2.247351437807083,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2044550308957696,
+      "step": 350
+    },
+    {
+      "epoch": 0.009,
+      "grad_norm": 33.0,
+      "grad_norm_var": 36.1150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6707,
+      "loss/crossentropy": 2.2508403569459916,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20586735829710961,
+      "step": 360
+    },
+    {
+      "epoch": 0.00925,
+      "grad_norm": 27.875,
+      "grad_norm_var": 42.02805989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.787,
+      "loss/crossentropy": 2.2524363905191422,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20639074668288232,
+      "step": 370
+    },
+    {
+      "epoch": 0.0095,
+      "grad_norm": 29.625,
+      "grad_norm_var": 19.65625,
+      "learning_rate": 0.0001,
+      "loss": 7.7258,
+      "loss/crossentropy": 2.2901588469743728,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20220000632107257,
+      "step": 380
+    },
+    {
+      "epoch": 0.00975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 15.653059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6445,
+      "loss/crossentropy": 2.1256834477186204,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190180828794837,
+      "step": 390
+    },
+    {
+      "epoch": 0.01,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.7712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.7719,
+      "loss/crossentropy": 2.2237626880407335,
+      "loss/hidden": 3.484765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2264560218900442,
+      "step": 400
+    },
+    {
+      "epoch": 0.01025,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.9247395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.7631,
+      "loss/crossentropy": 2.060544753074646,
+      "loss/hidden": 3.470703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2202487275004387,
+      "step": 410
+    },
+    {
+      "epoch": 0.0105,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.6712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.8193,
+      "loss/crossentropy": 2.270045906305313,
+      "loss/hidden": 3.49921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22449300419539214,
+      "step": 420
+    },
+    {
+      "epoch": 0.01075,
+      "grad_norm": 40.0,
+      "grad_norm_var": 5.551822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7502,
+      "loss/crossentropy": 2.220390594005585,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20708895958960055,
+      "step": 430
+    },
+    {
+      "epoch": 0.011,
+      "grad_norm": 33.75,
+      "grad_norm_var": 6.554622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6386,
+      "loss/crossentropy": 2.0412577211856844,
+      "loss/hidden": 3.52109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20780573673546315,
+      "step": 440
+    },
+    {
+      "epoch": 0.01125,
+      "grad_norm": 39.5,
+      "grad_norm_var": 9.018684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.7522,
+      "loss/crossentropy": 2.085190561413765,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22059339918196202,
+      "step": 450
+    },
+    {
+      "epoch": 0.0115,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.703580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.604,
+      "loss/crossentropy": 1.9451062515378,
+      "loss/hidden": 3.527734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2082566052675247,
+      "step": 460
+    },
+    {
+      "epoch": 0.01175,
+      "grad_norm": 34.25,
+      "grad_norm_var": 1.4416666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7286,
+      "loss/crossentropy": 2.23127267062664,
+      "loss/hidden": 3.560546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22645943984389305,
+      "step": 470
+    },
+    {
+      "epoch": 0.012,
+      "grad_norm": 43.5,
+      "grad_norm_var": 61.358333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5857,
+      "loss/crossentropy": 2.0275631666183473,
+      "loss/hidden": 3.341015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19867698503658177,
+      "step": 480
+    },
+    {
+      "epoch": 0.01225,
+      "grad_norm": 30.5,
+      "grad_norm_var": 62.425455729166664,
+      "learning_rate": 0.0001,
+      "loss": 7.7023,
+      "loss/crossentropy": 2.1564531326293945,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21295207217335702,
+      "step": 490
+    },
+    {
+      "epoch": 0.0125,
+      "grad_norm": 39.0,
+      "grad_norm_var": 7.580208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.8035,
+      "loss/crossentropy": 2.1716607972979545,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21046865545213223,
+      "step": 500
+    },
+    {
+      "epoch": 0.01275,
+      "grad_norm": 35.25,
+      "grad_norm_var": 5.121809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.764,
+      "loss/crossentropy": 2.181091034412384,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20339491367340087,
+      "step": 510
+    },
+    {
+      "epoch": 0.013,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.820572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6112,
+      "loss/crossentropy": 2.012446442246437,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20432959645986556,
+      "step": 520
+    },
+    {
+      "epoch": 0.01325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.92890625,
+      "learning_rate": 0.0001,
+      "loss": 7.73,
+      "loss/crossentropy": 2.0826023176312445,
+      "loss/hidden": 3.51171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21190985422581435,
+      "step": 530
+    },
+    {
+      "epoch": 0.0135,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.0229166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7527,
+      "loss/crossentropy": 2.191486455500126,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1943425141274929,
+      "step": 540
+    },
+    {
+      "epoch": 0.01375,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.983268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.7035,
+      "loss/crossentropy": 2.0664332896471023,
+      "loss/hidden": 3.475,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20404404532164336,
+      "step": 550
+    },
+    {
+      "epoch": 0.014,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.381705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5679,
+      "loss/crossentropy": 2.0470397621393204,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21556914187967777,
+      "step": 560
+    },
+    {
+      "epoch": 0.01425,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.508072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7909,
+      "loss/crossentropy": 2.21784293949604,
+      "loss/hidden": 3.4671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20632803943008185,
+      "step": 570
+    },
+    {
+      "epoch": 0.0145,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.6837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6449,
+      "loss/crossentropy": 2.291385439038277,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19818231668323277,
+      "step": 580
+    },
+    {
+      "epoch": 0.01475,
+      "grad_norm": 30.875,
+      "grad_norm_var": 106.909375,
+      "learning_rate": 0.0001,
+      "loss": 7.7009,
+      "loss/crossentropy": 2.102944087982178,
+      "loss/hidden": 3.532421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20003035496920346,
+      "step": 590
+    },
+    {
+      "epoch": 0.015,
+      "grad_norm": 33.0,
+      "grad_norm_var": 11.816080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5907,
+      "loss/crossentropy": 2.239013722538948,
+      "loss/hidden": 3.471484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22811597101390363,
+      "step": 600
+    },
+    {
+      "epoch": 0.01525,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.2811848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6578,
+      "loss/crossentropy": 2.1826944231986998,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19526719450950622,
+      "step": 610
+    },
+    {
+      "epoch": 0.0155,
+      "grad_norm": 32.0,
+      "grad_norm_var": 29.54140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6592,
+      "loss/crossentropy": 2.0491475805640222,
+      "loss/hidden": 3.47109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19470291025936604,
+      "step": 620
+    },
+    {
+      "epoch": 0.01575,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.458268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.7237,
+      "loss/crossentropy": 2.171899539232254,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19583625346422195,
+      "step": 630
+    },
+    {
+      "epoch": 0.016,
+      "grad_norm": 41.5,
+      "grad_norm_var": 835.4395182291667,
+      "learning_rate": 0.0001,
+      "loss": 7.6493,
+      "loss/crossentropy": 2.019927790760994,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889604590833187,
+      "step": 640
+    },
+    {
+      "epoch": 0.01625,
+      "grad_norm": 33.25,
+      "grad_norm_var": 816.1582682291667,
+      "learning_rate": 0.0001,
+      "loss": 7.7628,
+      "loss/crossentropy": 2.146556834876537,
+      "loss/hidden": 3.53046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21917179077863694,
+      "step": 650
+    },
+    {
+      "epoch": 0.0165,
+      "grad_norm": 34.0,
+      "grad_norm_var": 13.574739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6199,
+      "loss/crossentropy": 2.2131199680268763,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20602547163143753,
+      "step": 660
+    },
+    {
+      "epoch": 0.01675,
+      "grad_norm": 34.75,
+      "grad_norm_var": 28.6125,
+      "learning_rate": 0.0001,
+      "loss": 7.6113,
+      "loss/crossentropy": 2.0343705236911775,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18553176671266555,
+      "step": 670
+    },
+    {
+      "epoch": 0.017,
+      "grad_norm": 31.625,
+      "grad_norm_var": 19.8369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6443,
+      "loss/crossentropy": 2.1528817296028135,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20593271851539613,
+      "step": 680
+    },
+    {
+      "epoch": 0.01725,
+      "grad_norm": 36.25,
+      "grad_norm_var": 5.97265625,
+      "learning_rate": 0.0001,
+      "loss": 7.7476,
+      "loss/crossentropy": 2.2202903479337692,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2035485502332449,
+      "step": 690
+    },
+    {
+      "epoch": 0.0175,
+      "grad_norm": 31.5,
+      "grad_norm_var": 7.5431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.621,
+      "loss/crossentropy": 2.0744683638215067,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2023961789906025,
+      "step": 700
+    },
+    {
+      "epoch": 0.01775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 26.912239583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.7058,
+      "loss/crossentropy": 2.1098077327013014,
+      "loss/hidden": 3.48359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20867060720920563,
+      "step": 710
+    },
+    {
+      "epoch": 0.018,
+      "grad_norm": 34.0,
+      "grad_norm_var": 5.291080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5224,
+      "loss/crossentropy": 2.0663078971207143,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1999094202183187,
+      "step": 720
+    },
+    {
+      "epoch": 0.01825,
+      "grad_norm": 36.75,
+      "grad_norm_var": 34.112239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5679,
+      "loss/crossentropy": 2.264209559559822,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.213060562312603,
+      "step": 730
+    },
+    {
+      "epoch": 0.0185,
+      "grad_norm": 31.875,
+      "grad_norm_var": 31.756705729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.6714,
+      "loss/crossentropy": 2.128282290697098,
+      "loss/hidden": 3.494921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21533375550061465,
+      "step": 740
+    },
+    {
+      "epoch": 0.01875,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.314518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6825,
+      "loss/crossentropy": 2.0806978911161425,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19196727648377418,
+      "step": 750
+    },
+    {
+      "epoch": 0.019,
+      "grad_norm": 31.625,
+      "grad_norm_var": 9.520833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6058,
+      "loss/crossentropy": 2.2315777271986006,
+      "loss/hidden": 3.484765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21225934717804193,
+      "step": 760
+    },
+    {
+      "epoch": 0.01925,
+      "grad_norm": 33.25,
+      "grad_norm_var": 29.381705729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.8377,
+      "loss/crossentropy": 2.200978134572506,
+      "loss/hidden": 3.52890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22721426151692867,
+      "step": 770
+    },
+    {
+      "epoch": 0.0195,
+      "grad_norm": 32.75,
+      "grad_norm_var": 11.095247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.7305,
+      "loss/crossentropy": 2.2799030035734176,
+      "loss/hidden": 3.512890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2418980894610286,
+      "step": 780
+    },
+    {
+      "epoch": 0.01975,
+      "grad_norm": 46.5,
+      "grad_norm_var": 22.12265625,
+      "learning_rate": 0.0001,
+      "loss": 7.663,
+      "loss/crossentropy": 2.0916543275117876,
+      "loss/hidden": 3.4765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20638740565627814,
+      "step": 790
+    },
+    {
+      "epoch": 0.02,
+      "grad_norm": 53.75,
+      "grad_norm_var": 2.3053503983968586e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6743,
+      "loss/crossentropy": 2.2010276943445204,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2011238183826208,
+      "step": 800
+    },
+    {
+      "epoch": 0.02025,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.3053503962269005e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.7365,
+      "loss/crossentropy": 2.2651585280895232,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22051467839628458,
+      "step": 810
+    },
+    {
+      "epoch": 0.0205,
+      "grad_norm": 32.25,
+      "grad_norm_var": 69.03854166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5788,
+      "loss/crossentropy": 2.1270318403840065,
+      "loss/hidden": 3.47109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20940354652702808,
+      "step": 820
+    },
+    {
+      "epoch": 0.02075,
+      "grad_norm": 33.0,
+      "grad_norm_var": 8.0603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6237,
+      "loss/crossentropy": 2.1726820170879364,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20644128862768413,
+      "step": 830
+    },
+    {
+      "epoch": 0.021,
+      "grad_norm": 33.5,
+      "grad_norm_var": 9.085416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.455,
+      "loss/crossentropy": 2.166350546479225,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20701032150536774,
+      "step": 840
+    },
+    {
+      "epoch": 0.02125,
+      "grad_norm": 37.75,
+      "grad_norm_var": 6.362239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5725,
+      "loss/crossentropy": 2.095318245887756,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20427223704755307,
+      "step": 850
+    },
+    {
+      "epoch": 0.0215,
+      "grad_norm": 29.625,
+      "grad_norm_var": 15.934830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5728,
+      "loss/crossentropy": 2.1999073296785356,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2012161746621132,
+      "step": 860
+    },
+    {
+      "epoch": 0.02175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.681184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6168,
+      "loss/crossentropy": 2.2957967817783356,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20650502648204566,
+      "step": 870
+    },
+    {
+      "epoch": 0.022,
+      "grad_norm": 30.0,
+      "grad_norm_var": 10.9181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6929,
+      "loss/crossentropy": 2.2093090921640397,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18699637930840254,
+      "step": 880
+    },
+    {
+      "epoch": 0.02225,
+      "grad_norm": 35.0,
+      "grad_norm_var": 11.192122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5316,
+      "loss/crossentropy": 2.0586251467466354,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21464286223053933,
+      "step": 890
+    },
+    {
+      "epoch": 0.0225,
+      "grad_norm": 48.0,
+      "grad_norm_var": 22.92265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5857,
+      "loss/crossentropy": 2.1724458605051042,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19797445200383662,
+      "step": 900
+    },
+    {
+      "epoch": 0.02275,
+      "grad_norm": 33.0,
+      "grad_norm_var": 38.4837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6184,
+      "loss/crossentropy": 2.15934486836195,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19419073052704333,
+      "step": 910
+    },
+    {
+      "epoch": 0.023,
+      "grad_norm": 31.0,
+      "grad_norm_var": 29.8869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6209,
+      "loss/crossentropy": 2.1601561695337295,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20290262177586554,
+      "step": 920
+    },
+    {
+      "epoch": 0.02325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.5143229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5969,
+      "loss/crossentropy": 2.184823766350746,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2016730338335037,
+      "step": 930
+    },
+    {
+      "epoch": 0.0235,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.5764973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5772,
+      "loss/crossentropy": 2.2380657255649568,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19638443663716315,
+      "step": 940
+    },
+    {
+      "epoch": 0.02375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.199934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6348,
+      "loss/crossentropy": 2.183762513846159,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19700763542205096,
+      "step": 950
+    },
+    {
+      "epoch": 0.024,
+      "grad_norm": 33.0,
+      "grad_norm_var": 20.1666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6351,
+      "loss/crossentropy": 2.1142914414405825,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2082229983061552,
+      "step": 960
+    },
+    {
+      "epoch": 0.02425,
+      "grad_norm": 31.5,
+      "grad_norm_var": 18.738541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6302,
+      "loss/crossentropy": 2.315229868888855,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20684304945170878,
+      "step": 970
+    },
+    {
+      "epoch": 0.0245,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.501822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6816,
+      "loss/crossentropy": 2.094369947910309,
+      "loss/hidden": 3.59921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2428071454167366,
+      "step": 980
+    },
+    {
+      "epoch": 0.02475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6289,
+      "loss/crossentropy": 2.183109185099602,
+      "loss/hidden": 3.5,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21173047311604024,
+      "step": 990
+    },
+    {
+      "epoch": 0.025,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.9593098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.6417,
+      "loss/crossentropy": 2.1311034083366396,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1998819222673774,
+      "step": 1000
+    },
+    {
+      "epoch": 0.02525,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.510872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6299,
+      "loss/crossentropy": 2.1297010451555254,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20842864252626897,
+      "step": 1010
+    },
+    {
+      "epoch": 0.0255,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.9893229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6593,
+      "loss/crossentropy": 2.2224902719259263,
+      "loss/hidden": 3.509375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21827217563986778,
+      "step": 1020
+    },
+    {
+      "epoch": 0.02575,
+      "grad_norm": 38.25,
+      "grad_norm_var": 6.010416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6712,
+      "loss/crossentropy": 2.1976612359285355,
+      "loss/hidden": 3.506640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.210753770545125,
+      "step": 1030
+    },
+    {
+      "epoch": 0.026,
+      "grad_norm": 33.5,
+      "grad_norm_var": 9.435872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.748,
+      "loss/crossentropy": 2.1889317661523817,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2081079863011837,
+      "step": 1040
+    },
+    {
+      "epoch": 0.02625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 6.206184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5803,
+      "loss/crossentropy": 2.0802227064967154,
+      "loss/hidden": 3.46796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19740718584507705,
+      "step": 1050
+    },
+    {
+      "epoch": 0.0265,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.3770182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6098,
+      "loss/crossentropy": 1.969551184773445,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1954148853197694,
+      "step": 1060
+    },
+    {
+      "epoch": 0.02675,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.09375,
+      "learning_rate": 0.0001,
+      "loss": 7.7826,
+      "loss/crossentropy": 2.160974936187267,
+      "loss/hidden": 3.581640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2183740811422467,
+      "step": 1070
+    },
+    {
+      "epoch": 0.027,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.6520182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6186,
+      "loss/crossentropy": 2.179084411263466,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19020346291363238,
+      "step": 1080
+    },
+    {
+      "epoch": 0.02725,
+      "grad_norm": 56.5,
+      "grad_norm_var": 48.91868489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.7756,
+      "loss/crossentropy": 2.1456103891134264,
+      "loss/hidden": 3.5,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2178689869120717,
+      "step": 1090
+    },
+    {
+      "epoch": 0.0275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 42.19557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6714,
+      "loss/crossentropy": 2.2156156271696092,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013509316369891,
+      "step": 1100
+    },
+    {
+      "epoch": 0.02775,
+      "grad_norm": 31.625,
+      "grad_norm_var": 29.3775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6833,
+      "loss/crossentropy": 2.0683740943670275,
+      "loss/hidden": 3.5046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21320818569511174,
+      "step": 1110
+    },
+    {
+      "epoch": 0.028,
+      "grad_norm": 27.5,
+      "grad_norm_var": 35.483333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.7302,
+      "loss/crossentropy": 2.098052313923836,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19890500828623772,
+      "step": 1120
+    },
+    {
+      "epoch": 0.02825,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.8775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6716,
+      "loss/crossentropy": 2.0999813921749593,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18992104195058346,
+      "step": 1130
+    },
+    {
+      "epoch": 0.0285,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5401,
+      "loss/crossentropy": 2.07411085665226,
+      "loss/hidden": 3.433984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2018830729648471,
+      "step": 1140
+    },
+    {
+      "epoch": 0.02875,
+      "grad_norm": 31.125,
+      "grad_norm_var": 18.053580729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.645,
+      "loss/crossentropy": 2.0945233553647995,
+      "loss/hidden": 3.4859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21366582233458759,
+      "step": 1150
+    },
+    {
+      "epoch": 0.029,
+      "grad_norm": 31.25,
+      "grad_norm_var": 16.978125,
+      "learning_rate": 0.0001,
+      "loss": 7.6514,
+      "loss/crossentropy": 2.0980678737163543,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19850811325013637,
+      "step": 1160
+    },
+    {
+      "epoch": 0.02925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 30.2556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6021,
+      "loss/crossentropy": 2.155895306169987,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19983574748039246,
+      "step": 1170
+    },
+    {
+      "epoch": 0.0295,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.733072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6343,
+      "loss/crossentropy": 2.1906268298625946,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19584416709840297,
+      "step": 1180
+    },
+    {
+      "epoch": 0.02975,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.89765625,
+      "learning_rate": 0.0001,
+      "loss": 7.7077,
+      "loss/crossentropy": 2.163237012922764,
+      "loss/hidden": 3.562109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21741114580072463,
+      "step": 1190
+    },
+    {
+      "epoch": 0.03,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.6639973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6695,
+      "loss/crossentropy": 2.1346954315900804,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19178961254656315,
+      "step": 1200
+    },
+    {
+      "epoch": 0.03025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 24.9125,
+      "learning_rate": 0.0001,
+      "loss": 7.6108,
+      "loss/crossentropy": 2.2493597716093063,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.209975734166801,
+      "step": 1210
+    },
+    {
+      "epoch": 0.0305,
+      "grad_norm": 30.75,
+      "grad_norm_var": 5.566080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5902,
+      "loss/crossentropy": 2.046968361735344,
+      "loss/hidden": 3.451171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867401722818613,
+      "step": 1220
+    },
+    {
+      "epoch": 0.03075,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.396875,
+      "learning_rate": 0.0001,
+      "loss": 7.6659,
+      "loss/crossentropy": 2.0429708033800127,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18792454693466426,
+      "step": 1230
+    },
+    {
+      "epoch": 0.031,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.8827473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5425,
+      "loss/crossentropy": 2.124686148762703,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23028194047510625,
+      "step": 1240
+    },
+    {
+      "epoch": 0.03125,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.220833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.7766,
+      "loss/crossentropy": 2.1431034594774245,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1852023523300886,
+      "step": 1250
+    },
+    {
+      "epoch": 0.0315,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.3080729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6547,
+      "loss/crossentropy": 2.208024913072586,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19009452145546674,
+      "step": 1260
+    },
+    {
+      "epoch": 0.03175,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.7249348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.6331,
+      "loss/crossentropy": 2.0774734795093535,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19277823474258185,
+      "step": 1270
+    },
+    {
+      "epoch": 0.032,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.6393229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6147,
+      "loss/crossentropy": 2.244540962576866,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20150573179125786,
+      "step": 1280
+    },
+    {
+      "epoch": 0.03225,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.2280598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6724,
+      "loss/crossentropy": 2.1218355029821394,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19727950319647788,
+      "step": 1290
+    },
+    {
+      "epoch": 0.0325,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.6113932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5881,
+      "loss/crossentropy": 2.048227934539318,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19401397118344904,
+      "step": 1300
+    },
+    {
+      "epoch": 0.03275,
+      "grad_norm": 35.5,
+      "grad_norm_var": 2.8384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6481,
+      "loss/crossentropy": 2.0217724472284315,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19338970091193913,
+      "step": 1310
+    },
+    {
+      "epoch": 0.033,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.3671223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5734,
+      "loss/crossentropy": 2.123840129375458,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19704403057694436,
+      "step": 1320
+    },
+    {
+      "epoch": 0.03325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.8207682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5379,
+      "loss/crossentropy": 2.1691371381282805,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20379403475672006,
+      "step": 1330
+    },
+    {
+      "epoch": 0.0335,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.162239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5824,
+      "loss/crossentropy": 2.0320975854992867,
+      "loss/hidden": 3.490625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.204788769595325,
+      "step": 1340
+    },
+    {
+      "epoch": 0.03375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 29.69140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6835,
+      "loss/crossentropy": 2.1799038141965865,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21532316971570253,
+      "step": 1350
+    },
+    {
+      "epoch": 0.034,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.4837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6461,
+      "loss/crossentropy": 2.075017270445824,
+      "loss/hidden": 3.483984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21185822309926153,
+      "step": 1360
+    },
+    {
+      "epoch": 0.03425,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.6806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6084,
+      "loss/crossentropy": 2.2061389327049254,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19864549599587916,
+      "step": 1370
+    },
+    {
+      "epoch": 0.0345,
+      "grad_norm": 46.25,
+      "grad_norm_var": 16.428125,
+      "learning_rate": 0.0001,
+      "loss": 7.5773,
+      "loss/crossentropy": 2.1411263316869737,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18519967906177043,
+      "step": 1380
+    },
+    {
+      "epoch": 0.03475,
+      "grad_norm": 33.5,
+      "grad_norm_var": 312.2525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6651,
+      "loss/crossentropy": 2.1760765284299852,
+      "loss/hidden": 3.457421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19441114887595176,
+      "step": 1390
+    },
+    {
+      "epoch": 0.035,
+      "grad_norm": 30.375,
+      "grad_norm_var": 7.5697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6384,
+      "loss/crossentropy": 2.148920811712742,
+      "loss/hidden": 3.503515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20845398511737584,
+      "step": 1400
+    },
+    {
+      "epoch": 0.03525,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.2301432291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6503,
+      "loss/crossentropy": 2.2241507709026336,
+      "loss/hidden": 3.478515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013774536550045,
+      "step": 1410
+    },
+    {
+      "epoch": 0.0355,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.67265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6107,
+      "loss/crossentropy": 2.151585003733635,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19790932536125183,
+      "step": 1420
+    },
+    {
+      "epoch": 0.03575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.192643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6089,
+      "loss/crossentropy": 2.106749549508095,
+      "loss/hidden": 3.555078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20702882390469313,
+      "step": 1430
+    },
+    {
+      "epoch": 0.036,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.8622395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6081,
+      "loss/crossentropy": 2.0727885022759436,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21057205237448215,
+      "step": 1440
+    },
+    {
+      "epoch": 0.03625,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.32890625,
+      "learning_rate": 0.0001,
+      "loss": 7.7005,
+      "loss/crossentropy": 2.3014074742794035,
+      "loss/hidden": 3.35859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20723759960383176,
+      "step": 1450
+    },
+    {
+      "epoch": 0.0365,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.240625,
+      "learning_rate": 0.0001,
+      "loss": 7.5677,
+      "loss/crossentropy": 2.12650800794363,
+      "loss/hidden": 3.44609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20236929692327976,
+      "step": 1460
+    },
+    {
+      "epoch": 0.03675,
+      "grad_norm": 31.0,
+      "grad_norm_var": 8.195768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.7299,
+      "loss/crossentropy": 2.187662351131439,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20965678989887238,
+      "step": 1470
+    },
+    {
+      "epoch": 0.037,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.1666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5942,
+      "loss/crossentropy": 2.174471014738083,
+      "loss/hidden": 3.534765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21107099391520023,
+      "step": 1480
+    },
+    {
+      "epoch": 0.03725,
+      "grad_norm": 40.5,
+      "grad_norm_var": 7.112239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5362,
+      "loss/crossentropy": 2.069592148065567,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19253854881972074,
+      "step": 1490
+    },
+    {
+      "epoch": 0.0375,
+      "grad_norm": 33.25,
+      "grad_norm_var": 6.4869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5941,
+      "loss/crossentropy": 2.0679068714380264,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20371587071567773,
+      "step": 1500
+    },
+    {
+      "epoch": 0.03775,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.5686848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.6557,
+      "loss/crossentropy": 2.1364961892366408,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20447015166282653,
+      "step": 1510
+    },
+    {
+      "epoch": 0.038,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.6738932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.622,
+      "loss/crossentropy": 2.1275009989738463,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18973923586308955,
+      "step": 1520
+    },
+    {
+      "epoch": 0.03825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.3177083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.7366,
+      "loss/crossentropy": 2.212426933646202,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21735910680145026,
+      "step": 1530
+    },
+    {
+      "epoch": 0.0385,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.7348307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6356,
+      "loss/crossentropy": 2.1074824020266534,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19886015299707652,
+      "step": 1540
+    },
+    {
+      "epoch": 0.03875,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.234830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.7567,
+      "loss/crossentropy": 2.177932971715927,
+      "loss/hidden": 3.5609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2221821215003729,
+      "step": 1550
+    },
+    {
+      "epoch": 0.039,
+      "grad_norm": 36.25,
+      "grad_norm_var": 2.8791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5766,
+      "loss/crossentropy": 2.0759232968091963,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19086231291294098,
+      "step": 1560
+    },
+    {
+      "epoch": 0.03925,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.7046223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.537,
+      "loss/crossentropy": 2.2487298890948297,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19235554365441204,
+      "step": 1570
+    },
+    {
+      "epoch": 0.0395,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.5927083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5167,
+      "loss/crossentropy": 2.2023983120918276,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20318181458860635,
+      "step": 1580
+    },
+    {
+      "epoch": 0.03975,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.4436848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5528,
+      "loss/crossentropy": 1.9812082558870316,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19151438660919667,
+      "step": 1590
+    },
+    {
+      "epoch": 0.04,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.6098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5232,
+      "loss/crossentropy": 2.0910235196352005,
+      "loss/hidden": 3.4640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20488944984972476,
+      "step": 1600
+    },
+    {
+      "epoch": 0.04025,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.5854166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7078,
+      "loss/crossentropy": 2.079045358300209,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20172932054847478,
+      "step": 1610
+    },
+    {
+      "epoch": 0.0405,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.4368798010809257e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5441,
+      "loss/crossentropy": 2.146761792898178,
+      "loss/hidden": 3.562890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20237026009708642,
+      "step": 1620
+    },
+    {
+      "epoch": 0.04075,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.4368798024404393e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.7162,
+      "loss/crossentropy": 2.1575208425521852,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914537126198411,
+      "step": 1630
+    },
+    {
+      "epoch": 0.041,
+      "grad_norm": 29.0,
+      "grad_norm_var": 37.940625,
+      "learning_rate": 0.0001,
+      "loss": 7.5673,
+      "loss/crossentropy": 2.2059059768915175,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920375470072031,
+      "step": 1640
+    },
+    {
+      "epoch": 0.04125,
+      "grad_norm": 30.75,
+      "grad_norm_var": 61.631184895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.5788,
+      "loss/crossentropy": 2.1531882882118225,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192771671153605,
+      "step": 1650
+    },
+    {
+      "epoch": 0.0415,
+      "grad_norm": 35.25,
+      "grad_norm_var": 47.353125,
+      "learning_rate": 0.0001,
+      "loss": 7.5891,
+      "loss/crossentropy": 2.1217811673879625,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19420051630586385,
+      "step": 1660
+    },
+    {
+      "epoch": 0.04175,
+      "grad_norm": 29.875,
+      "grad_norm_var": 21.580989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6361,
+      "loss/crossentropy": 2.0970900297164916,
+      "loss/hidden": 3.48828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21125762071460485,
+      "step": 1670
+    },
+    {
+      "epoch": 0.042,
+      "grad_norm": 36.5,
+      "grad_norm_var": 72.77916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7045,
+      "loss/crossentropy": 2.1662445843219755,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21249269619584082,
+      "step": 1680
+    },
+    {
+      "epoch": 0.04225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 72.42473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5754,
+      "loss/crossentropy": 2.1745404630899428,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20168567337095739,
+      "step": 1690
+    },
+    {
+      "epoch": 0.0425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 9.149739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5555,
+      "loss/crossentropy": 2.126581160724163,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1933064443990588,
+      "step": 1700
+    },
+    {
+      "epoch": 0.04275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 10.1369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5377,
+      "loss/crossentropy": 2.1726800590753554,
+      "loss/hidden": 3.441796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20911512654274703,
+      "step": 1710
+    },
+    {
+      "epoch": 0.043,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.314322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6106,
+      "loss/crossentropy": 1.9999253153800964,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18493321686983108,
+      "step": 1720
+    },
+    {
+      "epoch": 0.04325,
+      "grad_norm": 38.0,
+      "grad_norm_var": 6.673958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.7146,
+      "loss/crossentropy": 2.1180618047714233,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2076649811118841,
+      "step": 1730
+    },
+    {
+      "epoch": 0.0435,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.339322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5464,
+      "loss/crossentropy": 2.0583778262138366,
+      "loss/hidden": 3.322265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18418019600212573,
+      "step": 1740
+    },
+    {
+      "epoch": 0.04375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.749739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5528,
+      "loss/crossentropy": 2.2259044647216797,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2031107559800148,
+      "step": 1750
+    },
+    {
+      "epoch": 0.044,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.1389973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5321,
+      "loss/crossentropy": 2.124616578221321,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2149519257247448,
+      "step": 1760
+    },
+    {
+      "epoch": 0.04425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 6.442708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.63,
+      "loss/crossentropy": 2.1418310686945916,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19125681575387715,
+      "step": 1770
+    },
+    {
+      "epoch": 0.0445,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.011458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5307,
+      "loss/crossentropy": 2.144363935291767,
+      "loss/hidden": 3.440625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19334549438208343,
+      "step": 1780
+    },
+    {
+      "epoch": 0.04475,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.6559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.627,
+      "loss/crossentropy": 2.2538854971528055,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2010333575308323,
+      "step": 1790
+    },
+    {
+      "epoch": 0.045,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6471,
+      "loss/crossentropy": 2.091226762533188,
+      "loss/hidden": 3.450390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2152847982943058,
+      "step": 1800
+    },
+    {
+      "epoch": 0.04525,
+      "grad_norm": 28.875,
+      "grad_norm_var": 18.03515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5626,
+      "loss/crossentropy": 2.295862782001495,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20952691733837128,
+      "step": 1810
+    },
+    {
+      "epoch": 0.0455,
+      "grad_norm": 30.75,
+      "grad_norm_var": 18.4541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6872,
+      "loss/crossentropy": 2.0833657890558244,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2070673793554306,
+      "step": 1820
+    },
+    {
+      "epoch": 0.04575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.812239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.647,
+      "loss/crossentropy": 2.1021256439387797,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18996517434716226,
+      "step": 1830
+    },
+    {
+      "epoch": 0.046,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.7009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5203,
+      "loss/crossentropy": 2.2004275381565095,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.187726416811347,
+      "step": 1840
+    },
+    {
+      "epoch": 0.04625,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.067643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6201,
+      "loss/crossentropy": 2.1705893486738206,
+      "loss/hidden": 3.522265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21116435080766677,
+      "step": 1850
+    },
+    {
+      "epoch": 0.0465,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.030989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6235,
+      "loss/crossentropy": 2.1204057022929192,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19528388790786266,
+      "step": 1860
+    },
+    {
+      "epoch": 0.04675,
+      "grad_norm": 45.0,
+      "grad_norm_var": 15.6603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4562,
+      "loss/crossentropy": 2.154220977425575,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19214881088119,
+      "step": 1870
+    },
+    {
+      "epoch": 0.047,
+      "grad_norm": 32.5,
+      "grad_norm_var": 15.614518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6258,
+      "loss/crossentropy": 2.1000912792980673,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20185065623372794,
+      "step": 1880
+    },
+    {
+      "epoch": 0.04725,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.931184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4863,
+      "loss/crossentropy": 2.2270338363945483,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19286383930593728,
+      "step": 1890
+    },
+    {
+      "epoch": 0.0475,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.311393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6001,
+      "loss/crossentropy": 2.1998794853687285,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19646506551653148,
+      "step": 1900
+    },
+    {
+      "epoch": 0.04775,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.577018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.653,
+      "loss/crossentropy": 2.104053999483585,
+      "loss/hidden": 3.48203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201510801166296,
+      "step": 1910
+    },
+    {
+      "epoch": 0.048,
+      "grad_norm": 30.5,
+      "grad_norm_var": 17.68125,
+      "learning_rate": 0.0001,
+      "loss": 7.5273,
+      "loss/crossentropy": 2.1184800997376443,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19248204957693815,
+      "step": 1920
+    },
+    {
+      "epoch": 0.04825,
+      "grad_norm": 33.5,
+      "grad_norm_var": 1.3,
+      "learning_rate": 0.0001,
+      "loss": 7.655,
+      "loss/crossentropy": 2.1734499007463457,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1961166137829423,
+      "step": 1930
+    },
+    {
+      "epoch": 0.0485,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.5205729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6186,
+      "loss/crossentropy": 2.237542712688446,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20616979897022247,
+      "step": 1940
+    },
+    {
+      "epoch": 0.04875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.810416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5436,
+      "loss/crossentropy": 2.1811093270778654,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1928004425019026,
+      "step": 1950
+    },
+    {
+      "epoch": 0.049,
+      "grad_norm": 34.25,
+      "grad_norm_var": 1.5181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6175,
+      "loss/crossentropy": 2.045266591012478,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21615136358886958,
+      "step": 1960
+    },
+    {
+      "epoch": 0.04925,
+      "grad_norm": 33.5,
+      "grad_norm_var": 26.225455729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.7083,
+      "loss/crossentropy": 2.086462992429733,
+      "loss/hidden": 3.469921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2192224683240056,
+      "step": 1970
+    },
+    {
+      "epoch": 0.0495,
+      "grad_norm": 29.75,
+      "grad_norm_var": 8.340625,
+      "learning_rate": 0.0001,
+      "loss": 7.5032,
+      "loss/crossentropy": 2.0052025958895685,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18473264537751674,
+      "step": 1980
+    },
+    {
+      "epoch": 0.04975,
+      "grad_norm": 44.5,
+      "grad_norm_var": 4635.33515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5567,
+      "loss/crossentropy": 2.04089385792613,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19560968028381467,
+      "step": 1990
+    },
+    {
+      "epoch": 0.05,
+      "grad_norm": 51.0,
+      "grad_norm_var": 4555.746875,
+      "learning_rate": 0.0001,
+      "loss": 7.6029,
+      "loss/crossentropy": 2.2241372987627983,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2209097046405077,
+      "step": 2000
+    },
+    {
+      "epoch": 0.05025,
+      "grad_norm": 32.5,
+      "grad_norm_var": 49.11608072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.5781,
+      "loss/crossentropy": 2.2160742044448853,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19539013858884574,
+      "step": 2010
+    },
+    {
+      "epoch": 0.0505,
+      "grad_norm": 28.625,
+      "grad_norm_var": 73.70201822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.6375,
+      "loss/crossentropy": 2.094863271713257,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20763578601181507,
+      "step": 2020
+    },
+    {
+      "epoch": 0.05075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 22.494791666666668,
+      "learning_rate": 0.0001,
+      "loss": 7.6164,
+      "loss/crossentropy": 2.121490868926048,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19190637897700072,
+      "step": 2030
+    },
+    {
+      "epoch": 0.051,
+      "grad_norm": 34.5,
+      "grad_norm_var": 20.549739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5717,
+      "loss/crossentropy": 2.0418698236346247,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18076165094971658,
+      "step": 2040
+    },
+    {
+      "epoch": 0.05125,
+      "grad_norm": 29.125,
+      "grad_norm_var": 14.407291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5867,
+      "loss/crossentropy": 2.1989556729793547,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2011225748807192,
+      "step": 2050
+    },
+    {
+      "epoch": 0.0515,
+      "grad_norm": 32.75,
+      "grad_norm_var": 17.120572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6813,
+      "loss/crossentropy": 2.073545518517494,
+      "loss/hidden": 3.546484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22323863469064237,
+      "step": 2060
+    },
+    {
+      "epoch": 0.05175,
+      "grad_norm": 42.25,
+      "grad_norm_var": 76.27291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6323,
+      "loss/crossentropy": 2.1424515694379807,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19750071745365858,
+      "step": 2070
+    },
+    {
+      "epoch": 0.052,
+      "grad_norm": 30.625,
+      "grad_norm_var": 74.925,
+      "learning_rate": 0.0001,
+      "loss": 7.6198,
+      "loss/crossentropy": 2.006666135787964,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19839615989476442,
+      "step": 2080
+    },
+    {
+      "epoch": 0.05225,
+      "grad_norm": 41.5,
+      "grad_norm_var": 30.869791666666668,
+      "learning_rate": 0.0001,
+      "loss": 7.5671,
+      "loss/crossentropy": 2.2227450221776963,
+      "loss/hidden": 3.451171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20443473970517517,
+      "step": 2090
+    },
+    {
+      "epoch": 0.0525,
+      "grad_norm": 34.75,
+      "grad_norm_var": 24.609309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5928,
+      "loss/crossentropy": 2.186553081870079,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18928833175450563,
+      "step": 2100
+    },
+    {
+      "epoch": 0.05275,
+      "grad_norm": 29.125,
+      "grad_norm_var": 18.37265625,
+      "learning_rate": 0.0001,
+      "loss": 7.7389,
+      "loss/crossentropy": 2.0680640071630476,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201955908536911,
+      "step": 2110
+    },
+    {
+      "epoch": 0.053,
+      "grad_norm": 33.5,
+      "grad_norm_var": 9.276497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6148,
+      "loss/crossentropy": 2.1981600403785704,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1901057105511427,
+      "step": 2120
+    },
+    {
+      "epoch": 0.05325,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.308072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.48,
+      "loss/crossentropy": 2.2574460208415985,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19361322987824678,
+      "step": 2130
+    },
+    {
+      "epoch": 0.0535,
+      "grad_norm": 29.625,
+      "grad_norm_var": 8.5900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6644,
+      "loss/crossentropy": 2.154197073727846,
+      "loss/hidden": 3.50625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2086074635386467,
+      "step": 2140
+    },
+    {
+      "epoch": 0.05375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 8.333072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5279,
+      "loss/crossentropy": 2.1245498836040495,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19000006280839443,
+      "step": 2150
+    },
+    {
+      "epoch": 0.054,
+      "grad_norm": 31.375,
+      "grad_norm_var": 10.856184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.7098,
+      "loss/crossentropy": 2.1968549311161043,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21285793352872134,
+      "step": 2160
+    },
+    {
+      "epoch": 0.05425,
+      "grad_norm": 34.25,
+      "grad_norm_var": 17.820572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.7095,
+      "loss/crossentropy": 2.2533976465463637,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20727334953844548,
+      "step": 2170
+    },
+    {
+      "epoch": 0.0545,
+      "grad_norm": 36.0,
+      "grad_norm_var": 22.128580729166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5933,
+      "loss/crossentropy": 2.227231651544571,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19693543761968613,
+      "step": 2180
+    },
+    {
+      "epoch": 0.05475,
+      "grad_norm": 29.125,
+      "grad_norm_var": 11.466666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5348,
+      "loss/crossentropy": 2.1581582985818386,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19432583590969443,
+      "step": 2190
+    },
+    {
+      "epoch": 0.055,
+      "grad_norm": 37.25,
+      "grad_norm_var": 10.514518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6264,
+      "loss/crossentropy": 2.0908095851540565,
+      "loss/hidden": 3.39296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20568534098565577,
+      "step": 2200
+    },
+    {
+      "epoch": 0.05525,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.6494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.671,
+      "loss/crossentropy": 2.0827252097427844,
+      "loss/hidden": 3.54921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20747530292719601,
+      "step": 2210
+    },
+    {
+      "epoch": 0.0555,
+      "grad_norm": 31.875,
+      "grad_norm_var": 47.234375,
+      "learning_rate": 0.0001,
+      "loss": 7.5823,
+      "loss/crossentropy": 2.2815075665712357,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20698099359869956,
+      "step": 2220
+    },
+    {
+      "epoch": 0.05575,
+      "grad_norm": 37.5,
+      "grad_norm_var": 30.277018229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5812,
+      "loss/crossentropy": 2.2282338082790374,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19244133178144693,
+      "step": 2230
+    },
+    {
+      "epoch": 0.056,
+      "grad_norm": 31.0,
+      "grad_norm_var": 27.68515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6485,
+      "loss/crossentropy": 2.184544026851654,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1996760057285428,
+      "step": 2240
+    },
+    {
+      "epoch": 0.05625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.223372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6003,
+      "loss/crossentropy": 2.1572179198265076,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20082181300967933,
+      "step": 2250
+    },
+    {
+      "epoch": 0.0565,
+      "grad_norm": 63.5,
+      "grad_norm_var": 240.61295572916666,
+      "learning_rate": 0.0001,
+      "loss": 7.6129,
+      "loss/crossentropy": 2.1616927281022074,
+      "loss/hidden": 3.494140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21196384327486156,
+      "step": 2260
+    },
+    {
+      "epoch": 0.05675,
+      "grad_norm": 29.875,
+      "grad_norm_var": 276.98932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5683,
+      "loss/crossentropy": 2.1882350742816925,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2052389807999134,
+      "step": 2270
+    },
+    {
+      "epoch": 0.057,
+      "grad_norm": 33.5,
+      "grad_norm_var": 11.6681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6076,
+      "loss/crossentropy": 2.1808354407548904,
+      "loss/hidden": 3.540625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21027475781738758,
+      "step": 2280
+    },
+    {
+      "epoch": 0.05725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6872395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5295,
+      "loss/crossentropy": 2.1488232225179673,
+      "loss/hidden": 3.496875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2111368477344513,
+      "step": 2290
+    },
+    {
+      "epoch": 0.0575,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.0603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6875,
+      "loss/crossentropy": 2.1594634115695954,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20478933807462454,
+      "step": 2300
+    },
+    {
+      "epoch": 0.05775,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.408072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6636,
+      "loss/crossentropy": 2.0932443618774412,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21733616031706332,
+      "step": 2310
+    },
+    {
+      "epoch": 0.058,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.5416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5737,
+      "loss/crossentropy": 2.1831247925758364,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1989585768431425,
+      "step": 2320
+    },
+    {
+      "epoch": 0.05825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2393229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5009,
+      "loss/crossentropy": 2.2847615987062455,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20479252003133297,
+      "step": 2330
+    },
+    {
+      "epoch": 0.0585,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.374934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.7313,
+      "loss/crossentropy": 2.1928456306457518,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20380632691085337,
+      "step": 2340
+    },
+    {
+      "epoch": 0.05875,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.7697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5476,
+      "loss/crossentropy": 2.1438629984855653,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19849517289549112,
+      "step": 2350
+    },
+    {
+      "epoch": 0.059,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.8893229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6545,
+      "loss/crossentropy": 2.1132273241877555,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21593583207577466,
+      "step": 2360
+    },
+    {
+      "epoch": 0.05925,
+      "grad_norm": 38.0,
+      "grad_norm_var": 3.9934895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.6438,
+      "loss/crossentropy": 2.1729103833436967,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.204007020406425,
+      "step": 2370
+    },
+    {
+      "epoch": 0.0595,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.180989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5261,
+      "loss/crossentropy": 2.1972236961126326,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21200564429163932,
+      "step": 2380
+    },
+    {
+      "epoch": 0.05975,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.9854166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6233,
+      "loss/crossentropy": 2.130857673287392,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2043182048946619,
+      "step": 2390
+    },
+    {
+      "epoch": 0.06,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.552018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5626,
+      "loss/crossentropy": 2.1780240714550017,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19501971434801818,
+      "step": 2400
+    },
+    {
+      "epoch": 0.06025,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.4978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5495,
+      "loss/crossentropy": 2.075847564637661,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19856880297884344,
+      "step": 2410
+    },
+    {
+      "epoch": 0.0605,
+      "grad_norm": 30.375,
+      "grad_norm_var": 8.763541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5161,
+      "loss/crossentropy": 1.960635770857334,
+      "loss/hidden": 3.521484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2014876109547913,
+      "step": 2420
+    },
+    {
+      "epoch": 0.06075,
+      "grad_norm": 30.375,
+      "grad_norm_var": 5.85390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6453,
+      "loss/crossentropy": 2.235329329967499,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22043778821825982,
+      "step": 2430
+    },
+    {
+      "epoch": 0.061,
+      "grad_norm": 35.0,
+      "grad_norm_var": 2.928465629875169e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5153,
+      "loss/crossentropy": 2.0534446865320204,
+      "loss/hidden": 3.469140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20287865065038205,
+      "step": 2440
+    },
+    {
+      "epoch": 0.06125,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.928465629810996e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5862,
+      "loss/crossentropy": 2.1208725392818453,
+      "loss/hidden": 3.534765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22640016246587039,
+      "step": 2450
+    },
+    {
+      "epoch": 0.0615,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.5614583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5751,
+      "loss/crossentropy": 2.206133508682251,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19071156904101372,
+      "step": 2460
+    },
+    {
+      "epoch": 0.06175,
+      "grad_norm": 28.625,
+      "grad_norm_var": 7.667643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5604,
+      "loss/crossentropy": 2.227986590564251,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19392532519996167,
+      "step": 2470
+    },
+    {
+      "epoch": 0.062,
+      "grad_norm": 34.0,
+      "grad_norm_var": 5.81015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6077,
+      "loss/crossentropy": 2.1948474526405333,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2092638686299324,
+      "step": 2480
+    },
+    {
+      "epoch": 0.06225,
+      "grad_norm": 34.5,
+      "grad_norm_var": 7.656184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4338,
+      "loss/crossentropy": 1.935844713449478,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184383431263268,
+      "step": 2490
+    },
+    {
+      "epoch": 0.0625,
+      "grad_norm": 42.75,
+      "grad_norm_var": 17.27890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5645,
+      "loss/crossentropy": 2.2957077413797378,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2223764518275857,
+      "step": 2500
+    },
+    {
+      "epoch": 0.06275,
+      "grad_norm": 30.125,
+      "grad_norm_var": 31.021809895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.5539,
+      "loss/crossentropy": 2.2163919866085053,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2129704337567091,
+      "step": 2510
+    },
+    {
+      "epoch": 0.063,
+      "grad_norm": 33.25,
+      "grad_norm_var": 40.36764322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.5664,
+      "loss/crossentropy": 2.08260739967227,
+      "loss/hidden": 3.4765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20597089193761348,
+      "step": 2520
+    },
+    {
+      "epoch": 0.06325,
+      "grad_norm": 35.5,
+      "grad_norm_var": 34.153125,
+      "learning_rate": 0.0001,
+      "loss": 7.559,
+      "loss/crossentropy": 2.0747475802898405,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1930603832937777,
+      "step": 2530
+    },
+    {
+      "epoch": 0.0635,
+      "grad_norm": 31.25,
+      "grad_norm_var": 12.302018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.609,
+      "loss/crossentropy": 2.1542711734771727,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19414089974015952,
+      "step": 2540
+    },
+    {
+      "epoch": 0.06375,
+      "grad_norm": 29.125,
+      "grad_norm_var": 34.333072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5103,
+      "loss/crossentropy": 2.1697633042931557,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18199986461549997,
+      "step": 2550
+    },
+    {
+      "epoch": 0.064,
+      "grad_norm": 30.75,
+      "grad_norm_var": 21.8322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6335,
+      "loss/crossentropy": 2.20294488966465,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20988359525799752,
+      "step": 2560
+    },
+    {
+      "epoch": 0.06425,
+      "grad_norm": 36.5,
+      "grad_norm_var": 21.339322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.665,
+      "loss/crossentropy": 2.1367219746112824,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19768773801624775,
+      "step": 2570
+    },
+    {
+      "epoch": 0.0645,
+      "grad_norm": 42.5,
+      "grad_norm_var": 30.690625,
+      "learning_rate": 0.0001,
+      "loss": 7.6476,
+      "loss/crossentropy": 2.2177338540554046,
+      "loss/hidden": 3.605078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22041462864726782,
+      "step": 2580
+    },
+    {
+      "epoch": 0.06475,
+      "grad_norm": 34.0,
+      "grad_norm_var": 31.164583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6158,
+      "loss/crossentropy": 2.2423059731721877,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21196699403226377,
+      "step": 2590
+    },
+    {
+      "epoch": 0.065,
+      "grad_norm": 34.25,
+      "grad_norm_var": 23.593684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5107,
+      "loss/crossentropy": 2.241020438075066,
+      "loss/hidden": 3.25859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18191274981945754,
+      "step": 2600
+    },
+    {
+      "epoch": 0.06525,
+      "grad_norm": 28.75,
+      "grad_norm_var": 30.975455729166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5492,
+      "loss/crossentropy": 2.052956056594849,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2021762602031231,
+      "step": 2610
+    },
+    {
+      "epoch": 0.0655,
+      "grad_norm": 32.5,
+      "grad_norm_var": 31.032291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.503,
+      "loss/crossentropy": 2.17630957365036,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19023955501616002,
+      "step": 2620
+    },
+    {
+      "epoch": 0.06575,
+      "grad_norm": 34.5,
+      "grad_norm_var": 6.083072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5728,
+      "loss/crossentropy": 2.178831994533539,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1946109678596258,
+      "step": 2630
+    },
+    {
+      "epoch": 0.066,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.01640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5928,
+      "loss/crossentropy": 2.0746659457683565,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19469854161143302,
+      "step": 2640
+    },
+    {
+      "epoch": 0.06625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 21.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6008,
+      "loss/crossentropy": 2.202178010344505,
+      "loss/hidden": 3.44609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2072868559509516,
+      "step": 2650
+    },
+    {
+      "epoch": 0.0665,
+      "grad_norm": 31.375,
+      "grad_norm_var": 24.743489583333332,
+      "learning_rate": 0.0001,
+      "loss": 7.5532,
+      "loss/crossentropy": 2.0579031884670256,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18248077742755414,
+      "step": 2660
+    },
+    {
+      "epoch": 0.06675,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.5181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6173,
+      "loss/crossentropy": 2.0323975652456285,
+      "loss/hidden": 3.47265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20279558952897786,
+      "step": 2670
+    },
+    {
+      "epoch": 0.067,
+      "grad_norm": 30.125,
+      "grad_norm_var": 9.7931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6202,
+      "loss/crossentropy": 2.170803511887789,
+      "loss/hidden": 3.448828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21887149531394243,
+      "step": 2680
+    },
+    {
+      "epoch": 0.06725,
+      "grad_norm": 28.75,
+      "grad_norm_var": 6.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6042,
+      "loss/crossentropy": 2.129136848449707,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21714741103351115,
+      "step": 2690
+    },
+    {
+      "epoch": 0.0675,
+      "grad_norm": 28.5,
+      "grad_norm_var": 5.76640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5307,
+      "loss/crossentropy": 2.2333458453416823,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19759367052465676,
+      "step": 2700
+    },
+    {
+      "epoch": 0.06775,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.925,
+      "learning_rate": 0.0001,
+      "loss": 7.5684,
+      "loss/crossentropy": 2.1930347234010696,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19194095116108656,
+      "step": 2710
+    },
+    {
+      "epoch": 0.068,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.0291666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6177,
+      "loss/crossentropy": 2.1858380883932114,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18830202352255582,
+      "step": 2720
+    },
+    {
+      "epoch": 0.06825,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.19765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6231,
+      "loss/crossentropy": 2.1380992412567137,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19919480197131634,
+      "step": 2730
+    },
+    {
+      "epoch": 0.0685,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.5872395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.7166,
+      "loss/crossentropy": 2.299161267280579,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21277474984526634,
+      "step": 2740
+    },
+    {
+      "epoch": 0.06875,
+      "grad_norm": 28.25,
+      "grad_norm_var": 10.6525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6385,
+      "loss/crossentropy": 2.0595856219530106,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20196862574666738,
+      "step": 2750
+    },
+    {
+      "epoch": 0.069,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.1885416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5306,
+      "loss/crossentropy": 2.1047334015369414,
+      "loss/hidden": 3.527734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20494798701256514,
+      "step": 2760
+    },
+    {
+      "epoch": 0.06925,
+      "grad_norm": 35.5,
+      "grad_norm_var": 2.5136418842603423e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5758,
+      "loss/crossentropy": 1.9750292956829072,
+      "loss/hidden": 3.49453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1999529790133238,
+      "step": 2770
+    },
+    {
+      "epoch": 0.0695,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.513641880435452e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5901,
+      "loss/crossentropy": 2.1417267471551895,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20058641098439695,
+      "step": 2780
+    },
+    {
+      "epoch": 0.06975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 155.97024739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.5485,
+      "loss/crossentropy": 2.1048891723155974,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19070767909288405,
+      "step": 2790
+    },
+    {
+      "epoch": 0.07,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.499739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5996,
+      "loss/crossentropy": 2.176823168247938,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20276176873594523,
+      "step": 2800
+    },
+    {
+      "epoch": 0.07025,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.684309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4426,
+      "loss/crossentropy": 2.041793665289879,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17560106106102466,
+      "step": 2810
+    },
+    {
+      "epoch": 0.0705,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.005989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5811,
+      "loss/crossentropy": 2.1094699330627917,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19798500649631023,
+      "step": 2820
+    },
+    {
+      "epoch": 0.07075,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.320572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5475,
+      "loss/crossentropy": 2.1311514347791674,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21825175136327743,
+      "step": 2830
+    },
+    {
+      "epoch": 0.071,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.7549465266724797e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5823,
+      "loss/crossentropy": 2.120433983206749,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20115374326705932,
+      "step": 2840
+    },
+    {
+      "epoch": 0.07125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.7549465268743306e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5444,
+      "loss/crossentropy": 2.1518867701292037,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19730570819228888,
+      "step": 2850
+    },
+    {
+      "epoch": 0.0715,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.317122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6362,
+      "loss/crossentropy": 2.137280356884003,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20695031639188527,
+      "step": 2860
+    },
+    {
+      "epoch": 0.07175,
+      "grad_norm": 34.0,
+      "grad_norm_var": 157.31015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6643,
+      "loss/crossentropy": 2.067914080619812,
+      "loss/hidden": 3.5015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22938031535595654,
+      "step": 2870
+    },
+    {
+      "epoch": 0.072,
+      "grad_norm": 32.5,
+      "grad_norm_var": 151.77057291666668,
+      "learning_rate": 0.0001,
+      "loss": 7.5762,
+      "loss/crossentropy": 2.0962916165590286,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19928023852407933,
+      "step": 2880
+    },
+    {
+      "epoch": 0.07225,
+      "grad_norm": 35.75,
+      "grad_norm_var": 104.17962239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.7066,
+      "loss/crossentropy": 2.1184468276798727,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1974171632900834,
+      "step": 2890
+    },
+    {
+      "epoch": 0.0725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 116.02493489583334,
+      "learning_rate": 0.0001,
+      "loss": 7.6075,
+      "loss/crossentropy": 2.1381339877843857,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19655965138226747,
+      "step": 2900
+    },
+    {
+      "epoch": 0.07275,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9551432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5247,
+      "loss/crossentropy": 2.141963595151901,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19582534320652484,
+      "step": 2910
+    },
+    {
+      "epoch": 0.073,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.3014973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5588,
+      "loss/crossentropy": 2.077046422660351,
+      "loss/hidden": 3.465234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20551967658102513,
+      "step": 2920
+    },
+    {
+      "epoch": 0.07325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 8.734375,
+      "learning_rate": 0.0001,
+      "loss": 7.6195,
+      "loss/crossentropy": 2.0110410653054713,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1805876674130559,
+      "step": 2930
+    },
+    {
+      "epoch": 0.0735,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4838.416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6623,
+      "loss/crossentropy": 2.1130379527807235,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2008265011012554,
+      "step": 2940
+    },
+    {
+      "epoch": 0.07375,
+      "grad_norm": 39.0,
+      "grad_norm_var": 57.90807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.59,
+      "loss/crossentropy": 2.1667733818292616,
+      "loss/hidden": 3.3109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19806477334350348,
+      "step": 2950
+    },
+    {
+      "epoch": 0.074,
+      "grad_norm": 31.875,
+      "grad_norm_var": 31.228059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5131,
+      "loss/crossentropy": 2.2757086992263793,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197940625064075,
+      "step": 2960
+    },
+    {
+      "epoch": 0.07425,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.996809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5841,
+      "loss/crossentropy": 2.143115535378456,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20177703239023687,
+      "step": 2970
+    },
+    {
+      "epoch": 0.0745,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.1546223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5363,
+      "loss/crossentropy": 2.2850147604942324,
+      "loss/hidden": 3.283984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19140432458370923,
+      "step": 2980
+    },
+    {
+      "epoch": 0.07475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.310416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5339,
+      "loss/crossentropy": 2.104042625427246,
+      "loss/hidden": 3.490625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.211691821180284,
+      "step": 2990
+    },
+    {
+      "epoch": 0.075,
+      "grad_norm": 32.5,
+      "grad_norm_var": 8.506705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5448,
+      "loss/crossentropy": 2.2270253866910936,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20328052509576083,
+      "step": 3000
+    },
+    {
+      "epoch": 0.07525,
+      "grad_norm": 46.75,
+      "grad_norm_var": 17.10390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6114,
+      "loss/crossentropy": 2.2623429775238035,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19479246698319913,
+      "step": 3010
+    },
+    {
+      "epoch": 0.0755,
+      "grad_norm": 31.75,
+      "grad_norm_var": 18.981184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5739,
+      "loss/crossentropy": 2.1012531995773314,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20767469964921476,
+      "step": 3020
+    },
+    {
+      "epoch": 0.07575,
+      "grad_norm": 45.25,
+      "grad_norm_var": 15.55,
+      "learning_rate": 0.0001,
+      "loss": 7.6153,
+      "loss/crossentropy": 2.090057593584061,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1894347405061126,
+      "step": 3030
+    },
+    {
+      "epoch": 0.076,
+      "grad_norm": 32.0,
+      "grad_norm_var": 29.878125,
+      "learning_rate": 0.0001,
+      "loss": 7.6585,
+      "loss/crossentropy": 2.200800988078117,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19270651414990425,
+      "step": 3040
+    },
+    {
+      "epoch": 0.07625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 10.167708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5696,
+      "loss/crossentropy": 2.0458697080612183,
+      "loss/hidden": 3.53046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20314501021057368,
+      "step": 3050
+    },
+    {
+      "epoch": 0.0765,
+      "grad_norm": 30.375,
+      "grad_norm_var": 10.0181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5747,
+      "loss/crossentropy": 2.2118399769067763,
+      "loss/hidden": 3.482421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22158339023590087,
+      "step": 3060
+    },
+    {
+      "epoch": 0.07675,
+      "grad_norm": 32.25,
+      "grad_norm_var": 0.9650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5631,
+      "loss/crossentropy": 2.1812553733587263,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.217013025470078,
+      "step": 3070
+    },
+    {
+      "epoch": 0.077,
+      "grad_norm": 32.75,
+      "grad_norm_var": 0.9171223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.578,
+      "loss/crossentropy": 2.181536224484444,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19952490609139203,
+      "step": 3080
+    },
+    {
+      "epoch": 0.07725,
+      "grad_norm": 31.5,
+      "grad_norm_var": 19.015559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4878,
+      "loss/crossentropy": 2.122265163064003,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19191155321896075,
+      "step": 3090
+    },
+    {
+      "epoch": 0.0775,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.199739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.7471,
+      "loss/crossentropy": 2.210584083199501,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21269479542970657,
+      "step": 3100
+    },
+    {
+      "epoch": 0.07775,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.490625,
+      "learning_rate": 0.0001,
+      "loss": 7.6586,
+      "loss/crossentropy": 2.090894425660372,
+      "loss/hidden": 3.46328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19853799045085907,
+      "step": 3110
+    },
+    {
+      "epoch": 0.078,
+      "grad_norm": 31.375,
+      "grad_norm_var": 27.370833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5818,
+      "loss/crossentropy": 2.044330509006977,
+      "loss/hidden": 3.50078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20974230151623488,
+      "step": 3120
+    },
+    {
+      "epoch": 0.07825,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.688997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.654,
+      "loss/crossentropy": 2.1380004197359086,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20455153118818997,
+      "step": 3130
+    },
+    {
+      "epoch": 0.0785,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.715625,
+      "learning_rate": 0.0001,
+      "loss": 7.6551,
+      "loss/crossentropy": 2.1576705113053323,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20728315506130457,
+      "step": 3140
+    },
+    {
+      "epoch": 0.07875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.2864583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5478,
+      "loss/crossentropy": 2.233546493947506,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20845879297703504,
+      "step": 3150
+    },
+    {
+      "epoch": 0.079,
+      "grad_norm": 34.25,
+      "grad_norm_var": 15.4728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6264,
+      "loss/crossentropy": 2.0873878210783006,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19576258175075054,
+      "step": 3160
+    },
+    {
+      "epoch": 0.07925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.12265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5192,
+      "loss/crossentropy": 2.132347696274519,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1958464809693396,
+      "step": 3170
+    },
+    {
+      "epoch": 0.0795,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.370247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5169,
+      "loss/crossentropy": 2.2280534476041796,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19412651136517525,
+      "step": 3180
+    },
+    {
+      "epoch": 0.07975,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.989518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6576,
+      "loss/crossentropy": 2.138143754005432,
+      "loss/hidden": 3.58828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23302078600972892,
+      "step": 3190
+    },
+    {
+      "epoch": 0.08,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.858072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5707,
+      "loss/crossentropy": 2.1512865126132965,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957532402127981,
+      "step": 3200
+    },
+    {
+      "epoch": 0.08025,
+      "grad_norm": 39.0,
+      "grad_norm_var": 34.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6814,
+      "loss/crossentropy": 2.113873428106308,
+      "loss/hidden": 3.52890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22486987188458443,
+      "step": 3210
+    },
+    {
+      "epoch": 0.0805,
+      "grad_norm": 30.625,
+      "grad_norm_var": 33.244205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6542,
+      "loss/crossentropy": 2.153007471561432,
+      "loss/hidden": 3.5078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20893471594899893,
+      "step": 3220
+    },
+    {
+      "epoch": 0.08075,
+      "grad_norm": 31.25,
+      "grad_norm_var": 619.4145182291667,
+      "learning_rate": 0.0001,
+      "loss": 7.5091,
+      "loss/crossentropy": 2.169908273220062,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20535510070621968,
+      "step": 3230
+    },
+    {
+      "epoch": 0.081,
+      "grad_norm": 30.0,
+      "grad_norm_var": 599.9247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5568,
+      "loss/crossentropy": 2.048283484578133,
+      "loss/hidden": 3.593359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21250182073563337,
+      "step": 3240
+    },
+    {
+      "epoch": 0.08125,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.982291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5569,
+      "loss/crossentropy": 2.0828478574752807,
+      "loss/hidden": 3.496484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20137296654284,
+      "step": 3250
+    },
+    {
+      "epoch": 0.0815,
+      "grad_norm": 28.25,
+      "grad_norm_var": 17.9900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5869,
+      "loss/crossentropy": 2.28429861664772,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21088685300201176,
+      "step": 3260
+    },
+    {
+      "epoch": 0.08175,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.5587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6037,
+      "loss/crossentropy": 2.1336950808763504,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20057348478585482,
+      "step": 3270
+    },
+    {
+      "epoch": 0.082,
+      "grad_norm": 33.75,
+      "grad_norm_var": 34.57076822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.6129,
+      "loss/crossentropy": 2.0662791609764097,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22612145096063613,
+      "step": 3280
+    },
+    {
+      "epoch": 0.08225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 18.8181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5469,
+      "loss/crossentropy": 2.069541847705841,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1844556663185358,
+      "step": 3290
+    },
+    {
+      "epoch": 0.0825,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.0712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6025,
+      "loss/crossentropy": 2.2499852567911147,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21941267363727093,
+      "step": 3300
+    },
+    {
+      "epoch": 0.08275,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.29140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6334,
+      "loss/crossentropy": 2.1277933359146117,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19795072823762894,
+      "step": 3310
+    },
+    {
+      "epoch": 0.083,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.41015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6326,
+      "loss/crossentropy": 2.214772176742554,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20066177062690257,
+      "step": 3320
+    },
+    {
+      "epoch": 0.08325,
+      "grad_norm": 29.875,
+      "grad_norm_var": 18.768489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.569,
+      "loss/crossentropy": 2.093920087814331,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2021244278177619,
+      "step": 3330
+    },
+    {
+      "epoch": 0.0835,
+      "grad_norm": 30.25,
+      "grad_norm_var": 18.695833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5041,
+      "loss/crossentropy": 2.0556397035717966,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17018966227769852,
+      "step": 3340
+    },
+    {
+      "epoch": 0.08375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.854622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5837,
+      "loss/crossentropy": 2.0227382972836496,
+      "loss/hidden": 3.4890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21123163159936667,
+      "step": 3350
+    },
+    {
+      "epoch": 0.084,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.894791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5374,
+      "loss/crossentropy": 2.146658593416214,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18999559991061687,
+      "step": 3360
+    },
+    {
+      "epoch": 0.08425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.5759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6197,
+      "loss/crossentropy": 1.9798088841140271,
+      "loss/hidden": 3.569921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20031734639778734,
+      "step": 3370
+    },
+    {
+      "epoch": 0.0845,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6436848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5738,
+      "loss/crossentropy": 2.1442878276109694,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20664554908871652,
+      "step": 3380
+    },
+    {
+      "epoch": 0.08475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.9875,
+      "learning_rate": 0.0001,
+      "loss": 7.4965,
+      "loss/crossentropy": 2.150037130713463,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19511928483843805,
+      "step": 3390
+    },
+    {
+      "epoch": 0.085,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.2122395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5151,
+      "loss/crossentropy": 2.0501646161079408,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19615819547325372,
+      "step": 3400
+    },
+    {
+      "epoch": 0.08525,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.720833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6432,
+      "loss/crossentropy": 2.2328430742025374,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19816372059285642,
+      "step": 3410
+    },
+    {
+      "epoch": 0.0855,
+      "grad_norm": 31.375,
+      "grad_norm_var": 7.519791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.665,
+      "loss/crossentropy": 2.1567111521959306,
+      "loss/hidden": 3.500390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19839788805693387,
+      "step": 3420
+    },
+    {
+      "epoch": 0.08575,
+      "grad_norm": 40.25,
+      "grad_norm_var": 7.808072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5919,
+      "loss/crossentropy": 2.2465982705354692,
+      "loss/hidden": 3.46328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21547308284789324,
+      "step": 3430
+    },
+    {
+      "epoch": 0.086,
+      "grad_norm": 33.0,
+      "grad_norm_var": 8.09765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6239,
+      "loss/crossentropy": 2.1680932968854902,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21484404131770135,
+      "step": 3440
+    },
+    {
+      "epoch": 0.08625,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.070572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4802,
+      "loss/crossentropy": 2.050510385632515,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18613745234906673,
+      "step": 3450
+    },
+    {
+      "epoch": 0.0865,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.2143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5311,
+      "loss/crossentropy": 2.0699805982410906,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1896037317812443,
+      "step": 3460
+    },
+    {
+      "epoch": 0.08675,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.651822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.596,
+      "loss/crossentropy": 2.0576356425881386,
+      "loss/hidden": 3.522265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20867941789329053,
+      "step": 3470
+    },
+    {
+      "epoch": 0.087,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.998372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4738,
+      "loss/crossentropy": 2.2489717990159988,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18595699593424797,
+      "step": 3480
+    },
+    {
+      "epoch": 0.08725,
+      "grad_norm": 28.5,
+      "grad_norm_var": 5.2056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5246,
+      "loss/crossentropy": 2.178911143541336,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.204429741948843,
+      "step": 3490
+    },
+    {
+      "epoch": 0.0875,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.8134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4901,
+      "loss/crossentropy": 2.1816289871931076,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20369651466608046,
+      "step": 3500
+    },
+    {
+      "epoch": 0.08775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 59.895572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6056,
+      "loss/crossentropy": 2.124045217037201,
+      "loss/hidden": 3.562890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2361205333843827,
+      "step": 3510
+    },
+    {
+      "epoch": 0.088,
+      "grad_norm": 31.375,
+      "grad_norm_var": 58.15358072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.6185,
+      "loss/crossentropy": 2.0576027542352677,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19163726801052688,
+      "step": 3520
+    },
+    {
+      "epoch": 0.08825,
+      "grad_norm": 30.875,
+      "grad_norm_var": 19.622330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.608,
+      "loss/crossentropy": 2.2166375398635862,
+      "loss/hidden": 3.430078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20556784830987454,
+      "step": 3530
+    },
+    {
+      "epoch": 0.0885,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.76875,
+      "learning_rate": 0.0001,
+      "loss": 7.5655,
+      "loss/crossentropy": 2.215288892388344,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21560241151601076,
+      "step": 3540
+    },
+    {
+      "epoch": 0.08875,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.316666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4909,
+      "loss/crossentropy": 2.0867604553699493,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1992955395951867,
+      "step": 3550
+    },
+    {
+      "epoch": 0.089,
+      "grad_norm": 33.25,
+      "grad_norm_var": 286.81555989583336,
+      "learning_rate": 0.0001,
+      "loss": 7.716,
+      "loss/crossentropy": 2.2209325939416886,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19783683270215988,
+      "step": 3560
+    },
+    {
+      "epoch": 0.08925,
+      "grad_norm": 34.75,
+      "grad_norm_var": 286.8197916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5905,
+      "loss/crossentropy": 2.098256954550743,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184653827175498,
+      "step": 3570
+    },
+    {
+      "epoch": 0.0895,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.868489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.625,
+      "loss/crossentropy": 2.0555992782115937,
+      "loss/hidden": 3.45546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19005871675908564,
+      "step": 3580
+    },
+    {
+      "epoch": 0.08975,
+      "grad_norm": 37.0,
+      "grad_norm_var": 7.585416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.542,
+      "loss/crossentropy": 2.0350914053618907,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192917075753212,
+      "step": 3590
+    },
+    {
+      "epoch": 0.09,
+      "grad_norm": 34.25,
+      "grad_norm_var": 26.8322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6028,
+      "loss/crossentropy": 2.249291920661926,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20490463990718127,
+      "step": 3600
+    },
+    {
+      "epoch": 0.09025,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.9457682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5738,
+      "loss/crossentropy": 2.201851597428322,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20158183835446836,
+      "step": 3610
+    },
+    {
+      "epoch": 0.0905,
+      "grad_norm": 30.125,
+      "grad_norm_var": 14.446809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5562,
+      "loss/crossentropy": 2.188534340262413,
+      "loss/hidden": 3.491796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22296689171344042,
+      "step": 3620
+    },
+    {
+      "epoch": 0.09075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 15.792643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5509,
+      "loss/crossentropy": 2.1134337186813354,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19732800796627997,
+      "step": 3630
+    },
+    {
+      "epoch": 0.091,
+      "grad_norm": 29.875,
+      "grad_norm_var": 14.366080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6349,
+      "loss/crossentropy": 2.0555363953113557,
+      "loss/hidden": 3.4328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19681458938866853,
+      "step": 3640
+    },
+    {
+      "epoch": 0.09125,
+      "grad_norm": 33.25,
+      "grad_norm_var": 13.740559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6513,
+      "loss/crossentropy": 2.2402344048023224,
+      "loss/hidden": 3.46171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20878477580845356,
+      "step": 3650
+    },
+    {
+      "epoch": 0.0915,
+      "grad_norm": 32.5,
+      "grad_norm_var": 30.0150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5723,
+      "loss/crossentropy": 2.14640394449234,
+      "loss/hidden": 3.453515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20020358953624964,
+      "step": 3660
+    },
+    {
+      "epoch": 0.09175,
+      "grad_norm": 36.0,
+      "grad_norm_var": 49.2619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5782,
+      "loss/crossentropy": 2.2063133299350737,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19907979741692544,
+      "step": 3670
+    },
+    {
+      "epoch": 0.092,
+      "grad_norm": 32.0,
+      "grad_norm_var": 15.4400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5501,
+      "loss/crossentropy": 2.1933626160025597,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19394716806709766,
+      "step": 3680
+    },
+    {
+      "epoch": 0.09225,
+      "grad_norm": 28.375,
+      "grad_norm_var": 13.159375,
+      "learning_rate": 0.0001,
+      "loss": 7.5258,
+      "loss/crossentropy": 2.0778674989938737,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19343881569802762,
+      "step": 3690
+    },
+    {
+      "epoch": 0.0925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 15.096809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4597,
+      "loss/crossentropy": 2.0565507017076015,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19122497290372847,
+      "step": 3700
+    },
+    {
+      "epoch": 0.09275,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.729622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5714,
+      "loss/crossentropy": 2.082234078645706,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2102882768958807,
+      "step": 3710
+    },
+    {
+      "epoch": 0.093,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.6254557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6155,
+      "loss/crossentropy": 2.2312920093536377,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18575988691300155,
+      "step": 3720
+    },
+    {
+      "epoch": 0.09325,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.6181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5344,
+      "loss/crossentropy": 2.2397233605384828,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20670964010059834,
+      "step": 3730
+    },
+    {
+      "epoch": 0.0935,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.3124348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5839,
+      "loss/crossentropy": 2.2065307170152666,
+      "loss/hidden": 3.348046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20138480551540852,
+      "step": 3740
+    },
+    {
+      "epoch": 0.09375,
+      "grad_norm": 36.25,
+      "grad_norm_var": 9.07265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6468,
+      "loss/crossentropy": 2.192644628882408,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20751077253371478,
+      "step": 3750
+    },
+    {
+      "epoch": 0.094,
+      "grad_norm": 32.25,
+      "grad_norm_var": 18.11015625,
+      "learning_rate": 0.0001,
+      "loss": 7.731,
+      "loss/crossentropy": 2.205154886841774,
+      "loss/hidden": 3.38984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20712865255773066,
+      "step": 3760
+    },
+    {
+      "epoch": 0.09425,
+      "grad_norm": 30.75,
+      "grad_norm_var": 30.259309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4267,
+      "loss/crossentropy": 2.0174816213548183,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1771852731704712,
+      "step": 3770
+    },
+    {
+      "epoch": 0.0945,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.687239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5078,
+      "loss/crossentropy": 2.142752841114998,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2145843595266342,
+      "step": 3780
+    },
+    {
+      "epoch": 0.09475,
+      "grad_norm": 41.5,
+      "grad_norm_var": 28.512239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5586,
+      "loss/crossentropy": 2.2184203058481216,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2169014524668455,
+      "step": 3790
+    },
+    {
+      "epoch": 0.095,
+      "grad_norm": 32.5,
+      "grad_norm_var": 29.637955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5796,
+      "loss/crossentropy": 2.0424424298107624,
+      "loss/hidden": 3.440234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19471338465809823,
+      "step": 3800
+    },
+    {
+      "epoch": 0.09525,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.846809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4697,
+      "loss/crossentropy": 2.2178969264030455,
+      "loss/hidden": 3.215625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18599100317806005,
+      "step": 3810
+    },
+    {
+      "epoch": 0.0955,
+      "grad_norm": 35.75,
+      "grad_norm_var": 5.571809895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5899,
+      "loss/crossentropy": 2.147325333952904,
+      "loss/hidden": 3.373046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19797458127141,
+      "step": 3820
+    },
+    {
+      "epoch": 0.09575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 19.242708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4524,
+      "loss/crossentropy": 2.222689136862755,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20618323031812907,
+      "step": 3830
+    },
+    {
+      "epoch": 0.096,
+      "grad_norm": 31.125,
+      "grad_norm_var": 6.659309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5371,
+      "loss/crossentropy": 2.045217160880566,
+      "loss/hidden": 3.50703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21105091590434313,
+      "step": 3840
+    },
+    {
+      "epoch": 0.09625,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.6624348958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4832,
+      "loss/crossentropy": 2.1246922612190247,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19520843997597695,
+      "step": 3850
+    },
+    {
+      "epoch": 0.0965,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.840625,
+      "learning_rate": 0.0001,
+      "loss": 7.5247,
+      "loss/crossentropy": 2.1101455599069596,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19655006285756826,
+      "step": 3860
+    },
+    {
+      "epoch": 0.09675,
+      "grad_norm": 31.375,
+      "grad_norm_var": 4.0759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.7123,
+      "loss/crossentropy": 2.1925098091363906,
+      "loss/hidden": 3.57265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21303071565926074,
+      "step": 3870
+    },
+    {
+      "epoch": 0.097,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.940559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6194,
+      "loss/crossentropy": 2.0254321210086346,
+      "loss/hidden": 3.504296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21549067068845035,
+      "step": 3880
+    },
+    {
+      "epoch": 0.09725,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.6895833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5341,
+      "loss/crossentropy": 2.18448192179203,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18761022239923478,
+      "step": 3890
+    },
+    {
+      "epoch": 0.0975,
+      "grad_norm": 32.5,
+      "grad_norm_var": 14.326822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6378,
+      "loss/crossentropy": 2.0941652059555054,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20138136427849532,
+      "step": 3900
+    },
+    {
+      "epoch": 0.09775,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.1186848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6588,
+      "loss/crossentropy": 2.09155390933156,
+      "loss/hidden": 3.487109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2071656842716038,
+      "step": 3910
+    },
+    {
+      "epoch": 0.098,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.729622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6663,
+      "loss/crossentropy": 2.1858886659145353,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20006632767617702,
+      "step": 3920
+    },
+    {
+      "epoch": 0.09825,
+      "grad_norm": 40.25,
+      "grad_norm_var": 15.908072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5567,
+      "loss/crossentropy": 2.2354005187749864,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20157534964382648,
+      "step": 3930
+    },
+    {
+      "epoch": 0.0985,
+      "grad_norm": 31.625,
+      "grad_norm_var": 17.745768229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5189,
+      "loss/crossentropy": 2.015377716720104,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868499366566539,
+      "step": 3940
+    },
+    {
+      "epoch": 0.09875,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.66640625,
+      "learning_rate": 0.0001,
+      "loss": 7.533,
+      "loss/crossentropy": 2.0591939449310304,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20258171651512386,
+      "step": 3950
+    },
+    {
+      "epoch": 0.099,
+      "grad_norm": 41.25,
+      "grad_norm_var": 3.3442041663272433e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4832,
+      "loss/crossentropy": 2.1117863088846205,
+      "loss/hidden": 3.559375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2428264247253537,
+      "step": 3960
+    },
+    {
+      "epoch": 0.09925,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.3442041639803904e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6253,
+      "loss/crossentropy": 2.1176558315753935,
+      "loss/hidden": 3.55546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22726768516004087,
+      "step": 3970
+    },
+    {
+      "epoch": 0.0995,
+      "grad_norm": 31.25,
+      "grad_norm_var": 19.5384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4413,
+      "loss/crossentropy": 2.088257111608982,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19193989606574177,
+      "step": 3980
+    },
+    {
+      "epoch": 0.09975,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.28125,
+      "learning_rate": 0.0001,
+      "loss": 7.6017,
+      "loss/crossentropy": 2.134918417036533,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19602251183241606,
+      "step": 3990
+    },
+    {
+      "epoch": 0.1,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.172916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5138,
+      "loss/crossentropy": 2.1410045489668845,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20588791109621524,
+      "step": 4000
+    },
+    {
+      "epoch": 0.10025,
+      "grad_norm": 30.0,
+      "grad_norm_var": 11.4306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5549,
+      "loss/crossentropy": 2.241489386558533,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20121449399739505,
+      "step": 4010
+    },
+    {
+      "epoch": 0.1005,
+      "grad_norm": 33.75,
+      "grad_norm_var": 5.220247395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6037,
+      "loss/crossentropy": 2.177129751443863,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.195396139472723,
+      "step": 4020
+    },
+    {
+      "epoch": 0.10075,
+      "grad_norm": 31.375,
+      "grad_norm_var": 25.279622395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.5495,
+      "loss/crossentropy": 2.1074128076434135,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18998505976051092,
+      "step": 4030
+    },
+    {
+      "epoch": 0.101,
+      "grad_norm": 30.375,
+      "grad_norm_var": 12.564518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4628,
+      "loss/crossentropy": 2.031256873905659,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19107761420309544,
+      "step": 4040
+    },
+    {
+      "epoch": 0.10125,
+      "grad_norm": 29.25,
+      "grad_norm_var": 8.242643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5906,
+      "loss/crossentropy": 2.2593255966901777,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20071447864174843,
+      "step": 4050
+    },
+    {
+      "epoch": 0.1015,
+      "grad_norm": 28.875,
+      "grad_norm_var": 7.72890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5008,
+      "loss/crossentropy": 2.1623566120862963,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20114662442356349,
+      "step": 4060
+    },
+    {
+      "epoch": 0.10175,
+      "grad_norm": 28.75,
+      "grad_norm_var": 2.5291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4901,
+      "loss/crossentropy": 2.1303680926561355,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19380100946873427,
+      "step": 4070
+    },
+    {
+      "epoch": 0.102,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.9583333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6067,
+      "loss/crossentropy": 2.180470046401024,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20080227889120578,
+      "step": 4080
+    },
+    {
+      "epoch": 0.10225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 43.7791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3949,
+      "loss/crossentropy": 2.04951853454113,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938589910045266,
+      "step": 4090
+    },
+    {
+      "epoch": 0.1025,
+      "grad_norm": 31.5,
+      "grad_norm_var": 40.233333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5,
+      "loss/crossentropy": 1.9436089858412742,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772780598141253,
+      "step": 4100
+    },
+    {
+      "epoch": 0.10275,
+      "grad_norm": 30.75,
+      "grad_norm_var": 6.14765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5926,
+      "loss/crossentropy": 2.1081591993570328,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20672952029854058,
+      "step": 4110
+    },
+    {
+      "epoch": 0.103,
+      "grad_norm": 31.875,
+      "grad_norm_var": 11.225455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6334,
+      "loss/crossentropy": 2.0973087579011915,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20208985283970832,
+      "step": 4120
+    },
+    {
+      "epoch": 0.10325,
+      "grad_norm": 29.375,
+      "grad_norm_var": 25.839583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4644,
+      "loss/crossentropy": 2.205477836728096,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19515758529305458,
+      "step": 4130
+    },
+    {
+      "epoch": 0.1035,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.6567057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4746,
+      "loss/crossentropy": 2.1042226657271383,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850555408746004,
+      "step": 4140
+    },
+    {
+      "epoch": 0.10375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.5952473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5805,
+      "loss/crossentropy": 2.1080187141895292,
+      "loss/hidden": 3.481640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20642356667667627,
+      "step": 4150
+    },
+    {
+      "epoch": 0.104,
+      "grad_norm": 34.75,
+      "grad_norm_var": 12.11875,
+      "learning_rate": 0.0001,
+      "loss": 7.6864,
+      "loss/crossentropy": 2.1868012815713884,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19968394786119462,
+      "step": 4160
+    },
+    {
+      "epoch": 0.10425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 11.406184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6461,
+      "loss/crossentropy": 2.0963368862867355,
+      "loss/hidden": 3.4609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21419077794998884,
+      "step": 4170
+    },
+    {
+      "epoch": 0.1045,
+      "grad_norm": 31.0,
+      "grad_norm_var": 11.670833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6029,
+      "loss/crossentropy": 2.100097879767418,
+      "loss/hidden": 3.35390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18958222791552543,
+      "step": 4180
+    },
+    {
+      "epoch": 0.10475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.664518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5954,
+      "loss/crossentropy": 2.2243005722761153,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19990855641663074,
+      "step": 4190
+    },
+    {
+      "epoch": 0.105,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.333072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6671,
+      "loss/crossentropy": 2.250749832391739,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19082491770386695,
+      "step": 4200
+    },
+    {
+      "epoch": 0.10525,
+      "grad_norm": 29.25,
+      "grad_norm_var": 10.568489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6118,
+      "loss/crossentropy": 2.155411234498024,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2127680890262127,
+      "step": 4210
+    },
+    {
+      "epoch": 0.1055,
+      "grad_norm": 34.25,
+      "grad_norm_var": 8.6556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4391,
+      "loss/crossentropy": 2.037487879395485,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17465929109603168,
+      "step": 4220
+    },
+    {
+      "epoch": 0.10575,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.183072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6537,
+      "loss/crossentropy": 2.067080709338188,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857584908604622,
+      "step": 4230
+    },
+    {
+      "epoch": 0.106,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.4171223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.608,
+      "loss/crossentropy": 2.279679241776466,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.206354571133852,
+      "step": 4240
+    },
+    {
+      "epoch": 0.10625,
+      "grad_norm": 31.75,
+      "grad_norm_var": 31.7681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5259,
+      "loss/crossentropy": 2.1007855504751207,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1838985349982977,
+      "step": 4250
+    },
+    {
+      "epoch": 0.1065,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.5104166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5786,
+      "loss/crossentropy": 2.1117694169282912,
+      "loss/hidden": 3.465625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2134232448413968,
+      "step": 4260
+    },
+    {
+      "epoch": 0.10675,
+      "grad_norm": 33.25,
+      "grad_norm_var": 5.585416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5424,
+      "loss/crossentropy": 2.1906253546476364,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19627480674535036,
+      "step": 4270
+    },
+    {
+      "epoch": 0.107,
+      "grad_norm": 33.75,
+      "grad_norm_var": 16.780208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.7039,
+      "loss/crossentropy": 2.167680537700653,
+      "loss/hidden": 3.469921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21632006093859674,
+      "step": 4280
+    },
+    {
+      "epoch": 0.10725,
+      "grad_norm": 32.75,
+      "grad_norm_var": 14.226497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6197,
+      "loss/crossentropy": 2.2265933483839033,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20777842812240124,
+      "step": 4290
+    },
+    {
+      "epoch": 0.1075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.06640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5019,
+      "loss/crossentropy": 2.0171881064772608,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1733700342476368,
+      "step": 4300
+    },
+    {
+      "epoch": 0.10775,
+      "grad_norm": 32.75,
+      "grad_norm_var": 39.37473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6624,
+      "loss/crossentropy": 2.052209459245205,
+      "loss/hidden": 3.487109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2088288875296712,
+      "step": 4310
+    },
+    {
+      "epoch": 0.108,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.460724588873515e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5801,
+      "loss/crossentropy": 2.096518099308014,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2093063434585929,
+      "step": 4320
+    },
+    {
+      "epoch": 0.10825,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.4607245888865874e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4793,
+      "loss/crossentropy": 2.1164163142442702,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19930963944643737,
+      "step": 4330
+    },
+    {
+      "epoch": 0.1085,
+      "grad_norm": 33.5,
+      "grad_norm_var": 46.587239583333336,
+      "learning_rate": 0.0001,
+      "loss": 7.5915,
+      "loss/crossentropy": 2.1831407219171526,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21474861968308687,
+      "step": 4340
+    },
+    {
+      "epoch": 0.10875,
+      "grad_norm": 31.0,
+      "grad_norm_var": 17.534375,
+      "learning_rate": 0.0001,
+      "loss": 7.4743,
+      "loss/crossentropy": 2.1157304018735887,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2028682116419077,
+      "step": 4350
+    },
+    {
+      "epoch": 0.109,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.282291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6076,
+      "loss/crossentropy": 1.980335572361946,
+      "loss/hidden": 3.541796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20984734632074833,
+      "step": 4360
+    },
+    {
+      "epoch": 0.10925,
+      "grad_norm": 31.25,
+      "grad_norm_var": 30.812955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5349,
+      "loss/crossentropy": 2.107056123018265,
+      "loss/hidden": 3.25859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18440892472863196,
+      "step": 4370
+    },
+    {
+      "epoch": 0.1095,
+      "grad_norm": 30.375,
+      "grad_norm_var": 21.462239583333332,
+      "learning_rate": 0.0001,
+      "loss": 7.5459,
+      "loss/crossentropy": 2.2860846698284147,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1911198776215315,
+      "step": 4380
+    },
+    {
+      "epoch": 0.10975,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.008333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.59,
+      "loss/crossentropy": 2.150055022537708,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20779079720377922,
+      "step": 4390
+    },
+    {
+      "epoch": 0.11,
+      "grad_norm": 37.25,
+      "grad_norm_var": 24.933268229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.6256,
+      "loss/crossentropy": 2.225931641459465,
+      "loss/hidden": 3.569921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23082431070506573,
+      "step": 4400
+    },
+    {
+      "epoch": 0.11025,
+      "grad_norm": 29.375,
+      "grad_norm_var": 26.770572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5585,
+      "loss/crossentropy": 2.1318808451294897,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.205277425237,
+      "step": 4410
+    },
+    {
+      "epoch": 0.1105,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.009830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5037,
+      "loss/crossentropy": 2.108688759803772,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19271691460162402,
+      "step": 4420
+    },
+    {
+      "epoch": 0.11075,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.85390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5599,
+      "loss/crossentropy": 2.1696368783712385,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19752040579915048,
+      "step": 4430
+    },
+    {
+      "epoch": 0.111,
+      "grad_norm": 30.25,
+      "grad_norm_var": 18.476822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5321,
+      "loss/crossentropy": 2.167906680703163,
+      "loss/hidden": 3.48828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2087454443797469,
+      "step": 4440
+    },
+    {
+      "epoch": 0.11125,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.6634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6168,
+      "loss/crossentropy": 2.0555444791913033,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17809300348162652,
+      "step": 4450
+    },
+    {
+      "epoch": 0.1115,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.618489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.669,
+      "loss/crossentropy": 2.2362961381673814,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21373681984841825,
+      "step": 4460
+    },
+    {
+      "epoch": 0.11175,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.8311848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.469,
+      "loss/crossentropy": 2.207303923368454,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21020539589226245,
+      "step": 4470
+    },
+    {
+      "epoch": 0.112,
+      "grad_norm": 66.0,
+      "grad_norm_var": 76.79973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5099,
+      "loss/crossentropy": 2.170773930847645,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20324019938707352,
+      "step": 4480
+    },
+    {
+      "epoch": 0.11225,
+      "grad_norm": 35.0,
+      "grad_norm_var": 76.34895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5149,
+      "loss/crossentropy": 2.047274041175842,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18544426914304496,
+      "step": 4490
+    },
+    {
+      "epoch": 0.1125,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.081705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5745,
+      "loss/crossentropy": 2.1270942091941833,
+      "loss/hidden": 3.311328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19201683439314365,
+      "step": 4500
+    },
+    {
+      "epoch": 0.11275,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.6192057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.632,
+      "loss/crossentropy": 2.0815075978636743,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20613169986754656,
+      "step": 4510
+    },
+    {
+      "epoch": 0.113,
+      "grad_norm": 33.25,
+      "grad_norm_var": 166.34212239583334,
+      "learning_rate": 0.0001,
+      "loss": 7.5383,
+      "loss/crossentropy": 2.144553080201149,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19253196399658917,
+      "step": 4520
+    },
+    {
+      "epoch": 0.11325,
+      "grad_norm": 35.75,
+      "grad_norm_var": 1.734519148252968e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5876,
+      "loss/crossentropy": 2.212298333644867,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19719784446060656,
+      "step": 4530
+    },
+    {
+      "epoch": 0.1135,
+      "grad_norm": 30.625,
+      "grad_norm_var": 32.881184895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.4563,
+      "loss/crossentropy": 2.095240616798401,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19202221632003785,
+      "step": 4540
+    },
+    {
+      "epoch": 0.11375,
+      "grad_norm": 31.625,
+      "grad_norm_var": 87.4853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5184,
+      "loss/crossentropy": 2.134147650748491,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19104634067043663,
+      "step": 4550
+    },
+    {
+      "epoch": 0.114,
+      "grad_norm": 32.25,
+      "grad_norm_var": 78.95618489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4668,
+      "loss/crossentropy": 2.1294006586074827,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1853051505982876,
+      "step": 4560
+    },
+    {
+      "epoch": 0.11425,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.706705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5114,
+      "loss/crossentropy": 2.0886638939380644,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20859680250287055,
+      "step": 4570
+    },
+    {
+      "epoch": 0.1145,
+      "grad_norm": 35.5,
+      "grad_norm_var": 5.934830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.56,
+      "loss/crossentropy": 2.1498399868607523,
+      "loss/hidden": 3.50703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18966037435457112,
+      "step": 4580
+    },
+    {
+      "epoch": 0.11475,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.358268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6013,
+      "loss/crossentropy": 2.199784816801548,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18477672804147005,
+      "step": 4590
+    },
+    {
+      "epoch": 0.115,
+      "grad_norm": 31.375,
+      "grad_norm_var": 17.598372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5806,
+      "loss/crossentropy": 2.165570431947708,
+      "loss/hidden": 3.505859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2113142903894186,
+      "step": 4600
+    },
+    {
+      "epoch": 0.11525,
+      "grad_norm": 31.75,
+      "grad_norm_var": 11.908268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6265,
+      "loss/crossentropy": 2.07996127307415,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21436248067766428,
+      "step": 4610
+    },
+    {
+      "epoch": 0.1155,
+      "grad_norm": 28.125,
+      "grad_norm_var": 4.520833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.57,
+      "loss/crossentropy": 2.1484976023435594,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1913412597030401,
+      "step": 4620
+    },
+    {
+      "epoch": 0.11575,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.1875,
+      "learning_rate": 0.0001,
+      "loss": 7.4701,
+      "loss/crossentropy": 2.0956287920475005,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19558896012604238,
+      "step": 4630
+    },
+    {
+      "epoch": 0.116,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.106184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4538,
+      "loss/crossentropy": 2.1474092990159988,
+      "loss/hidden": 3.45234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2039597311988473,
+      "step": 4640
+    },
+    {
+      "epoch": 0.11625,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.520768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6657,
+      "loss/crossentropy": 2.1128339886665346,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21120875477790832,
+      "step": 4650
+    },
+    {
+      "epoch": 0.1165,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.133072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5441,
+      "loss/crossentropy": 2.1643586844205855,
+      "loss/hidden": 3.4765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20440610516816377,
+      "step": 4660
+    },
+    {
+      "epoch": 0.11675,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.504622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6051,
+      "loss/crossentropy": 2.195269528031349,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20059011913836003,
+      "step": 4670
+    },
+    {
+      "epoch": 0.117,
+      "grad_norm": 30.0,
+      "grad_norm_var": 31.235416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4526,
+      "loss/crossentropy": 2.082709529995918,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19617473538964986,
+      "step": 4680
+    },
+    {
+      "epoch": 0.11725,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.45625,
+      "learning_rate": 0.0001,
+      "loss": 7.5593,
+      "loss/crossentropy": 2.1473275452852247,
+      "loss/hidden": 3.481640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20148523468524218,
+      "step": 4690
+    },
+    {
+      "epoch": 0.1175,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.145768229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4609,
+      "loss/crossentropy": 2.14631325006485,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20976583026349543,
+      "step": 4700
+    },
+    {
+      "epoch": 0.11775,
+      "grad_norm": 32.5,
+      "grad_norm_var": 33.20826822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.5877,
+      "loss/crossentropy": 2.1093395471572878,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20207433141767978,
+      "step": 4710
+    },
+    {
+      "epoch": 0.118,
+      "grad_norm": 30.75,
+      "grad_norm_var": 10.86875,
+      "learning_rate": 0.0001,
+      "loss": 7.5437,
+      "loss/crossentropy": 2.1447531282901764,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1963239062577486,
+      "step": 4720
+    },
+    {
+      "epoch": 0.11825,
+      "grad_norm": 34.5,
+      "grad_norm_var": 4.473372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6248,
+      "loss/crossentropy": 2.1142423778772352,
+      "loss/hidden": 3.283203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19059138614684343,
+      "step": 4730
+    },
+    {
+      "epoch": 0.1185,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.098372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5631,
+      "loss/crossentropy": 2.07417613863945,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19420854579657315,
+      "step": 4740
+    },
+    {
+      "epoch": 0.11875,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.734375,
+      "learning_rate": 0.0001,
+      "loss": 7.6499,
+      "loss/crossentropy": 2.0995796024799347,
+      "loss/hidden": 3.518359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21285411342978477,
+      "step": 4750
+    },
+    {
+      "epoch": 0.119,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.7228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5854,
+      "loss/crossentropy": 2.169099047780037,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22147531677037477,
+      "step": 4760
+    },
+    {
+      "epoch": 0.11925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 34.799739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.435,
+      "loss/crossentropy": 2.0821361050009726,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17594938166439533,
+      "step": 4770
+    },
+    {
+      "epoch": 0.1195,
+      "grad_norm": 33.75,
+      "grad_norm_var": 23.00390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6152,
+      "loss/crossentropy": 2.183753404021263,
+      "loss/hidden": 3.364453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19579742290079594,
+      "step": 4780
+    },
+    {
+      "epoch": 0.11975,
+      "grad_norm": 29.25,
+      "grad_norm_var": 23.6541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5295,
+      "loss/crossentropy": 2.1669270396232605,
+      "loss/hidden": 3.505859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21605710163712502,
+      "step": 4790
+    },
+    {
+      "epoch": 0.12,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.624739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6351,
+      "loss/crossentropy": 2.116096779704094,
+      "loss/hidden": 3.4625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21523005720227956,
+      "step": 4800
+    },
+    {
+      "epoch": 0.12025,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.009375,
+      "learning_rate": 0.0001,
+      "loss": 7.5683,
+      "loss/crossentropy": 2.051844981312752,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17574754767119885,
+      "step": 4810
+    },
+    {
+      "epoch": 0.1205,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.0087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5892,
+      "loss/crossentropy": 2.2223299980163573,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19802077617496253,
+      "step": 4820
+    },
+    {
+      "epoch": 0.12075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.1080729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5951,
+      "loss/crossentropy": 2.1982655793428423,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2064062263816595,
+      "step": 4830
+    },
+    {
+      "epoch": 0.121,
+      "grad_norm": 30.0,
+      "grad_norm_var": 12.412434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5254,
+      "loss/crossentropy": 1.9753010511398315,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19706314485520124,
+      "step": 4840
+    },
+    {
+      "epoch": 0.12125,
+      "grad_norm": 30.125,
+      "grad_norm_var": 13.1447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5351,
+      "loss/crossentropy": 2.0562877766788006,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17667807769030333,
+      "step": 4850
+    },
+    {
+      "epoch": 0.1215,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.465625,
+      "learning_rate": 0.0001,
+      "loss": 7.6024,
+      "loss/crossentropy": 2.1578008987009527,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18234786652028562,
+      "step": 4860
+    },
+    {
+      "epoch": 0.12175,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.343489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4726,
+      "loss/crossentropy": 2.1330083698034286,
+      "loss/hidden": 3.333203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18738476932048798,
+      "step": 4870
+    },
+    {
+      "epoch": 0.122,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.2322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.605,
+      "loss/crossentropy": 2.1732089832425117,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19949225690215827,
+      "step": 4880
+    },
+    {
+      "epoch": 0.12225,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.238997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5382,
+      "loss/crossentropy": 2.188927575945854,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19776681065559387,
+      "step": 4890
+    },
+    {
+      "epoch": 0.1225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 15.819791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5728,
+      "loss/crossentropy": 2.202793037891388,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1992744604125619,
+      "step": 4900
+    },
+    {
+      "epoch": 0.12275,
+      "grad_norm": 30.25,
+      "grad_norm_var": 16.1775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.426,
+      "loss/crossentropy": 2.177696394920349,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19968770015984774,
+      "step": 4910
+    },
+    {
+      "epoch": 0.123,
+      "grad_norm": 35.5,
+      "grad_norm_var": 7.707291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.7485,
+      "loss/crossentropy": 2.040875867009163,
+      "loss/hidden": 3.45234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1929738214239478,
+      "step": 4920
+    },
+    {
+      "epoch": 0.12325,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.771809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5517,
+      "loss/crossentropy": 2.199721184372902,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20861553251743317,
+      "step": 4930
+    },
+    {
+      "epoch": 0.1235,
+      "grad_norm": 28.75,
+      "grad_norm_var": 4.235416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5576,
+      "loss/crossentropy": 2.133353302627802,
+      "loss/hidden": 3.45546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21631606128066777,
+      "step": 4940
+    },
+    {
+      "epoch": 0.12375,
+      "grad_norm": 40.0,
+      "grad_norm_var": 15.1650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5129,
+      "loss/crossentropy": 2.2380867928266523,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2078275766223669,
+      "step": 4950
+    },
+    {
+      "epoch": 0.124,
+      "grad_norm": 32.5,
+      "grad_norm_var": 9.342643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4421,
+      "loss/crossentropy": 2.1563473463058473,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1852986102923751,
+      "step": 4960
+    },
+    {
+      "epoch": 0.12425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.3916666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5503,
+      "loss/crossentropy": 2.124380439519882,
+      "loss/hidden": 3.480859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20270956568419934,
+      "step": 4970
+    },
+    {
+      "epoch": 0.1245,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.837434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4857,
+      "loss/crossentropy": 2.0482941284775733,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18650466352701187,
+      "step": 4980
+    },
+    {
+      "epoch": 0.12475,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.0858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5446,
+      "loss/crossentropy": 2.160063311457634,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18700905814766883,
+      "step": 4990
+    },
+    {
+      "epoch": 0.125,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.533333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5333,
+      "loss/crossentropy": 2.056824396550655,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18965724110603333,
+      "step": 5000
+    },
+    {
+      "epoch": 0.12525,
+      "grad_norm": 32.5,
+      "grad_norm_var": 32.25807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5864,
+      "loss/crossentropy": 2.2407308876514436,
+      "loss/hidden": 3.51328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22782632596790792,
+      "step": 5010
+    },
+    {
+      "epoch": 0.1255,
+      "grad_norm": 31.375,
+      "grad_norm_var": 32.7353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5671,
+      "loss/crossentropy": 2.18543721139431,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19135277662426234,
+      "step": 5020
+    },
+    {
+      "epoch": 0.12575,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.814322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.578,
+      "loss/crossentropy": 2.1418938025832177,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20001544915139674,
+      "step": 5030
+    },
+    {
+      "epoch": 0.126,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.3238932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5594,
+      "loss/crossentropy": 2.1838817209005357,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2107716018334031,
+      "step": 5040
+    },
+    {
+      "epoch": 0.12625,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.47890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5275,
+      "loss/crossentropy": 2.2816754072904586,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201267159730196,
+      "step": 5050
+    },
+    {
+      "epoch": 0.1265,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.2087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6079,
+      "loss/crossentropy": 2.0686806365847588,
+      "loss/hidden": 3.603125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2172813605517149,
+      "step": 5060
+    },
+    {
+      "epoch": 0.12675,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.19765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5561,
+      "loss/crossentropy": 2.1800751775503158,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20199469216167926,
+      "step": 5070
+    },
+    {
+      "epoch": 0.127,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.8643229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4702,
+      "loss/crossentropy": 2.2943209201097488,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1958466824144125,
+      "step": 5080
+    },
+    {
+      "epoch": 0.12725,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.9030598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4383,
+      "loss/crossentropy": 2.291040873527527,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18960105255246162,
+      "step": 5090
+    },
+    {
+      "epoch": 0.1275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.4504557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6139,
+      "loss/crossentropy": 2.116188834607601,
+      "loss/hidden": 3.5015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19662482757121325,
+      "step": 5100
+    },
+    {
+      "epoch": 0.12775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.5395182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5074,
+      "loss/crossentropy": 2.200083887577057,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19970641760155558,
+      "step": 5110
+    },
+    {
+      "epoch": 0.128,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.990625,
+      "learning_rate": 0.0001,
+      "loss": 7.5949,
+      "loss/crossentropy": 2.080552561581135,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19426519125699998,
+      "step": 5120
+    },
+    {
+      "epoch": 0.12825,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.200455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5036,
+      "loss/crossentropy": 2.138309660553932,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20698665007948874,
+      "step": 5130
+    },
+    {
+      "epoch": 0.1285,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.530989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6081,
+      "loss/crossentropy": 2.0686957597732545,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22128727175295354,
+      "step": 5140
+    },
+    {
+      "epoch": 0.12875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.120833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5236,
+      "loss/crossentropy": 2.0967576891183852,
+      "loss/hidden": 3.46640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19716181065887212,
+      "step": 5150
+    },
+    {
+      "epoch": 0.129,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.8468098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5679,
+      "loss/crossentropy": 2.172739614546299,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1919636652804911,
+      "step": 5160
+    },
+    {
+      "epoch": 0.12925,
+      "grad_norm": 43.75,
+      "grad_norm_var": 17.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5027,
+      "loss/crossentropy": 2.1427868396043777,
+      "loss/hidden": 3.51953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20654744990170001,
+      "step": 5170
+    },
+    {
+      "epoch": 0.1295,
+      "grad_norm": 33.25,
+      "grad_norm_var": 61.1962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.52,
+      "loss/crossentropy": 2.1186717480421065,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938714198768139,
+      "step": 5180
+    },
+    {
+      "epoch": 0.12975,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.07890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6231,
+      "loss/crossentropy": 2.072993017733097,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18984810579568148,
+      "step": 5190
+    },
+    {
+      "epoch": 0.13,
+      "grad_norm": 36.25,
+      "grad_norm_var": 20.777083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.597,
+      "loss/crossentropy": 2.206225660443306,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2030201606452465,
+      "step": 5200
+    },
+    {
+      "epoch": 0.13025,
+      "grad_norm": 43.75,
+      "grad_norm_var": 24.43515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5075,
+      "loss/crossentropy": 2.228325179219246,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20732430163770915,
+      "step": 5210
+    },
+    {
+      "epoch": 0.1305,
+      "grad_norm": 30.5,
+      "grad_norm_var": 11.99765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5295,
+      "loss/crossentropy": 2.129982355237007,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19593548215925694,
+      "step": 5220
+    },
+    {
+      "epoch": 0.13075,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.292708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5261,
+      "loss/crossentropy": 2.1533152967691422,
+      "loss/hidden": 3.3296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19487107992172242,
+      "step": 5230
+    },
+    {
+      "epoch": 0.131,
+      "grad_norm": 30.875,
+      "grad_norm_var": 12.989583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5566,
+      "loss/crossentropy": 2.195969894528389,
+      "loss/hidden": 3.48046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2189876638352871,
+      "step": 5240
+    },
+    {
+      "epoch": 0.13125,
+      "grad_norm": 33.25,
+      "grad_norm_var": 22.575,
+      "learning_rate": 0.0001,
+      "loss": 7.482,
+      "loss/crossentropy": 2.074573493748903,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19280508980154992,
+      "step": 5250
+    },
+    {
+      "epoch": 0.1315,
+      "grad_norm": 29.125,
+      "grad_norm_var": 15.599739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5395,
+      "loss/crossentropy": 2.227403500676155,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20514250732958317,
+      "step": 5260
+    },
+    {
+      "epoch": 0.13175,
+      "grad_norm": 31.625,
+      "grad_norm_var": 113.15618489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.5209,
+      "loss/crossentropy": 2.055370827019215,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18272479642182587,
+      "step": 5270
+    },
+    {
+      "epoch": 0.132,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4798,
+      "loss/crossentropy": 2.1356967806816103,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19737351574003698,
+      "step": 5280
+    },
+    {
+      "epoch": 0.13225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.879622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4715,
+      "loss/crossentropy": 2.177875056862831,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20798433478921652,
+      "step": 5290
+    },
+    {
+      "epoch": 0.1325,
+      "grad_norm": 29.5,
+      "grad_norm_var": 8.760872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5419,
+      "loss/crossentropy": 2.182955250144005,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18478553090244532,
+      "step": 5300
+    },
+    {
+      "epoch": 0.13275,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.260416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5816,
+      "loss/crossentropy": 2.0469735309481623,
+      "loss/hidden": 3.55859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18724320270121098,
+      "step": 5310
+    },
+    {
+      "epoch": 0.133,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.8801432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.7815,
+      "loss/crossentropy": 2.1846155911684035,
+      "loss/hidden": 3.499609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21609773077070712,
+      "step": 5320
+    },
+    {
+      "epoch": 0.13325,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.584375,
+      "learning_rate": 0.0001,
+      "loss": 7.7067,
+      "loss/crossentropy": 2.0075640469789504,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20185804851353167,
+      "step": 5330
+    },
+    {
+      "epoch": 0.1335,
+      "grad_norm": 27.375,
+      "grad_norm_var": 6.120247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5345,
+      "loss/crossentropy": 2.058630608022213,
+      "loss/hidden": 3.421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19536744449287652,
+      "step": 5340
+    },
+    {
+      "epoch": 0.13375,
+      "grad_norm": 52.5,
+      "grad_norm_var": 30.692643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6655,
+      "loss/crossentropy": 2.050611114501953,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19592730849981307,
+      "step": 5350
+    },
+    {
+      "epoch": 0.134,
+      "grad_norm": 31.125,
+      "grad_norm_var": 28.90390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5509,
+      "loss/crossentropy": 2.1665500849485397,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19354073759168386,
+      "step": 5360
+    },
+    {
+      "epoch": 0.13425,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.9931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6343,
+      "loss/crossentropy": 2.1793171644210814,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20293663591146469,
+      "step": 5370
+    },
+    {
+      "epoch": 0.1345,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.8098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6,
+      "loss/crossentropy": 2.152762657403946,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21694996021687984,
+      "step": 5380
+    },
+    {
+      "epoch": 0.13475,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.4301432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5121,
+      "loss/crossentropy": 2.083320555835962,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18746816255152227,
+      "step": 5390
+    },
+    {
+      "epoch": 0.135,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.109830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4172,
+      "loss/crossentropy": 2.0639937698841093,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18618469405919313,
+      "step": 5400
+    },
+    {
+      "epoch": 0.13525,
+      "grad_norm": 31.125,
+      "grad_norm_var": 12.337434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5749,
+      "loss/crossentropy": 2.068472331762314,
+      "loss/hidden": 3.524609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18952864613384007,
+      "step": 5410
+    },
+    {
+      "epoch": 0.1355,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.9541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5385,
+      "loss/crossentropy": 2.085490897297859,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854228163138032,
+      "step": 5420
+    },
+    {
+      "epoch": 0.13575,
+      "grad_norm": 30.0,
+      "grad_norm_var": 7.678125,
+      "learning_rate": 0.0001,
+      "loss": 7.5583,
+      "loss/crossentropy": 2.1747709423303605,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20177022367715836,
+      "step": 5430
+    },
+    {
+      "epoch": 0.136,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.6541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4911,
+      "loss/crossentropy": 2.167823739349842,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2082101447507739,
+      "step": 5440
+    },
+    {
+      "epoch": 0.13625,
+      "grad_norm": 30.625,
+      "grad_norm_var": 11.2087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.639,
+      "loss/crossentropy": 2.1367167800664904,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20548735409975052,
+      "step": 5450
+    },
+    {
+      "epoch": 0.1365,
+      "grad_norm": 28.375,
+      "grad_norm_var": 15.654166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4951,
+      "loss/crossentropy": 2.109811532497406,
+      "loss/hidden": 3.4640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1994595667347312,
+      "step": 5460
+    },
+    {
+      "epoch": 0.13675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.463997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5263,
+      "loss/crossentropy": 2.2288592010736465,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20487434454262257,
+      "step": 5470
+    },
+    {
+      "epoch": 0.137,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.8400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5681,
+      "loss/crossentropy": 2.0951829612255097,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20034745894372463,
+      "step": 5480
+    },
+    {
+      "epoch": 0.13725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.299934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4772,
+      "loss/crossentropy": 2.022165683656931,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18401878620497883,
+      "step": 5490
+    },
+    {
+      "epoch": 0.1375,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.205989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6869,
+      "loss/crossentropy": 2.095045933127403,
+      "loss/hidden": 3.609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21982598043978213,
+      "step": 5500
+    },
+    {
+      "epoch": 0.13775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.177018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4483,
+      "loss/crossentropy": 2.07875951230526,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898285737261176,
+      "step": 5510
+    },
+    {
+      "epoch": 0.138,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.4181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5258,
+      "loss/crossentropy": 2.194947564601898,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21397264376282693,
+      "step": 5520
+    },
+    {
+      "epoch": 0.13825,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.5541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6574,
+      "loss/crossentropy": 2.2282156944274902,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19613583907485008,
+      "step": 5530
+    },
+    {
+      "epoch": 0.1385,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.927018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5361,
+      "loss/crossentropy": 2.1092587068676947,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1894394876435399,
+      "step": 5540
+    },
+    {
+      "epoch": 0.13875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 6.080989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6223,
+      "loss/crossentropy": 2.189313694834709,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.210250511392951,
+      "step": 5550
+    },
+    {
+      "epoch": 0.139,
+      "grad_norm": 41.25,
+      "grad_norm_var": 8.976497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4876,
+      "loss/crossentropy": 2.046143325418234,
+      "loss/hidden": 3.457421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19476603530347347,
+      "step": 5560
+    },
+    {
+      "epoch": 0.13925,
+      "grad_norm": 31.25,
+      "grad_norm_var": 9.076497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5338,
+      "loss/crossentropy": 2.1882025837898254,
+      "loss/hidden": 3.45,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2058427443727851,
+      "step": 5570
+    },
+    {
+      "epoch": 0.1395,
+      "grad_norm": 30.25,
+      "grad_norm_var": 182.91875,
+      "learning_rate": 0.0001,
+      "loss": 7.6344,
+      "loss/crossentropy": 2.160731779038906,
+      "loss/hidden": 3.5171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22695957981050013,
+      "step": 5580
+    },
+    {
+      "epoch": 0.13975,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.140559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5419,
+      "loss/crossentropy": 1.8924851581454276,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1802460763603449,
+      "step": 5590
+    },
+    {
+      "epoch": 0.14,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.084830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5498,
+      "loss/crossentropy": 2.242933538556099,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22144901510328055,
+      "step": 5600
+    },
+    {
+      "epoch": 0.14025,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.5622395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5133,
+      "loss/crossentropy": 2.132666201889515,
+      "loss/hidden": 3.492578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19677439089864493,
+      "step": 5610
+    },
+    {
+      "epoch": 0.1405,
+      "grad_norm": 30.5,
+      "grad_norm_var": 14.234375,
+      "learning_rate": 0.0001,
+      "loss": 7.6226,
+      "loss/crossentropy": 2.0671760708093645,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20731164887547493,
+      "step": 5620
+    },
+    {
+      "epoch": 0.14075,
+      "grad_norm": 35.5,
+      "grad_norm_var": 14.44375,
+      "learning_rate": 0.0001,
+      "loss": 7.5063,
+      "loss/crossentropy": 2.060420323908329,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18428542967885733,
+      "step": 5630
+    },
+    {
+      "epoch": 0.141,
+      "grad_norm": 38.5,
+      "grad_norm_var": 19.1509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5322,
+      "loss/crossentropy": 2.2572733104228973,
+      "loss/hidden": 3.291015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19184892270714043,
+      "step": 5640
+    },
+    {
+      "epoch": 0.14125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 20.373893229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5473,
+      "loss/crossentropy": 2.1559308364987375,
+      "loss/hidden": 3.237109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18201812207698823,
+      "step": 5650
+    },
+    {
+      "epoch": 0.1415,
+      "grad_norm": 28.5,
+      "grad_norm_var": 10.033268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4529,
+      "loss/crossentropy": 2.1479016572237013,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20532293617725372,
+      "step": 5660
+    },
+    {
+      "epoch": 0.14175,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.6332682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6001,
+      "loss/crossentropy": 2.140682426095009,
+      "loss/hidden": 3.321484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18856783863157034,
+      "step": 5670
+    },
+    {
+      "epoch": 0.142,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.84140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6348,
+      "loss/crossentropy": 2.25071659386158,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2029257183894515,
+      "step": 5680
+    },
+    {
+      "epoch": 0.14225,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.643684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5323,
+      "loss/crossentropy": 2.149606391787529,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2141027996316552,
+      "step": 5690
+    },
+    {
+      "epoch": 0.1425,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.32890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5615,
+      "loss/crossentropy": 2.2273970007896424,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22348886616528035,
+      "step": 5700
+    },
+    {
+      "epoch": 0.14275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.8285807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4366,
+      "loss/crossentropy": 2.2010063380002975,
+      "loss/hidden": 3.289453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18594364672899247,
+      "step": 5710
+    },
+    {
+      "epoch": 0.143,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.6020182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.508,
+      "loss/crossentropy": 2.0483784288167954,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2091974811628461,
+      "step": 5720
+    },
+    {
+      "epoch": 0.14325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.6744140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6646,
+      "loss/crossentropy": 2.025853230059147,
+      "loss/hidden": 3.589453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2227605242282152,
+      "step": 5730
+    },
+    {
+      "epoch": 0.1435,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.8327473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6064,
+      "loss/crossentropy": 2.075891287624836,
+      "loss/hidden": 3.38359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1890827091410756,
+      "step": 5740
+    },
+    {
+      "epoch": 0.14375,
+      "grad_norm": 31.625,
+      "grad_norm_var": 13.943489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4294,
+      "loss/crossentropy": 2.0676268830895426,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20394732002168894,
+      "step": 5750
+    },
+    {
+      "epoch": 0.144,
+      "grad_norm": 30.0,
+      "grad_norm_var": 14.658072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5958,
+      "loss/crossentropy": 2.1388431686908005,
+      "loss/hidden": 3.499609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21454429486766458,
+      "step": 5760
+    },
+    {
+      "epoch": 0.14425,
+      "grad_norm": 32.0,
+      "grad_norm_var": 13.958072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5342,
+      "loss/crossentropy": 2.0246446818113326,
+      "loss/hidden": 3.3140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18158567249774932,
+      "step": 5770
+    },
+    {
+      "epoch": 0.1445,
+      "grad_norm": 28.375,
+      "grad_norm_var": 14.469205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5141,
+      "loss/crossentropy": 2.103305173665285,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842075905762613,
+      "step": 5780
+    },
+    {
+      "epoch": 0.14475,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.160872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.6213,
+      "loss/crossentropy": 2.096319726109505,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19167389422655107,
+      "step": 5790
+    },
+    {
+      "epoch": 0.145,
+      "grad_norm": 35.25,
+      "grad_norm_var": 3.4916666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4461,
+      "loss/crossentropy": 2.0800597339868547,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18269639350473882,
+      "step": 5800
+    },
+    {
+      "epoch": 0.14525,
+      "grad_norm": 45.5,
+      "grad_norm_var": 22.3306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4855,
+      "loss/crossentropy": 2.036512078344822,
+      "loss/hidden": 3.462890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1910469362512231,
+      "step": 5810
+    },
+    {
+      "epoch": 0.1455,
+      "grad_norm": 33.0,
+      "grad_norm_var": 23.284309895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.4838,
+      "loss/crossentropy": 2.1922324389219283,
+      "loss/hidden": 3.48671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2080824663862586,
+      "step": 5820
+    },
+    {
+      "epoch": 0.14575,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.7171223958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4657,
+      "loss/crossentropy": 2.0417690485715867,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19891292043030262,
+      "step": 5830
+    },
+    {
+      "epoch": 0.146,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5864,
+      "loss/crossentropy": 2.12301287651062,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19781142249703407,
+      "step": 5840
+    },
+    {
+      "epoch": 0.14625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 6.574934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6264,
+      "loss/crossentropy": 2.1600560665130617,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22095333114266397,
+      "step": 5850
+    },
+    {
+      "epoch": 0.1465,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.3822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5231,
+      "loss/crossentropy": 2.185663253068924,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2029089467599988,
+      "step": 5860
+    },
+    {
+      "epoch": 0.14675,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.020247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4938,
+      "loss/crossentropy": 2.052996274828911,
+      "loss/hidden": 3.477734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19343450702726842,
+      "step": 5870
+    },
+    {
+      "epoch": 0.147,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.112239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5855,
+      "loss/crossentropy": 2.2062906324863434,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20249157436192036,
+      "step": 5880
+    },
+    {
+      "epoch": 0.14725,
+      "grad_norm": 36.25,
+      "grad_norm_var": 2.4330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5644,
+      "loss/crossentropy": 2.0493671208620072,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19113321453332902,
+      "step": 5890
+    },
+    {
+      "epoch": 0.1475,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.6968098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5941,
+      "loss/crossentropy": 2.20456420481205,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20127719175070524,
+      "step": 5900
+    },
+    {
+      "epoch": 0.14775,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.8268229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4958,
+      "loss/crossentropy": 2.2630672723054888,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20422303341329098,
+      "step": 5910
+    },
+    {
+      "epoch": 0.148,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.45,
+      "learning_rate": 0.0001,
+      "loss": 7.4893,
+      "loss/crossentropy": 2.130161929130554,
+      "loss/hidden": 3.442578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2029714561998844,
+      "step": 5920
+    },
+    {
+      "epoch": 0.14825,
+      "grad_norm": 31.375,
+      "grad_norm_var": 0.9728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6087,
+      "loss/crossentropy": 2.2219777315855027,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20015078466385602,
+      "step": 5930
+    },
+    {
+      "epoch": 0.1485,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.9514973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5908,
+      "loss/crossentropy": 2.2077976912260056,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1996950874105096,
+      "step": 5940
+    },
+    {
+      "epoch": 0.14875,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.3619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5402,
+      "loss/crossentropy": 2.1866880118846894,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19317448064684867,
+      "step": 5950
+    },
+    {
+      "epoch": 0.149,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.6968098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4202,
+      "loss/crossentropy": 2.1943173080682756,
+      "loss/hidden": 3.2390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18432049825787544,
+      "step": 5960
+    },
+    {
+      "epoch": 0.14925,
+      "grad_norm": 30.75,
+      "grad_norm_var": 16.936393229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5056,
+      "loss/crossentropy": 2.17142014503479,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1876060863956809,
+      "step": 5970
+    },
+    {
+      "epoch": 0.1495,
+      "grad_norm": 31.5,
+      "grad_norm_var": 4.03515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5442,
+      "loss/crossentropy": 2.2047942698001863,
+      "loss/hidden": 3.369140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1988227294757962,
+      "step": 5980
+    },
+    {
+      "epoch": 0.14975,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.332291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5693,
+      "loss/crossentropy": 2.281469625234604,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19730366189032794,
+      "step": 5990
+    },
+    {
+      "epoch": 0.15,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.7603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4556,
+      "loss/crossentropy": 2.256136628985405,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971954844892025,
+      "step": 6000
+    },
+    {
+      "epoch": 0.15025,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.1285807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5146,
+      "loss/crossentropy": 2.105835199356079,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19158641248941422,
+      "step": 6010
+    },
+    {
+      "epoch": 0.1505,
+      "grad_norm": 40.0,
+      "grad_norm_var": 18.542708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4687,
+      "loss/crossentropy": 2.095852518081665,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18887621480971575,
+      "step": 6020
+    },
+    {
+      "epoch": 0.15075,
+      "grad_norm": 30.5,
+      "grad_norm_var": 19.169791666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6293,
+      "loss/crossentropy": 2.0981243371963503,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22096464578062297,
+      "step": 6030
+    },
+    {
+      "epoch": 0.151,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.716080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5935,
+      "loss/crossentropy": 2.212895154953003,
+      "loss/hidden": 3.510546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22050899863243104,
+      "step": 6040
+    },
+    {
+      "epoch": 0.15125,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.042122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5,
+      "loss/crossentropy": 2.1280829131603243,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18497859146445988,
+      "step": 6050
+    },
+    {
+      "epoch": 0.1515,
+      "grad_norm": 29.25,
+      "grad_norm_var": 47.15807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5469,
+      "loss/crossentropy": 2.1731619790196417,
+      "loss/hidden": 3.4578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22194090783596038,
+      "step": 6060
+    },
+    {
+      "epoch": 0.15175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.218489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4378,
+      "loss/crossentropy": 2.047077566385269,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1876080146059394,
+      "step": 6070
+    },
+    {
+      "epoch": 0.152,
+      "grad_norm": 33.0,
+      "grad_norm_var": 5.45,
+      "learning_rate": 0.0001,
+      "loss": 7.584,
+      "loss/crossentropy": 2.1896773248910906,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19448568094521762,
+      "step": 6080
+    },
+    {
+      "epoch": 0.15225,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.7634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6047,
+      "loss/crossentropy": 2.1123378753662108,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1931258851662278,
+      "step": 6090
+    },
+    {
+      "epoch": 0.1525,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.5488932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5301,
+      "loss/crossentropy": 2.2152541011571882,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19427606835961342,
+      "step": 6100
+    },
+    {
+      "epoch": 0.15275,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.7421223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5936,
+      "loss/crossentropy": 2.109275272488594,
+      "loss/hidden": 3.531640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2041168488562107,
+      "step": 6110
+    },
+    {
+      "epoch": 0.153,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.723893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5107,
+      "loss/crossentropy": 2.0911407291889192,
+      "loss/hidden": 3.403515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19639678560197354,
+      "step": 6120
+    },
+    {
+      "epoch": 0.15325,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.2561848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4788,
+      "loss/crossentropy": 2.1557929456233977,
+      "loss/hidden": 3.5140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2058223508298397,
+      "step": 6130
+    },
+    {
+      "epoch": 0.1535,
+      "grad_norm": 30.0,
+      "grad_norm_var": 84.3337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5656,
+      "loss/crossentropy": 2.0974055036902426,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19299208922311664,
+      "step": 6140
+    },
+    {
+      "epoch": 0.15375,
+      "grad_norm": 30.625,
+      "grad_norm_var": 85.70149739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4607,
+      "loss/crossentropy": 2.0354842752218247,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2099142510443926,
+      "step": 6150
+    },
+    {
+      "epoch": 0.154,
+      "grad_norm": 29.75,
+      "grad_norm_var": 9.5134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4221,
+      "loss/crossentropy": 1.9916995614767075,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17939293198287487,
+      "step": 6160
+    },
+    {
+      "epoch": 0.15425,
+      "grad_norm": 33.5,
+      "grad_norm_var": 10.60390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5836,
+      "loss/crossentropy": 2.097960978746414,
+      "loss/hidden": 3.353515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1908213123679161,
+      "step": 6170
+    },
+    {
+      "epoch": 0.1545,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.0942057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5073,
+      "loss/crossentropy": 2.253797325491905,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19871626775711776,
+      "step": 6180
+    },
+    {
+      "epoch": 0.15475,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.3900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5189,
+      "loss/crossentropy": 2.1756670624017715,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19654417987912892,
+      "step": 6190
+    },
+    {
+      "epoch": 0.155,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.4181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3147,
+      "loss/crossentropy": 2.1110276162624357,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1801340376958251,
+      "step": 6200
+    },
+    {
+      "epoch": 0.15525,
+      "grad_norm": 28.5,
+      "grad_norm_var": 1.6223307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6015,
+      "loss/crossentropy": 2.1547244489192963,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18071307614445686,
+      "step": 6210
+    },
+    {
+      "epoch": 0.1555,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.36875,
+      "learning_rate": 0.0001,
+      "loss": 7.5559,
+      "loss/crossentropy": 2.1889162242412565,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.185367326810956,
+      "step": 6220
+    },
+    {
+      "epoch": 0.15575,
+      "grad_norm": 38.75,
+      "grad_norm_var": 25.381184895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.5323,
+      "loss/crossentropy": 2.023681116104126,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19041543006896972,
+      "step": 6230
+    },
+    {
+      "epoch": 0.156,
+      "grad_norm": 31.5,
+      "grad_norm_var": 12.545833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6086,
+      "loss/crossentropy": 2.260247975587845,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19489852003753186,
+      "step": 6240
+    },
+    {
+      "epoch": 0.15625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 6.642122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5419,
+      "loss/crossentropy": 2.1487904608249666,
+      "loss/hidden": 3.37421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1957969294860959,
+      "step": 6250
+    },
+    {
+      "epoch": 0.1565,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.493489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.588,
+      "loss/crossentropy": 2.279403430223465,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20682295374572277,
+      "step": 6260
+    },
+    {
+      "epoch": 0.15675,
+      "grad_norm": 30.125,
+      "grad_norm_var": 8.14765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5894,
+      "loss/crossentropy": 2.1407866299152376,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20337907522916793,
+      "step": 6270
+    },
+    {
+      "epoch": 0.157,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1314.8994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6872,
+      "loss/crossentropy": 2.1802447110414507,
+      "loss/hidden": 3.540625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22727628983557224,
+      "step": 6280
+    },
+    {
+      "epoch": 0.15725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1378.0379557291667,
+      "learning_rate": 0.0001,
+      "loss": 7.5596,
+      "loss/crossentropy": 2.1761491730809213,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18902392424643039,
+      "step": 6290
+    },
+    {
+      "epoch": 0.1575,
+      "grad_norm": 42.25,
+      "grad_norm_var": 146.03020833333332,
+      "learning_rate": 0.0001,
+      "loss": 7.6132,
+      "loss/crossentropy": 2.106969301402569,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2025001836940646,
+      "step": 6300
+    },
+    {
+      "epoch": 0.15775,
+      "grad_norm": 29.25,
+      "grad_norm_var": 14.539322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.514,
+      "loss/crossentropy": 2.0654825627803803,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19361731074750424,
+      "step": 6310
+    },
+    {
+      "epoch": 0.158,
+      "grad_norm": 31.75,
+      "grad_norm_var": 8.153580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5587,
+      "loss/crossentropy": 2.025081543624401,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2247559091076255,
+      "step": 6320
+    },
+    {
+      "epoch": 0.15825,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.0431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4657,
+      "loss/crossentropy": 2.136493813991547,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19502629730850457,
+      "step": 6330
+    },
+    {
+      "epoch": 0.1585,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.189322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5454,
+      "loss/crossentropy": 2.176800549030304,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.201618373952806,
+      "step": 6340
+    },
+    {
+      "epoch": 0.15875,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.97265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5116,
+      "loss/crossentropy": 2.2223973661661147,
+      "loss/hidden": 3.43359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2056947773322463,
+      "step": 6350
+    },
+    {
+      "epoch": 0.159,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.0580729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5133,
+      "loss/crossentropy": 2.147826671600342,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19723634477704763,
+      "step": 6360
+    },
+    {
+      "epoch": 0.15925,
+      "grad_norm": 35.0,
+      "grad_norm_var": 18.772916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.679,
+      "loss/crossentropy": 2.079222206771374,
+      "loss/hidden": 3.47890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2045620433986187,
+      "step": 6370
+    },
+    {
+      "epoch": 0.1595,
+      "grad_norm": 28.75,
+      "grad_norm_var": 20.639518229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.5245,
+      "loss/crossentropy": 2.163569709658623,
+      "loss/hidden": 3.45546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20100143328309059,
+      "step": 6380
+    },
+    {
+      "epoch": 0.15975,
+      "grad_norm": 36.75,
+      "grad_norm_var": 18.972916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5705,
+      "loss/crossentropy": 2.0975339651107787,
+      "loss/hidden": 3.435546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19847188219428064,
+      "step": 6390
+    },
+    {
+      "epoch": 0.16,
+      "grad_norm": 31.375,
+      "grad_norm_var": 64.11770833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6051,
+      "loss/crossentropy": 2.0802886128425597,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20291686560958624,
+      "step": 6400
+    },
+    {
+      "epoch": 0.16025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 74.28958333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4922,
+      "loss/crossentropy": 2.1224244251847266,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.194674601405859,
+      "step": 6410
+    },
+    {
+      "epoch": 0.1605,
+      "grad_norm": 62.5,
+      "grad_norm_var": 69.86041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6584,
+      "loss/crossentropy": 2.1936039954423903,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19804685264825822,
+      "step": 6420
+    },
+    {
+      "epoch": 0.16075,
+      "grad_norm": 36.5,
+      "grad_norm_var": 66.04576822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4329,
+      "loss/crossentropy": 2.103436988592148,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19700358025729656,
+      "step": 6430
+    },
+    {
+      "epoch": 0.161,
+      "grad_norm": 28.875,
+      "grad_norm_var": 8.187239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5397,
+      "loss/crossentropy": 2.107574874162674,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1931281227618456,
+      "step": 6440
+    },
+    {
+      "epoch": 0.16125,
+      "grad_norm": 32.75,
+      "grad_norm_var": 20.245768229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5289,
+      "loss/crossentropy": 2.251564306020737,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20565700754523278,
+      "step": 6450
+    },
+    {
+      "epoch": 0.1615,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.187239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4563,
+      "loss/crossentropy": 2.0432717867195604,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17879284229129552,
+      "step": 6460
+    },
+    {
+      "epoch": 0.16175,
+      "grad_norm": 40.0,
+      "grad_norm_var": 8.5416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5052,
+      "loss/crossentropy": 2.130405417084694,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868636028841138,
+      "step": 6470
+    },
+    {
+      "epoch": 0.162,
+      "grad_norm": 30.625,
+      "grad_norm_var": 8.284375,
+      "learning_rate": 0.0001,
+      "loss": 7.4387,
+      "loss/crossentropy": 2.1113599717617033,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20160141140222548,
+      "step": 6480
+    },
+    {
+      "epoch": 0.16225,
+      "grad_norm": 30.5,
+      "grad_norm_var": 10.103125,
+      "learning_rate": 0.0001,
+      "loss": 7.4812,
+      "loss/crossentropy": 2.177107959985733,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19870908968150616,
+      "step": 6490
+    },
+    {
+      "epoch": 0.1625,
+      "grad_norm": 29.0,
+      "grad_norm_var": 12.816080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5634,
+      "loss/crossentropy": 2.2469130218029023,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22473043091595174,
+      "step": 6500
+    },
+    {
+      "epoch": 0.16275,
+      "grad_norm": 62.75,
+      "grad_norm_var": 67.71555989583334,
+      "learning_rate": 0.0001,
+      "loss": 7.4504,
+      "loss/crossentropy": 2.1134796291589737,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20341392308473588,
+      "step": 6510
+    },
+    {
+      "epoch": 0.163,
+      "grad_norm": 38.5,
+      "grad_norm_var": 132.73723958333332,
+      "learning_rate": 0.0001,
+      "loss": 7.6099,
+      "loss/crossentropy": 2.0655704945325852,
+      "loss/hidden": 3.45625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2016578745096922,
+      "step": 6520
+    },
+    {
+      "epoch": 0.16325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 117.9541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.668,
+      "loss/crossentropy": 2.171022225916386,
+      "loss/hidden": 3.575,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23951137959957122,
+      "step": 6530
+    },
+    {
+      "epoch": 0.1635,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.513541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5344,
+      "loss/crossentropy": 2.1631700932979583,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19134688563644886,
+      "step": 6540
+    },
+    {
+      "epoch": 0.16375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 10.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5036,
+      "loss/crossentropy": 2.154293045401573,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2056902015581727,
+      "step": 6550
+    },
+    {
+      "epoch": 0.164,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.233268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4959,
+      "loss/crossentropy": 2.077804160118103,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18005848359316587,
+      "step": 6560
+    },
+    {
+      "epoch": 0.16425,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.2343098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3855,
+      "loss/crossentropy": 2.238145849108696,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20202569160610437,
+      "step": 6570
+    },
+    {
+      "epoch": 0.1645,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.284375,
+      "learning_rate": 0.0001,
+      "loss": 7.5453,
+      "loss/crossentropy": 2.1169906362891195,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1980356415733695,
+      "step": 6580
+    },
+    {
+      "epoch": 0.16475,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.6848307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5151,
+      "loss/crossentropy": 2.165386658906937,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19807699434459208,
+      "step": 6590
+    },
+    {
+      "epoch": 0.165,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.905989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4242,
+      "loss/crossentropy": 2.2100414454936983,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21563767176121473,
+      "step": 6600
+    },
+    {
+      "epoch": 0.16525,
+      "grad_norm": 28.25,
+      "grad_norm_var": 21.574934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4757,
+      "loss/crossentropy": 2.1352153360843658,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.176172699034214,
+      "step": 6610
+    },
+    {
+      "epoch": 0.1655,
+      "grad_norm": 32.0,
+      "grad_norm_var": 9.960872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5596,
+      "loss/crossentropy": 2.1773680597543716,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18952815048396587,
+      "step": 6620
+    },
+    {
+      "epoch": 0.16575,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.380143229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.56,
+      "loss/crossentropy": 2.1314379185438157,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19869489334523677,
+      "step": 6630
+    },
+    {
+      "epoch": 0.166,
+      "grad_norm": 32.75,
+      "grad_norm_var": 65.24765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5285,
+      "loss/crossentropy": 2.1807857722043993,
+      "loss/hidden": 3.44453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20406192895025016,
+      "step": 6640
+    },
+    {
+      "epoch": 0.16625,
+      "grad_norm": 36.75,
+      "grad_norm_var": 61.514322916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.5774,
+      "loss/crossentropy": 2.1761706352233885,
+      "loss/hidden": 3.490625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22523548007011412,
+      "step": 6650
+    },
+    {
+      "epoch": 0.1665,
+      "grad_norm": 40.0,
+      "grad_norm_var": 10.858072916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4837,
+      "loss/crossentropy": 2.0422482162714006,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19435476139187813,
+      "step": 6660
+    },
+    {
+      "epoch": 0.16675,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.977083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6686,
+      "loss/crossentropy": 2.182368849217892,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20404126346111298,
+      "step": 6670
+    },
+    {
+      "epoch": 0.167,
+      "grad_norm": 32.0,
+      "grad_norm_var": 11.715625,
+      "learning_rate": 0.0001,
+      "loss": 7.5378,
+      "loss/crossentropy": 2.246191081404686,
+      "loss/hidden": 3.28359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18508785720914603,
+      "step": 6680
+    },
+    {
+      "epoch": 0.16725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.314322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5093,
+      "loss/crossentropy": 2.1791825108230114,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18074298240244388,
+      "step": 6690
+    },
+    {
+      "epoch": 0.1675,
+      "grad_norm": 30.75,
+      "grad_norm_var": 8.992708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5574,
+      "loss/crossentropy": 2.1036971658468246,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981421067379415,
+      "step": 6700
+    },
+    {
+      "epoch": 0.16775,
+      "grad_norm": 29.5,
+      "grad_norm_var": 14.267643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3344,
+      "loss/crossentropy": 2.001809497177601,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1832934280857444,
+      "step": 6710
+    },
+    {
+      "epoch": 0.168,
+      "grad_norm": 61.25,
+      "grad_norm_var": 63.89375,
+      "learning_rate": 0.0001,
+      "loss": 7.5556,
+      "loss/crossentropy": 2.1391955494880674,
+      "loss/hidden": 3.49375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20318875089287758,
+      "step": 6720
+    },
+    {
+      "epoch": 0.16825,
+      "grad_norm": 33.75,
+      "grad_norm_var": 56.9791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.556,
+      "loss/crossentropy": 2.155173195898533,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19701551645994186,
+      "step": 6730
+    },
+    {
+      "epoch": 0.1685,
+      "grad_norm": 30.5,
+      "grad_norm_var": 9.07890625,
+      "learning_rate": 0.0001,
+      "loss": 7.444,
+      "loss/crossentropy": 2.1643510669469834,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19851317517459394,
+      "step": 6740
+    },
+    {
+      "epoch": 0.16875,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.964518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4711,
+      "loss/crossentropy": 2.1824121534824372,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20471413098275662,
+      "step": 6750
+    },
+    {
+      "epoch": 0.169,
+      "grad_norm": 32.75,
+      "grad_norm_var": 6.1009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4304,
+      "loss/crossentropy": 2.0600946068763735,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18951213210821152,
+      "step": 6760
+    },
+    {
+      "epoch": 0.16925,
+      "grad_norm": 34.25,
+      "grad_norm_var": 6.034309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5587,
+      "loss/crossentropy": 2.1303545042872427,
+      "loss/hidden": 3.32578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19376144148409366,
+      "step": 6770
+    },
+    {
+      "epoch": 0.1695,
+      "grad_norm": 35.75,
+      "grad_norm_var": 8.519205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4825,
+      "loss/crossentropy": 2.223961615562439,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993227696046233,
+      "step": 6780
+    },
+    {
+      "epoch": 0.16975,
+      "grad_norm": 52.5,
+      "grad_norm_var": 31.4837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4688,
+      "loss/crossentropy": 2.1645863845944406,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19716022843495012,
+      "step": 6790
+    },
+    {
+      "epoch": 0.17,
+      "grad_norm": 32.75,
+      "grad_norm_var": 29.0400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5153,
+      "loss/crossentropy": 2.251170714199543,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21400584764778613,
+      "step": 6800
+    },
+    {
+      "epoch": 0.17025,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.945833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4593,
+      "loss/crossentropy": 2.1392437756061553,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19613203704357146,
+      "step": 6810
+    },
+    {
+      "epoch": 0.1705,
+      "grad_norm": 31.5,
+      "grad_norm_var": 9.330989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4589,
+      "loss/crossentropy": 2.1152432590723036,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1848454039543867,
+      "step": 6820
+    },
+    {
+      "epoch": 0.17075,
+      "grad_norm": 33.5,
+      "grad_norm_var": 15.155208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5984,
+      "loss/crossentropy": 2.208776918053627,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20963791087269784,
+      "step": 6830
+    },
+    {
+      "epoch": 0.171,
+      "grad_norm": 31.0,
+      "grad_norm_var": 14.967708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5348,
+      "loss/crossentropy": 2.1368553161621096,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2036329936236143,
+      "step": 6840
+    },
+    {
+      "epoch": 0.17125,
+      "grad_norm": 33.25,
+      "grad_norm_var": 11.615625,
+      "learning_rate": 0.0001,
+      "loss": 7.4543,
+      "loss/crossentropy": 2.2001425683498383,
+      "loss/hidden": 3.291015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18403711169958115,
+      "step": 6850
+    },
+    {
+      "epoch": 0.1715,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.045247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4924,
+      "loss/crossentropy": 2.0965635985136033,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117785301059484,
+      "step": 6860
+    },
+    {
+      "epoch": 0.17175,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.587434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4879,
+      "loss/crossentropy": 2.106214761734009,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19950247537344695,
+      "step": 6870
+    },
+    {
+      "epoch": 0.172,
+      "grad_norm": 30.75,
+      "grad_norm_var": 8.289518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4561,
+      "loss/crossentropy": 2.1776740878820418,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19133987911045552,
+      "step": 6880
+    },
+    {
+      "epoch": 0.17225,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.77265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4422,
+      "loss/crossentropy": 2.082959216833115,
+      "loss/hidden": 3.544921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2135216325521469,
+      "step": 6890
+    },
+    {
+      "epoch": 0.1725,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.1556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.6053,
+      "loss/crossentropy": 2.2018980890512467,
+      "loss/hidden": 3.495703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21013570427894593,
+      "step": 6900
+    },
+    {
+      "epoch": 0.17275,
+      "grad_norm": 39.5,
+      "grad_norm_var": 6.5322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5447,
+      "loss/crossentropy": 2.2833516895771027,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19603817090392112,
+      "step": 6910
+    },
+    {
+      "epoch": 0.173,
+      "grad_norm": 31.125,
+      "grad_norm_var": 12.677018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4817,
+      "loss/crossentropy": 2.166572627425194,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19017961733043193,
+      "step": 6920
+    },
+    {
+      "epoch": 0.17325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.6572,
+      "loss/crossentropy": 2.1658833861351012,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22312654759734868,
+      "step": 6930
+    },
+    {
+      "epoch": 0.1735,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.5747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4706,
+      "loss/crossentropy": 2.1533073887228964,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20075356736779212,
+      "step": 6940
+    },
+    {
+      "epoch": 0.17375,
+      "grad_norm": 33.0,
+      "grad_norm_var": 7.611393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6678,
+      "loss/crossentropy": 2.2583969831466675,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2209668504074216,
+      "step": 6950
+    },
+    {
+      "epoch": 0.174,
+      "grad_norm": 30.0,
+      "grad_norm_var": 11.0087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3805,
+      "loss/crossentropy": 2.1416700780391693,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20209096949547528,
+      "step": 6960
+    },
+    {
+      "epoch": 0.17425,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.171875,
+      "learning_rate": 0.0001,
+      "loss": 7.5472,
+      "loss/crossentropy": 2.1253055185079575,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19859984703361988,
+      "step": 6970
+    },
+    {
+      "epoch": 0.1745,
+      "grad_norm": 30.0,
+      "grad_norm_var": 6.6603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5648,
+      "loss/crossentropy": 2.182023701816797,
+      "loss/hidden": 3.29296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18740688636898994,
+      "step": 6980
+    },
+    {
+      "epoch": 0.17475,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.4947916666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4586,
+      "loss/crossentropy": 2.1650559276342394,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1944490736350417,
+      "step": 6990
+    },
+    {
+      "epoch": 0.175,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.6587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.569,
+      "loss/crossentropy": 2.177471086382866,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053987860679628,
+      "step": 7000
+    },
+    {
+      "epoch": 0.17525,
+      "grad_norm": 32.5,
+      "grad_norm_var": 26.516080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4753,
+      "loss/crossentropy": 2.154606765508652,
+      "loss/hidden": 3.423828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19910989850759506,
+      "step": 7010
+    },
+    {
+      "epoch": 0.1755,
+      "grad_norm": 34.5,
+      "grad_norm_var": 39.670572916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.631,
+      "loss/crossentropy": 2.197327023744583,
+      "loss/hidden": 3.541796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21263533756136893,
+      "step": 7020
+    },
+    {
+      "epoch": 0.17575,
+      "grad_norm": 32.25,
+      "grad_norm_var": 19.970833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5712,
+      "loss/crossentropy": 2.143507385253906,
+      "loss/hidden": 3.41796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20819550901651382,
+      "step": 7030
+    },
+    {
+      "epoch": 0.176,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.755989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6156,
+      "loss/crossentropy": 2.160289117693901,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2082687295973301,
+      "step": 7040
+    },
+    {
+      "epoch": 0.17625,
+      "grad_norm": 34.0,
+      "grad_norm_var": 1.6035807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5106,
+      "loss/crossentropy": 2.096612122654915,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2005817864090204,
+      "step": 7050
+    },
+    {
+      "epoch": 0.1765,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.7921223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4678,
+      "loss/crossentropy": 2.150402194261551,
+      "loss/hidden": 3.31953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18803946152329445,
+      "step": 7060
+    },
+    {
+      "epoch": 0.17675,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.7754557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5601,
+      "loss/crossentropy": 2.111455664038658,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1900298461318016,
+      "step": 7070
+    },
+    {
+      "epoch": 0.177,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.0497395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5888,
+      "loss/crossentropy": 2.172677582502365,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1932773429900408,
+      "step": 7080
+    },
+    {
+      "epoch": 0.17725,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.131705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.527,
+      "loss/crossentropy": 2.2066776901483536,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20891736317425966,
+      "step": 7090
+    },
+    {
+      "epoch": 0.1775,
+      "grad_norm": 33.0,
+      "grad_norm_var": 5.386713240151209e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6693,
+      "loss/crossentropy": 2.1923015132546424,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2022488683462143,
+      "step": 7100
+    },
+    {
+      "epoch": 0.17775,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.9284656292120494e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.557,
+      "loss/crossentropy": 2.0847490578889847,
+      "loss/hidden": 3.552734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2148862171918154,
+      "step": 7110
+    },
+    {
+      "epoch": 0.178,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.3369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4864,
+      "loss/crossentropy": 2.1249582156538964,
+      "loss/hidden": 3.296484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19168496485799552,
+      "step": 7120
+    },
+    {
+      "epoch": 0.17825,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.878125,
+      "learning_rate": 0.0001,
+      "loss": 7.36,
+      "loss/crossentropy": 2.173662793636322,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978321423754096,
+      "step": 7130
+    },
+    {
+      "epoch": 0.1785,
+      "grad_norm": 34.5,
+      "grad_norm_var": 3.9723307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5663,
+      "loss/crossentropy": 2.1832578271627425,
+      "loss/hidden": 3.4484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22951248846948147,
+      "step": 7140
+    },
+    {
+      "epoch": 0.17875,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.022330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5795,
+      "loss/crossentropy": 2.1229765862226486,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857854513451457,
+      "step": 7150
+    },
+    {
+      "epoch": 0.179,
+      "grad_norm": 30.25,
+      "grad_norm_var": 42.5978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4524,
+      "loss/crossentropy": 2.1213574737310408,
+      "loss/hidden": 3.4671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22352912444621326,
+      "step": 7160
+    },
+    {
+      "epoch": 0.17925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 44.90774739583333,
+      "learning_rate": 0.0001,
+      "loss": 7.5702,
+      "loss/crossentropy": 2.2016272962093355,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19035741928964853,
+      "step": 7170
+    },
+    {
+      "epoch": 0.1795,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.08515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5485,
+      "loss/crossentropy": 2.1106868594884873,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19449420645833015,
+      "step": 7180
+    },
+    {
+      "epoch": 0.17975,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.414322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4434,
+      "loss/crossentropy": 2.1368062049150467,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1976731013506651,
+      "step": 7190
+    },
+    {
+      "epoch": 0.18,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.5483723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4368,
+      "loss/crossentropy": 2.1658909559249877,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897917227819562,
+      "step": 7200
+    },
+    {
+      "epoch": 0.18025,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.39765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5235,
+      "loss/crossentropy": 2.0003643825650217,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881322082132101,
+      "step": 7210
+    },
+    {
+      "epoch": 0.1805,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.2369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5351,
+      "loss/crossentropy": 2.1550097078084947,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885005658492446,
+      "step": 7220
+    },
+    {
+      "epoch": 0.18075,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.6809895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3615,
+      "loss/crossentropy": 2.089699313044548,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18596011865884066,
+      "step": 7230
+    },
+    {
+      "epoch": 0.181,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.780989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.521,
+      "loss/crossentropy": 2.1315871119499206,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2100736267864704,
+      "step": 7240
+    },
+    {
+      "epoch": 0.18125,
+      "grad_norm": 34.75,
+      "grad_norm_var": 4.068684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4489,
+      "loss/crossentropy": 2.0509297475218773,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21798528153449298,
+      "step": 7250
+    },
+    {
+      "epoch": 0.1815,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.546875,
+      "learning_rate": 0.0001,
+      "loss": 7.5147,
+      "loss/crossentropy": 2.097806680202484,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1954873636364937,
+      "step": 7260
+    },
+    {
+      "epoch": 0.18175,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.9270833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5335,
+      "loss/crossentropy": 2.1867170676589014,
+      "loss/hidden": 3.471484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20327441468834878,
+      "step": 7270
+    },
+    {
+      "epoch": 0.182,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.0434895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5098,
+      "loss/crossentropy": 2.059905408322811,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984263988211751,
+      "step": 7280
+    },
+    {
+      "epoch": 0.18225,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.0004557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6258,
+      "loss/crossentropy": 2.19571368098259,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20168307051062584,
+      "step": 7290
+    },
+    {
+      "epoch": 0.1825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.7051432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5923,
+      "loss/crossentropy": 2.12765002399683,
+      "loss/hidden": 3.518359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19315951522439717,
+      "step": 7300
+    },
+    {
+      "epoch": 0.18275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.746875,
+      "learning_rate": 0.0001,
+      "loss": 7.5356,
+      "loss/crossentropy": 2.2317890375852585,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19801790416240692,
+      "step": 7310
+    },
+    {
+      "epoch": 0.183,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.206184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4805,
+      "loss/crossentropy": 2.115889000892639,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20432772096246482,
+      "step": 7320
+    },
+    {
+      "epoch": 0.18325,
+      "grad_norm": 29.875,
+      "grad_norm_var": 7.782291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5582,
+      "loss/crossentropy": 2.20442833006382,
+      "loss/hidden": 3.488671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20083660595119,
+      "step": 7330
+    },
+    {
+      "epoch": 0.1835,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.6947916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6294,
+      "loss/crossentropy": 2.210834649205208,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19781918153166772,
+      "step": 7340
+    },
+    {
+      "epoch": 0.18375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.7791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5733,
+      "loss/crossentropy": 2.094112278521061,
+      "loss/hidden": 3.41171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1902735486626625,
+      "step": 7350
+    },
+    {
+      "epoch": 0.184,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.3796223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4645,
+      "loss/crossentropy": 2.1219489932060243,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1875064203515649,
+      "step": 7360
+    },
+    {
+      "epoch": 0.18425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.262955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5375,
+      "loss/crossentropy": 2.1768325984478,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20637014620006083,
+      "step": 7370
+    },
+    {
+      "epoch": 0.1845,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.51015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5972,
+      "loss/crossentropy": 2.088278591632843,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19868529513478278,
+      "step": 7380
+    },
+    {
+      "epoch": 0.18475,
+      "grad_norm": 28.125,
+      "grad_norm_var": 7.376041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4234,
+      "loss/crossentropy": 2.162215715646744,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20086740422993898,
+      "step": 7390
+    },
+    {
+      "epoch": 0.185,
+      "grad_norm": 34.0,
+      "grad_norm_var": 141.65201822916666,
+      "learning_rate": 0.0001,
+      "loss": 7.4958,
+      "loss/crossentropy": 2.076205277442932,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904180521145463,
+      "step": 7400
+    },
+    {
+      "epoch": 0.18525,
+      "grad_norm": 31.125,
+      "grad_norm_var": 138.825,
+      "learning_rate": 0.0001,
+      "loss": 7.574,
+      "loss/crossentropy": 2.1355942092835902,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2197154738008976,
+      "step": 7410
+    },
+    {
+      "epoch": 0.1855,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.405847189186238e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4496,
+      "loss/crossentropy": 2.152878683805466,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18649988640099763,
+      "step": 7420
+    },
+    {
+      "epoch": 0.18575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.40584718949382e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4482,
+      "loss/crossentropy": 2.100820633769035,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1907477645203471,
+      "step": 7430
+    },
+    {
+      "epoch": 0.186,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.299934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.436,
+      "loss/crossentropy": 2.1752007991075515,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19189766012132167,
+      "step": 7440
+    },
+    {
+      "epoch": 0.18625,
+      "grad_norm": 30.25,
+      "grad_norm_var": 5.689322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5691,
+      "loss/crossentropy": 2.242092598974705,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19479812681674957,
+      "step": 7450
+    },
+    {
+      "epoch": 0.1865,
+      "grad_norm": 30.625,
+      "grad_norm_var": 5.6369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.6412,
+      "loss/crossentropy": 2.1058298379182814,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22841914147138595,
+      "step": 7460
+    },
+    {
+      "epoch": 0.18675,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.435872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3797,
+      "loss/crossentropy": 2.0960675440728664,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850858678109944,
+      "step": 7470
+    },
+    {
+      "epoch": 0.187,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.34765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5432,
+      "loss/crossentropy": 2.180506870150566,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20674382094293833,
+      "step": 7480
+    },
+    {
+      "epoch": 0.18725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5242,
+      "loss/crossentropy": 2.101006045937538,
+      "loss/hidden": 3.503125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19774735551327466,
+      "step": 7490
+    },
+    {
+      "epoch": 0.1875,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.7791666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5166,
+      "loss/crossentropy": 2.069665388762951,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19810831565409898,
+      "step": 7500
+    },
+    {
+      "epoch": 0.18775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.07890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4959,
+      "loss/crossentropy": 2.0420530915260313,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18399292901158332,
+      "step": 7510
+    },
+    {
+      "epoch": 0.188,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.0587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5325,
+      "loss/crossentropy": 2.1312338694930077,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19504649545997382,
+      "step": 7520
+    },
+    {
+      "epoch": 0.18825,
+      "grad_norm": 43.0,
+      "grad_norm_var": 9.308333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6141,
+      "loss/crossentropy": 2.2894584849476813,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20183086302131414,
+      "step": 7530
+    },
+    {
+      "epoch": 0.1885,
+      "grad_norm": 34.0,
+      "grad_norm_var": 10.281705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5206,
+      "loss/crossentropy": 2.1557983949780466,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19000910818576813,
+      "step": 7540
+    },
+    {
+      "epoch": 0.18875,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.21015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5249,
+      "loss/crossentropy": 2.124994584918022,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19348742607980968,
+      "step": 7550
+    },
+    {
+      "epoch": 0.189,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.334830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4858,
+      "loss/crossentropy": 2.154696786403656,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19012149404734374,
+      "step": 7560
+    },
+    {
+      "epoch": 0.18925,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.9872395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.487,
+      "loss/crossentropy": 2.1938005700707435,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19971045572310686,
+      "step": 7570
+    },
+    {
+      "epoch": 0.1895,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.5624176423088947e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4698,
+      "loss/crossentropy": 2.120267179608345,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19198095686733724,
+      "step": 7580
+    },
+    {
+      "epoch": 0.18975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.756184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4743,
+      "loss/crossentropy": 2.1569736182689665,
+      "loss/hidden": 3.53984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21590765863656997,
+      "step": 7590
+    },
+    {
+      "epoch": 0.19,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.9166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4344,
+      "loss/crossentropy": 2.1558305487036704,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18447545487433673,
+      "step": 7600
+    },
+    {
+      "epoch": 0.19025,
+      "grad_norm": 31.375,
+      "grad_norm_var": 4.1025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4889,
+      "loss/crossentropy": 2.0612099058926105,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19961078523192555,
+      "step": 7610
+    },
+    {
+      "epoch": 0.1905,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.3087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6558,
+      "loss/crossentropy": 2.04723744392395,
+      "loss/hidden": 3.5171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22622217051684856,
+      "step": 7620
+    },
+    {
+      "epoch": 0.19075,
+      "grad_norm": 29.25,
+      "grad_norm_var": 5.770833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4133,
+      "loss/crossentropy": 2.1897004157304765,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826148485764861,
+      "step": 7630
+    },
+    {
+      "epoch": 0.191,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.260416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4913,
+      "loss/crossentropy": 2.157025161385536,
+      "loss/hidden": 3.5171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22510031294077634,
+      "step": 7640
+    },
+    {
+      "epoch": 0.19125,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.2270182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.517,
+      "loss/crossentropy": 2.090579715371132,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889673013240099,
+      "step": 7650
+    },
+    {
+      "epoch": 0.1915,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.0770833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6179,
+      "loss/crossentropy": 2.1681595921516417,
+      "loss/hidden": 3.555078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2365275712683797,
+      "step": 7660
+    },
+    {
+      "epoch": 0.19175,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.4280598958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5348,
+      "loss/crossentropy": 2.207126745581627,
+      "loss/hidden": 3.34921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19211382810026406,
+      "step": 7670
+    },
+    {
+      "epoch": 0.192,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.796875,
+      "learning_rate": 0.0001,
+      "loss": 7.553,
+      "loss/crossentropy": 2.161148224771023,
+      "loss/hidden": 3.52734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22304172217845916,
+      "step": 7680
+    },
+    {
+      "epoch": 0.19225,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.8889973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5009,
+      "loss/crossentropy": 2.0679882526397706,
+      "loss/hidden": 3.46015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2086691778153181,
+      "step": 7690
+    },
+    {
+      "epoch": 0.1925,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.2552083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5372,
+      "loss/crossentropy": 2.0694913983345034,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19288846384733915,
+      "step": 7700
+    },
+    {
+      "epoch": 0.19275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.8072916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5833,
+      "loss/crossentropy": 2.2242089927196504,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1934451697394252,
+      "step": 7710
+    },
+    {
+      "epoch": 0.193,
+      "grad_norm": 34.5,
+      "grad_norm_var": 12.787955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6402,
+      "loss/crossentropy": 2.157019394636154,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19785581808537245,
+      "step": 7720
+    },
+    {
+      "epoch": 0.19325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 12.268489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3766,
+      "loss/crossentropy": 2.226413035392761,
+      "loss/hidden": 3.40625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20680981017649175,
+      "step": 7730
+    },
+    {
+      "epoch": 0.1935,
+      "grad_norm": 35.5,
+      "grad_norm_var": 6.34140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5761,
+      "loss/crossentropy": 2.0320019692182543,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19264598488807677,
+      "step": 7740
+    },
+    {
+      "epoch": 0.19375,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.897916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.453,
+      "loss/crossentropy": 2.0598455399274824,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851517017930746,
+      "step": 7750
+    },
+    {
+      "epoch": 0.194,
+      "grad_norm": 29.0,
+      "grad_norm_var": 6.7025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4278,
+      "loss/crossentropy": 2.1927490830421448,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20509992372244595,
+      "step": 7760
+    },
+    {
+      "epoch": 0.19425,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.700455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4874,
+      "loss/crossentropy": 2.154856140911579,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20458675045520067,
+      "step": 7770
+    },
+    {
+      "epoch": 0.1945,
+      "grad_norm": 37.5,
+      "grad_norm_var": 6.71015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5452,
+      "loss/crossentropy": 2.1202862530946733,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18985650166869164,
+      "step": 7780
+    },
+    {
+      "epoch": 0.19475,
+      "grad_norm": 36.25,
+      "grad_norm_var": 6.918684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5427,
+      "loss/crossentropy": 2.1549450784921644,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18924216069281102,
+      "step": 7790
+    },
+    {
+      "epoch": 0.195,
+      "grad_norm": 35.25,
+      "grad_norm_var": 6.2291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4733,
+      "loss/crossentropy": 2.0557655200362204,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1919045101851225,
+      "step": 7800
+    },
+    {
+      "epoch": 0.19525,
+      "grad_norm": 46.75,
+      "grad_norm_var": 50.040625,
+      "learning_rate": 0.0001,
+      "loss": 7.4911,
+      "loss/crossentropy": 2.171146012097597,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20546058733016254,
+      "step": 7810
+    },
+    {
+      "epoch": 0.1955,
+      "grad_norm": 33.25,
+      "grad_norm_var": 15.828580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5692,
+      "loss/crossentropy": 2.1003697514533997,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20100659225136042,
+      "step": 7820
+    },
+    {
+      "epoch": 0.19575,
+      "grad_norm": 36.0,
+      "grad_norm_var": 29.226497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5267,
+      "loss/crossentropy": 2.1091859377920628,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19987453976646066,
+      "step": 7830
+    },
+    {
+      "epoch": 0.196,
+      "grad_norm": 48.5,
+      "grad_norm_var": 42.45201822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.111352452635765,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18483758307993411,
+      "step": 7840
+    },
+    {
+      "epoch": 0.19625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 25.65390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5415,
+      "loss/crossentropy": 2.0419967979192735,
+      "loss/hidden": 3.464453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2173704007640481,
+      "step": 7850
+    },
+    {
+      "epoch": 0.1965,
+      "grad_norm": 30.875,
+      "grad_norm_var": 21.546875,
+      "learning_rate": 0.0001,
+      "loss": 7.4951,
+      "loss/crossentropy": 2.084785957634449,
+      "loss/hidden": 3.304296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17884304635226728,
+      "step": 7860
+    },
+    {
+      "epoch": 0.19675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 18.153059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4509,
+      "loss/crossentropy": 2.1023138776421546,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20335167981684207,
+      "step": 7870
+    },
+    {
+      "epoch": 0.197,
+      "grad_norm": 36.25,
+      "grad_norm_var": 12.093489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5467,
+      "loss/crossentropy": 2.171536546945572,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21221924368292094,
+      "step": 7880
+    },
+    {
+      "epoch": 0.19725,
+      "grad_norm": 28.75,
+      "grad_norm_var": 9.983333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3785,
+      "loss/crossentropy": 2.1119479715824125,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18313394635915756,
+      "step": 7890
+    },
+    {
+      "epoch": 0.1975,
+      "grad_norm": 37.25,
+      "grad_norm_var": 11.928580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4945,
+      "loss/crossentropy": 1.9945223838090897,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19973532520234585,
+      "step": 7900
+    },
+    {
+      "epoch": 0.19775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 11.839583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4542,
+      "loss/crossentropy": 2.1550634860992433,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19187356438487768,
+      "step": 7910
+    },
+    {
+      "epoch": 0.198,
+      "grad_norm": 31.5,
+      "grad_norm_var": 12.584309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4245,
+      "loss/crossentropy": 2.092045524716377,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18641792330890894,
+      "step": 7920
+    },
+    {
+      "epoch": 0.19825,
+      "grad_norm": 28.875,
+      "grad_norm_var": 12.959375,
+      "learning_rate": 0.0001,
+      "loss": 7.4616,
+      "loss/crossentropy": 2.1302111998200415,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2105752557516098,
+      "step": 7930
+    },
+    {
+      "epoch": 0.1985,
+      "grad_norm": 30.75,
+      "grad_norm_var": 14.9431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4869,
+      "loss/crossentropy": 2.1739411771297457,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19759192429482936,
+      "step": 7940
+    },
+    {
+      "epoch": 0.19875,
+      "grad_norm": 33.25,
+      "grad_norm_var": 12.703580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4435,
+      "loss/crossentropy": 2.0400329381227493,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18711388055235148,
+      "step": 7950
+    },
+    {
+      "epoch": 0.199,
+      "grad_norm": 40.0,
+      "grad_norm_var": 10.454622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3235,
+      "loss/crossentropy": 2.1173346668481825,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.200854654237628,
+      "step": 7960
+    },
+    {
+      "epoch": 0.19925,
+      "grad_norm": 31.125,
+      "grad_norm_var": 18.376822916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.551,
+      "loss/crossentropy": 2.256546127796173,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2098003163933754,
+      "step": 7970
+    },
+    {
+      "epoch": 0.1995,
+      "grad_norm": 30.625,
+      "grad_norm_var": 11.159375,
+      "learning_rate": 0.0001,
+      "loss": 7.5863,
+      "loss/crossentropy": 2.0988403081893923,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2052627831697464,
+      "step": 7980
+    },
+    {
+      "epoch": 0.19975,
+      "grad_norm": 34.75,
+      "grad_norm_var": 3.539322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5501,
+      "loss/crossentropy": 2.157486143708229,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19415573356673121,
+      "step": 7990
+    },
+    {
+      "epoch": 0.2,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.282291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5487,
+      "loss/crossentropy": 2.204603946208954,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19015649259090422,
+      "step": 8000
+    },
+    {
+      "epoch": 0.20025,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.59375,
+      "learning_rate": 0.0001,
+      "loss": 7.4113,
+      "loss/crossentropy": 2.1252332285046576,
+      "loss/hidden": 3.287109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1751571564003825,
+      "step": 8010
+    },
+    {
+      "epoch": 0.2005,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.498,
+      "loss/crossentropy": 2.048645743727684,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1930268170312047,
+      "step": 8020
+    },
+    {
+      "epoch": 0.20075,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.3822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4496,
+      "loss/crossentropy": 2.1435810789465903,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19256617035716772,
+      "step": 8030
+    },
+    {
+      "epoch": 0.201,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.8837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5991,
+      "loss/crossentropy": 2.06428082883358,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18795610759407283,
+      "step": 8040
+    },
+    {
+      "epoch": 0.20125,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.08515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5106,
+      "loss/crossentropy": 2.2008739590644835,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1890289282426238,
+      "step": 8050
+    },
+    {
+      "epoch": 0.2015,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.7708333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4302,
+      "loss/crossentropy": 2.1280903786420824,
+      "loss/hidden": 3.511328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20874056480824948,
+      "step": 8060
+    },
+    {
+      "epoch": 0.20175,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.2035807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4529,
+      "loss/crossentropy": 2.1864554077386855,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19697167426347734,
+      "step": 8070
+    },
+    {
+      "epoch": 0.202,
+      "grad_norm": 35.0,
+      "grad_norm_var": 2.278125,
+      "learning_rate": 0.0001,
+      "loss": 7.5744,
+      "loss/crossentropy": 2.211556833982468,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20347684845328332,
+      "step": 8080
+    },
+    {
+      "epoch": 0.20225,
+      "grad_norm": 28.875,
+      "grad_norm_var": 33.37890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5018,
+      "loss/crossentropy": 2.185184845328331,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2022322118282318,
+      "step": 8090
+    },
+    {
+      "epoch": 0.2025,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.3931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5218,
+      "loss/crossentropy": 2.1828758120536804,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1962041450664401,
+      "step": 8100
+    },
+    {
+      "epoch": 0.20275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.9488932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4982,
+      "loss/crossentropy": 2.159992370009422,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20945403575897217,
+      "step": 8110
+    },
+    {
+      "epoch": 0.203,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.748958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4657,
+      "loss/crossentropy": 2.1400886207818983,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19184730034321545,
+      "step": 8120
+    },
+    {
+      "epoch": 0.20325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.0733723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5559,
+      "loss/crossentropy": 2.112087991833687,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20091654695570468,
+      "step": 8130
+    },
+    {
+      "epoch": 0.2035,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.921875,
+      "learning_rate": 0.0001,
+      "loss": 7.5644,
+      "loss/crossentropy": 2.139628532528877,
+      "loss/hidden": 3.52265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19385650400072335,
+      "step": 8140
+    },
+    {
+      "epoch": 0.20375,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.379166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5002,
+      "loss/crossentropy": 2.091212958097458,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2040353151038289,
+      "step": 8150
+    },
+    {
+      "epoch": 0.204,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.2228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4494,
+      "loss/crossentropy": 2.2567614316940308,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18545747231692075,
+      "step": 8160
+    },
+    {
+      "epoch": 0.20425,
+      "grad_norm": 31.125,
+      "grad_norm_var": 229.56666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4626,
+      "loss/crossentropy": 2.1603645354509355,
+      "loss/hidden": 3.539453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21300703901797532,
+      "step": 8170
+    },
+    {
+      "epoch": 0.2045,
+      "grad_norm": 28.625,
+      "grad_norm_var": 4.2806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5117,
+      "loss/crossentropy": 2.1204755783081053,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18535650707781315,
+      "step": 8180
+    },
+    {
+      "epoch": 0.20475,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.1072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5637,
+      "loss/crossentropy": 2.1639575958251953,
+      "loss/hidden": 3.398828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20935236364603044,
+      "step": 8190
+    },
+    {
+      "epoch": 0.205,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.2494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3019,
+      "loss/crossentropy": 2.056389382481575,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874819153919816,
+      "step": 8200
+    },
+    {
+      "epoch": 0.20525,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.274934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5162,
+      "loss/crossentropy": 2.1393348813056945,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20481376871466636,
+      "step": 8210
+    },
+    {
+      "epoch": 0.2055,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.553125,
+      "learning_rate": 0.0001,
+      "loss": 7.5046,
+      "loss/crossentropy": 2.1898138865828516,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1965622954070568,
+      "step": 8220
+    },
+    {
+      "epoch": 0.20575,
+      "grad_norm": 34.5,
+      "grad_norm_var": 2.899825551450536e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6057,
+      "loss/crossentropy": 2.139435464143753,
+      "loss/hidden": 3.51328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20182280670851468,
+      "step": 8230
+    },
+    {
+      "epoch": 0.206,
+      "grad_norm": 37.75,
+      "grad_norm_var": 2.899825549761839e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5603,
+      "loss/crossentropy": 2.248063361644745,
+      "loss/hidden": 3.44609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21399107333272696,
+      "step": 8240
+    },
+    {
+      "epoch": 0.20625,
+      "grad_norm": 33.5,
+      "grad_norm_var": 114.23118489583334,
+      "learning_rate": 0.0001,
+      "loss": 7.4545,
+      "loss/crossentropy": 2.140778873860836,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18022785130888225,
+      "step": 8250
+    },
+    {
+      "epoch": 0.2065,
+      "grad_norm": 30.5,
+      "grad_norm_var": 110.5072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5364,
+      "loss/crossentropy": 2.259766247868538,
+      "loss/hidden": 3.312109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19342643208801746,
+      "step": 8260
+    },
+    {
+      "epoch": 0.20675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 0.7666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6531,
+      "loss/crossentropy": 2.158949288725853,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19808744061738254,
+      "step": 8270
+    },
+    {
+      "epoch": 0.207,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.4311848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6092,
+      "loss/crossentropy": 2.111227548122406,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20227425117045642,
+      "step": 8280
+    },
+    {
+      "epoch": 0.20725,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.56640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5372,
+      "loss/crossentropy": 2.0839753821492195,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18574036825448276,
+      "step": 8290
+    },
+    {
+      "epoch": 0.2075,
+      "grad_norm": 31.375,
+      "grad_norm_var": 0.8384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.536,
+      "loss/crossentropy": 2.1465844094753264,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2083981443196535,
+      "step": 8300
+    },
+    {
+      "epoch": 0.20775,
+      "grad_norm": 32.5,
+      "grad_norm_var": 20.8416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4722,
+      "loss/crossentropy": 2.1032681286334993,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19353157542645932,
+      "step": 8310
+    },
+    {
+      "epoch": 0.208,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.1395833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4453,
+      "loss/crossentropy": 2.092882976680994,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20083120074123145,
+      "step": 8320
+    },
+    {
+      "epoch": 0.20825,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.42265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4658,
+      "loss/crossentropy": 2.047010327875614,
+      "loss/hidden": 3.26484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17298451280221344,
+      "step": 8330
+    },
+    {
+      "epoch": 0.2085,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.8114583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6227,
+      "loss/crossentropy": 2.1705314934253694,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2104344118386507,
+      "step": 8340
+    },
+    {
+      "epoch": 0.20875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 15.13515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5292,
+      "loss/crossentropy": 2.251467025279999,
+      "loss/hidden": 3.44296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.206053901091218,
+      "step": 8350
+    },
+    {
+      "epoch": 0.209,
+      "grad_norm": 30.25,
+      "grad_norm_var": 9.75,
+      "learning_rate": 0.0001,
+      "loss": 7.5297,
+      "loss/crossentropy": 2.2416324824094773,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18910795077681541,
+      "step": 8360
+    },
+    {
+      "epoch": 0.20925,
+      "grad_norm": 49.75,
+      "grad_norm_var": 177.46145833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4604,
+      "loss/crossentropy": 1.9895980581641197,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19730695113539695,
+      "step": 8370
+    },
+    {
+      "epoch": 0.2095,
+      "grad_norm": 56.25,
+      "grad_norm_var": 334.9145182291667,
+      "learning_rate": 0.0001,
+      "loss": 7.5698,
+      "loss/crossentropy": 2.2018245279788973,
+      "loss/hidden": 3.38828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21164307594299317,
+      "step": 8380
+    },
+    {
+      "epoch": 0.20975,
+      "grad_norm": 32.75,
+      "grad_norm_var": 72.22890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5902,
+      "loss/crossentropy": 2.1519288808107375,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19217228647321463,
+      "step": 8390
+    },
+    {
+      "epoch": 0.21,
+      "grad_norm": 30.0,
+      "grad_norm_var": 7.153059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.451,
+      "loss/crossentropy": 2.1779698938131333,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1994494464248419,
+      "step": 8400
+    },
+    {
+      "epoch": 0.21025,
+      "grad_norm": 32.0,
+      "grad_norm_var": 7.124739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5019,
+      "loss/crossentropy": 2.140475983917713,
+      "loss/hidden": 3.3734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20144703481346368,
+      "step": 8410
+    },
+    {
+      "epoch": 0.2105,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.2639973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4268,
+      "loss/crossentropy": 2.2145755022764204,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19062853194773197,
+      "step": 8420
+    },
+    {
+      "epoch": 0.21075,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.620247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5149,
+      "loss/crossentropy": 2.3047141253948213,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20336541533470154,
+      "step": 8430
+    },
+    {
+      "epoch": 0.211,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.7514973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5319,
+      "loss/crossentropy": 2.232821634411812,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21666326355189086,
+      "step": 8440
+    },
+    {
+      "epoch": 0.21125,
+      "grad_norm": 33.5,
+      "grad_norm_var": 1.49765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5718,
+      "loss/crossentropy": 2.092620450258255,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20793931670486926,
+      "step": 8450
+    },
+    {
+      "epoch": 0.2115,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.1306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4344,
+      "loss/crossentropy": 2.151240213960409,
+      "loss/hidden": 3.40078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2046731175854802,
+      "step": 8460
+    },
+    {
+      "epoch": 0.21175,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.81640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5581,
+      "loss/crossentropy": 2.1520677715539933,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19603268206119537,
+      "step": 8470
+    },
+    {
+      "epoch": 0.212,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.092643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5382,
+      "loss/crossentropy": 2.1510326713323593,
+      "loss/hidden": 3.33359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18077819589525462,
+      "step": 8480
+    },
+    {
+      "epoch": 0.21225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.753059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4642,
+      "loss/crossentropy": 2.0811485938727854,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18939817287027835,
+      "step": 8490
+    },
+    {
+      "epoch": 0.2125,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.855989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5158,
+      "loss/crossentropy": 2.0734725266695024,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18070020200684667,
+      "step": 8500
+    },
+    {
+      "epoch": 0.21275,
+      "grad_norm": 29.25,
+      "grad_norm_var": 5.5462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4723,
+      "loss/crossentropy": 2.144814722239971,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19649898763746024,
+      "step": 8510
+    },
+    {
+      "epoch": 0.213,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.2309895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3777,
+      "loss/crossentropy": 2.0877407863736153,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19544295389205218,
+      "step": 8520
+    },
+    {
+      "epoch": 0.21325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6583333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5179,
+      "loss/crossentropy": 2.1433625385165214,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19359017722308636,
+      "step": 8530
+    },
+    {
+      "epoch": 0.2135,
+      "grad_norm": 34.25,
+      "grad_norm_var": 1.6510416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4658,
+      "loss/crossentropy": 2.221453693509102,
+      "loss/hidden": 3.43125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20547619834542274,
+      "step": 8540
+    },
+    {
+      "epoch": 0.21375,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.5791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5163,
+      "loss/crossentropy": 2.213938871026039,
+      "loss/hidden": 3.328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20213278364390136,
+      "step": 8550
+    },
+    {
+      "epoch": 0.214,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.784375,
+      "learning_rate": 0.0001,
+      "loss": 7.6361,
+      "loss/crossentropy": 2.133354830741882,
+      "loss/hidden": 3.512109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.207733928412199,
+      "step": 8560
+    },
+    {
+      "epoch": 0.21425,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.219205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5661,
+      "loss/crossentropy": 2.0587344974279405,
+      "loss/hidden": 3.550390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22256299555301667,
+      "step": 8570
+    },
+    {
+      "epoch": 0.2145,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.805989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5549,
+      "loss/crossentropy": 2.0949306935071945,
+      "loss/hidden": 3.4796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19931861553341151,
+      "step": 8580
+    },
+    {
+      "epoch": 0.21475,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.095247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.478,
+      "loss/crossentropy": 2.1063201159238814,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20397470220923425,
+      "step": 8590
+    },
+    {
+      "epoch": 0.215,
+      "grad_norm": 33.5,
+      "grad_norm_var": 28.859375,
+      "learning_rate": 0.0001,
+      "loss": 7.5632,
+      "loss/crossentropy": 2.177382430434227,
+      "loss/hidden": 3.46875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2048982411623001,
+      "step": 8600
+    },
+    {
+      "epoch": 0.21525,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.7955729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5325,
+      "loss/crossentropy": 2.0719266816973687,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19100363925099373,
+      "step": 8610
+    },
+    {
+      "epoch": 0.2155,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.7372395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5877,
+      "loss/crossentropy": 2.147695118188858,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19447963647544383,
+      "step": 8620
+    },
+    {
+      "epoch": 0.21575,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.406184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5509,
+      "loss/crossentropy": 2.1029836043715475,
+      "loss/hidden": 3.45,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19613635465502738,
+      "step": 8630
+    },
+    {
+      "epoch": 0.216,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.32265625,
+      "learning_rate": 0.0001,
+      "loss": 7.555,
+      "loss/crossentropy": 2.1027876287698746,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18709390722215175,
+      "step": 8640
+    },
+    {
+      "epoch": 0.21625,
+      "grad_norm": 31.75,
+      "grad_norm_var": 7.4978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.481,
+      "loss/crossentropy": 2.078503981232643,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19765212927013637,
+      "step": 8650
+    },
+    {
+      "epoch": 0.2165,
+      "grad_norm": 28.375,
+      "grad_norm_var": 6.212434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5951,
+      "loss/crossentropy": 2.1425766557455064,
+      "loss/hidden": 3.522265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2285369848832488,
+      "step": 8660
+    },
+    {
+      "epoch": 0.21675,
+      "grad_norm": 37.0,
+      "grad_norm_var": 5.02890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4619,
+      "loss/crossentropy": 2.094095268845558,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18905023839324714,
+      "step": 8670
+    },
+    {
+      "epoch": 0.217,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.205989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6255,
+      "loss/crossentropy": 2.0726657152175902,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19401923939585686,
+      "step": 8680
+    },
+    {
+      "epoch": 0.21725,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.8240281960963725e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.476,
+      "loss/crossentropy": 2.1063093028962614,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18630194589495658,
+      "step": 8690
+    },
+    {
+      "epoch": 0.2175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 9.2947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6565,
+      "loss/crossentropy": 2.188694643974304,
+      "loss/hidden": 3.561328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21225023418664932,
+      "step": 8700
+    },
+    {
+      "epoch": 0.21775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.644791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4697,
+      "loss/crossentropy": 2.1252513483166693,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19880297109484674,
+      "step": 8710
+    },
+    {
+      "epoch": 0.218,
+      "grad_norm": 30.875,
+      "grad_norm_var": 11.028059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.1700349181890486,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20139609836041927,
+      "step": 8720
+    },
+    {
+      "epoch": 0.21825,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.4189153379885778e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.616,
+      "loss/crossentropy": 2.032346047461033,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1754102895502001,
+      "step": 8730
+    },
+    {
+      "epoch": 0.2185,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.25390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5088,
+      "loss/crossentropy": 2.1633499920368195,
+      "loss/hidden": 3.448828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1996068175882101,
+      "step": 8740
+    },
+    {
+      "epoch": 0.21875,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.6122395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5088,
+      "loss/crossentropy": 2.1443722933530807,
+      "loss/hidden": 3.501171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2017217763699591,
+      "step": 8750
+    },
+    {
+      "epoch": 0.219,
+      "grad_norm": 32.25,
+      "grad_norm_var": 14.970572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5224,
+      "loss/crossentropy": 2.176077055931091,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2048908168449998,
+      "step": 8760
+    },
+    {
+      "epoch": 0.21925,
+      "grad_norm": 31.0,
+      "grad_norm_var": 14.087955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5417,
+      "loss/crossentropy": 2.183764386177063,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20730934757739305,
+      "step": 8770
+    },
+    {
+      "epoch": 0.2195,
+      "grad_norm": 46.5,
+      "grad_norm_var": 23.55390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5464,
+      "loss/crossentropy": 2.0760655224323274,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18335200790315867,
+      "step": 8780
+    },
+    {
+      "epoch": 0.21975,
+      "grad_norm": 31.125,
+      "grad_norm_var": 16.1369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5001,
+      "loss/crossentropy": 2.215297505259514,
+      "loss/hidden": 3.33671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053485784679652,
+      "step": 8790
+    },
+    {
+      "epoch": 0.22,
+      "grad_norm": 41.75,
+      "grad_norm_var": 10.250455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5664,
+      "loss/crossentropy": 2.1419076189398765,
+      "loss/hidden": 3.390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1872857511974871,
+      "step": 8800
+    },
+    {
+      "epoch": 0.22025,
+      "grad_norm": 31.625,
+      "grad_norm_var": 13.099934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5152,
+      "loss/crossentropy": 2.136802351474762,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19524938464164734,
+      "step": 8810
+    },
+    {
+      "epoch": 0.2205,
+      "grad_norm": 29.0,
+      "grad_norm_var": 18.0634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.6215,
+      "loss/crossentropy": 2.231374368071556,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18930297438055277,
+      "step": 8820
+    },
+    {
+      "epoch": 0.22075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 18.358333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5354,
+      "loss/crossentropy": 2.2225961208343508,
+      "loss/hidden": 3.41328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19855661746114492,
+      "step": 8830
+    },
+    {
+      "epoch": 0.221,
+      "grad_norm": 31.625,
+      "grad_norm_var": 18.312239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6216,
+      "loss/crossentropy": 2.1328855097293853,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20832530297338964,
+      "step": 8840
+    },
+    {
+      "epoch": 0.22125,
+      "grad_norm": 28.375,
+      "grad_norm_var": 8.085416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3697,
+      "loss/crossentropy": 2.0526474595069883,
+      "loss/hidden": 3.5015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21284359116107227,
+      "step": 8850
+    },
+    {
+      "epoch": 0.2215,
+      "grad_norm": 37.25,
+      "grad_norm_var": 9.942708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4967,
+      "loss/crossentropy": 2.1749773651361464,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20905132126063108,
+      "step": 8860
+    },
+    {
+      "epoch": 0.22175,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.322330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.503,
+      "loss/crossentropy": 2.108053085952997,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20318386573344469,
+      "step": 8870
+    },
+    {
+      "epoch": 0.222,
+      "grad_norm": 34.0,
+      "grad_norm_var": 7.685416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3618,
+      "loss/crossentropy": 2.1234827637672424,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18484860006719828,
+      "step": 8880
+    },
+    {
+      "epoch": 0.22225,
+      "grad_norm": 29.75,
+      "grad_norm_var": 8.220247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.393,
+      "loss/crossentropy": 2.099182015657425,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19038616605103015,
+      "step": 8890
+    },
+    {
+      "epoch": 0.2225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 7.569205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4768,
+      "loss/crossentropy": 2.148054042458534,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2065868368372321,
+      "step": 8900
+    },
+    {
+      "epoch": 0.22275,
+      "grad_norm": 28.75,
+      "grad_norm_var": 6.009375,
+      "learning_rate": 0.0001,
+      "loss": 7.5001,
+      "loss/crossentropy": 2.126410482823849,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.210067143663764,
+      "step": 8910
+    },
+    {
+      "epoch": 0.223,
+      "grad_norm": 37.5,
+      "grad_norm_var": 7.870572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4678,
+      "loss/crossentropy": 2.1258621901273727,
+      "loss/hidden": 3.428515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19669083636254073,
+      "step": 8920
+    },
+    {
+      "epoch": 0.22325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 8.908268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5039,
+      "loss/crossentropy": 2.1579898312687873,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18557702358812095,
+      "step": 8930
+    },
+    {
+      "epoch": 0.2235,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.448372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5477,
+      "loss/crossentropy": 2.2186635404825212,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19810736775398255,
+      "step": 8940
+    },
+    {
+      "epoch": 0.22375,
+      "grad_norm": 36.25,
+      "grad_norm_var": 7.304166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5522,
+      "loss/crossentropy": 2.1860512644052505,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18810464218258857,
+      "step": 8950
+    },
+    {
+      "epoch": 0.224,
+      "grad_norm": 35.75,
+      "grad_norm_var": 4.498372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5105,
+      "loss/crossentropy": 2.244818753004074,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19993619099259377,
+      "step": 8960
+    },
+    {
+      "epoch": 0.22425,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.34375,
+      "learning_rate": 0.0001,
+      "loss": 7.6077,
+      "loss/crossentropy": 2.1209320187568665,
+      "loss/hidden": 3.47421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2167982429265976,
+      "step": 8970
+    },
+    {
+      "epoch": 0.2245,
+      "grad_norm": 31.0,
+      "grad_norm_var": 18.992122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5306,
+      "loss/crossentropy": 2.008339713513851,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19748742282390594,
+      "step": 8980
+    },
+    {
+      "epoch": 0.22475,
+      "grad_norm": 29.5,
+      "grad_norm_var": 20.906705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5837,
+      "loss/crossentropy": 2.245701877772808,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117865581065416,
+      "step": 8990
+    },
+    {
+      "epoch": 0.225,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.1119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5251,
+      "loss/crossentropy": 2.050145834684372,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19412143267691134,
+      "step": 9000
+    },
+    {
+      "epoch": 0.22525,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.176822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5851,
+      "loss/crossentropy": 2.190397572517395,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20755842179059983,
+      "step": 9010
+    },
+    {
+      "epoch": 0.2255,
+      "grad_norm": 30.75,
+      "grad_norm_var": 8.977083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4264,
+      "loss/crossentropy": 2.284222900867462,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21356079690158367,
+      "step": 9020
+    },
+    {
+      "epoch": 0.22575,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.0639973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5988,
+      "loss/crossentropy": 2.1513551443815233,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19452885985374452,
+      "step": 9030
+    },
+    {
+      "epoch": 0.226,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.7830729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4915,
+      "loss/crossentropy": 2.2022497206926346,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1973773717880249,
+      "step": 9040
+    },
+    {
+      "epoch": 0.22625,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.584830729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5443,
+      "loss/crossentropy": 2.1286503240466117,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18610758930444718,
+      "step": 9050
+    },
+    {
+      "epoch": 0.2265,
+      "grad_norm": 27.75,
+      "grad_norm_var": 3.345572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5659,
+      "loss/crossentropy": 2.1127908319234847,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19702097605913876,
+      "step": 9060
+    },
+    {
+      "epoch": 0.22675,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.716080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5575,
+      "loss/crossentropy": 2.0748134687542916,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19112884253263474,
+      "step": 9070
+    },
+    {
+      "epoch": 0.227,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.0764973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6302,
+      "loss/crossentropy": 2.1477744698524477,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.198454162850976,
+      "step": 9080
+    },
+    {
+      "epoch": 0.22725,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.3197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4827,
+      "loss/crossentropy": 2.1293207883834837,
+      "loss/hidden": 3.55625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21182843409478663,
+      "step": 9090
+    },
+    {
+      "epoch": 0.2275,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.1333333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5027,
+      "loss/crossentropy": 2.145271519571543,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18950196234509348,
+      "step": 9100
+    },
+    {
+      "epoch": 0.22775,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.9337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5571,
+      "loss/crossentropy": 2.1665233090519904,
+      "loss/hidden": 3.48046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20093025229871272,
+      "step": 9110
+    },
+    {
+      "epoch": 0.228,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.9718098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4992,
+      "loss/crossentropy": 2.119904951751232,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19110188409686088,
+      "step": 9120
+    },
+    {
+      "epoch": 0.22825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.21875,
+      "learning_rate": 0.0001,
+      "loss": 7.5861,
+      "loss/crossentropy": 2.169039398431778,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18969318978488445,
+      "step": 9130
+    },
+    {
+      "epoch": 0.2285,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.0134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4457,
+      "loss/crossentropy": 2.202808904647827,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21587605867534876,
+      "step": 9140
+    },
+    {
+      "epoch": 0.22875,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.9014973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5869,
+      "loss/crossentropy": 2.188230502605438,
+      "loss/hidden": 3.32578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20162757411599158,
+      "step": 9150
+    },
+    {
+      "epoch": 0.229,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.6268229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5041,
+      "loss/crossentropy": 2.166210785508156,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19896902665495872,
+      "step": 9160
+    },
+    {
+      "epoch": 0.22925,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.3582682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4795,
+      "loss/crossentropy": 2.1638981848955154,
+      "loss/hidden": 3.28046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18999446034431458,
+      "step": 9170
+    },
+    {
+      "epoch": 0.2295,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.6177083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4529,
+      "loss/crossentropy": 2.170040412247181,
+      "loss/hidden": 3.399609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22052502054721118,
+      "step": 9180
+    },
+    {
+      "epoch": 0.22975,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.2400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.114721930027008,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19851209484040738,
+      "step": 9190
+    },
+    {
+      "epoch": 0.23,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.2457682291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5862,
+      "loss/crossentropy": 2.1541184708476067,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1960083631798625,
+      "step": 9200
+    },
+    {
+      "epoch": 0.23025,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.676041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4881,
+      "loss/crossentropy": 2.1309454582631586,
+      "loss/hidden": 3.355078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19452987853437662,
+      "step": 9210
+    },
+    {
+      "epoch": 0.2305,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.671875,
+      "learning_rate": 0.0001,
+      "loss": 7.4938,
+      "loss/crossentropy": 2.18781658411026,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2072685670107603,
+      "step": 9220
+    },
+    {
+      "epoch": 0.23075,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.2556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4355,
+      "loss/crossentropy": 2.1591438576579094,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18837961312383414,
+      "step": 9230
+    },
+    {
+      "epoch": 0.231,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.2087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4499,
+      "loss/crossentropy": 2.133663722872734,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18666253574192523,
+      "step": 9240
+    },
+    {
+      "epoch": 0.23125,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.1806640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5916,
+      "loss/crossentropy": 2.12740980386734,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19293057937175034,
+      "step": 9250
+    },
+    {
+      "epoch": 0.2315,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.1947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5307,
+      "loss/crossentropy": 2.1379271537065505,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19501187074929477,
+      "step": 9260
+    },
+    {
+      "epoch": 0.23175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.8264973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4995,
+      "loss/crossentropy": 2.220619598031044,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19177700616419316,
+      "step": 9270
+    },
+    {
+      "epoch": 0.232,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.855208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.461,
+      "loss/crossentropy": 2.1386048540472986,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117407780140638,
+      "step": 9280
+    },
+    {
+      "epoch": 0.23225,
+      "grad_norm": 33.75,
+      "grad_norm_var": 3.1122395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5308,
+      "loss/crossentropy": 2.201369822025299,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18732962422072888,
+      "step": 9290
+    },
+    {
+      "epoch": 0.2325,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.4559895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 2.172528564929962,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19425926376134156,
+      "step": 9300
+    },
+    {
+      "epoch": 0.23275,
+      "grad_norm": 27.5,
+      "grad_norm_var": 2.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4503,
+      "loss/crossentropy": 2.021423862874508,
+      "loss/hidden": 3.48125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20009233951568603,
+      "step": 9310
+    },
+    {
+      "epoch": 0.233,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.4125,
+      "learning_rate": 0.0001,
+      "loss": 7.5898,
+      "loss/crossentropy": 2.197957542538643,
+      "loss/hidden": 3.426953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21913623586297035,
+      "step": 9320
+    },
+    {
+      "epoch": 0.23325,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.70390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5196,
+      "loss/crossentropy": 2.00966841802001,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21357689071446656,
+      "step": 9330
+    },
+    {
+      "epoch": 0.2335,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.5629557291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5156,
+      "loss/crossentropy": 2.200144296884537,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19718017429113388,
+      "step": 9340
+    },
+    {
+      "epoch": 0.23375,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.6541666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4899,
+      "loss/crossentropy": 1.991941288113594,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1878137281164527,
+      "step": 9350
+    },
+    {
+      "epoch": 0.234,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.6934895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5461,
+      "loss/crossentropy": 2.0645473077893257,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20249587260186672,
+      "step": 9360
+    },
+    {
+      "epoch": 0.23425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.3462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4449,
+      "loss/crossentropy": 2.203160837292671,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20446450710296632,
+      "step": 9370
+    },
+    {
+      "epoch": 0.2345,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.572330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.451,
+      "loss/crossentropy": 2.1672566562891005,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22198531460016965,
+      "step": 9380
+    },
+    {
+      "epoch": 0.23475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.42265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6011,
+      "loss/crossentropy": 2.0229784891009333,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18394104316830634,
+      "step": 9390
+    },
+    {
+      "epoch": 0.235,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.330208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4942,
+      "loss/crossentropy": 2.1677876338362694,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22315683960914612,
+      "step": 9400
+    },
+    {
+      "epoch": 0.23525,
+      "grad_norm": 31.375,
+      "grad_norm_var": 19.045572916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.5569,
+      "loss/crossentropy": 2.2457904130220414,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20503853298723698,
+      "step": 9410
+    },
+    {
+      "epoch": 0.2355,
+      "grad_norm": 31.0,
+      "grad_norm_var": 18.926822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4461,
+      "loss/crossentropy": 2.142396827042103,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1910883378237486,
+      "step": 9420
+    },
+    {
+      "epoch": 0.23575,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.741666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.475,
+      "loss/crossentropy": 2.063981272280216,
+      "loss/hidden": 3.4921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20397737752646208,
+      "step": 9430
+    },
+    {
+      "epoch": 0.236,
+      "grad_norm": 30.0,
+      "grad_norm_var": 9.062434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4995,
+      "loss/crossentropy": 1.9866131611168385,
+      "loss/hidden": 3.516015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19697492588311433,
+      "step": 9440
+    },
+    {
+      "epoch": 0.23625,
+      "grad_norm": 40.25,
+      "grad_norm_var": 12.64140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5931,
+      "loss/crossentropy": 2.168704579770565,
+      "loss/hidden": 3.46796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19889446310698985,
+      "step": 9450
+    },
+    {
+      "epoch": 0.2365,
+      "grad_norm": 29.25,
+      "grad_norm_var": 51.16223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4255,
+      "loss/crossentropy": 2.241490375995636,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19278081078082324,
+      "step": 9460
+    },
+    {
+      "epoch": 0.23675,
+      "grad_norm": 31.125,
+      "grad_norm_var": 49.08118489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.44,
+      "loss/crossentropy": 2.0530834168195726,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1845833698287606,
+      "step": 9470
+    },
+    {
+      "epoch": 0.237,
+      "grad_norm": 35.25,
+      "grad_norm_var": 5.190559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4893,
+      "loss/crossentropy": 2.1680017322301866,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19126822520047426,
+      "step": 9480
+    },
+    {
+      "epoch": 0.23725,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.1728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5044,
+      "loss/crossentropy": 2.1760893553495406,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1902306279167533,
+      "step": 9490
+    },
+    {
+      "epoch": 0.2375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.610872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5567,
+      "loss/crossentropy": 2.169127979874611,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897792138159275,
+      "step": 9500
+    },
+    {
+      "epoch": 0.23775,
+      "grad_norm": 34.25,
+      "grad_norm_var": 31.978059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5548,
+      "loss/crossentropy": 2.0474624037742615,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897742820903659,
+      "step": 9510
+    },
+    {
+      "epoch": 0.238,
+      "grad_norm": 28.875,
+      "grad_norm_var": 33.4025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.46,
+      "loss/crossentropy": 2.167206625640392,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971887955442071,
+      "step": 9520
+    },
+    {
+      "epoch": 0.23825,
+      "grad_norm": 38.0,
+      "grad_norm_var": 10.677018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4721,
+      "loss/crossentropy": 2.160684567689896,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2026047335937619,
+      "step": 9530
+    },
+    {
+      "epoch": 0.2385,
+      "grad_norm": 33.25,
+      "grad_norm_var": 6.915559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4439,
+      "loss/crossentropy": 2.101367045938969,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18591827657073737,
+      "step": 9540
+    },
+    {
+      "epoch": 0.23875,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.285416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6017,
+      "loss/crossentropy": 2.144346782565117,
+      "loss/hidden": 3.458203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22905603051185608,
+      "step": 9550
+    },
+    {
+      "epoch": 0.239,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.608333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.391,
+      "loss/crossentropy": 2.2163674563169478,
+      "loss/hidden": 3.491015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20480377711355685,
+      "step": 9560
+    },
+    {
+      "epoch": 0.23925,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.554166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4754,
+      "loss/crossentropy": 2.05806076079607,
+      "loss/hidden": 3.43046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1894669458270073,
+      "step": 9570
+    },
+    {
+      "epoch": 0.2395,
+      "grad_norm": 35.75,
+      "grad_norm_var": 6.116666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.64,
+      "loss/crossentropy": 2.2116071820259093,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20237535443156957,
+      "step": 9580
+    },
+    {
+      "epoch": 0.23975,
+      "grad_norm": 27.625,
+      "grad_norm_var": 6.3166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4412,
+      "loss/crossentropy": 2.0997579991817474,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18707791212946176,
+      "step": 9590
+    },
+    {
+      "epoch": 0.24,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.391666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5587,
+      "loss/crossentropy": 2.1704364523291586,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20182078052312136,
+      "step": 9600
+    },
+    {
+      "epoch": 0.24025,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.959375,
+      "learning_rate": 0.0001,
+      "loss": 7.5494,
+      "loss/crossentropy": 2.1063621580600738,
+      "loss/hidden": 3.4671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2080043438822031,
+      "step": 9610
+    },
+    {
+      "epoch": 0.2405,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.5872395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4507,
+      "loss/crossentropy": 2.1176367297768595,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19296143669635057,
+      "step": 9620
+    },
+    {
+      "epoch": 0.24075,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.3082682291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5035,
+      "loss/crossentropy": 2.216022199392319,
+      "loss/hidden": 3.486328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2059534139931202,
+      "step": 9630
+    },
+    {
+      "epoch": 0.241,
+      "grad_norm": 31.5,
+      "grad_norm_var": 10.676041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5155,
+      "loss/crossentropy": 2.209742319583893,
+      "loss/hidden": 3.3421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19856051169335842,
+      "step": 9640
+    },
+    {
+      "epoch": 0.24125,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.10390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4983,
+      "loss/crossentropy": 2.040982872247696,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19508841075003147,
+      "step": 9650
+    },
+    {
+      "epoch": 0.2415,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.1697916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4469,
+      "loss/crossentropy": 2.1563887119293215,
+      "loss/hidden": 3.2921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18813634980469943,
+      "step": 9660
+    },
+    {
+      "epoch": 0.24175,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.5768229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5616,
+      "loss/crossentropy": 2.056604099273682,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19395184125751258,
+      "step": 9670
+    },
+    {
+      "epoch": 0.242,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.067643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4244,
+      "loss/crossentropy": 2.174482125043869,
+      "loss/hidden": 3.484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20453137308359146,
+      "step": 9680
+    },
+    {
+      "epoch": 0.24225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.905989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5003,
+      "loss/crossentropy": 2.119745084643364,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20908331871032715,
+      "step": 9690
+    },
+    {
+      "epoch": 0.2425,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.4785807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3362,
+      "loss/crossentropy": 2.122311297059059,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18319036178290843,
+      "step": 9700
+    },
+    {
+      "epoch": 0.24275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.9541666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4396,
+      "loss/crossentropy": 1.9478351414203643,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20451803579926492,
+      "step": 9710
+    },
+    {
+      "epoch": 0.243,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.3997395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4967,
+      "loss/crossentropy": 2.0708641201257705,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053229298442603,
+      "step": 9720
+    },
+    {
+      "epoch": 0.24325,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.635872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5016,
+      "loss/crossentropy": 2.0913224294781685,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1940496889874339,
+      "step": 9730
+    },
+    {
+      "epoch": 0.2435,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.3291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5077,
+      "loss/crossentropy": 2.221079145371914,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1924398820847273,
+      "step": 9740
+    },
+    {
+      "epoch": 0.24375,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.56875,
+      "learning_rate": 0.0001,
+      "loss": 7.4463,
+      "loss/crossentropy": 2.148006671667099,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20756886284798384,
+      "step": 9750
+    },
+    {
+      "epoch": 0.244,
+      "grad_norm": 30.25,
+      "grad_norm_var": 8.2978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4544,
+      "loss/crossentropy": 2.2184116363525392,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19135836903005837,
+      "step": 9760
+    },
+    {
+      "epoch": 0.24425,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.690625,
+      "learning_rate": 0.0001,
+      "loss": 7.4095,
+      "loss/crossentropy": 2.1005363285541536,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19765366949141025,
+      "step": 9770
+    },
+    {
+      "epoch": 0.2445,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.5327473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5002,
+      "loss/crossentropy": 2.167180609703064,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19258719477802516,
+      "step": 9780
+    },
+    {
+      "epoch": 0.24475,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.4697916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3885,
+      "loss/crossentropy": 1.9792610332369804,
+      "loss/hidden": 3.3671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18384026363492012,
+      "step": 9790
+    },
+    {
+      "epoch": 0.245,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.5447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5547,
+      "loss/crossentropy": 2.0596527442336083,
+      "loss/hidden": 3.569921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20697965249419212,
+      "step": 9800
+    },
+    {
+      "epoch": 0.24525,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.042122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4887,
+      "loss/crossentropy": 2.132730546593666,
+      "loss/hidden": 3.498828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19533955454826354,
+      "step": 9810
+    },
+    {
+      "epoch": 0.2455,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.31015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5278,
+      "loss/crossentropy": 2.151031643152237,
+      "loss/hidden": 3.450390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19322178810834884,
+      "step": 9820
+    },
+    {
+      "epoch": 0.24575,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.6059895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5534,
+      "loss/crossentropy": 2.143332117795944,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897228866815567,
+      "step": 9830
+    },
+    {
+      "epoch": 0.246,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.4395182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5657,
+      "loss/crossentropy": 2.1097486332058906,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20945656076073646,
+      "step": 9840
+    },
+    {
+      "epoch": 0.24625,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.9452473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5013,
+      "loss/crossentropy": 2.065919445455074,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18999069584533573,
+      "step": 9850
+    },
+    {
+      "epoch": 0.2465,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.2322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5501,
+      "loss/crossentropy": 2.183481493592262,
+      "loss/hidden": 3.504296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21892244052141904,
+      "step": 9860
+    },
+    {
+      "epoch": 0.24675,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.4926432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.478,
+      "loss/crossentropy": 2.1966081470251084,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20018710754811764,
+      "step": 9870
+    },
+    {
+      "epoch": 0.247,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.2244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5127,
+      "loss/crossentropy": 1.9914133831858636,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19236233234405517,
+      "step": 9880
+    },
+    {
+      "epoch": 0.24725,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.8900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5007,
+      "loss/crossentropy": 2.1605743557214736,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1983661765232682,
+      "step": 9890
+    },
+    {
+      "epoch": 0.2475,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.0666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5307,
+      "loss/crossentropy": 2.0920628547668456,
+      "loss/hidden": 3.404296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19224398005753757,
+      "step": 9900
+    },
+    {
+      "epoch": 0.24775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 28.910416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4845,
+      "loss/crossentropy": 2.13549183011055,
+      "loss/hidden": 3.435546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2024377616122365,
+      "step": 9910
+    },
+    {
+      "epoch": 0.248,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.2122395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5567,
+      "loss/crossentropy": 2.077779620885849,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914847193285823,
+      "step": 9920
+    },
+    {
+      "epoch": 0.24825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.3416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.6034,
+      "loss/crossentropy": 2.24170383810997,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21483709737658502,
+      "step": 9930
+    },
+    {
+      "epoch": 0.2485,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.9525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5277,
+      "loss/crossentropy": 2.0574432730674745,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19156391881406307,
+      "step": 9940
+    },
+    {
+      "epoch": 0.24875,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.6181640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4432,
+      "loss/crossentropy": 2.231716367602348,
+      "loss/hidden": 3.271484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19992320109158754,
+      "step": 9950
+    },
+    {
+      "epoch": 0.249,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.496809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5233,
+      "loss/crossentropy": 2.2805121034383773,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21222201287746428,
+      "step": 9960
+    },
+    {
+      "epoch": 0.24925,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.6143229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5022,
+      "loss/crossentropy": 2.1643686085939406,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2017618851736188,
+      "step": 9970
+    },
+    {
+      "epoch": 0.2495,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.9947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5351,
+      "loss/crossentropy": 2.052297804504633,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19462402295321227,
+      "step": 9980
+    },
+    {
+      "epoch": 0.24975,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.8958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4501,
+      "loss/crossentropy": 2.1721100926399233,
+      "loss/hidden": 3.3234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20133505929261447,
+      "step": 9990
+    },
+    {
+      "epoch": 0.25,
+      "grad_norm": 28.0,
+      "grad_norm_var": 2.2645833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5168,
+      "loss/crossentropy": 2.0236786626279355,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18019386120140551,
+      "step": 10000
+    },
+    {
+      "epoch": 0.25025,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.1801432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6,
+      "loss/crossentropy": 2.179265484213829,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21681770533323289,
+      "step": 10010
+    },
+    {
+      "epoch": 0.2505,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.2884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4947,
+      "loss/crossentropy": 2.238062459230423,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1930771939456463,
+      "step": 10020
+    },
+    {
+      "epoch": 0.25075,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.2957682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4869,
+      "loss/crossentropy": 2.016241580247879,
+      "loss/hidden": 3.40703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18652069568634033,
+      "step": 10030
+    },
+    {
+      "epoch": 0.251,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.47265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4921,
+      "loss/crossentropy": 2.1150885075330734,
+      "loss/hidden": 3.416015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19066351260989906,
+      "step": 10040
+    },
+    {
+      "epoch": 0.25125,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.8410807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4968,
+      "loss/crossentropy": 2.028704561293125,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19649554397910834,
+      "step": 10050
+    },
+    {
+      "epoch": 0.2515,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.282747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5077,
+      "loss/crossentropy": 2.1834514826536178,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19737422596663237,
+      "step": 10060
+    },
+    {
+      "epoch": 0.25175,
+      "grad_norm": 30.875,
+      "grad_norm_var": 4.7541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5032,
+      "loss/crossentropy": 2.2020941093564033,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20502295531332493,
+      "step": 10070
+    },
+    {
+      "epoch": 0.252,
+      "grad_norm": 34.25,
+      "grad_norm_var": 5.939322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4971,
+      "loss/crossentropy": 2.1619419783353804,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19227985218167304,
+      "step": 10080
+    },
+    {
+      "epoch": 0.25225,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.762434895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5093,
+      "loss/crossentropy": 2.045739908516407,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1985299138352275,
+      "step": 10090
+    },
+    {
+      "epoch": 0.2525,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.0645182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5485,
+      "loss/crossentropy": 2.1392029881477357,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22696843333542346,
+      "step": 10100
+    },
+    {
+      "epoch": 0.25275,
+      "grad_norm": 41.25,
+      "grad_norm_var": 8.16640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5099,
+      "loss/crossentropy": 2.112062802910805,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20055703315883874,
+      "step": 10110
+    },
+    {
+      "epoch": 0.253,
+      "grad_norm": 32.75,
+      "grad_norm_var": 8.356705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5021,
+      "loss/crossentropy": 2.081932318210602,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885841501876712,
+      "step": 10120
+    },
+    {
+      "epoch": 0.25325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.154166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5364,
+      "loss/crossentropy": 2.1130091354250906,
+      "loss/hidden": 3.321875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19095001947134732,
+      "step": 10130
+    },
+    {
+      "epoch": 0.2535,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.7160807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5311,
+      "loss/crossentropy": 2.0348104894161225,
+      "loss/hidden": 3.458203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20060632489621638,
+      "step": 10140
+    },
+    {
+      "epoch": 0.25375,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.8551432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5424,
+      "loss/crossentropy": 2.1739980176091196,
+      "loss/hidden": 3.361328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19488737732172012,
+      "step": 10150
+    },
+    {
+      "epoch": 0.254,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.4330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5232,
+      "loss/crossentropy": 2.1551436215639113,
+      "loss/hidden": 3.422265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20004159081727266,
+      "step": 10160
+    },
+    {
+      "epoch": 0.25425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.01015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5319,
+      "loss/crossentropy": 2.100159841030836,
+      "loss/hidden": 3.535546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19967089565470814,
+      "step": 10170
+    },
+    {
+      "epoch": 0.2545,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.896875,
+      "learning_rate": 0.0001,
+      "loss": 7.4974,
+      "loss/crossentropy": 2.1067795649170877,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1976080872118473,
+      "step": 10180
+    },
+    {
+      "epoch": 0.25475,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.3619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5954,
+      "loss/crossentropy": 2.0253565564751623,
+      "loss/hidden": 3.473828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19032611772418023,
+      "step": 10190
+    },
+    {
+      "epoch": 0.255,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.9875,
+      "learning_rate": 0.0001,
+      "loss": 7.4835,
+      "loss/crossentropy": 2.0224084883928297,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18663749676197766,
+      "step": 10200
+    },
+    {
+      "epoch": 0.25525,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.7973307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5863,
+      "loss/crossentropy": 2.126485800743103,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2074130540713668,
+      "step": 10210
+    },
+    {
+      "epoch": 0.2555,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.6093098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.6372,
+      "loss/crossentropy": 2.0780230551958083,
+      "loss/hidden": 3.435546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1949790535494685,
+      "step": 10220
+    },
+    {
+      "epoch": 0.25575,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.7247395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4954,
+      "loss/crossentropy": 2.0865010380744935,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19579052459448576,
+      "step": 10230
+    },
+    {
+      "epoch": 0.256,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.9025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5297,
+      "loss/crossentropy": 2.06137825101614,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20053859017789363,
+      "step": 10240
+    },
+    {
+      "epoch": 0.25625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 5.632291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5228,
+      "loss/crossentropy": 2.102967083454132,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1902999484911561,
+      "step": 10250
+    },
+    {
+      "epoch": 0.2565,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.2010416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4742,
+      "loss/crossentropy": 2.0118507161736487,
+      "loss/hidden": 3.489453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20069226585328578,
+      "step": 10260
+    },
+    {
+      "epoch": 0.25675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.0728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5045,
+      "loss/crossentropy": 2.161550810933113,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856826152652502,
+      "step": 10270
+    },
+    {
+      "epoch": 0.257,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4474,
+      "loss/crossentropy": 2.209408235549927,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19467408265918493,
+      "step": 10280
+    },
+    {
+      "epoch": 0.25725,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.925,
+      "learning_rate": 0.0001,
+      "loss": 7.5425,
+      "loss/crossentropy": 2.1496171414852143,
+      "loss/hidden": 3.2546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18077059462666512,
+      "step": 10290
+    },
+    {
+      "epoch": 0.2575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.8108723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4462,
+      "loss/crossentropy": 2.2561030745506288,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20665224064141513,
+      "step": 10300
+    },
+    {
+      "epoch": 0.25775,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.11015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3919,
+      "loss/crossentropy": 2.1672320902347564,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19321133568882942,
+      "step": 10310
+    },
+    {
+      "epoch": 0.258,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.8114583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4607,
+      "loss/crossentropy": 2.1774761766195296,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1832532085478306,
+      "step": 10320
+    },
+    {
+      "epoch": 0.25825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 3.006705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5093,
+      "loss/crossentropy": 2.08485603928566,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20309378914535045,
+      "step": 10330
+    },
+    {
+      "epoch": 0.2585,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.8994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5961,
+      "loss/crossentropy": 2.1903068631887437,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19475248865783215,
+      "step": 10340
+    },
+    {
+      "epoch": 0.25875,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.6723307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5157,
+      "loss/crossentropy": 2.1206096917390824,
+      "loss/hidden": 3.40546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19404957443475723,
+      "step": 10350
+    },
+    {
+      "epoch": 0.259,
+      "grad_norm": 35.75,
+      "grad_norm_var": 7.144791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5988,
+      "loss/crossentropy": 2.1614494152367114,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19104234427213668,
+      "step": 10360
+    },
+    {
+      "epoch": 0.25925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.8650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5585,
+      "loss/crossentropy": 2.2361265033483506,
+      "loss/hidden": 3.293359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18775968980044128,
+      "step": 10370
+    },
+    {
+      "epoch": 0.2595,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.8166666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5804,
+      "loss/crossentropy": 2.086420488357544,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19417675212025642,
+      "step": 10380
+    },
+    {
+      "epoch": 0.25975,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.603125,
+      "learning_rate": 0.0001,
+      "loss": 7.4606,
+      "loss/crossentropy": 2.0714549839496614,
+      "loss/hidden": 3.44453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21479819286614657,
+      "step": 10390
+    },
+    {
+      "epoch": 0.26,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.0738932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6124,
+      "loss/crossentropy": 2.0951902747154234,
+      "loss/hidden": 3.381640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19614790324121714,
+      "step": 10400
+    },
+    {
+      "epoch": 0.26025,
+      "grad_norm": 33.75,
+      "grad_norm_var": 1.9416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.535,
+      "loss/crossentropy": 2.1505746722221373,
+      "loss/hidden": 3.413671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20192769896239043,
+      "step": 10410
+    },
+    {
+      "epoch": 0.2605,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.5260416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.6095,
+      "loss/crossentropy": 2.249527391791344,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21959280855953695,
+      "step": 10420
+    },
+    {
+      "epoch": 0.26075,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.6025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5459,
+      "loss/crossentropy": 2.190070366859436,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18353844359517096,
+      "step": 10430
+    },
+    {
+      "epoch": 0.261,
+      "grad_norm": 33.5,
+      "grad_norm_var": 4.64375,
+      "learning_rate": 0.0001,
+      "loss": 7.5007,
+      "loss/crossentropy": 2.148374534398317,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19743462903425096,
+      "step": 10440
+    },
+    {
+      "epoch": 0.26125,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.2,
+      "learning_rate": 0.0001,
+      "loss": 7.5689,
+      "loss/crossentropy": 2.2100040167570114,
+      "loss/hidden": 3.42109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22015094980597497,
+      "step": 10450
+    },
+    {
+      "epoch": 0.2615,
+      "grad_norm": 54.5,
+      "grad_norm_var": 40.2072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5953,
+      "loss/crossentropy": 2.0877301573753355,
+      "loss/hidden": 3.48515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2032545803114772,
+      "step": 10460
+    },
+    {
+      "epoch": 0.26175,
+      "grad_norm": 29.375,
+      "grad_norm_var": 67.60807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.55,
+      "loss/crossentropy": 2.1474430203437804,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1959912359714508,
+      "step": 10470
+    },
+    {
+      "epoch": 0.262,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.184309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5324,
+      "loss/crossentropy": 2.2022393375635145,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18873728141188623,
+      "step": 10480
+    },
+    {
+      "epoch": 0.26225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.621809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5054,
+      "loss/crossentropy": 2.038270506262779,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18390222508460283,
+      "step": 10490
+    },
+    {
+      "epoch": 0.2625,
+      "grad_norm": 39.5,
+      "grad_norm_var": 6.564583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4524,
+      "loss/crossentropy": 2.112567237019539,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18860076945275067,
+      "step": 10500
+    },
+    {
+      "epoch": 0.26275,
+      "grad_norm": 28.5,
+      "grad_norm_var": 8.082747395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4385,
+      "loss/crossentropy": 2.172071099281311,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1897763295099139,
+      "step": 10510
+    },
+    {
+      "epoch": 0.263,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.86015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5004,
+      "loss/crossentropy": 2.2221229881048203,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19093570075929164,
+      "step": 10520
+    },
+    {
+      "epoch": 0.26325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.7697916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5274,
+      "loss/crossentropy": 2.121767780184746,
+      "loss/hidden": 3.421484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20014208517968654,
+      "step": 10530
+    },
+    {
+      "epoch": 0.2635,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.1059895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6305,
+      "loss/crossentropy": 2.0758725464344026,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21838635317981242,
+      "step": 10540
+    },
+    {
+      "epoch": 0.26375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.246875,
+      "learning_rate": 0.0001,
+      "loss": 7.5622,
+      "loss/crossentropy": 2.1207674235105514,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19129925426095723,
+      "step": 10550
+    },
+    {
+      "epoch": 0.264,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.0947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4523,
+      "loss/crossentropy": 2.0643288612365724,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1818947261199355,
+      "step": 10560
+    },
+    {
+      "epoch": 0.26425,
+      "grad_norm": 36.0,
+      "grad_norm_var": 4.464322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.467,
+      "loss/crossentropy": 2.168118804693222,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19394876565784216,
+      "step": 10570
+    },
+    {
+      "epoch": 0.2645,
+      "grad_norm": 29.875,
+      "grad_norm_var": 5.934830729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5136,
+      "loss/crossentropy": 2.1245740592479705,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18961659409105777,
+      "step": 10580
+    },
+    {
+      "epoch": 0.26475,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.565559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4291,
+      "loss/crossentropy": 2.166120108962059,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18523282706737518,
+      "step": 10590
+    },
+    {
+      "epoch": 0.265,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.995833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5069,
+      "loss/crossentropy": 2.200042313337326,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1952303361147642,
+      "step": 10600
+    },
+    {
+      "epoch": 0.26525,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.0082682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4612,
+      "loss/crossentropy": 2.0848385438323023,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19754947163164616,
+      "step": 10610
+    },
+    {
+      "epoch": 0.2655,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.125,
+      "learning_rate": 0.0001,
+      "loss": 7.507,
+      "loss/crossentropy": 2.0842891573905944,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18975816927850248,
+      "step": 10620
+    },
+    {
+      "epoch": 0.26575,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.1405598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5813,
+      "loss/crossentropy": 2.133824145793915,
+      "loss/hidden": 3.3890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20222157016396522,
+      "step": 10630
+    },
+    {
+      "epoch": 0.266,
+      "grad_norm": 30.0,
+      "grad_norm_var": 18.200455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5877,
+      "loss/crossentropy": 2.1643952041864396,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19804936908185483,
+      "step": 10640
+    },
+    {
+      "epoch": 0.26625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 399.8686848958333,
+      "learning_rate": 0.0001,
+      "loss": 7.6453,
+      "loss/crossentropy": 2.049100285768509,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20171927530318498,
+      "step": 10650
+    },
+    {
+      "epoch": 0.2665,
+      "grad_norm": 32.0,
+      "grad_norm_var": 7.462955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5871,
+      "loss/crossentropy": 2.086188179254532,
+      "loss/hidden": 3.53359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2075439091771841,
+      "step": 10660
+    },
+    {
+      "epoch": 0.26675,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.7143229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5477,
+      "loss/crossentropy": 2.1709790498018267,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18898459300398826,
+      "step": 10670
+    },
+    {
+      "epoch": 0.267,
+      "grad_norm": 35.5,
+      "grad_norm_var": 8.525,
+      "learning_rate": 0.0001,
+      "loss": 7.5583,
+      "loss/crossentropy": 2.191944640874863,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18916468657553195,
+      "step": 10680
+    },
+    {
+      "epoch": 0.26725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 9.723372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6552,
+      "loss/crossentropy": 2.153936019539833,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21916016209870576,
+      "step": 10690
+    },
+    {
+      "epoch": 0.2675,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.5327473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4995,
+      "loss/crossentropy": 2.144308550655842,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18704238552600144,
+      "step": 10700
+    },
+    {
+      "epoch": 0.26775,
+      "grad_norm": 29.375,
+      "grad_norm_var": 14.4875,
+      "learning_rate": 0.0001,
+      "loss": 7.6781,
+      "loss/crossentropy": 2.1939100988209246,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19588800482451915,
+      "step": 10710
+    },
+    {
+      "epoch": 0.268,
+      "grad_norm": 32.5,
+      "grad_norm_var": 21.031705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5024,
+      "loss/crossentropy": 2.0107320092618464,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1809020394459367,
+      "step": 10720
+    },
+    {
+      "epoch": 0.26825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 12.817708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5567,
+      "loss/crossentropy": 2.21520319879055,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18951662816107273,
+      "step": 10730
+    },
+    {
+      "epoch": 0.2685,
+      "grad_norm": 30.625,
+      "grad_norm_var": 31.861393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5161,
+      "loss/crossentropy": 2.1877179771661757,
+      "loss/hidden": 3.265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1967765673995018,
+      "step": 10740
+    },
+    {
+      "epoch": 0.26875,
+      "grad_norm": 34.25,
+      "grad_norm_var": 30.971875,
+      "learning_rate": 0.0001,
+      "loss": 7.549,
+      "loss/crossentropy": 2.1257426261901857,
+      "loss/hidden": 3.3078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18805306870490313,
+      "step": 10750
+    },
+    {
+      "epoch": 0.269,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.608072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5008,
+      "loss/crossentropy": 2.0466638416051866,
+      "loss/hidden": 3.316796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18291978891938926,
+      "step": 10760
+    },
+    {
+      "epoch": 0.26925,
+      "grad_norm": 32.5,
+      "grad_norm_var": 15.839518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6738,
+      "loss/crossentropy": 2.177478575706482,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2091924037784338,
+      "step": 10770
+    },
+    {
+      "epoch": 0.2695,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.19765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4797,
+      "loss/crossentropy": 2.054281398653984,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19896940067410468,
+      "step": 10780
+    },
+    {
+      "epoch": 0.26975,
+      "grad_norm": 28.0,
+      "grad_norm_var": 4.278580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5147,
+      "loss/crossentropy": 2.2530731678009035,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1928473612293601,
+      "step": 10790
+    },
+    {
+      "epoch": 0.27,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.033072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5229,
+      "loss/crossentropy": 2.218820883333683,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18697738237679004,
+      "step": 10800
+    },
+    {
+      "epoch": 0.27025,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.622916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6507,
+      "loss/crossentropy": 2.100504669547081,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19196046218276025,
+      "step": 10810
+    },
+    {
+      "epoch": 0.2705,
+      "grad_norm": 33.0,
+      "grad_norm_var": 11.7212890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6045,
+      "loss/crossentropy": 2.0671754188835623,
+      "loss/hidden": 3.466796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.195990754943341,
+      "step": 10820
+    },
+    {
+      "epoch": 0.27075,
+      "grad_norm": 29.875,
+      "grad_norm_var": 11.322916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5369,
+      "loss/crossentropy": 2.0837562650442125,
+      "loss/hidden": 3.39453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18965425249189138,
+      "step": 10830
+    },
+    {
+      "epoch": 0.271,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.9462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.6529,
+      "loss/crossentropy": 2.1033942848443985,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21247205398976804,
+      "step": 10840
+    },
+    {
+      "epoch": 0.27125,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.7796223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5122,
+      "loss/crossentropy": 2.2221016943454743,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20388032998889685,
+      "step": 10850
+    },
+    {
+      "epoch": 0.2715,
+      "grad_norm": 33.5,
+      "grad_norm_var": 1.4080729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6508,
+      "loss/crossentropy": 2.129495047032833,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22404770106077193,
+      "step": 10860
+    },
+    {
+      "epoch": 0.27175,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.537239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4812,
+      "loss/crossentropy": 2.085365228354931,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21701993197202682,
+      "step": 10870
+    },
+    {
+      "epoch": 0.272,
+      "grad_norm": 30.25,
+      "grad_norm_var": 78.16640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 2.0185007080435753,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1912636984139681,
+      "step": 10880
+    },
+    {
+      "epoch": 0.27225,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.4389973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4235,
+      "loss/crossentropy": 2.0273312851786613,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935686475597322,
+      "step": 10890
+    },
+    {
+      "epoch": 0.2725,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.676041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.56,
+      "loss/crossentropy": 2.1340904593467713,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21148081459105014,
+      "step": 10900
+    },
+    {
+      "epoch": 0.27275,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.169205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4542,
+      "loss/crossentropy": 2.1168949902057648,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19148083217442036,
+      "step": 10910
+    },
+    {
+      "epoch": 0.273,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.99765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5497,
+      "loss/crossentropy": 2.1183649614453315,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1918137041851878,
+      "step": 10920
+    },
+    {
+      "epoch": 0.27325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.7768229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6078,
+      "loss/crossentropy": 2.0465846031904222,
+      "loss/hidden": 3.434375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21847008895128966,
+      "step": 10930
+    },
+    {
+      "epoch": 0.2735,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.513997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.421,
+      "loss/crossentropy": 2.1968978524208067,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2054054219275713,
+      "step": 10940
+    },
+    {
+      "epoch": 0.27375,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.718489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5254,
+      "loss/crossentropy": 2.284649354219437,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1995658937841654,
+      "step": 10950
+    },
+    {
+      "epoch": 0.274,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5708,
+      "loss/crossentropy": 2.141575986146927,
+      "loss/hidden": 3.32109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19440446924418212,
+      "step": 10960
+    },
+    {
+      "epoch": 0.27425,
+      "grad_norm": 42.5,
+      "grad_norm_var": 4.468696696373851e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5234,
+      "loss/crossentropy": 2.131410190463066,
+      "loss/hidden": 3.311328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1760883742943406,
+      "step": 10970
+    },
+    {
+      "epoch": 0.2745,
+      "grad_norm": 29.5,
+      "grad_norm_var": 4.468696695713248e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5372,
+      "loss/crossentropy": 2.1476279973983763,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.184328780695796,
+      "step": 10980
+    },
+    {
+      "epoch": 0.27475,
+      "grad_norm": 31.0,
+      "grad_norm_var": 8.8931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5383,
+      "loss/crossentropy": 2.1352312207221984,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19372365307062864,
+      "step": 10990
+    },
+    {
+      "epoch": 0.275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 7.148372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4628,
+      "loss/crossentropy": 2.0489871561527253,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18995484188199044,
+      "step": 11000
+    },
+    {
+      "epoch": 0.27525,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.6264973958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4443,
+      "loss/crossentropy": 2.1590236008167265,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19273458905518054,
+      "step": 11010
+    },
+    {
+      "epoch": 0.2755,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.8848307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4967,
+      "loss/crossentropy": 2.1544351994991304,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21255392767488956,
+      "step": 11020
+    },
+    {
+      "epoch": 0.27575,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.5893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5015,
+      "loss/crossentropy": 2.018085864186287,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19312111511826516,
+      "step": 11030
+    },
+    {
+      "epoch": 0.276,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.8645833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4226,
+      "loss/crossentropy": 2.105189894139767,
+      "loss/hidden": 3.291796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17890707962214947,
+      "step": 11040
+    },
+    {
+      "epoch": 0.27625,
+      "grad_norm": 36.25,
+      "grad_norm_var": 64.57493489583334,
+      "learning_rate": 0.0001,
+      "loss": 7.6245,
+      "loss/crossentropy": 2.0516477465629577,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2067173158749938,
+      "step": 11050
+    },
+    {
+      "epoch": 0.2765,
+      "grad_norm": 31.5,
+      "grad_norm_var": 21.5666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5657,
+      "loss/crossentropy": 2.1685635060071946,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20545070786029102,
+      "step": 11060
+    },
+    {
+      "epoch": 0.27675,
+      "grad_norm": 31.0,
+      "grad_norm_var": 18.304622395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5266,
+      "loss/crossentropy": 2.168036562204361,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20288394428789616,
+      "step": 11070
+    },
+    {
+      "epoch": 0.277,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.6587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.637,
+      "loss/crossentropy": 2.175144538283348,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2135882146656513,
+      "step": 11080
+    },
+    {
+      "epoch": 0.27725,
+      "grad_norm": 31.75,
+      "grad_norm_var": 4.45625,
+      "learning_rate": 0.0001,
+      "loss": 7.5509,
+      "loss/crossentropy": 2.025794816017151,
+      "loss/hidden": 3.460546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19755149651318787,
+      "step": 11090
+    },
+    {
+      "epoch": 0.2775,
+      "grad_norm": 27.5,
+      "grad_norm_var": 3.9025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5019,
+      "loss/crossentropy": 2.1192509084939957,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885302910581231,
+      "step": 11100
+    },
+    {
+      "epoch": 0.27775,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.4514973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.508,
+      "loss/crossentropy": 2.0708123981952666,
+      "loss/hidden": 3.515234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21488154623657466,
+      "step": 11110
+    },
+    {
+      "epoch": 0.278,
+      "grad_norm": 29.375,
+      "grad_norm_var": 6.984375,
+      "learning_rate": 0.0001,
+      "loss": 7.5729,
+      "loss/crossentropy": 2.196261405944824,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19242637380957603,
+      "step": 11120
+    },
+    {
+      "epoch": 0.27825,
+      "grad_norm": 32.0,
+      "grad_norm_var": 13.114322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.501,
+      "loss/crossentropy": 2.0530719697475432,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19876221660524607,
+      "step": 11130
+    },
+    {
+      "epoch": 0.2785,
+      "grad_norm": 29.625,
+      "grad_norm_var": 13.57890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5154,
+      "loss/crossentropy": 2.1174173802137375,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20414691772311927,
+      "step": 11140
+    },
+    {
+      "epoch": 0.27875,
+      "grad_norm": 29.25,
+      "grad_norm_var": 44.18932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5442,
+      "loss/crossentropy": 2.2288032442331316,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2252635546028614,
+      "step": 11150
+    },
+    {
+      "epoch": 0.279,
+      "grad_norm": 30.75,
+      "grad_norm_var": 45.85182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5199,
+      "loss/crossentropy": 2.2578700333833694,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21590106561779976,
+      "step": 11160
+    },
+    {
+      "epoch": 0.27925,
+      "grad_norm": 29.0,
+      "grad_norm_var": 65.8634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4833,
+      "loss/crossentropy": 2.15876952111721,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21022603176534177,
+      "step": 11170
+    },
+    {
+      "epoch": 0.2795,
+      "grad_norm": 33.5,
+      "grad_norm_var": 41.5337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5426,
+      "loss/crossentropy": 2.2054150819778444,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1938659494742751,
+      "step": 11180
+    },
+    {
+      "epoch": 0.27975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 37.57265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5083,
+      "loss/crossentropy": 2.191286247968674,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19200536869466306,
+      "step": 11190
+    },
+    {
+      "epoch": 0.28,
+      "grad_norm": 30.875,
+      "grad_norm_var": 36.5697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3998,
+      "loss/crossentropy": 2.240337911248207,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19107160083949565,
+      "step": 11200
+    },
+    {
+      "epoch": 0.28025,
+      "grad_norm": 42.25,
+      "grad_norm_var": 37.92701822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4565,
+      "loss/crossentropy": 2.21739219725132,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2041622471064329,
+      "step": 11210
+    },
+    {
+      "epoch": 0.2805,
+      "grad_norm": 30.0,
+      "grad_norm_var": 14.87890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4073,
+      "loss/crossentropy": 2.1360912755131722,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19229222014546393,
+      "step": 11220
+    },
+    {
+      "epoch": 0.28075,
+      "grad_norm": 36.5,
+      "grad_norm_var": 21.35625,
+      "learning_rate": 0.0001,
+      "loss": 7.445,
+      "loss/crossentropy": 2.2468197375535963,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20960316248238087,
+      "step": 11230
+    },
+    {
+      "epoch": 0.281,
+      "grad_norm": 30.5,
+      "grad_norm_var": 24.69765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4355,
+      "loss/crossentropy": 2.2591811805963515,
+      "loss/hidden": 3.298828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18697103951126337,
+      "step": 11240
+    },
+    {
+      "epoch": 0.28125,
+      "grad_norm": 29.75,
+      "grad_norm_var": 10.726822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5225,
+      "loss/crossentropy": 2.2241678714752195,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19806783739477396,
+      "step": 11250
+    },
+    {
+      "epoch": 0.2815,
+      "grad_norm": 28.25,
+      "grad_norm_var": 11.403059895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4736,
+      "loss/crossentropy": 2.28880070745945,
+      "loss/hidden": 3.35703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19563225321471692,
+      "step": 11260
+    },
+    {
+      "epoch": 0.28175,
+      "grad_norm": 33.0,
+      "grad_norm_var": 16.953580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5007,
+      "loss/crossentropy": 2.1222257822752,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20468491613864898,
+      "step": 11270
+    },
+    {
+      "epoch": 0.282,
+      "grad_norm": 32.0,
+      "grad_norm_var": 16.693684895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.4601,
+      "loss/crossentropy": 2.0914298713207247,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20181928034871816,
+      "step": 11280
+    },
+    {
+      "epoch": 0.28225,
+      "grad_norm": 32.5,
+      "grad_norm_var": 14.206705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4363,
+      "loss/crossentropy": 2.058793123066425,
+      "loss/hidden": 3.3578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1982463788241148,
+      "step": 11290
+    },
+    {
+      "epoch": 0.2825,
+      "grad_norm": 31.75,
+      "grad_norm_var": 8.758333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4243,
+      "loss/crossentropy": 2.0935801014304163,
+      "loss/hidden": 3.3703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1959468424320221,
+      "step": 11300
+    },
+    {
+      "epoch": 0.28275,
+      "grad_norm": 28.125,
+      "grad_norm_var": 9.334309895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4081,
+      "loss/crossentropy": 2.156096602976322,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856603730469942,
+      "step": 11310
+    },
+    {
+      "epoch": 0.283,
+      "grad_norm": 30.375,
+      "grad_norm_var": 7.953059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4932,
+      "loss/crossentropy": 2.0780559942126273,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20045752841979264,
+      "step": 11320
+    },
+    {
+      "epoch": 0.28325,
+      "grad_norm": 38.25,
+      "grad_norm_var": 9.335872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4376,
+      "loss/crossentropy": 2.203425186872482,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1887454966083169,
+      "step": 11330
+    },
+    {
+      "epoch": 0.2835,
+      "grad_norm": 31.25,
+      "grad_norm_var": 7.339322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5769,
+      "loss/crossentropy": 2.139732484519482,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.200992492120713,
+      "step": 11340
+    },
+    {
+      "epoch": 0.28375,
+      "grad_norm": 29.375,
+      "grad_norm_var": 8.889583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4588,
+      "loss/crossentropy": 2.1970251157879828,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20897436924278737,
+      "step": 11350
+    },
+    {
+      "epoch": 0.284,
+      "grad_norm": 29.625,
+      "grad_norm_var": 10.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.556,
+      "loss/crossentropy": 2.188813480734825,
+      "loss/hidden": 3.49453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2200198858976364,
+      "step": 11360
+    },
+    {
+      "epoch": 0.28425,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.669205729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4272,
+      "loss/crossentropy": 2.0679334342479705,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19683563616126776,
+      "step": 11370
+    },
+    {
+      "epoch": 0.2845,
+      "grad_norm": 32.5,
+      "grad_norm_var": 6.563997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5581,
+      "loss/crossentropy": 2.29857756793499,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19470676891505717,
+      "step": 11380
+    },
+    {
+      "epoch": 0.28475,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.73515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4517,
+      "loss/crossentropy": 2.061806722730398,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20109398122876881,
+      "step": 11390
+    },
+    {
+      "epoch": 0.285,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.0205729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.668,
+      "loss/crossentropy": 2.0487837977707386,
+      "loss/hidden": 3.475390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2249346138909459,
+      "step": 11400
+    },
+    {
+      "epoch": 0.28525,
+      "grad_norm": 43.25,
+      "grad_norm_var": 3.0152302910081336e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5882,
+      "loss/crossentropy": 2.071147194504738,
+      "loss/hidden": 3.605078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19121263399720193,
+      "step": 11410
+    },
+    {
+      "epoch": 0.2855,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.0873121650067896e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6125,
+      "loss/crossentropy": 2.136325827240944,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19542958196252586,
+      "step": 11420
+    },
+    {
+      "epoch": 0.28575,
+      "grad_norm": 49.5,
+      "grad_norm_var": 3.85311092376286e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5728,
+      "loss/crossentropy": 2.061227947473526,
+      "loss/hidden": 3.62578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19353212472051382,
+      "step": 11430
+    },
+    {
+      "epoch": 0.286,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.8531109241063736e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4906,
+      "loss/crossentropy": 2.1435365855693815,
+      "loss/hidden": 3.394140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19920943658798934,
+      "step": 11440
+    },
+    {
+      "epoch": 0.28625,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.358268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4355,
+      "loss/crossentropy": 2.1882937461137772,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20106471050530672,
+      "step": 11450
+    },
+    {
+      "epoch": 0.2865,
+      "grad_norm": 32.75,
+      "grad_norm_var": 42.1759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5564,
+      "loss/crossentropy": 2.3196360021829605,
+      "loss/hidden": 3.483984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20747811421751977,
+      "step": 11460
+    },
+    {
+      "epoch": 0.28675,
+      "grad_norm": 29.25,
+      "grad_norm_var": 39.1962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5218,
+      "loss/crossentropy": 2.0665975011885167,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18888808842748403,
+      "step": 11470
+    },
+    {
+      "epoch": 0.287,
+      "grad_norm": 38.25,
+      "grad_norm_var": 257.53515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6489,
+      "loss/crossentropy": 2.0905568569898607,
+      "loss/hidden": 3.471484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21102672098204495,
+      "step": 11480
+    },
+    {
+      "epoch": 0.28725,
+      "grad_norm": 33.5,
+      "grad_norm_var": 254.4853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4738,
+      "loss/crossentropy": 2.091688461601734,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17688411958515643,
+      "step": 11490
+    },
+    {
+      "epoch": 0.2875,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.16875,
+      "learning_rate": 0.0001,
+      "loss": 7.5848,
+      "loss/crossentropy": 2.127404825389385,
+      "loss/hidden": 3.37265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2071824749931693,
+      "step": 11500
+    },
+    {
+      "epoch": 0.28775,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.8989583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4661,
+      "loss/crossentropy": 2.1476680278778075,
+      "loss/hidden": 3.431640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993227731436491,
+      "step": 11510
+    },
+    {
+      "epoch": 0.288,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.4770833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5997,
+      "loss/crossentropy": 2.168101379275322,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2063988609239459,
+      "step": 11520
+    },
+    {
+      "epoch": 0.28825,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.9166666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5253,
+      "loss/crossentropy": 2.1605693727731703,
+      "loss/hidden": 3.450390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20308145191520452,
+      "step": 11530
+    },
+    {
+      "epoch": 0.2885,
+      "grad_norm": 32.5,
+      "grad_norm_var": 5.273893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3857,
+      "loss/crossentropy": 2.2063374549150465,
+      "loss/hidden": 3.294921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1978879366070032,
+      "step": 11540
+    },
+    {
+      "epoch": 0.28875,
+      "grad_norm": 30.375,
+      "grad_norm_var": 376.1587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5441,
+      "loss/crossentropy": 2.1251575350761414,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21420632135123013,
+      "step": 11550
+    },
+    {
+      "epoch": 0.289,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.8624348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5367,
+      "loss/crossentropy": 2.152555100619793,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.209245034866035,
+      "step": 11560
+    },
+    {
+      "epoch": 0.28925,
+      "grad_norm": 27.125,
+      "grad_norm_var": 6.71875,
+      "learning_rate": 0.0001,
+      "loss": 7.4824,
+      "loss/crossentropy": 2.068640360236168,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873430425301194,
+      "step": 11570
+    },
+    {
+      "epoch": 0.2895,
+      "grad_norm": 31.875,
+      "grad_norm_var": 7.520768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4681,
+      "loss/crossentropy": 2.0883365333080293,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19091288559138775,
+      "step": 11580
+    },
+    {
+      "epoch": 0.28975,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.349739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4169,
+      "loss/crossentropy": 2.042503895610571,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20744232581928373,
+      "step": 11590
+    },
+    {
+      "epoch": 0.29,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.72265625,
+      "learning_rate": 0.0001,
+      "loss": 7.603,
+      "loss/crossentropy": 2.1102429538965226,
+      "loss/hidden": 3.500390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20691623724997044,
+      "step": 11600
+    },
+    {
+      "epoch": 0.29025,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.84140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5562,
+      "loss/crossentropy": 2.098295146226883,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20017402190715075,
+      "step": 11610
+    },
+    {
+      "epoch": 0.2905,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.2684895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5017,
+      "loss/crossentropy": 2.139223408699036,
+      "loss/hidden": 3.352734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20002739522606133,
+      "step": 11620
+    },
+    {
+      "epoch": 0.29075,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.357291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5172,
+      "loss/crossentropy": 2.1825324445962906,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20125013943761588,
+      "step": 11630
+    },
+    {
+      "epoch": 0.291,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.46015625,
+      "learning_rate": 0.0001,
+      "loss": 7.599,
+      "loss/crossentropy": 2.068669117987156,
+      "loss/hidden": 3.480078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19485772978514432,
+      "step": 11640
+    },
+    {
+      "epoch": 0.29125,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.86015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4215,
+      "loss/crossentropy": 2.089947706460953,
+      "loss/hidden": 3.319921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1832545857876539,
+      "step": 11650
+    },
+    {
+      "epoch": 0.2915,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.2494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4485,
+      "loss/crossentropy": 2.203251999616623,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20148768946528434,
+      "step": 11660
+    },
+    {
+      "epoch": 0.29175,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.1489583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4095,
+      "loss/crossentropy": 2.15777850151062,
+      "loss/hidden": 3.4890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20323845893144607,
+      "step": 11670
+    },
+    {
+      "epoch": 0.292,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.42890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4914,
+      "loss/crossentropy": 2.156290701031685,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21260349955409766,
+      "step": 11680
+    },
+    {
+      "epoch": 0.29225,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.283072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5226,
+      "loss/crossentropy": 2.218624970316887,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21999017260968684,
+      "step": 11690
+    },
+    {
+      "epoch": 0.2925,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.7957682291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.452,
+      "loss/crossentropy": 2.146151360869408,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19877625294029713,
+      "step": 11700
+    },
+    {
+      "epoch": 0.29275,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.1552083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5397,
+      "loss/crossentropy": 2.176427762210369,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.220915474742651,
+      "step": 11710
+    },
+    {
+      "epoch": 0.293,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.4330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4466,
+      "loss/crossentropy": 1.9928596645593644,
+      "loss/hidden": 3.371875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17805732283741235,
+      "step": 11720
+    },
+    {
+      "epoch": 0.29325,
+      "grad_norm": 36.0,
+      "grad_norm_var": 3.7087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4407,
+      "loss/crossentropy": 2.04853350520134,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19008731115609406,
+      "step": 11730
+    },
+    {
+      "epoch": 0.2935,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.638541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4562,
+      "loss/crossentropy": 2.1776413440704347,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19186513982713221,
+      "step": 11740
+    },
+    {
+      "epoch": 0.29375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.2343098958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.473,
+      "loss/crossentropy": 2.151021400094032,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19510295912623404,
+      "step": 11750
+    },
+    {
+      "epoch": 0.294,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.453125,
+      "learning_rate": 0.0001,
+      "loss": 7.6263,
+      "loss/crossentropy": 2.137827117741108,
+      "loss/hidden": 3.48359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.24175900630652905,
+      "step": 11760
+    },
+    {
+      "epoch": 0.29425,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.168684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4144,
+      "loss/crossentropy": 2.104247944802046,
+      "loss/hidden": 3.28984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19388842806220055,
+      "step": 11770
+    },
+    {
+      "epoch": 0.2945,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.7197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5832,
+      "loss/crossentropy": 2.234465000033379,
+      "loss/hidden": 3.4328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19337044432759284,
+      "step": 11780
+    },
+    {
+      "epoch": 0.29475,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.7729166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5465,
+      "loss/crossentropy": 2.253598779439926,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19784107208251953,
+      "step": 11790
+    },
+    {
+      "epoch": 0.295,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.7108723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5977,
+      "loss/crossentropy": 2.127879926562309,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19584042821079492,
+      "step": 11800
+    },
+    {
+      "epoch": 0.29525,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.9541666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5794,
+      "loss/crossentropy": 2.1220794349908827,
+      "loss/hidden": 3.451953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20339306239038707,
+      "step": 11810
+    },
+    {
+      "epoch": 0.2955,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.5384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5332,
+      "loss/crossentropy": 2.1137719243764876,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17815515641123056,
+      "step": 11820
+    },
+    {
+      "epoch": 0.29575,
+      "grad_norm": 36.0,
+      "grad_norm_var": 2.2046223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5875,
+      "loss/crossentropy": 2.060352697968483,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20077992398291827,
+      "step": 11830
+    },
+    {
+      "epoch": 0.296,
+      "grad_norm": 30.875,
+      "grad_norm_var": 451.95625,
+      "learning_rate": 0.0001,
+      "loss": 7.4559,
+      "loss/crossentropy": 2.0753053151071073,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18949195835739374,
+      "step": 11840
+    },
+    {
+      "epoch": 0.29625,
+      "grad_norm": 30.125,
+      "grad_norm_var": 464.97389322916666,
+      "learning_rate": 0.0001,
+      "loss": 7.5175,
+      "loss/crossentropy": 2.216064375638962,
+      "loss/hidden": 3.4078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19813688658177853,
+      "step": 11850
+    },
+    {
+      "epoch": 0.2965,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.785872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4496,
+      "loss/crossentropy": 2.151813006401062,
+      "loss/hidden": 3.473046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2109227592125535,
+      "step": 11860
+    },
+    {
+      "epoch": 0.29675,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.512955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.6473,
+      "loss/crossentropy": 2.2002500608563422,
+      "loss/hidden": 3.440234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22418888993561267,
+      "step": 11870
+    },
+    {
+      "epoch": 0.297,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.9947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5421,
+      "loss/crossentropy": 2.161974515020847,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20292557254433632,
+      "step": 11880
+    },
+    {
+      "epoch": 0.29725,
+      "grad_norm": 30.375,
+      "grad_norm_var": 4.8125,
+      "learning_rate": 0.0001,
+      "loss": 7.4689,
+      "loss/crossentropy": 2.1570896983146666,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18599064461886883,
+      "step": 11890
+    },
+    {
+      "epoch": 0.2975,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.1639973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.602,
+      "loss/crossentropy": 2.116308979690075,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20236566830426456,
+      "step": 11900
+    },
+    {
+      "epoch": 0.29775,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.478125,
+      "learning_rate": 0.0001,
+      "loss": 7.4508,
+      "loss/crossentropy": 2.0929686918854715,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18616765905171634,
+      "step": 11910
+    },
+    {
+      "epoch": 0.298,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.142643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5382,
+      "loss/crossentropy": 2.0649349838495255,
+      "loss/hidden": 3.491015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22931914012879134,
+      "step": 11920
+    },
+    {
+      "epoch": 0.29825,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.2009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5117,
+      "loss/crossentropy": 2.218345195055008,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19991735983639955,
+      "step": 11930
+    },
+    {
+      "epoch": 0.2985,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.8608723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4722,
+      "loss/crossentropy": 2.125768192112446,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20784178506582976,
+      "step": 11940
+    },
+    {
+      "epoch": 0.29875,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.0625,
+      "learning_rate": 0.0001,
+      "loss": 7.5416,
+      "loss/crossentropy": 2.1809418946504593,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20119600538164378,
+      "step": 11950
+    },
+    {
+      "epoch": 0.299,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.268684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4921,
+      "loss/crossentropy": 2.1606124877929687,
+      "loss/hidden": 3.34296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20459957644343377,
+      "step": 11960
+    },
+    {
+      "epoch": 0.29925,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.785872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4666,
+      "loss/crossentropy": 2.0705448195338247,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1915181243792176,
+      "step": 11970
+    },
+    {
+      "epoch": 0.2995,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.0718098958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5573,
+      "loss/crossentropy": 2.219760975241661,
+      "loss/hidden": 3.5890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2002726301550865,
+      "step": 11980
+    },
+    {
+      "epoch": 0.29975,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.840625,
+      "learning_rate": 0.0001,
+      "loss": 7.4501,
+      "loss/crossentropy": 2.1825525611639023,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1877422071993351,
+      "step": 11990
+    },
+    {
+      "epoch": 0.3,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.609375,
+      "learning_rate": 0.0001,
+      "loss": 7.568,
+      "loss/crossentropy": 2.034752905368805,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19768680389970542,
+      "step": 12000
+    },
+    {
+      "epoch": 0.30025,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.9072916666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4314,
+      "loss/crossentropy": 1.9573114350438119,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17485380116850138,
+      "step": 12010
+    },
+    {
+      "epoch": 0.3005,
+      "grad_norm": 33.0,
+      "grad_norm_var": 1.8921223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5809,
+      "loss/crossentropy": 2.2018728822469713,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1915513090789318,
+      "step": 12020
+    },
+    {
+      "epoch": 0.30075,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.051041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4786,
+      "loss/crossentropy": 2.121686166524887,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17941332049667835,
+      "step": 12030
+    },
+    {
+      "epoch": 0.301,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.1660807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.146714040637016,
+      "loss/hidden": 3.348828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18955296482890843,
+      "step": 12040
+    },
+    {
+      "epoch": 0.30125,
+      "grad_norm": 29.875,
+      "grad_norm_var": 4.1916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4308,
+      "loss/crossentropy": 2.1516218155622484,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19275461547076703,
+      "step": 12050
+    },
+    {
+      "epoch": 0.3015,
+      "grad_norm": 29.0,
+      "grad_norm_var": 4.0244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4391,
+      "loss/crossentropy": 2.082157927751541,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2028081189841032,
+      "step": 12060
+    },
+    {
+      "epoch": 0.30175,
+      "grad_norm": 8153726976.0,
+      "grad_norm_var": 4.1552039435295524e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5841,
+      "loss/crossentropy": 2.155959886312485,
+      "loss/hidden": 3.412890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21119905970990657,
+      "step": 12070
+    },
+    {
+      "epoch": 0.302,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.1552039419327805e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4881,
+      "loss/crossentropy": 2.175818865001202,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19981482364237307,
+      "step": 12080
+    },
+    {
+      "epoch": 0.30225,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.1306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5543,
+      "loss/crossentropy": 2.153113231062889,
+      "loss/hidden": 3.305078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18751353081315755,
+      "step": 12090
+    },
+    {
+      "epoch": 0.3025,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.55,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.113540416955948,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18914419934153556,
+      "step": 12100
+    },
+    {
+      "epoch": 0.30275,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.664322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5051,
+      "loss/crossentropy": 2.0748536720871926,
+      "loss/hidden": 3.541015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21318843495100737,
+      "step": 12110
+    },
+    {
+      "epoch": 0.303,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.9830729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5366,
+      "loss/crossentropy": 2.1331150621175765,
+      "loss/hidden": 3.38671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19423545580357313,
+      "step": 12120
+    },
+    {
+      "epoch": 0.30325,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.4400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4711,
+      "loss/crossentropy": 2.220665395259857,
+      "loss/hidden": 3.468359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21431318037211894,
+      "step": 12130
+    },
+    {
+      "epoch": 0.3035,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.9233723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5293,
+      "loss/crossentropy": 2.076695668697357,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19359768629074098,
+      "step": 12140
+    },
+    {
+      "epoch": 0.30375,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.037239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.502,
+      "loss/crossentropy": 2.007551699131727,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18256566096097232,
+      "step": 12150
+    },
+    {
+      "epoch": 0.304,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.1858723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5602,
+      "loss/crossentropy": 2.2614489823579786,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21236035842448472,
+      "step": 12160
+    },
+    {
+      "epoch": 0.30425,
+      "grad_norm": 29.125,
+      "grad_norm_var": 4.1150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4919,
+      "loss/crossentropy": 2.102987366914749,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19789319038391112,
+      "step": 12170
+    },
+    {
+      "epoch": 0.3045,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.8921223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5418,
+      "loss/crossentropy": 2.1469786316156387,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20192387700080872,
+      "step": 12180
+    },
+    {
+      "epoch": 0.30475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.7893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4362,
+      "loss/crossentropy": 2.068286693096161,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19657166097313167,
+      "step": 12190
+    },
+    {
+      "epoch": 0.305,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.6239583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4976,
+      "loss/crossentropy": 2.0656373471021654,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264612033963204,
+      "step": 12200
+    },
+    {
+      "epoch": 0.30525,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.4760416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5079,
+      "loss/crossentropy": 2.0986109912395476,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18751255813986062,
+      "step": 12210
+    },
+    {
+      "epoch": 0.3055,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.784309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5652,
+      "loss/crossentropy": 2.128317725658417,
+      "loss/hidden": 3.52890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19654161781072615,
+      "step": 12220
+    },
+    {
+      "epoch": 0.30575,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.205989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4884,
+      "loss/crossentropy": 2.1307172656059263,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1911188881844282,
+      "step": 12230
+    },
+    {
+      "epoch": 0.306,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.9622395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5641,
+      "loss/crossentropy": 2.106894627213478,
+      "loss/hidden": 3.4125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19271711856126786,
+      "step": 12240
+    },
+    {
+      "epoch": 0.30625,
+      "grad_norm": 29.25,
+      "grad_norm_var": 14.451822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5277,
+      "loss/crossentropy": 2.09238311201334,
+      "loss/hidden": 3.2703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898360250517726,
+      "step": 12250
+    },
+    {
+      "epoch": 0.3065,
+      "grad_norm": 30.375,
+      "grad_norm_var": 44.141666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5351,
+      "loss/crossentropy": 2.189305357635021,
+      "loss/hidden": 3.52421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2375885730609298,
+      "step": 12260
+    },
+    {
+      "epoch": 0.30675,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.3309895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.518,
+      "loss/crossentropy": 2.0894197657704354,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18371071619912982,
+      "step": 12270
+    },
+    {
+      "epoch": 0.307,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.203125,
+      "learning_rate": 0.0001,
+      "loss": 7.5372,
+      "loss/crossentropy": 2.2316055417060854,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1975148657336831,
+      "step": 12280
+    },
+    {
+      "epoch": 0.30725,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.3671223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.499,
+      "loss/crossentropy": 2.168405243754387,
+      "loss/hidden": 3.463671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2092548543587327,
+      "step": 12290
+    },
+    {
+      "epoch": 0.3075,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.1372395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4549,
+      "loss/crossentropy": 2.1289770871400835,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1914392326027155,
+      "step": 12300
+    },
+    {
+      "epoch": 0.30775,
+      "grad_norm": 31.75,
+      "grad_norm_var": 15.703125,
+      "learning_rate": 0.0001,
+      "loss": 7.4803,
+      "loss/crossentropy": 1.9998106390237809,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981509406119585,
+      "step": 12310
+    },
+    {
+      "epoch": 0.308,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.53125,
+      "learning_rate": 0.0001,
+      "loss": 7.5639,
+      "loss/crossentropy": 2.1574712097644806,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19861921556293965,
+      "step": 12320
+    },
+    {
+      "epoch": 0.30825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.6462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5662,
+      "loss/crossentropy": 2.1368797808885573,
+      "loss/hidden": 3.36484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19260376282036304,
+      "step": 12330
+    },
+    {
+      "epoch": 0.3085,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.4353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5165,
+      "loss/crossentropy": 2.157291880249977,
+      "loss/hidden": 3.326171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19906458482146264,
+      "step": 12340
+    },
+    {
+      "epoch": 0.30875,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.84765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5003,
+      "loss/crossentropy": 2.198679545521736,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20154633428901433,
+      "step": 12350
+    },
+    {
+      "epoch": 0.309,
+      "grad_norm": 30.75,
+      "grad_norm_var": 6.72890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3993,
+      "loss/crossentropy": 2.0015364930033686,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17583242971450092,
+      "step": 12360
+    },
+    {
+      "epoch": 0.30925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 6.652083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5535,
+      "loss/crossentropy": 2.0551007747650147,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17546763848513364,
+      "step": 12370
+    },
+    {
+      "epoch": 0.3095,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.09140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5495,
+      "loss/crossentropy": 2.1019306644797324,
+      "loss/hidden": 3.47734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21077220756560563,
+      "step": 12380
+    },
+    {
+      "epoch": 0.30975,
+      "grad_norm": 28.625,
+      "grad_norm_var": 5.5947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5257,
+      "loss/crossentropy": 2.059135194122791,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19150259550660848,
+      "step": 12390
+    },
+    {
+      "epoch": 0.31,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.9872395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5651,
+      "loss/crossentropy": 2.1537068828940393,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20102579537779092,
+      "step": 12400
+    },
+    {
+      "epoch": 0.31025,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.4660807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4731,
+      "loss/crossentropy": 2.1616740792989733,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19620374999940396,
+      "step": 12410
+    },
+    {
+      "epoch": 0.3105,
+      "grad_norm": 28.5,
+      "grad_norm_var": 19.721809895833335,
+      "learning_rate": 0.0001,
+      "loss": 7.5764,
+      "loss/crossentropy": 2.094822147488594,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18539618086069823,
+      "step": 12420
+    },
+    {
+      "epoch": 0.31075,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.62265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5147,
+      "loss/crossentropy": 2.061335799098015,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2312490541487932,
+      "step": 12430
+    },
+    {
+      "epoch": 0.311,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.849739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.47,
+      "loss/crossentropy": 2.0725695550441743,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18395131248980762,
+      "step": 12440
+    },
+    {
+      "epoch": 0.31125,
+      "grad_norm": 31.625,
+      "grad_norm_var": 37.8478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5891,
+      "loss/crossentropy": 2.177759502083063,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19881557505577802,
+      "step": 12450
+    },
+    {
+      "epoch": 0.3115,
+      "grad_norm": 30.0,
+      "grad_norm_var": 37.426822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5683,
+      "loss/crossentropy": 2.215823370218277,
+      "loss/hidden": 3.49296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21547670513391495,
+      "step": 12460
+    },
+    {
+      "epoch": 0.31175,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.113997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5074,
+      "loss/crossentropy": 2.152268370985985,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19586993865668773,
+      "step": 12470
+    },
+    {
+      "epoch": 0.312,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.90625,
+      "learning_rate": 0.0001,
+      "loss": 7.4881,
+      "loss/crossentropy": 2.1033642858266832,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19163406621664764,
+      "step": 12480
+    },
+    {
+      "epoch": 0.31225,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.8848307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4933,
+      "loss/crossentropy": 2.1959578454494477,
+      "loss/hidden": 3.286328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19511055387556553,
+      "step": 12490
+    },
+    {
+      "epoch": 0.3125,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.41875,
+      "learning_rate": 0.0001,
+      "loss": 7.4391,
+      "loss/crossentropy": 2.195823746919632,
+      "loss/hidden": 3.2265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17207753676921128,
+      "step": 12500
+    },
+    {
+      "epoch": 0.31275,
+      "grad_norm": 29.0,
+      "grad_norm_var": 3.8358723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5033,
+      "loss/crossentropy": 2.0963750213384627,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17908218652009963,
+      "step": 12510
+    },
+    {
+      "epoch": 0.313,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.308333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4239,
+      "loss/crossentropy": 1.9604337126016618,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18109768088907002,
+      "step": 12520
+    },
+    {
+      "epoch": 0.31325,
+      "grad_norm": 33.5,
+      "grad_norm_var": 5.990625,
+      "learning_rate": 0.0001,
+      "loss": 7.48,
+      "loss/crossentropy": 2.1750424414873124,
+      "loss/hidden": 3.444921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19859890583902598,
+      "step": 12530
+    },
+    {
+      "epoch": 0.3135,
+      "grad_norm": 34.75,
+      "grad_norm_var": 2.2145182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5841,
+      "loss/crossentropy": 2.2009086132049562,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20133890733122825,
+      "step": 12540
+    },
+    {
+      "epoch": 0.31375,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.0931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4803,
+      "loss/crossentropy": 2.3103501707315446,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2070220198482275,
+      "step": 12550
+    },
+    {
+      "epoch": 0.314,
+      "grad_norm": 32.75,
+      "grad_norm_var": 20.39140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5026,
+      "loss/crossentropy": 2.0587240263819693,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18765954226255416,
+      "step": 12560
+    },
+    {
+      "epoch": 0.31425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.32265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5031,
+      "loss/crossentropy": 2.194789284467697,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19717307426035405,
+      "step": 12570
+    },
+    {
+      "epoch": 0.3145,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.5416666666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4895,
+      "loss/crossentropy": 2.2054634511470796,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1899452781304717,
+      "step": 12580
+    },
+    {
+      "epoch": 0.31475,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.279166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4928,
+      "loss/crossentropy": 2.1559681832790374,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18954651188105345,
+      "step": 12590
+    },
+    {
+      "epoch": 0.315,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.5337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4422,
+      "loss/crossentropy": 2.254222446680069,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19737180043011904,
+      "step": 12600
+    },
+    {
+      "epoch": 0.31525,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.3681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4784,
+      "loss/crossentropy": 2.146516689658165,
+      "loss/hidden": 3.467578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20649947822093964,
+      "step": 12610
+    },
+    {
+      "epoch": 0.3155,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.608268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4844,
+      "loss/crossentropy": 2.0874697089195253,
+      "loss/hidden": 3.467578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19496283829212188,
+      "step": 12620
+    },
+    {
+      "epoch": 0.31575,
+      "grad_norm": 31.0,
+      "grad_norm_var": 13.2681640625,
+      "learning_rate": 0.0001,
+      "loss": 7.55,
+      "loss/crossentropy": 2.1777540147304535,
+      "loss/hidden": 3.41015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20055020954459907,
+      "step": 12630
+    },
+    {
+      "epoch": 0.316,
+      "grad_norm": 29.75,
+      "grad_norm_var": 13.5791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4849,
+      "loss/crossentropy": 2.1960367292165754,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19755847714841365,
+      "step": 12640
+    },
+    {
+      "epoch": 0.31625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.7603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4898,
+      "loss/crossentropy": 2.1551318049430845,
+      "loss/hidden": 3.428515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2070561083033681,
+      "step": 12650
+    },
+    {
+      "epoch": 0.3165,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.3018229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4701,
+      "loss/crossentropy": 2.1339357078075407,
+      "loss/hidden": 3.471875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19703674577176572,
+      "step": 12660
+    },
+    {
+      "epoch": 0.31675,
+      "grad_norm": 29.5,
+      "grad_norm_var": 11.317643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4713,
+      "loss/crossentropy": 2.0686328157782556,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1992340425029397,
+      "step": 12670
+    },
+    {
+      "epoch": 0.317,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.022330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3718,
+      "loss/crossentropy": 2.1510003715753556,
+      "loss/hidden": 3.285546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1883657006546855,
+      "step": 12680
+    },
+    {
+      "epoch": 0.31725,
+      "grad_norm": 31.875,
+      "grad_norm_var": 1.9809895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4673,
+      "loss/crossentropy": 2.0955928951501845,
+      "loss/hidden": 3.341796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19225707659497857,
+      "step": 12690
+    },
+    {
+      "epoch": 0.3175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.220572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4942,
+      "loss/crossentropy": 2.178106242418289,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19251996818929912,
+      "step": 12700
+    },
+    {
+      "epoch": 0.31775,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.0405598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.419,
+      "loss/crossentropy": 2.323311746120453,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18311577010899782,
+      "step": 12710
+    },
+    {
+      "epoch": 0.318,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.255143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3814,
+      "loss/crossentropy": 2.166349285840988,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19064488224685192,
+      "step": 12720
+    },
+    {
+      "epoch": 0.31825,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.2400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4573,
+      "loss/crossentropy": 2.0821331262588503,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19217007402330638,
+      "step": 12730
+    },
+    {
+      "epoch": 0.3185,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.31640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3945,
+      "loss/crossentropy": 2.2248596966266634,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1899226889014244,
+      "step": 12740
+    },
+    {
+      "epoch": 0.31875,
+      "grad_norm": 28.25,
+      "grad_norm_var": 1.87890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5347,
+      "loss/crossentropy": 2.2382041692733763,
+      "loss/hidden": 3.28671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984447505325079,
+      "step": 12750
+    },
+    {
+      "epoch": 0.319,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.644791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.566,
+      "loss/crossentropy": 2.1457022726535797,
+      "loss/hidden": 3.439453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2007513264194131,
+      "step": 12760
+    },
+    {
+      "epoch": 0.31925,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.006705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3134,
+      "loss/crossentropy": 2.0300011321902276,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18935967702418566,
+      "step": 12770
+    },
+    {
+      "epoch": 0.3195,
+      "grad_norm": 31.25,
+      "grad_norm_var": 21.826041666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5243,
+      "loss/crossentropy": 2.1192230358719826,
+      "loss/hidden": 3.566015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23796151615679265,
+      "step": 12780
+    },
+    {
+      "epoch": 0.31975,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.03125,
+      "learning_rate": 0.0001,
+      "loss": 7.4538,
+      "loss/crossentropy": 2.1349218785762787,
+      "loss/hidden": 3.291796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18145886678248643,
+      "step": 12790
+    },
+    {
+      "epoch": 0.32,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.998893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5129,
+      "loss/crossentropy": 2.0976530820131303,
+      "loss/hidden": 3.359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19899471253156661,
+      "step": 12800
+    },
+    {
+      "epoch": 0.32025,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.8134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3964,
+      "loss/crossentropy": 2.1131363533437253,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19826100897043944,
+      "step": 12810
+    },
+    {
+      "epoch": 0.3205,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.3947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3552,
+      "loss/crossentropy": 2.1666725143790244,
+      "loss/hidden": 3.25625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17309209685772659,
+      "step": 12820
+    },
+    {
+      "epoch": 0.32075,
+      "grad_norm": 35.25,
+      "grad_norm_var": 6.137434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4774,
+      "loss/crossentropy": 2.0877491027116775,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19852426294237374,
+      "step": 12830
+    },
+    {
+      "epoch": 0.321,
+      "grad_norm": 31.25,
+      "grad_norm_var": 31.881184895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.5434,
+      "loss/crossentropy": 2.0780150018632413,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19723174571990967,
+      "step": 12840
+    },
+    {
+      "epoch": 0.32125,
+      "grad_norm": 30.625,
+      "grad_norm_var": 7.818489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.393,
+      "loss/crossentropy": 2.260560867190361,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2075773360207677,
+      "step": 12850
+    },
+    {
+      "epoch": 0.3215,
+      "grad_norm": 32.0,
+      "grad_norm_var": 10.083072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3997,
+      "loss/crossentropy": 2.1417742699384688,
+      "loss/hidden": 3.273828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17287890873849393,
+      "step": 12860
+    },
+    {
+      "epoch": 0.32175,
+      "grad_norm": 27.75,
+      "grad_norm_var": 2.2705729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3752,
+      "loss/crossentropy": 2.131768229603767,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19949521739035844,
+      "step": 12870
+    },
+    {
+      "epoch": 0.322,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.3863932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4512,
+      "loss/crossentropy": 2.266862842440605,
+      "loss/hidden": 3.28359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19267170149832963,
+      "step": 12880
+    },
+    {
+      "epoch": 0.32225,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.132291666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.534,
+      "loss/crossentropy": 2.0951140992343427,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2013495948165655,
+      "step": 12890
+    },
+    {
+      "epoch": 0.3225,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.995833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4035,
+      "loss/crossentropy": 2.2091780215501786,
+      "loss/hidden": 3.327734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2002504987642169,
+      "step": 12900
+    },
+    {
+      "epoch": 0.32275,
+      "grad_norm": 30.375,
+      "grad_norm_var": 0.6301432291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4548,
+      "loss/crossentropy": 2.1198602825403214,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18528146538883447,
+      "step": 12910
+    },
+    {
+      "epoch": 0.323,
+      "grad_norm": 28.875,
+      "grad_norm_var": 0.8921223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4588,
+      "loss/crossentropy": 2.18964803814888,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19150267392396927,
+      "step": 12920
+    },
+    {
+      "epoch": 0.32325,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.7434895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5432,
+      "loss/crossentropy": 2.1617533951997756,
+      "loss/hidden": 3.349609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20485697444528342,
+      "step": 12930
+    },
+    {
+      "epoch": 0.3235,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.9655598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5072,
+      "loss/crossentropy": 1.9271393403410912,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19526469726115464,
+      "step": 12940
+    },
+    {
+      "epoch": 0.32375,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.6333333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5258,
+      "loss/crossentropy": 2.0988502144813537,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18954623490571976,
+      "step": 12950
+    },
+    {
+      "epoch": 0.324,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.6372395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4828,
+      "loss/crossentropy": 2.0505402773618697,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19810300767421724,
+      "step": 12960
+    },
+    {
+      "epoch": 0.32425,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.749739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5484,
+      "loss/crossentropy": 2.0827158600091935,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18706083726137876,
+      "step": 12970
+    },
+    {
+      "epoch": 0.3245,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.0462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.1275136440992357,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18214173689484597,
+      "step": 12980
+    },
+    {
+      "epoch": 0.32475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 8.691080729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.2110335171222686,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21523924693465232,
+      "step": 12990
+    },
+    {
+      "epoch": 0.325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.461393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4891,
+      "loss/crossentropy": 2.147765000164509,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19382500275969505,
+      "step": 13000
+    },
+    {
+      "epoch": 0.32525,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.283072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4506,
+      "loss/crossentropy": 2.0027987882494926,
+      "loss/hidden": 3.44375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18850265592336654,
+      "step": 13010
+    },
+    {
+      "epoch": 0.3255,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.158333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4449,
+      "loss/crossentropy": 2.239279788732529,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2110375264659524,
+      "step": 13020
+    },
+    {
+      "epoch": 0.32575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 5.57265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3912,
+      "loss/crossentropy": 2.1745992332696913,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2059601791203022,
+      "step": 13030
+    },
+    {
+      "epoch": 0.326,
+      "grad_norm": 29.125,
+      "grad_norm_var": 8.201822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4196,
+      "loss/crossentropy": 2.1236157566308975,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19811425134539604,
+      "step": 13040
+    },
+    {
+      "epoch": 0.32625,
+      "grad_norm": 29.125,
+      "grad_norm_var": 14.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3974,
+      "loss/crossentropy": 2.1588226526975633,
+      "loss/hidden": 3.43515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2041918309405446,
+      "step": 13050
+    },
+    {
+      "epoch": 0.3265,
+      "grad_norm": 33.25,
+      "grad_norm_var": 14.5853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4069,
+      "loss/crossentropy": 2.1836913257837294,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884943839162588,
+      "step": 13060
+    },
+    {
+      "epoch": 0.32675,
+      "grad_norm": 28.875,
+      "grad_norm_var": 15.664518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5712,
+      "loss/crossentropy": 2.173456999659538,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20252696424722672,
+      "step": 13070
+    },
+    {
+      "epoch": 0.327,
+      "grad_norm": 32.0,
+      "grad_norm_var": 13.63515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4706,
+      "loss/crossentropy": 2.222018560767174,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19077708926051856,
+      "step": 13080
+    },
+    {
+      "epoch": 0.32725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 8.099934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5034,
+      "loss/crossentropy": 2.2448195964097977,
+      "loss/hidden": 3.5734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.241958281211555,
+      "step": 13090
+    },
+    {
+      "epoch": 0.3275,
+      "grad_norm": 28.625,
+      "grad_norm_var": 10.601822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4343,
+      "loss/crossentropy": 2.231740725040436,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1907974945381284,
+      "step": 13100
+    },
+    {
+      "epoch": 0.32775,
+      "grad_norm": 35.25,
+      "grad_norm_var": 15.445247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.45,
+      "loss/crossentropy": 2.116320764273405,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17694039307534695,
+      "step": 13110
+    },
+    {
+      "epoch": 0.328,
+      "grad_norm": 35.25,
+      "grad_norm_var": 11.502018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5445,
+      "loss/crossentropy": 2.099977496266365,
+      "loss/hidden": 3.52265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20624227696098388,
+      "step": 13120
+    },
+    {
+      "epoch": 0.32825,
+      "grad_norm": 30.75,
+      "grad_norm_var": 7.521875,
+      "learning_rate": 0.0001,
+      "loss": 7.4769,
+      "loss/crossentropy": 2.1053822576999663,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19216917939484118,
+      "step": 13130
+    },
+    {
+      "epoch": 0.3285,
+      "grad_norm": 34.75,
+      "grad_norm_var": 14.083268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4462,
+      "loss/crossentropy": 1.9960962146520616,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18692312780767678,
+      "step": 13140
+    },
+    {
+      "epoch": 0.32875,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.630208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4762,
+      "loss/crossentropy": 2.1613951712846755,
+      "loss/hidden": 3.537109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20941741671413183,
+      "step": 13150
+    },
+    {
+      "epoch": 0.329,
+      "grad_norm": 33.5,
+      "grad_norm_var": 6.3900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4808,
+      "loss/crossentropy": 2.189284533262253,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19653193168342115,
+      "step": 13160
+    },
+    {
+      "epoch": 0.32925,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.8489583333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4326,
+      "loss/crossentropy": 2.1141976565122604,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18286358639597894,
+      "step": 13170
+    },
+    {
+      "epoch": 0.3295,
+      "grad_norm": 31.25,
+      "grad_norm_var": 5.464583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3723,
+      "loss/crossentropy": 2.1913035601377486,
+      "loss/hidden": 3.272265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1858003603294492,
+      "step": 13180
+    },
+    {
+      "epoch": 0.32975,
+      "grad_norm": 29.375,
+      "grad_norm_var": 17.029622395833332,
+      "learning_rate": 0.0001,
+      "loss": 7.4759,
+      "loss/crossentropy": 1.9873070642352104,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1818276308476925,
+      "step": 13190
+    },
+    {
+      "epoch": 0.33,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.840559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4491,
+      "loss/crossentropy": 2.0309854298830032,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18652353323996068,
+      "step": 13200
+    },
+    {
+      "epoch": 0.33025,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.896875,
+      "learning_rate": 0.0001,
+      "loss": 7.4246,
+      "loss/crossentropy": 2.1724058866500853,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19039318859577178,
+      "step": 13210
+    },
+    {
+      "epoch": 0.3305,
+      "grad_norm": 33.25,
+      "grad_norm_var": 3.903580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5702,
+      "loss/crossentropy": 2.144900679588318,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19142087129876018,
+      "step": 13220
+    },
+    {
+      "epoch": 0.33075,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.763997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4713,
+      "loss/crossentropy": 2.1755593717098236,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18869520910084248,
+      "step": 13230
+    },
+    {
+      "epoch": 0.331,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.674934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3374,
+      "loss/crossentropy": 2.2121481001377106,
+      "loss/hidden": 3.27421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18241031859070062,
+      "step": 13240
+    },
+    {
+      "epoch": 0.33125,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.4494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3846,
+      "loss/crossentropy": 2.166096642613411,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18299505580216646,
+      "step": 13250
+    },
+    {
+      "epoch": 0.3315,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.7018229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4058,
+      "loss/crossentropy": 2.0837877124547957,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18499783985316753,
+      "step": 13260
+    },
+    {
+      "epoch": 0.33175,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.5603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4647,
+      "loss/crossentropy": 2.114008131623268,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1959222162142396,
+      "step": 13270
+    },
+    {
+      "epoch": 0.332,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.910872395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.511,
+      "loss/crossentropy": 2.2248871475458145,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18698529433459044,
+      "step": 13280
+    },
+    {
+      "epoch": 0.33225,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.077018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4879,
+      "loss/crossentropy": 2.203841781616211,
+      "loss/hidden": 3.472265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2007746297866106,
+      "step": 13290
+    },
+    {
+      "epoch": 0.3325,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.0483723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5087,
+      "loss/crossentropy": 2.1073865801095963,
+      "loss/hidden": 3.45,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2040329247713089,
+      "step": 13300
+    },
+    {
+      "epoch": 0.33275,
+      "grad_norm": 31.5,
+      "grad_norm_var": 5.128580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.425,
+      "loss/crossentropy": 2.1313794389367104,
+      "loss/hidden": 3.41875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1976473730057478,
+      "step": 13310
+    },
+    {
+      "epoch": 0.333,
+      "grad_norm": 28.625,
+      "grad_norm_var": 5.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.3878,
+      "loss/crossentropy": 2.0418414741754534,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19092906154692174,
+      "step": 13320
+    },
+    {
+      "epoch": 0.33325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.6229166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4116,
+      "loss/crossentropy": 2.1283623784780503,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854261551052332,
+      "step": 13330
+    },
+    {
+      "epoch": 0.3335,
+      "grad_norm": 30.375,
+      "grad_norm_var": 6.0994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4184,
+      "loss/crossentropy": 2.084540565311909,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18820476066321135,
+      "step": 13340
+    },
+    {
+      "epoch": 0.33375,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.408072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4188,
+      "loss/crossentropy": 2.133148857951164,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17596916407346724,
+      "step": 13350
+    },
+    {
+      "epoch": 0.334,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.331184895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4688,
+      "loss/crossentropy": 2.070942518115044,
+      "loss/hidden": 3.3,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17796068042516708,
+      "step": 13360
+    },
+    {
+      "epoch": 0.33425,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.38515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5483,
+      "loss/crossentropy": 2.1084318548440932,
+      "loss/hidden": 3.355859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19390486590564252,
+      "step": 13370
+    },
+    {
+      "epoch": 0.3345,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.348372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4769,
+      "loss/crossentropy": 2.0068695291876795,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18838447090238333,
+      "step": 13380
+    },
+    {
+      "epoch": 0.33475,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.5145833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4633,
+      "loss/crossentropy": 2.118716311454773,
+      "loss/hidden": 3.44140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20167391393333672,
+      "step": 13390
+    },
+    {
+      "epoch": 0.335,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.232291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4333,
+      "loss/crossentropy": 2.1667033195495606,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19795446433126926,
+      "step": 13400
+    },
+    {
+      "epoch": 0.33525,
+      "grad_norm": 28.375,
+      "grad_norm_var": 3.226822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.489,
+      "loss/crossentropy": 2.1369444996118547,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19194683507084848,
+      "step": 13410
+    },
+    {
+      "epoch": 0.3355,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.664322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.514,
+      "loss/crossentropy": 2.1737091183662414,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18577092699706554,
+      "step": 13420
+    },
+    {
+      "epoch": 0.33575,
+      "grad_norm": 30.625,
+      "grad_norm_var": 2.653059895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5082,
+      "loss/crossentropy": 2.1547589361667634,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893642207607627,
+      "step": 13430
+    },
+    {
+      "epoch": 0.336,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.721809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4585,
+      "loss/crossentropy": 2.228061503171921,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19817217160016298,
+      "step": 13440
+    },
+    {
+      "epoch": 0.33625,
+      "grad_norm": 32.0,
+      "grad_norm_var": 4.395768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4325,
+      "loss/crossentropy": 2.0584416151046754,
+      "loss/hidden": 3.33828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18820574525743722,
+      "step": 13450
+    },
+    {
+      "epoch": 0.3365,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.1978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5845,
+      "loss/crossentropy": 2.1853963419795037,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20326036512851714,
+      "step": 13460
+    },
+    {
+      "epoch": 0.33675,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.5145833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4957,
+      "loss/crossentropy": 2.058859869837761,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18476009368896484,
+      "step": 13470
+    },
+    {
+      "epoch": 0.337,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.189322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3814,
+      "loss/crossentropy": 2.09515139721334,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18534991908818482,
+      "step": 13480
+    },
+    {
+      "epoch": 0.33725,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.5666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4269,
+      "loss/crossentropy": 2.1185129195451737,
+      "loss/hidden": 3.427734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19515257980674505,
+      "step": 13490
+    },
+    {
+      "epoch": 0.3375,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.351822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5122,
+      "loss/crossentropy": 2.138470843434334,
+      "loss/hidden": 3.424609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19903527870774268,
+      "step": 13500
+    },
+    {
+      "epoch": 0.33775,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.41875,
+      "learning_rate": 0.0001,
+      "loss": 7.6212,
+      "loss/crossentropy": 2.1452796325087546,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1878256807103753,
+      "step": 13510
+    },
+    {
+      "epoch": 0.338,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.1223307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5111,
+      "loss/crossentropy": 2.1782995223999024,
+      "loss/hidden": 3.423828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19461661092936994,
+      "step": 13520
+    },
+    {
+      "epoch": 0.33825,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.5809895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4255,
+      "loss/crossentropy": 2.184438717365265,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2046054745092988,
+      "step": 13530
+    },
+    {
+      "epoch": 0.3385,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.736458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4071,
+      "loss/crossentropy": 2.082030822336674,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1870069233700633,
+      "step": 13540
+    },
+    {
+      "epoch": 0.33875,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.645247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.0698161378502844,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18587249666452407,
+      "step": 13550
+    },
+    {
+      "epoch": 0.339,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.824934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5059,
+      "loss/crossentropy": 2.1267208635807036,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20201743990182877,
+      "step": 13560
+    },
+    {
+      "epoch": 0.33925,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.14765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5921,
+      "loss/crossentropy": 2.1734737694263457,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21583349518477918,
+      "step": 13570
+    },
+    {
+      "epoch": 0.3395,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.4989583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4379,
+      "loss/crossentropy": 2.214940372109413,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21127836983650922,
+      "step": 13580
+    },
+    {
+      "epoch": 0.33975,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.468489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5302,
+      "loss/crossentropy": 2.086373192071915,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1951220341026783,
+      "step": 13590
+    },
+    {
+      "epoch": 0.34,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.908333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4568,
+      "loss/crossentropy": 2.1543196618556975,
+      "loss/hidden": 3.358984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20275593940168618,
+      "step": 13600
+    },
+    {
+      "epoch": 0.34025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.83515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4797,
+      "loss/crossentropy": 2.0696166858077047,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18954422865062953,
+      "step": 13610
+    },
+    {
+      "epoch": 0.3405,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.9145182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5892,
+      "loss/crossentropy": 2.1566884085536002,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22052888944745064,
+      "step": 13620
+    },
+    {
+      "epoch": 0.34075,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.57890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5791,
+      "loss/crossentropy": 2.1455826848745345,
+      "loss/hidden": 3.46953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2092026699334383,
+      "step": 13630
+    },
+    {
+      "epoch": 0.341,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.464583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5429,
+      "loss/crossentropy": 2.208886668086052,
+      "loss/hidden": 3.29609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18887277618050574,
+      "step": 13640
+    },
+    {
+      "epoch": 0.34125,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.5747395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4508,
+      "loss/crossentropy": 2.013479822874069,
+      "loss/hidden": 3.3765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18337813336402178,
+      "step": 13650
+    },
+    {
+      "epoch": 0.3415,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.2509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5422,
+      "loss/crossentropy": 2.143598672747612,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2065559087321162,
+      "step": 13660
+    },
+    {
+      "epoch": 0.34175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.7927083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5533,
+      "loss/crossentropy": 2.2552105635404587,
+      "loss/hidden": 3.4890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19607291985303163,
+      "step": 13670
+    },
+    {
+      "epoch": 0.342,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.9983723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4126,
+      "loss/crossentropy": 2.1220908224582673,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19602216491475702,
+      "step": 13680
+    },
+    {
+      "epoch": 0.34225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.3827473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4261,
+      "loss/crossentropy": 2.045897740870714,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18407570132985712,
+      "step": 13690
+    },
+    {
+      "epoch": 0.3425,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.2030598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4016,
+      "loss/crossentropy": 2.1739032745361326,
+      "loss/hidden": 3.31796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1999795723706484,
+      "step": 13700
+    },
+    {
+      "epoch": 0.34275,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.0410807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4881,
+      "loss/crossentropy": 2.120612841844559,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18889743015170096,
+      "step": 13710
+    },
+    {
+      "epoch": 0.343,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.365625,
+      "learning_rate": 0.0001,
+      "loss": 7.5733,
+      "loss/crossentropy": 2.222416913509369,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20086848884820938,
+      "step": 13720
+    },
+    {
+      "epoch": 0.34325,
+      "grad_norm": 28.25,
+      "grad_norm_var": 2.4614583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5037,
+      "loss/crossentropy": 2.1907012492418287,
+      "loss/hidden": 3.347265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.190412163361907,
+      "step": 13730
+    },
+    {
+      "epoch": 0.3435,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.7978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5508,
+      "loss/crossentropy": 2.1012231409549713,
+      "loss/hidden": 3.46328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20470567271113396,
+      "step": 13740
+    },
+    {
+      "epoch": 0.34375,
+      "grad_norm": 29.5,
+      "grad_norm_var": 14.058072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4824,
+      "loss/crossentropy": 2.1705787777900696,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20006757453083993,
+      "step": 13750
+    },
+    {
+      "epoch": 0.344,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.3279592088337556e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3957,
+      "loss/crossentropy": 2.0967624366283415,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17579186484217643,
+      "step": 13760
+    },
+    {
+      "epoch": 0.34425,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.3279592092845036e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5046,
+      "loss/crossentropy": 2.0625901982188224,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18775402326136828,
+      "step": 13770
+    },
+    {
+      "epoch": 0.3445,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.1556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4818,
+      "loss/crossentropy": 2.18047796189785,
+      "loss/hidden": 3.250390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18007917553186417,
+      "step": 13780
+    },
+    {
+      "epoch": 0.34475,
+      "grad_norm": 31.875,
+      "grad_norm_var": 9.3416015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5204,
+      "loss/crossentropy": 2.1267434477806093,
+      "loss/hidden": 3.49453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20445957724004984,
+      "step": 13790
+    },
+    {
+      "epoch": 0.345,
+      "grad_norm": 31.0,
+      "grad_norm_var": 24.017643229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3979,
+      "loss/crossentropy": 2.1567360132932665,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19118222668766977,
+      "step": 13800
+    },
+    {
+      "epoch": 0.34525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.7625,
+      "learning_rate": 0.0001,
+      "loss": 7.3801,
+      "loss/crossentropy": 2.1303461611270906,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19525523390620947,
+      "step": 13810
+    },
+    {
+      "epoch": 0.3455,
+      "grad_norm": 32.0,
+      "grad_norm_var": 4.1275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5682,
+      "loss/crossentropy": 2.1890379935503006,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20919501278549432,
+      "step": 13820
+    },
+    {
+      "epoch": 0.34575,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.558268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3675,
+      "loss/crossentropy": 2.0688234619796275,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2000229710713029,
+      "step": 13830
+    },
+    {
+      "epoch": 0.346,
+      "grad_norm": 30.625,
+      "grad_norm_var": 6.05390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4978,
+      "loss/crossentropy": 2.0757779657840727,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.177880766056478,
+      "step": 13840
+    },
+    {
+      "epoch": 0.34625,
+      "grad_norm": 33.75,
+      "grad_norm_var": 6.74765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4687,
+      "loss/crossentropy": 2.2370450526475905,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18801292553544044,
+      "step": 13850
+    },
+    {
+      "epoch": 0.3465,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.43515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5391,
+      "loss/crossentropy": 2.2622280076146124,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970792453736067,
+      "step": 13860
+    },
+    {
+      "epoch": 0.34675,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.715559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4578,
+      "loss/crossentropy": 2.1148511469364166,
+      "loss/hidden": 3.391015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19618813470005989,
+      "step": 13870
+    },
+    {
+      "epoch": 0.347,
+      "grad_norm": 34.0,
+      "grad_norm_var": 2.9916015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4502,
+      "loss/crossentropy": 2.0950228199362755,
+      "loss/hidden": 3.466015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19036842305213214,
+      "step": 13880
+    },
+    {
+      "epoch": 0.34725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.252083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5902,
+      "loss/crossentropy": 2.127346533536911,
+      "loss/hidden": 3.479296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20207451861351727,
+      "step": 13890
+    },
+    {
+      "epoch": 0.3475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 96.4712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4647,
+      "loss/crossentropy": 2.1354104042053224,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19764424189925195,
+      "step": 13900
+    },
+    {
+      "epoch": 0.34775,
+      "grad_norm": 32.25,
+      "grad_norm_var": 49.921875,
+      "learning_rate": 0.0001,
+      "loss": 7.4506,
+      "loss/crossentropy": 2.0013342767953874,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18574044182896615,
+      "step": 13910
+    },
+    {
+      "epoch": 0.348,
+      "grad_norm": 31.75,
+      "grad_norm_var": 34.174739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.395,
+      "loss/crossentropy": 2.087017002701759,
+      "loss/hidden": 3.33515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1856970646418631,
+      "step": 13920
+    },
+    {
+      "epoch": 0.34825,
+      "grad_norm": 31.5,
+      "grad_norm_var": 33.81041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4898,
+      "loss/crossentropy": 2.2239955455064773,
+      "loss/hidden": 3.28359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19794288650155067,
+      "step": 13930
+    },
+    {
+      "epoch": 0.3485,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.250455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5032,
+      "loss/crossentropy": 2.1092188715934754,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.187043634057045,
+      "step": 13940
+    },
+    {
+      "epoch": 0.34875,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.587239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5959,
+      "loss/crossentropy": 2.134988930821419,
+      "loss/hidden": 3.579296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.24077722635120152,
+      "step": 13950
+    },
+    {
+      "epoch": 0.349,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.0708333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5289,
+      "loss/crossentropy": 2.1751519560813906,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20373264253139495,
+      "step": 13960
+    },
+    {
+      "epoch": 0.34925,
+      "grad_norm": 29.375,
+      "grad_norm_var": 2.91640625,
+      "learning_rate": 0.0001,
+      "loss": 7.3548,
+      "loss/crossentropy": 2.0371064260601996,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17776737520471214,
+      "step": 13970
+    },
+    {
+      "epoch": 0.3495,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.8893229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3986,
+      "loss/crossentropy": 2.0750357516109945,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18888103473000228,
+      "step": 13980
+    },
+    {
+      "epoch": 0.34975,
+      "grad_norm": 51.25,
+      "grad_norm_var": 3.1925595247275407e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6214,
+      "loss/crossentropy": 2.1400484412908556,
+      "loss/hidden": 3.57578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22168501131236554,
+      "step": 13990
+    },
+    {
+      "epoch": 0.35,
+      "grad_norm": 29.875,
+      "grad_norm_var": 34.92916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3981,
+      "loss/crossentropy": 2.1600725933909417,
+      "loss/hidden": 3.308203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18070472478866578,
+      "step": 14000
+    },
+    {
+      "epoch": 0.35025,
+      "grad_norm": 28.5,
+      "grad_norm_var": 79.35625,
+      "learning_rate": 0.0001,
+      "loss": 7.375,
+      "loss/crossentropy": 2.127374938130379,
+      "loss/hidden": 3.3015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1964468475431204,
+      "step": 14010
+    },
+    {
+      "epoch": 0.3505,
+      "grad_norm": 31.5,
+      "grad_norm_var": 11.227018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4185,
+      "loss/crossentropy": 2.064249020814896,
+      "loss/hidden": 3.448046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1939970314502716,
+      "step": 14020
+    },
+    {
+      "epoch": 0.35075,
+      "grad_norm": 28.125,
+      "grad_norm_var": 10.7962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3458,
+      "loss/crossentropy": 2.009464919567108,
+      "loss/hidden": 3.3625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16926231551915408,
+      "step": 14030
+    },
+    {
+      "epoch": 0.351,
+      "grad_norm": 33.0,
+      "grad_norm_var": 10.735416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4381,
+      "loss/crossentropy": 2.114129716157913,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18966517839580774,
+      "step": 14040
+    },
+    {
+      "epoch": 0.35125,
+      "grad_norm": 38.25,
+      "grad_norm_var": 11.432747395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4797,
+      "loss/crossentropy": 2.025939238071442,
+      "loss/hidden": 3.561328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21536036860197783,
+      "step": 14050
+    },
+    {
+      "epoch": 0.3515,
+      "grad_norm": 29.0,
+      "grad_norm_var": 10.072330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4484,
+      "loss/crossentropy": 2.067097157239914,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19576688967645167,
+      "step": 14060
+    },
+    {
+      "epoch": 0.35175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 8.662434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4599,
+      "loss/crossentropy": 2.1595573753118513,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18690041936933993,
+      "step": 14070
+    },
+    {
+      "epoch": 0.352,
+      "grad_norm": 29.875,
+      "grad_norm_var": 33.607747395833336,
+      "learning_rate": 0.0001,
+      "loss": 7.3965,
+      "loss/crossentropy": 2.2124534368515016,
+      "loss/hidden": 3.319921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18615365140140055,
+      "step": 14080
+    },
+    {
+      "epoch": 0.35225,
+      "grad_norm": 28.5,
+      "grad_norm_var": 34.65514322916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4262,
+      "loss/crossentropy": 2.1372614175081255,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18902692012488842,
+      "step": 14090
+    },
+    {
+      "epoch": 0.3525,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.59140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5356,
+      "loss/crossentropy": 2.1618730440735816,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19613058008253575,
+      "step": 14100
+    },
+    {
+      "epoch": 0.35275,
+      "grad_norm": 29.125,
+      "grad_norm_var": 5.245833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4814,
+      "loss/crossentropy": 2.159251680970192,
+      "loss/hidden": 3.4828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22244902290403842,
+      "step": 14110
+    },
+    {
+      "epoch": 0.353,
+      "grad_norm": 30.75,
+      "grad_norm_var": 5.6,
+      "learning_rate": 0.0001,
+      "loss": 7.3935,
+      "loss/crossentropy": 2.1684078365564345,
+      "loss/hidden": 3.32265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18789836578071117,
+      "step": 14120
+    },
+    {
+      "epoch": 0.35325,
+      "grad_norm": 28.875,
+      "grad_norm_var": 5.843489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.419,
+      "loss/crossentropy": 2.1756194710731505,
+      "loss/hidden": 3.27265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1782940372824669,
+      "step": 14130
+    },
+    {
+      "epoch": 0.3535,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.5005,
+      "loss/crossentropy": 2.1395913138985634,
+      "loss/hidden": 3.3875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19922848492860795,
+      "step": 14140
+    },
+    {
+      "epoch": 0.35375,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.9184895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3926,
+      "loss/crossentropy": 2.1444697260856627,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18964640200138091,
+      "step": 14150
+    },
+    {
+      "epoch": 0.354,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.342708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4802,
+      "loss/crossentropy": 2.085355066135526,
+      "loss/hidden": 3.382421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20059554176405073,
+      "step": 14160
+    },
+    {
+      "epoch": 0.35425,
+      "grad_norm": 30.375,
+      "grad_norm_var": 17.780989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3967,
+      "loss/crossentropy": 2.1756529211997986,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1973805770277977,
+      "step": 14170
+    },
+    {
+      "epoch": 0.3545,
+      "grad_norm": 35.0,
+      "grad_norm_var": 5.43125,
+      "learning_rate": 0.0001,
+      "loss": 7.5431,
+      "loss/crossentropy": 2.197963294386864,
+      "loss/hidden": 3.344140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20036561377346515,
+      "step": 14180
+    },
+    {
+      "epoch": 0.35475,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.4775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4884,
+      "loss/crossentropy": 2.1392128080129624,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19742392897605895,
+      "step": 14190
+    },
+    {
+      "epoch": 0.355,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.24765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4598,
+      "loss/crossentropy": 2.064734974503517,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21485688779503107,
+      "step": 14200
+    },
+    {
+      "epoch": 0.35525,
+      "grad_norm": 28.75,
+      "grad_norm_var": 3.8520182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3441,
+      "loss/crossentropy": 2.0820038333535194,
+      "loss/hidden": 3.346484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1940332455560565,
+      "step": 14210
+    },
+    {
+      "epoch": 0.3555,
+      "grad_norm": 29.0,
+      "grad_norm_var": 1.7738932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5072,
+      "loss/crossentropy": 2.1766203552484513,
+      "loss/hidden": 3.49296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23402530644088984,
+      "step": 14220
+    },
+    {
+      "epoch": 0.35575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.2108723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4591,
+      "loss/crossentropy": 2.1537226110696794,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20421480108052492,
+      "step": 14230
+    },
+    {
+      "epoch": 0.356,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.18515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4918,
+      "loss/crossentropy": 2.14600280970335,
+      "loss/hidden": 3.317578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18569285795092583,
+      "step": 14240
+    },
+    {
+      "epoch": 0.35625,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.1848307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3701,
+      "loss/crossentropy": 1.989667421579361,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17830973230302333,
+      "step": 14250
+    },
+    {
+      "epoch": 0.3565,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.1634765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4071,
+      "loss/crossentropy": 1.9839058618992567,
+      "loss/hidden": 3.392578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19238084005191922,
+      "step": 14260
+    },
+    {
+      "epoch": 0.35675,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.7497395833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5671,
+      "loss/crossentropy": 2.3518798291683196,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22251383662223817,
+      "step": 14270
+    },
+    {
+      "epoch": 0.357,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.5697916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4098,
+      "loss/crossentropy": 2.0802294224500657,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18012972828000784,
+      "step": 14280
+    },
+    {
+      "epoch": 0.35725,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.298958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5495,
+      "loss/crossentropy": 2.1887664824724196,
+      "loss/hidden": 3.4390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21481433548033238,
+      "step": 14290
+    },
+    {
+      "epoch": 0.3575,
+      "grad_norm": 28.625,
+      "grad_norm_var": 3.442122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4145,
+      "loss/crossentropy": 2.164269728958607,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18548591639846562,
+      "step": 14300
+    },
+    {
+      "epoch": 0.35775,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.4488932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4602,
+      "loss/crossentropy": 2.295723894238472,
+      "loss/hidden": 3.344921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19964433908462526,
+      "step": 14310
+    },
+    {
+      "epoch": 0.358,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.9473307291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3661,
+      "loss/crossentropy": 2.166135938465595,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2061729446053505,
+      "step": 14320
+    },
+    {
+      "epoch": 0.35825,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.8650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4077,
+      "loss/crossentropy": 2.2319384172558783,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20099803544580935,
+      "step": 14330
+    },
+    {
+      "epoch": 0.3585,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.536393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5508,
+      "loss/crossentropy": 2.1728119254112244,
+      "loss/hidden": 3.446875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20843122638761996,
+      "step": 14340
+    },
+    {
+      "epoch": 0.35875,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.5192057291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5253,
+      "loss/crossentropy": 2.2258054494857786,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2185509093105793,
+      "step": 14350
+    },
+    {
+      "epoch": 0.359,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.6666015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4171,
+      "loss/crossentropy": 2.195216101408005,
+      "loss/hidden": 3.258984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18341686353087425,
+      "step": 14360
+    },
+    {
+      "epoch": 0.35925,
+      "grad_norm": 30.25,
+      "grad_norm_var": 7.613541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4032,
+      "loss/crossentropy": 2.1008095175027846,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18845746479928493,
+      "step": 14370
+    },
+    {
+      "epoch": 0.3595,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.676041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4559,
+      "loss/crossentropy": 2.0987029120326044,
+      "loss/hidden": 3.4765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20877646040171385,
+      "step": 14380
+    },
+    {
+      "epoch": 0.35975,
+      "grad_norm": 30.75,
+      "grad_norm_var": 1.84765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4529,
+      "loss/crossentropy": 2.0258067578077315,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1862058535218239,
+      "step": 14390
+    },
+    {
+      "epoch": 0.36,
+      "grad_norm": 28.0,
+      "grad_norm_var": 12.8400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4557,
+      "loss/crossentropy": 2.126949782669544,
+      "loss/hidden": 3.340625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1885037848725915,
+      "step": 14400
+    },
+    {
+      "epoch": 0.36025,
+      "grad_norm": 40.25,
+      "grad_norm_var": 9.422330729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.455,
+      "loss/crossentropy": 2.1198273450136185,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1814750697463751,
+      "step": 14410
+    },
+    {
+      "epoch": 0.3605,
+      "grad_norm": 32.5,
+      "grad_norm_var": 439.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4941,
+      "loss/crossentropy": 2.2227646946907043,
+      "loss/hidden": 3.30703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1837354227900505,
+      "step": 14420
+    },
+    {
+      "epoch": 0.36075,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.0145833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3752,
+      "loss/crossentropy": 2.241603446006775,
+      "loss/hidden": 3.25,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18308813124895096,
+      "step": 14430
+    },
+    {
+      "epoch": 0.361,
+      "grad_norm": 29.875,
+      "grad_norm_var": 16.7587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.509,
+      "loss/crossentropy": 2.207683742046356,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20882598627358676,
+      "step": 14440
+    },
+    {
+      "epoch": 0.36125,
+      "grad_norm": 30.0,
+      "grad_norm_var": 31.902083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4703,
+      "loss/crossentropy": 2.086098091304302,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18962491042912005,
+      "step": 14450
+    },
+    {
+      "epoch": 0.3615,
+      "grad_norm": 30.75,
+      "grad_norm_var": 540.1921223958333,
+      "learning_rate": 0.0001,
+      "loss": 7.4866,
+      "loss/crossentropy": 2.1778491258621218,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21017426177859305,
+      "step": 14460
+    },
+    {
+      "epoch": 0.36175,
+      "grad_norm": 38.75,
+      "grad_norm_var": 1848.2552083333333,
+      "learning_rate": 0.0001,
+      "loss": 7.8459,
+      "loss/crossentropy": 2.2158830940723417,
+      "loss/hidden": 3.571875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20809865184128284,
+      "step": 14470
+    },
+    {
+      "epoch": 0.362,
+      "grad_norm": 27.0,
+      "grad_norm_var": 147.39108072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4835,
+      "loss/crossentropy": 2.1877679374068975,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1906817673705518,
+      "step": 14480
+    },
+    {
+      "epoch": 0.36225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.633072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5583,
+      "loss/crossentropy": 2.17199744284153,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19461486004292966,
+      "step": 14490
+    },
+    {
+      "epoch": 0.3625,
+      "grad_norm": 33.0,
+      "grad_norm_var": 18.4119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4614,
+      "loss/crossentropy": 2.115432745218277,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1788713935762644,
+      "step": 14500
+    },
+    {
+      "epoch": 0.36275,
+      "grad_norm": 30.5,
+      "grad_norm_var": 18.049739583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5055,
+      "loss/crossentropy": 2.1718098491430284,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18509582802653313,
+      "step": 14510
+    },
+    {
+      "epoch": 0.363,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.419205729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4428,
+      "loss/crossentropy": 2.1485825031995773,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1956698888912797,
+      "step": 14520
+    },
+    {
+      "epoch": 0.36325,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.57890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3589,
+      "loss/crossentropy": 2.121652737259865,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17529746759682893,
+      "step": 14530
+    },
+    {
+      "epoch": 0.3635,
+      "grad_norm": 29.375,
+      "grad_norm_var": 3.726822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4459,
+      "loss/crossentropy": 2.154259353876114,
+      "loss/hidden": 3.3515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19357640612870455,
+      "step": 14540
+    },
+    {
+      "epoch": 0.36375,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.4488932291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5765,
+      "loss/crossentropy": 2.2079395562410356,
+      "loss/hidden": 3.394921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19934294298291205,
+      "step": 14550
+    },
+    {
+      "epoch": 0.364,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.1207682291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5529,
+      "loss/crossentropy": 2.1579170167446136,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18290599882602693,
+      "step": 14560
+    },
+    {
+      "epoch": 0.36425,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.414322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5717,
+      "loss/crossentropy": 2.0264066070318223,
+      "loss/hidden": 3.466796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2068371519446373,
+      "step": 14570
+    },
+    {
+      "epoch": 0.3645,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.905143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4105,
+      "loss/crossentropy": 2.0989100337028503,
+      "loss/hidden": 3.289453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18423682264983654,
+      "step": 14580
+    },
+    {
+      "epoch": 0.36475,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.785872395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4701,
+      "loss/crossentropy": 2.172179477661848,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19203788777813316,
+      "step": 14590
+    },
+    {
+      "epoch": 0.365,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.4895833333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5094,
+      "loss/crossentropy": 2.132567846775055,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192837636731565,
+      "step": 14600
+    },
+    {
+      "epoch": 0.36525,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.6837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4883,
+      "loss/crossentropy": 2.184123566746712,
+      "loss/hidden": 3.371484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19760762453079223,
+      "step": 14610
+    },
+    {
+      "epoch": 0.3655,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.4098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5364,
+      "loss/crossentropy": 2.160658273100853,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19369914215058087,
+      "step": 14620
+    },
+    {
+      "epoch": 0.36575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 2.3684895833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5713,
+      "loss/crossentropy": 2.121895319223404,
+      "loss/hidden": 3.487109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20311131626367568,
+      "step": 14630
+    },
+    {
+      "epoch": 0.366,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.7067057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5138,
+      "loss/crossentropy": 2.2521799474954607,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21882211808115243,
+      "step": 14640
+    },
+    {
+      "epoch": 0.36625,
+      "grad_norm": 33.25,
+      "grad_norm_var": 7.970572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4956,
+      "loss/crossentropy": 2.0071460634469984,
+      "loss/hidden": 3.548046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984138570725918,
+      "step": 14650
+    },
+    {
+      "epoch": 0.3665,
+      "grad_norm": 34.5,
+      "grad_norm_var": 3.4879557291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.659,
+      "loss/crossentropy": 2.201695331931114,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21316100917756559,
+      "step": 14660
+    },
+    {
+      "epoch": 0.36675,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.5509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5,
+      "loss/crossentropy": 2.1117188185453415,
+      "loss/hidden": 3.49375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20876432172954082,
+      "step": 14670
+    },
+    {
+      "epoch": 0.367,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.667708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5143,
+      "loss/crossentropy": 2.190372434258461,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19909227322787046,
+      "step": 14680
+    },
+    {
+      "epoch": 0.36725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.52265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4599,
+      "loss/crossentropy": 2.126342362165451,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18715179907158017,
+      "step": 14690
+    },
+    {
+      "epoch": 0.3675,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.5072916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4414,
+      "loss/crossentropy": 2.2693688839673998,
+      "loss/hidden": 3.425390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20018572360277176,
+      "step": 14700
+    },
+    {
+      "epoch": 0.36775,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.044791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4719,
+      "loss/crossentropy": 2.181891369819641,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19805356562137605,
+      "step": 14710
+    },
+    {
+      "epoch": 0.368,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.7613932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5656,
+      "loss/crossentropy": 2.1095578551292418,
+      "loss/hidden": 3.434765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22850083969533444,
+      "step": 14720
+    },
+    {
+      "epoch": 0.36825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.870247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3893,
+      "loss/crossentropy": 1.992812941968441,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18111684722825885,
+      "step": 14730
+    },
+    {
+      "epoch": 0.3685,
+      "grad_norm": 31.125,
+      "grad_norm_var": 2.957291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.465,
+      "loss/crossentropy": 2.1468236982822417,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18966571036726237,
+      "step": 14740
+    },
+    {
+      "epoch": 0.36875,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.887434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5622,
+      "loss/crossentropy": 2.0730134949088095,
+      "loss/hidden": 3.478125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1993141021579504,
+      "step": 14750
+    },
+    {
+      "epoch": 0.369,
+      "grad_norm": 30.875,
+      "grad_norm_var": 519.0749348958333,
+      "learning_rate": 0.0001,
+      "loss": 7.6123,
+      "loss/crossentropy": 2.2167199581861494,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20820484217256308,
+      "step": 14760
+    },
+    {
+      "epoch": 0.36925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 512.0455729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4422,
+      "loss/crossentropy": 2.1445954963564873,
+      "loss/hidden": 3.280078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18091966006904842,
+      "step": 14770
+    },
+    {
+      "epoch": 0.3695,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.2997395833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5236,
+      "loss/crossentropy": 2.1166233003139494,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19354630559682845,
+      "step": 14780
+    },
+    {
+      "epoch": 0.36975,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.0389973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5297,
+      "loss/crossentropy": 2.1253889203071594,
+      "loss/hidden": 3.438671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18810804057866334,
+      "step": 14790
+    },
+    {
+      "epoch": 0.37,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.134375,
+      "learning_rate": 0.0001,
+      "loss": 7.494,
+      "loss/crossentropy": 2.1236992135643957,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21158513389527797,
+      "step": 14800
+    },
+    {
+      "epoch": 0.37025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.37265625,
+      "learning_rate": 0.0001,
+      "loss": 7.6228,
+      "loss/crossentropy": 2.1661381393671038,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19546456951647997,
+      "step": 14810
+    },
+    {
+      "epoch": 0.3705,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.5059895833333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5015,
+      "loss/crossentropy": 2.1351567402482035,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18865119516849518,
+      "step": 14820
+    },
+    {
+      "epoch": 0.37075,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.138541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6212,
+      "loss/crossentropy": 2.092208056151867,
+      "loss/hidden": 3.49140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21708083804696798,
+      "step": 14830
+    },
+    {
+      "epoch": 0.371,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.2087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5208,
+      "loss/crossentropy": 2.0589574337005616,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19297052770853043,
+      "step": 14840
+    },
+    {
+      "epoch": 0.37125,
+      "grad_norm": 29.375,
+      "grad_norm_var": 15.303580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5897,
+      "loss/crossentropy": 2.116301205009222,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19354041842743755,
+      "step": 14850
+    },
+    {
+      "epoch": 0.3715,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.428125,
+      "learning_rate": 0.0001,
+      "loss": 7.414,
+      "loss/crossentropy": 2.136660024523735,
+      "loss/hidden": 3.474609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19879130963236094,
+      "step": 14860
+    },
+    {
+      "epoch": 0.37175,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.9759765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5417,
+      "loss/crossentropy": 2.1285243958234785,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19073361102491618,
+      "step": 14870
+    },
+    {
+      "epoch": 0.372,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.7768229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.556,
+      "loss/crossentropy": 2.164254853129387,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20029650218784809,
+      "step": 14880
+    },
+    {
+      "epoch": 0.37225,
+      "grad_norm": 28.0,
+      "grad_norm_var": 3.5462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5072,
+      "loss/crossentropy": 2.140005186200142,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20111254062503575,
+      "step": 14890
+    },
+    {
+      "epoch": 0.3725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.7018229166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4189,
+      "loss/crossentropy": 2.122871424257755,
+      "loss/hidden": 3.459765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20454877093434334,
+      "step": 14900
+    },
+    {
+      "epoch": 0.37275,
+      "grad_norm": 29.875,
+      "grad_norm_var": 9.73515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5896,
+      "loss/crossentropy": 2.0703314900398255,
+      "loss/hidden": 3.471875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22251170128583908,
+      "step": 14910
+    },
+    {
+      "epoch": 0.373,
+      "grad_norm": 32.25,
+      "grad_norm_var": 9.148372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4195,
+      "loss/crossentropy": 1.9516542181372643,
+      "loss/hidden": 3.46796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18584211356937885,
+      "step": 14920
+    },
+    {
+      "epoch": 0.37325,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.926497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4303,
+      "loss/crossentropy": 2.207534837722778,
+      "loss/hidden": 3.295703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1845931550487876,
+      "step": 14930
+    },
+    {
+      "epoch": 0.3735,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.1087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4834,
+      "loss/crossentropy": 2.2537055611610413,
+      "loss/hidden": 3.315625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1929902819916606,
+      "step": 14940
+    },
+    {
+      "epoch": 0.37375,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.155989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5208,
+      "loss/crossentropy": 2.0970467746257784,
+      "loss/hidden": 3.53203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20525697618722916,
+      "step": 14950
+    },
+    {
+      "epoch": 0.374,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.9587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3624,
+      "loss/crossentropy": 2.203538399934769,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2022376826032996,
+      "step": 14960
+    },
+    {
+      "epoch": 0.37425,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.6488932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5996,
+      "loss/crossentropy": 2.242584636807442,
+      "loss/hidden": 3.362109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19289392810314893,
+      "step": 14970
+    },
+    {
+      "epoch": 0.3745,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.21875,
+      "learning_rate": 0.0001,
+      "loss": 7.4014,
+      "loss/crossentropy": 2.181095063686371,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19425397235900163,
+      "step": 14980
+    },
+    {
+      "epoch": 0.37475,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.5707682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.443,
+      "loss/crossentropy": 2.247698572278023,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19984396249055864,
+      "step": 14990
+    },
+    {
+      "epoch": 0.375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 11.702083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5215,
+      "loss/crossentropy": 2.2001713097095488,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19376100320369005,
+      "step": 15000
+    },
+    {
+      "epoch": 0.37525,
+      "grad_norm": 31.625,
+      "grad_norm_var": 5.531705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4883,
+      "loss/crossentropy": 2.219767712801695,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21262812037020923,
+      "step": 15010
+    },
+    {
+      "epoch": 0.3755,
+      "grad_norm": 30.25,
+      "grad_norm_var": 14.839322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.506,
+      "loss/crossentropy": 2.1578361958265306,
+      "loss/hidden": 3.31640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1888531982898712,
+      "step": 15020
+    },
+    {
+      "epoch": 0.37575,
+      "grad_norm": 34.25,
+      "grad_norm_var": 3.611458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4726,
+      "loss/crossentropy": 2.246231307089329,
+      "loss/hidden": 3.28515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20316961575299503,
+      "step": 15030
+    },
+    {
+      "epoch": 0.376,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.6830729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4463,
+      "loss/crossentropy": 2.2054271012544633,
+      "loss/hidden": 3.255859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17821243554353713,
+      "step": 15040
+    },
+    {
+      "epoch": 0.37625,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.6552083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4054,
+      "loss/crossentropy": 2.198264943063259,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1874470232054591,
+      "step": 15050
+    },
+    {
+      "epoch": 0.3765,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.098893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5948,
+      "loss/crossentropy": 2.125826768577099,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19506640397012234,
+      "step": 15060
+    },
+    {
+      "epoch": 0.37675,
+      "grad_norm": 31.625,
+      "grad_norm_var": 12.9009765625,
+      "learning_rate": 0.0001,
+      "loss": 7.403,
+      "loss/crossentropy": 2.1264645174145698,
+      "loss/hidden": 3.194921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1758699558675289,
+      "step": 15070
+    },
+    {
+      "epoch": 0.377,
+      "grad_norm": 30.25,
+      "grad_norm_var": 12.3072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4164,
+      "loss/crossentropy": 2.0446790009737015,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656185436993838,
+      "step": 15080
+    },
+    {
+      "epoch": 0.37725,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.7770182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5228,
+      "loss/crossentropy": 2.2252503603696825,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1904544487595558,
+      "step": 15090
+    },
+    {
+      "epoch": 0.3775,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.2988932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5016,
+      "loss/crossentropy": 2.09442842900753,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18910849764943122,
+      "step": 15100
+    },
+    {
+      "epoch": 0.37775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.792708333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4727,
+      "loss/crossentropy": 2.2667126506567,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20034724567085505,
+      "step": 15110
+    },
+    {
+      "epoch": 0.378,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.4447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4735,
+      "loss/crossentropy": 2.1522352784872054,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19715874120593072,
+      "step": 15120
+    },
+    {
+      "epoch": 0.37825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.1738932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5987,
+      "loss/crossentropy": 2.2373314648866653,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20188613571226596,
+      "step": 15130
+    },
+    {
+      "epoch": 0.3785,
+      "grad_norm": 31.625,
+      "grad_norm_var": 3.076041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4819,
+      "loss/crossentropy": 2.0119318321347235,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1882522501051426,
+      "step": 15140
+    },
+    {
+      "epoch": 0.37875,
+      "grad_norm": 33.0,
+      "grad_norm_var": 3.067643229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6136,
+      "loss/crossentropy": 2.169758787751198,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19593703038990498,
+      "step": 15150
+    },
+    {
+      "epoch": 0.379,
+      "grad_norm": 32.0,
+      "grad_norm_var": 11.815625,
+      "learning_rate": 0.0001,
+      "loss": 7.5821,
+      "loss/crossentropy": 2.184238949418068,
+      "loss/hidden": 3.395703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19631746597588062,
+      "step": 15160
+    },
+    {
+      "epoch": 0.37925,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.576822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5062,
+      "loss/crossentropy": 2.1800487637519836,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1907597079873085,
+      "step": 15170
+    },
+    {
+      "epoch": 0.3795,
+      "grad_norm": 32.75,
+      "grad_norm_var": 7.9431640625,
+      "learning_rate": 0.0001,
+      "loss": 7.448,
+      "loss/crossentropy": 2.036279045790434,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1857758706435561,
+      "step": 15180
+    },
+    {
+      "epoch": 0.37975,
+      "grad_norm": 31.25,
+      "grad_norm_var": 8.451822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4471,
+      "loss/crossentropy": 2.133337301015854,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19323884584009648,
+      "step": 15190
+    },
+    {
+      "epoch": 0.38,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.470572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4197,
+      "loss/crossentropy": 2.166276270151138,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19621124118566513,
+      "step": 15200
+    },
+    {
+      "epoch": 0.38025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 3.7270182291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5059,
+      "loss/crossentropy": 2.2187248289585115,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18203979916870594,
+      "step": 15210
+    },
+    {
+      "epoch": 0.3805,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.8229166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4983,
+      "loss/crossentropy": 2.1355438262224196,
+      "loss/hidden": 3.312890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1705906637944281,
+      "step": 15220
+    },
+    {
+      "epoch": 0.38075,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.8587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4817,
+      "loss/crossentropy": 2.20551737844944,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19247971139848233,
+      "step": 15230
+    },
+    {
+      "epoch": 0.381,
+      "grad_norm": 31.375,
+      "grad_norm_var": 1.3666666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3827,
+      "loss/crossentropy": 2.0013756826519966,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1825490690767765,
+      "step": 15240
+    },
+    {
+      "epoch": 0.38125,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.2139973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4722,
+      "loss/crossentropy": 2.040935108065605,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.185430452786386,
+      "step": 15250
+    },
+    {
+      "epoch": 0.3815,
+      "grad_norm": 35.25,
+      "grad_norm_var": 2.1119140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4966,
+      "loss/crossentropy": 2.115219935774803,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19823081977665424,
+      "step": 15260
+    },
+    {
+      "epoch": 0.38175,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.3947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3673,
+      "loss/crossentropy": 2.157605493068695,
+      "loss/hidden": 3.278125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17537485528737307,
+      "step": 15270
+    },
+    {
+      "epoch": 0.382,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.5833333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3478,
+      "loss/crossentropy": 2.1017936944961546,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18104431293904782,
+      "step": 15280
+    },
+    {
+      "epoch": 0.38225,
+      "grad_norm": 6207569920.0,
+      "grad_norm_var": 5.670665197791722e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.6066,
+      "loss/crossentropy": 2.139920362830162,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1915447361767292,
+      "step": 15290
+    },
+    {
+      "epoch": 0.3825,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.40837024446275e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.3699,
+      "loss/crossentropy": 2.1570042625069616,
+      "loss/hidden": 3.28125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18620338048785925,
+      "step": 15300
+    },
+    {
+      "epoch": 0.38275,
+      "grad_norm": 31.625,
+      "grad_norm_var": 0.7934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3904,
+      "loss/crossentropy": 2.0324711576104164,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18631240744143723,
+      "step": 15310
+    },
+    {
+      "epoch": 0.383,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.9098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4851,
+      "loss/crossentropy": 2.094607950001955,
+      "loss/hidden": 3.5671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19497194588184358,
+      "step": 15320
+    },
+    {
+      "epoch": 0.38325,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.1030598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4493,
+      "loss/crossentropy": 2.131368662416935,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19036126751452684,
+      "step": 15330
+    },
+    {
+      "epoch": 0.3835,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.9983723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4902,
+      "loss/crossentropy": 2.0336036890745164,
+      "loss/hidden": 3.414453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20238237082958221,
+      "step": 15340
+    },
+    {
+      "epoch": 0.38375,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.334375,
+      "learning_rate": 0.0001,
+      "loss": 7.4585,
+      "loss/crossentropy": 2.0801710695028306,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19465604964643718,
+      "step": 15350
+    },
+    {
+      "epoch": 0.384,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.2041666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4369,
+      "loss/crossentropy": 2.10711809694767,
+      "loss/hidden": 3.389453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19854038041085004,
+      "step": 15360
+    },
+    {
+      "epoch": 0.38425,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.4447265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5833,
+      "loss/crossentropy": 2.1751110404729843,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971418472006917,
+      "step": 15370
+    },
+    {
+      "epoch": 0.3845,
+      "grad_norm": 33.5,
+      "grad_norm_var": 2.231705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5016,
+      "loss/crossentropy": 2.1026677012443544,
+      "loss/hidden": 3.365625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1858460023999214,
+      "step": 15380
+    },
+    {
+      "epoch": 0.38475,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.7135416666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.483,
+      "loss/crossentropy": 2.140875709801912,
+      "loss/hidden": 3.306640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19617951191030442,
+      "step": 15390
+    },
+    {
+      "epoch": 0.385,
+      "grad_norm": 28.75,
+      "grad_norm_var": 4.5337890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4622,
+      "loss/crossentropy": 2.0783541426062584,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20269902739673853,
+      "step": 15400
+    },
+    {
+      "epoch": 0.38525,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.095572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5227,
+      "loss/crossentropy": 2.0219787105917932,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17434774748981,
+      "step": 15410
+    },
+    {
+      "epoch": 0.3855,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.75390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5115,
+      "loss/crossentropy": 2.303240016102791,
+      "loss/hidden": 3.345703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981578789651394,
+      "step": 15420
+    },
+    {
+      "epoch": 0.38575,
+      "grad_norm": 29.375,
+      "grad_norm_var": 1.784375,
+      "learning_rate": 0.0001,
+      "loss": 7.4007,
+      "loss/crossentropy": 2.116858586668968,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17753979042172433,
+      "step": 15430
+    },
+    {
+      "epoch": 0.386,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.3041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4919,
+      "loss/crossentropy": 2.2533510133624075,
+      "loss/hidden": 3.2890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19535556957125663,
+      "step": 15440
+    },
+    {
+      "epoch": 0.38625,
+      "grad_norm": 33.25,
+      "grad_norm_var": 21.045572916666668,
+      "learning_rate": 0.0001,
+      "loss": 7.4599,
+      "loss/crossentropy": 2.1751621633768083,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19173544868826867,
+      "step": 15450
+    },
+    {
+      "epoch": 0.3865,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.405847188086631e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5371,
+      "loss/crossentropy": 2.165526232123375,
+      "loss/hidden": 3.42890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20593719817698003,
+      "step": 15460
+    },
+    {
+      "epoch": 0.38675,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.1025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4792,
+      "loss/crossentropy": 2.254777705669403,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18779365364462136,
+      "step": 15470
+    },
+    {
+      "epoch": 0.387,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.3358723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4772,
+      "loss/crossentropy": 2.1895380139350893,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19807571023702622,
+      "step": 15480
+    },
+    {
+      "epoch": 0.38725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.5541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5285,
+      "loss/crossentropy": 2.1380341425538063,
+      "loss/hidden": 3.353125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21063561849296092,
+      "step": 15490
+    },
+    {
+      "epoch": 0.3875,
+      "grad_norm": 37.0,
+      "grad_norm_var": 13.8134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5225,
+      "loss/crossentropy": 2.0771276742219924,
+      "loss/hidden": 3.53203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21799675188958645,
+      "step": 15500
+    },
+    {
+      "epoch": 0.38775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 15.705989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4049,
+      "loss/crossentropy": 2.0267057463526728,
+      "loss/hidden": 3.55390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21528998501598834,
+      "step": 15510
+    },
+    {
+      "epoch": 0.388,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.25,
+      "learning_rate": 0.0001,
+      "loss": 7.5097,
+      "loss/crossentropy": 2.1733275309205053,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18399418219923974,
+      "step": 15520
+    },
+    {
+      "epoch": 0.38825,
+      "grad_norm": 33.0,
+      "grad_norm_var": 4.39140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4645,
+      "loss/crossentropy": 2.0331453289836645,
+      "loss/hidden": 3.499609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19905872382223605,
+      "step": 15530
+    },
+    {
+      "epoch": 0.3885,
+      "grad_norm": 28.375,
+      "grad_norm_var": 5.15390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4482,
+      "loss/crossentropy": 2.0666572242975234,
+      "loss/hidden": 3.497265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.200329645909369,
+      "step": 15540
+    },
+    {
+      "epoch": 0.38875,
+      "grad_norm": 37.0,
+      "grad_norm_var": 4.893489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5307,
+      "loss/crossentropy": 2.121652653813362,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18504574447870253,
+      "step": 15550
+    },
+    {
+      "epoch": 0.389,
+      "grad_norm": 30.875,
+      "grad_norm_var": 3.8202473958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4378,
+      "loss/crossentropy": 2.0938068509101866,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18647999074310065,
+      "step": 15560
+    },
+    {
+      "epoch": 0.38925,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.5309895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5078,
+      "loss/crossentropy": 2.016429315507412,
+      "loss/hidden": 3.4171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18244766257703304,
+      "step": 15570
+    },
+    {
+      "epoch": 0.3895,
+      "grad_norm": 32.25,
+      "grad_norm_var": 4.033268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5083,
+      "loss/crossentropy": 2.220339822769165,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20719899125397206,
+      "step": 15580
+    },
+    {
+      "epoch": 0.38975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.937239583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5042,
+      "loss/crossentropy": 2.1800031810998917,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941234938800335,
+      "step": 15590
+    },
+    {
+      "epoch": 0.39,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.6353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6156,
+      "loss/crossentropy": 2.226159965991974,
+      "loss/hidden": 3.499609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2090428164228797,
+      "step": 15600
+    },
+    {
+      "epoch": 0.39025,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.77890625,
+      "learning_rate": 0.0001,
+      "loss": 7.539,
+      "loss/crossentropy": 2.068866019695997,
+      "loss/hidden": 3.4515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1953371877782047,
+      "step": 15610
+    },
+    {
+      "epoch": 0.3905,
+      "grad_norm": 31.0,
+      "grad_norm_var": 54.42916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5415,
+      "loss/crossentropy": 2.2375737935304643,
+      "loss/hidden": 3.325,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20228417851030828,
+      "step": 15620
+    },
+    {
+      "epoch": 0.39075,
+      "grad_norm": 28.375,
+      "grad_norm_var": 67.47604166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5058,
+      "loss/crossentropy": 2.1122362539172173,
+      "loss/hidden": 3.385546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18648870382457972,
+      "step": 15630
+    },
+    {
+      "epoch": 0.391,
+      "grad_norm": 27.0,
+      "grad_norm_var": 32.35390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4127,
+      "loss/crossentropy": 2.062555433809757,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18710520956665277,
+      "step": 15640
+    },
+    {
+      "epoch": 0.39125,
+      "grad_norm": 30.75,
+      "grad_norm_var": 65.40983072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4815,
+      "loss/crossentropy": 2.185184709727764,
+      "loss/hidden": 3.340234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18476471938192846,
+      "step": 15650
+    },
+    {
+      "epoch": 0.3915,
+      "grad_norm": 33.0,
+      "grad_norm_var": 80.87858072916667,
+      "learning_rate": 0.0001,
+      "loss": 7.531,
+      "loss/crossentropy": 2.1845591366291046,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21412495020776987,
+      "step": 15660
+    },
+    {
+      "epoch": 0.39175,
+      "grad_norm": 33.75,
+      "grad_norm_var": 39.119791666666664,
+      "learning_rate": 0.0001,
+      "loss": 7.5411,
+      "loss/crossentropy": 2.1479857102036477,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2234236292541027,
+      "step": 15670
+    },
+    {
+      "epoch": 0.392,
+      "grad_norm": 42.25,
+      "grad_norm_var": 156.56666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4113,
+      "loss/crossentropy": 2.134408001601696,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18845739047974347,
+      "step": 15680
+    },
+    {
+      "epoch": 0.39225,
+      "grad_norm": 28.75,
+      "grad_norm_var": 175.39524739583334,
+      "learning_rate": 0.0001,
+      "loss": 7.4248,
+      "loss/crossentropy": 2.128613194823265,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2125804962590337,
+      "step": 15690
+    },
+    {
+      "epoch": 0.3925,
+      "grad_norm": 36.5,
+      "grad_norm_var": 23.85625,
+      "learning_rate": 0.0001,
+      "loss": 7.4262,
+      "loss/crossentropy": 2.224800446629524,
+      "loss/hidden": 3.396484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20871379896998404,
+      "step": 15700
+    },
+    {
+      "epoch": 0.39275,
+      "grad_norm": 35.25,
+      "grad_norm_var": 13.641080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3782,
+      "loss/crossentropy": 2.1991523414850236,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19195745941251516,
+      "step": 15710
+    },
+    {
+      "epoch": 0.393,
+      "grad_norm": 31.875,
+      "grad_norm_var": 11.7931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4026,
+      "loss/crossentropy": 2.1917616426944733,
+      "loss/hidden": 3.328515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19351839125156403,
+      "step": 15720
+    },
+    {
+      "epoch": 0.39325,
+      "grad_norm": 47.75,
+      "grad_norm_var": 26.167122395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.457,
+      "loss/crossentropy": 2.120897516608238,
+      "loss/hidden": 3.31484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19529615007340909,
+      "step": 15730
+    },
+    {
+      "epoch": 0.3935,
+      "grad_norm": 34.5,
+      "grad_norm_var": 26.758268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4009,
+      "loss/crossentropy": 2.1326122283935547,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17355557810515165,
+      "step": 15740
+    },
+    {
+      "epoch": 0.39375,
+      "grad_norm": 29.75,
+      "grad_norm_var": 10.8853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3956,
+      "loss/crossentropy": 2.0307308197021485,
+      "loss/hidden": 3.44921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20066507179290055,
+      "step": 15750
+    },
+    {
+      "epoch": 0.394,
+      "grad_norm": 36.0,
+      "grad_norm_var": 13.039518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3499,
+      "loss/crossentropy": 2.067159628868103,
+      "loss/hidden": 3.27734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1807652136310935,
+      "step": 15760
+    },
+    {
+      "epoch": 0.39425,
+      "grad_norm": 30.25,
+      "grad_norm_var": 6.055989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3641,
+      "loss/crossentropy": 2.072852502763271,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17794448863714935,
+      "step": 15770
+    },
+    {
+      "epoch": 0.3945,
+      "grad_norm": 30.875,
+      "grad_norm_var": 46.35598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.422,
+      "loss/crossentropy": 2.1680019706487657,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2028266169130802,
+      "step": 15780
+    },
+    {
+      "epoch": 0.39475,
+      "grad_norm": 27.625,
+      "grad_norm_var": 43.4603515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3911,
+      "loss/crossentropy": 2.16780024766922,
+      "loss/hidden": 3.379296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19396517537534236,
+      "step": 15790
+    },
+    {
+      "epoch": 0.395,
+      "grad_norm": 43.0,
+      "grad_norm_var": 39.1072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3941,
+      "loss/crossentropy": 2.175208270549774,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18382846023887395,
+      "step": 15800
+    },
+    {
+      "epoch": 0.39525,
+      "grad_norm": 44.5,
+      "grad_norm_var": 36.8900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4574,
+      "loss/crossentropy": 2.1617070853710176,
+      "loss/hidden": 3.303125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18401278592646123,
+      "step": 15810
+    },
+    {
+      "epoch": 0.3955,
+      "grad_norm": 28.0,
+      "grad_norm_var": 30.280989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.2727,
+      "loss/crossentropy": 2.063553161919117,
+      "loss/hidden": 3.24765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17086460664868355,
+      "step": 15820
+    },
+    {
+      "epoch": 0.39575,
+      "grad_norm": 33.0,
+      "grad_norm_var": 41.735416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3294,
+      "loss/crossentropy": 2.091352239251137,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19158628936856986,
+      "step": 15830
+    },
+    {
+      "epoch": 0.396,
+      "grad_norm": 28.75,
+      "grad_norm_var": 42.423893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3487,
+      "loss/crossentropy": 2.1424781650304796,
+      "loss/hidden": 3.333203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1843854110687971,
+      "step": 15840
+    },
+    {
+      "epoch": 0.39625,
+      "grad_norm": 34.0,
+      "grad_norm_var": 47.520572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4008,
+      "loss/crossentropy": 2.1899289660155774,
+      "loss/hidden": 3.303515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1849813165143132,
+      "step": 15850
+    },
+    {
+      "epoch": 0.3965,
+      "grad_norm": 27.875,
+      "grad_norm_var": 15.66015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3814,
+      "loss/crossentropy": 2.181295931339264,
+      "loss/hidden": 3.310546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18565880618989467,
+      "step": 15860
+    },
+    {
+      "epoch": 0.39675,
+      "grad_norm": 36.25,
+      "grad_norm_var": 17.7587890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3741,
+      "loss/crossentropy": 2.1227088302373884,
+      "loss/hidden": 3.32734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1942040240392089,
+      "step": 15870
+    },
+    {
+      "epoch": 0.397,
+      "grad_norm": 38.0,
+      "grad_norm_var": 25.970572916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4482,
+      "loss/crossentropy": 2.0236525177955627,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20845895111560822,
+      "step": 15880
+    },
+    {
+      "epoch": 0.39725,
+      "grad_norm": 40.25,
+      "grad_norm_var": 21.3728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3296,
+      "loss/crossentropy": 2.089261847734451,
+      "loss/hidden": 3.251171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18734547607600688,
+      "step": 15890
+    },
+    {
+      "epoch": 0.3975,
+      "grad_norm": 32.5,
+      "grad_norm_var": 10.598893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4169,
+      "loss/crossentropy": 2.0026068508625032,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18220478743314744,
+      "step": 15900
+    },
+    {
+      "epoch": 0.39775,
+      "grad_norm": 29.125,
+      "grad_norm_var": 9.208268229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.2837,
+      "loss/crossentropy": 2.237262162566185,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1820806261152029,
+      "step": 15910
+    },
+    {
+      "epoch": 0.398,
+      "grad_norm": 32.75,
+      "grad_norm_var": 50.037434895833336,
+      "learning_rate": 0.0001,
+      "loss": 7.4485,
+      "loss/crossentropy": 2.1642710946500303,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18893331224098803,
+      "step": 15920
+    },
+    {
+      "epoch": 0.39825,
+      "grad_norm": 29.125,
+      "grad_norm_var": 31.555208333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.287,
+      "loss/crossentropy": 2.1653111398220064,
+      "loss/hidden": 3.274609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1766128923743963,
+      "step": 15930
+    },
+    {
+      "epoch": 0.3985,
+      "grad_norm": 30.5,
+      "grad_norm_var": 42.138997395833336,
+      "learning_rate": 0.0001,
+      "loss": 7.3823,
+      "loss/crossentropy": 2.0916855938732626,
+      "loss/hidden": 3.320703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19688500016927718,
+      "step": 15940
+    },
+    {
+      "epoch": 0.39875,
+      "grad_norm": 37.5,
+      "grad_norm_var": 78.5375,
+      "learning_rate": 0.0001,
+      "loss": 7.2848,
+      "loss/crossentropy": 2.1427202939987184,
+      "loss/hidden": 3.48125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851636625826359,
+      "step": 15950
+    },
+    {
+      "epoch": 0.399,
+      "grad_norm": 38.75,
+      "grad_norm_var": 68.78125,
+      "learning_rate": 0.0001,
+      "loss": 7.3317,
+      "loss/crossentropy": 2.003783145546913,
+      "loss/hidden": 3.38125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19389798324555158,
+      "step": 15960
+    },
+    {
+      "epoch": 0.39925,
+      "grad_norm": 42.0,
+      "grad_norm_var": 32.01712239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4473,
+      "loss/crossentropy": 2.196589732170105,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970522753894329,
+      "step": 15970
+    },
+    {
+      "epoch": 0.3995,
+      "grad_norm": 41.25,
+      "grad_norm_var": 58.436458333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3478,
+      "loss/crossentropy": 2.1799759164452555,
+      "loss/hidden": 3.286328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18633378595113753,
+      "step": 15980
+    },
+    {
+      "epoch": 0.39975,
+      "grad_norm": 30.875,
+      "grad_norm_var": 59.228580729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3209,
+      "loss/crossentropy": 2.0315997526049614,
+      "loss/hidden": 3.39921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1865892570465803,
+      "step": 15990
+    },
+    {
+      "epoch": 0.4,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.465559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3845,
+      "loss/crossentropy": 2.1453574761748313,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17996564768254758,
+      "step": 16000
+    },
+    {
+      "epoch": 0.40025,
+      "grad_norm": 34.25,
+      "grad_norm_var": 10.563997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3941,
+      "loss/crossentropy": 2.1078505992889403,
+      "loss/hidden": 3.28984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18582884669303895,
+      "step": 16010
+    },
+    {
+      "epoch": 0.4005,
+      "grad_norm": 39.0,
+      "grad_norm_var": 12.724934895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.317,
+      "loss/crossentropy": 2.0792032331228256,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16934819463640452,
+      "step": 16020
+    },
+    {
+      "epoch": 0.40075,
+      "grad_norm": 39.75,
+      "grad_norm_var": 13.576041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3955,
+      "loss/crossentropy": 2.1093589305877685,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19456310402601956,
+      "step": 16030
+    },
+    {
+      "epoch": 0.401,
+      "grad_norm": 38.25,
+      "grad_norm_var": 10.633268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3139,
+      "loss/crossentropy": 2.064394600689411,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18512170277535916,
+      "step": 16040
+    },
+    {
+      "epoch": 0.40125,
+      "grad_norm": 36.25,
+      "grad_norm_var": 13.001497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4009,
+      "loss/crossentropy": 2.140559023618698,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.195547959767282,
+      "step": 16050
+    },
+    {
+      "epoch": 0.4015,
+      "grad_norm": 31.0,
+      "grad_norm_var": 14.090559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4652,
+      "loss/crossentropy": 2.1010533660650252,
+      "loss/hidden": 3.490234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20802920442074538,
+      "step": 16060
+    },
+    {
+      "epoch": 0.40175,
+      "grad_norm": 31.0,
+      "grad_norm_var": 11.946875,
+      "learning_rate": 0.0001,
+      "loss": 7.391,
+      "loss/crossentropy": 2.1016806095838545,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1986994445323944,
+      "step": 16070
+    },
+    {
+      "epoch": 0.402,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.3671223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4359,
+      "loss/crossentropy": 2.1667724169790743,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18537729047238827,
+      "step": 16080
+    },
+    {
+      "epoch": 0.40225,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.411393229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5328,
+      "loss/crossentropy": 2.1518414229154588,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19391167853027583,
+      "step": 16090
+    },
+    {
+      "epoch": 0.4025,
+      "grad_norm": 33.0,
+      "grad_norm_var": 25.363997395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.026207575201988,
+      "loss/hidden": 3.49453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19631834831088782,
+      "step": 16100
+    },
+    {
+      "epoch": 0.40275,
+      "grad_norm": 33.0,
+      "grad_norm_var": 28.158333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4104,
+      "loss/crossentropy": 2.1191447034478186,
+      "loss/hidden": 3.290625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17606843169778585,
+      "step": 16110
+    },
+    {
+      "epoch": 0.403,
+      "grad_norm": 33.25,
+      "grad_norm_var": 8.17265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4121,
+      "loss/crossentropy": 2.2775932759046555,
+      "loss/hidden": 3.32421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2066437665373087,
+      "step": 16120
+    },
+    {
+      "epoch": 0.40325,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.3990009838940193e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4655,
+      "loss/crossentropy": 2.1553472220897674,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1962028866633773,
+      "step": 16130
+    },
+    {
+      "epoch": 0.4035,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.3436848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4711,
+      "loss/crossentropy": 2.177881157398224,
+      "loss/hidden": 3.407421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19104315489530563,
+      "step": 16140
+    },
+    {
+      "epoch": 0.40375,
+      "grad_norm": 29.875,
+      "grad_norm_var": 6.098893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4871,
+      "loss/crossentropy": 2.1309458047151564,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20508808381855487,
+      "step": 16150
+    },
+    {
+      "epoch": 0.404,
+      "grad_norm": 32.0,
+      "grad_norm_var": 10.82890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4862,
+      "loss/crossentropy": 2.149015510082245,
+      "loss/hidden": 3.44765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21942622102797033,
+      "step": 16160
+    },
+    {
+      "epoch": 0.40425,
+      "grad_norm": 28.375,
+      "grad_norm_var": 8.847330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4679,
+      "loss/crossentropy": 2.19928492307663,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19909227825701237,
+      "step": 16170
+    },
+    {
+      "epoch": 0.4045,
+      "grad_norm": 27.25,
+      "grad_norm_var": 4.3994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4377,
+      "loss/crossentropy": 2.05338691174984,
+      "loss/hidden": 3.462890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20643874276429414,
+      "step": 16180
+    },
+    {
+      "epoch": 0.40475,
+      "grad_norm": 37.0,
+      "grad_norm_var": 6.540625,
+      "learning_rate": 0.0001,
+      "loss": 7.4132,
+      "loss/crossentropy": 2.093175410479307,
+      "loss/hidden": 3.48984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18920900514349343,
+      "step": 16190
+    },
+    {
+      "epoch": 0.405,
+      "grad_norm": 28.25,
+      "grad_norm_var": 6.5962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3788,
+      "loss/crossentropy": 2.074980768561363,
+      "loss/hidden": 3.301953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19456995707005262,
+      "step": 16200
+    },
+    {
+      "epoch": 0.40525,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.972916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4665,
+      "loss/crossentropy": 2.2107421159744263,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19778486154973507,
+      "step": 16210
+    },
+    {
+      "epoch": 0.4055,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.1791015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4696,
+      "loss/crossentropy": 2.180025112628937,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19855868369340895,
+      "step": 16220
+    },
+    {
+      "epoch": 0.40575,
+      "grad_norm": 30.625,
+      "grad_norm_var": 3.1150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4948,
+      "loss/crossentropy": 2.1638258934020995,
+      "loss/hidden": 3.515234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21891864482313395,
+      "step": 16230
+    },
+    {
+      "epoch": 0.406,
+      "grad_norm": 29.75,
+      "grad_norm_var": 9.517122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5008,
+      "loss/crossentropy": 2.145371067523956,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22061494514346122,
+      "step": 16240
+    },
+    {
+      "epoch": 0.40625,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.1572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4569,
+      "loss/crossentropy": 2.20810389816761,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20236902087926864,
+      "step": 16250
+    },
+    {
+      "epoch": 0.4065,
+      "grad_norm": 31.625,
+      "grad_norm_var": 7.5775390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4047,
+      "loss/crossentropy": 2.117805525660515,
+      "loss/hidden": 3.330859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1842184830456972,
+      "step": 16260
+    },
+    {
+      "epoch": 0.40675,
+      "grad_norm": 30.125,
+      "grad_norm_var": 10.220247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.203564515709877,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18627342265099286,
+      "step": 16270
+    },
+    {
+      "epoch": 0.407,
+      "grad_norm": 30.5,
+      "grad_norm_var": 11.406705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4088,
+      "loss/crossentropy": 2.0998367607593535,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264923337846994,
+      "step": 16280
+    },
+    {
+      "epoch": 0.40725,
+      "grad_norm": 30.375,
+      "grad_norm_var": 24.726497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.5606,
+      "loss/crossentropy": 2.192747640609741,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2141475934535265,
+      "step": 16290
+    },
+    {
+      "epoch": 0.4075,
+      "grad_norm": 30.25,
+      "grad_norm_var": 9.641080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5167,
+      "loss/crossentropy": 2.167286628484726,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19353726021945478,
+      "step": 16300
+    },
+    {
+      "epoch": 0.40775,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.417708333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4252,
+      "loss/crossentropy": 2.058204619586468,
+      "loss/hidden": 3.365234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1987381473183632,
+      "step": 16310
+    },
+    {
+      "epoch": 0.408,
+      "grad_norm": 32.25,
+      "grad_norm_var": 7.0306640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4876,
+      "loss/crossentropy": 2.1063947066664697,
+      "loss/hidden": 3.30078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19293669573962688,
+      "step": 16320
+    },
+    {
+      "epoch": 0.40825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.2030598958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5029,
+      "loss/crossentropy": 2.2337026357650758,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22151997741311788,
+      "step": 16330
+    },
+    {
+      "epoch": 0.4085,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.6958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4837,
+      "loss/crossentropy": 2.1971309810876845,
+      "loss/hidden": 3.376171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1970578793436289,
+      "step": 16340
+    },
+    {
+      "epoch": 0.40875,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.151497395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3711,
+      "loss/crossentropy": 2.0743446350097656,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19049311596900226,
+      "step": 16350
+    },
+    {
+      "epoch": 0.409,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.40625,
+      "learning_rate": 0.0001,
+      "loss": 7.4231,
+      "loss/crossentropy": 1.9305765472352505,
+      "loss/hidden": 3.370703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19289642302319407,
+      "step": 16360
+    },
+    {
+      "epoch": 0.40925,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.846809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4518,
+      "loss/crossentropy": 2.102352428436279,
+      "loss/hidden": 3.45859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2020608950406313,
+      "step": 16370
+    },
+    {
+      "epoch": 0.4095,
+      "grad_norm": 27.875,
+      "grad_norm_var": 12.543489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4805,
+      "loss/crossentropy": 2.1010148346424105,
+      "loss/hidden": 3.415625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20252777952700854,
+      "step": 16380
+    },
+    {
+      "epoch": 0.40975,
+      "grad_norm": 29.625,
+      "grad_norm_var": 8.712239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.435,
+      "loss/crossentropy": 2.087410292029381,
+      "loss/hidden": 3.54453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21575740706175567,
+      "step": 16390
+    },
+    {
+      "epoch": 0.41,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.9087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4989,
+      "loss/crossentropy": 2.2157984226942062,
+      "loss/hidden": 3.453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2107794165611267,
+      "step": 16400
+    },
+    {
+      "epoch": 0.41025,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.8372395833333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4437,
+      "loss/crossentropy": 2.141116687655449,
+      "loss/hidden": 3.380859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18751906510442495,
+      "step": 16410
+    },
+    {
+      "epoch": 0.4105,
+      "grad_norm": 31.125,
+      "grad_norm_var": 3.67890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5811,
+      "loss/crossentropy": 2.1563937693834303,
+      "loss/hidden": 3.36796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19754156917333604,
+      "step": 16420
+    },
+    {
+      "epoch": 0.41075,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.380989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4985,
+      "loss/crossentropy": 2.163114275038242,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18562516067177057,
+      "step": 16430
+    },
+    {
+      "epoch": 0.411,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.005143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5249,
+      "loss/crossentropy": 2.1273641526699065,
+      "loss/hidden": 3.374609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20784277468919754,
+      "step": 16440
+    },
+    {
+      "epoch": 0.41125,
+      "grad_norm": 30.625,
+      "grad_norm_var": 1.9393229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4749,
+      "loss/crossentropy": 2.107171121239662,
+      "loss/hidden": 3.35625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18700738586485385,
+      "step": 16450
+    },
+    {
+      "epoch": 0.4115,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.2863932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5479,
+      "loss/crossentropy": 2.060547313094139,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19373653624206782,
+      "step": 16460
+    },
+    {
+      "epoch": 0.41175,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.5625,
+      "learning_rate": 0.0001,
+      "loss": 7.5479,
+      "loss/crossentropy": 2.1868856728076933,
+      "loss/hidden": 3.36328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20505926571786404,
+      "step": 16470
+    },
+    {
+      "epoch": 0.412,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.034309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5314,
+      "loss/crossentropy": 2.07418372631073,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20410634595900773,
+      "step": 16480
+    },
+    {
+      "epoch": 0.41225,
+      "grad_norm": 29.375,
+      "grad_norm_var": 20.245247395833335,
+      "learning_rate": 0.0001,
+      "loss": 7.4788,
+      "loss/crossentropy": 2.1793991714715957,
+      "loss/hidden": 3.334375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.189803871512413,
+      "step": 16490
+    },
+    {
+      "epoch": 0.4125,
+      "grad_norm": 42.75,
+      "grad_norm_var": 32.5228515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5413,
+      "loss/crossentropy": 2.252856284379959,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19915211237967015,
+      "step": 16500
+    },
+    {
+      "epoch": 0.41275,
+      "grad_norm": 29.375,
+      "grad_norm_var": 11.610416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4184,
+      "loss/crossentropy": 2.1054195404052733,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19910655040293931,
+      "step": 16510
+    },
+    {
+      "epoch": 0.413,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.3958333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4706,
+      "loss/crossentropy": 2.171055743098259,
+      "loss/hidden": 3.357421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1864585865288973,
+      "step": 16520
+    },
+    {
+      "epoch": 0.41325,
+      "grad_norm": 31.75,
+      "grad_norm_var": 1.7192057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5641,
+      "loss/crossentropy": 2.087317834794521,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20189851969480516,
+      "step": 16530
+    },
+    {
+      "epoch": 0.4135,
+      "grad_norm": 30.875,
+      "grad_norm_var": 7.19140625,
+      "learning_rate": 0.0001,
+      "loss": 7.494,
+      "loss/crossentropy": 2.0119203999638557,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19246891103684902,
+      "step": 16540
+    },
+    {
+      "epoch": 0.41375,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.47890625,
+      "learning_rate": 0.0001,
+      "loss": 7.3923,
+      "loss/crossentropy": 2.1976596504449843,
+      "loss/hidden": 3.323828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1891266519203782,
+      "step": 16550
+    },
+    {
+      "epoch": 0.414,
+      "grad_norm": 33.25,
+      "grad_norm_var": 1.7124937379258301e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4992,
+      "loss/crossentropy": 2.1714249163866044,
+      "loss/hidden": 3.331640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19376156255602836,
+      "step": 16560
+    },
+    {
+      "epoch": 0.41425,
+      "grad_norm": 35.5,
+      "grad_norm_var": 27.289322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4315,
+      "loss/crossentropy": 2.062089595198631,
+      "loss/hidden": 3.403125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19295348990708588,
+      "step": 16570
+    },
+    {
+      "epoch": 0.4145,
+      "grad_norm": 28.625,
+      "grad_norm_var": 27.770572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4442,
+      "loss/crossentropy": 2.0051744118332864,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1870790995657444,
+      "step": 16580
+    },
+    {
+      "epoch": 0.41475,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.789322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4585,
+      "loss/crossentropy": 2.2442015290260313,
+      "loss/hidden": 3.33046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19160873014479876,
+      "step": 16590
+    },
+    {
+      "epoch": 0.415,
+      "grad_norm": 30.0,
+      "grad_norm_var": 0.3973307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3867,
+      "loss/crossentropy": 2.0700665429234504,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18396779522299767,
+      "step": 16600
+    },
+    {
+      "epoch": 0.41525,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.98515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4826,
+      "loss/crossentropy": 2.0733146488666536,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18656179327517747,
+      "step": 16610
+    },
+    {
+      "epoch": 0.4155,
+      "grad_norm": 53.5,
+      "grad_norm_var": 37.53515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4502,
+      "loss/crossentropy": 2.0442128658294676,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19807227458804846,
+      "step": 16620
+    },
+    {
+      "epoch": 0.41575,
+      "grad_norm": 29.5,
+      "grad_norm_var": 39.595768229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3489,
+      "loss/crossentropy": 1.989547997713089,
+      "loss/hidden": 3.39140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20996779929846526,
+      "step": 16630
+    },
+    {
+      "epoch": 0.416,
+      "grad_norm": 32.75,
+      "grad_norm_var": 1.8098307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5352,
+      "loss/crossentropy": 2.233032874763012,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19636076968163252,
+      "step": 16640
+    },
+    {
+      "epoch": 0.41625,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.8395182291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.553,
+      "loss/crossentropy": 2.121747186779976,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19236730560660362,
+      "step": 16650
+    },
+    {
+      "epoch": 0.4165,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.1705729166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.5506,
+      "loss/crossentropy": 2.1418200969696044,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20644566994160413,
+      "step": 16660
+    },
+    {
+      "epoch": 0.41675,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.7811848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3972,
+      "loss/crossentropy": 2.1746320933103562,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2003942009061575,
+      "step": 16670
+    },
+    {
+      "epoch": 0.417,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.2718098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5947,
+      "loss/crossentropy": 2.109486374258995,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18977900519967078,
+      "step": 16680
+    },
+    {
+      "epoch": 0.41725,
+      "grad_norm": 32.75,
+      "grad_norm_var": 2.0608723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5853,
+      "loss/crossentropy": 2.190330383181572,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2031781405210495,
+      "step": 16690
+    },
+    {
+      "epoch": 0.4175,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.530143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5202,
+      "loss/crossentropy": 2.239956870675087,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19103901870548726,
+      "step": 16700
+    },
+    {
+      "epoch": 0.41775,
+      "grad_norm": 30.5,
+      "grad_norm_var": 1.7889973958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5543,
+      "loss/crossentropy": 2.041231009364128,
+      "loss/hidden": 3.50625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20854433607310058,
+      "step": 16710
+    },
+    {
+      "epoch": 0.418,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.063541666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4644,
+      "loss/crossentropy": 2.0960741981863977,
+      "loss/hidden": 3.452734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19888665452599524,
+      "step": 16720
+    },
+    {
+      "epoch": 0.41825,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.901822916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4165,
+      "loss/crossentropy": 2.085656823962927,
+      "loss/hidden": 3.30390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18611110784113408,
+      "step": 16730
+    },
+    {
+      "epoch": 0.4185,
+      "grad_norm": 32.75,
+      "grad_norm_var": 4.4509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4439,
+      "loss/crossentropy": 2.05613434612751,
+      "loss/hidden": 3.418359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20244114883244038,
+      "step": 16740
+    },
+    {
+      "epoch": 0.41875,
+      "grad_norm": 27.875,
+      "grad_norm_var": 3.0869140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5106,
+      "loss/crossentropy": 2.124755159020424,
+      "loss/hidden": 3.334765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1944043705239892,
+      "step": 16750
+    },
+    {
+      "epoch": 0.419,
+      "grad_norm": 32.5,
+      "grad_norm_var": 17.409375,
+      "learning_rate": 0.0001,
+      "loss": 7.5009,
+      "loss/crossentropy": 2.17716204226017,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19449110236018896,
+      "step": 16760
+    },
+    {
+      "epoch": 0.41925,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.465559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4858,
+      "loss/crossentropy": 2.08994981944561,
+      "loss/hidden": 3.315234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1893147800117731,
+      "step": 16770
+    },
+    {
+      "epoch": 0.4195,
+      "grad_norm": 28.5,
+      "grad_norm_var": 5.300455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4615,
+      "loss/crossentropy": 2.2256261706352234,
+      "loss/hidden": 3.43828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.217317346483469,
+      "step": 16780
+    },
+    {
+      "epoch": 0.41975,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.66015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5383,
+      "loss/crossentropy": 2.1508295446634293,
+      "loss/hidden": 3.467578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21005581114441157,
+      "step": 16790
+    },
+    {
+      "epoch": 0.42,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.4184895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.1244804561138153,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19516022335737943,
+      "step": 16800
+    },
+    {
+      "epoch": 0.42025,
+      "grad_norm": 29.25,
+      "grad_norm_var": 2.6468098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3777,
+      "loss/crossentropy": 2.1440630443394184,
+      "loss/hidden": 3.34609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1946810379624367,
+      "step": 16810
+    },
+    {
+      "epoch": 0.4205,
+      "grad_norm": 28.875,
+      "grad_norm_var": 2.633072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4624,
+      "loss/crossentropy": 2.1416983366012574,
+      "loss/hidden": 3.465234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21523999786004425,
+      "step": 16820
+    },
+    {
+      "epoch": 0.42075,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.5103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4346,
+      "loss/crossentropy": 2.1387193992733957,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19993042033165692,
+      "step": 16830
+    },
+    {
+      "epoch": 0.421,
+      "grad_norm": 31.625,
+      "grad_norm_var": 1.046875,
+      "learning_rate": 0.0001,
+      "loss": 7.5782,
+      "loss/crossentropy": 2.1304811269044874,
+      "loss/hidden": 3.523828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20474526304751633,
+      "step": 16840
+    },
+    {
+      "epoch": 0.42125,
+      "grad_norm": 36.25,
+      "grad_norm_var": 2.7309405658680745e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5442,
+      "loss/crossentropy": 2.0812308952212333,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20068600717931986,
+      "step": 16850
+    },
+    {
+      "epoch": 0.4215,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.7309405660953006e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4532,
+      "loss/crossentropy": 2.193872679769993,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20241283550858497,
+      "step": 16860
+    },
+    {
+      "epoch": 0.42175,
+      "grad_norm": 29.125,
+      "grad_norm_var": 1.9760416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4989,
+      "loss/crossentropy": 2.083309184014797,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20324783604592084,
+      "step": 16870
+    },
+    {
+      "epoch": 0.422,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.5150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4589,
+      "loss/crossentropy": 2.085008730739355,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17911175787448883,
+      "step": 16880
+    },
+    {
+      "epoch": 0.42225,
+      "grad_norm": 29.125,
+      "grad_norm_var": 2.3650390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5188,
+      "loss/crossentropy": 2.216263073682785,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20870072897523642,
+      "step": 16890
+    },
+    {
+      "epoch": 0.4225,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.1958333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.1983559757471083,
+      "loss/hidden": 3.40390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19983625803142785,
+      "step": 16900
+    },
+    {
+      "epoch": 0.42275,
+      "grad_norm": 30.875,
+      "grad_norm_var": 2.29765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4603,
+      "loss/crossentropy": 2.1883961871266364,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1971004357561469,
+      "step": 16910
+    },
+    {
+      "epoch": 0.423,
+      "grad_norm": 29.75,
+      "grad_norm_var": 5.770833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.6179,
+      "loss/crossentropy": 2.219552181661129,
+      "loss/hidden": 3.284375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19232371896505357,
+      "step": 16920
+    },
+    {
+      "epoch": 0.42325,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.780989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4868,
+      "loss/crossentropy": 2.114982584118843,
+      "loss/hidden": 3.401171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19648698382079602,
+      "step": 16930
+    },
+    {
+      "epoch": 0.4235,
+      "grad_norm": 31.25,
+      "grad_norm_var": 2.4525390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4523,
+      "loss/crossentropy": 2.1595448046922683,
+      "loss/hidden": 3.449609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18362618330866098,
+      "step": 16940
+    },
+    {
+      "epoch": 0.42375,
+      "grad_norm": 37.0,
+      "grad_norm_var": 3.780989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5867,
+      "loss/crossentropy": 2.055322435498238,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20943177156150342,
+      "step": 16950
+    },
+    {
+      "epoch": 0.424,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.687955729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5252,
+      "loss/crossentropy": 2.138558331131935,
+      "loss/hidden": 3.42265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19545765966176987,
+      "step": 16960
+    },
+    {
+      "epoch": 0.42425,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.5143229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4497,
+      "loss/crossentropy": 2.080673611164093,
+      "loss/hidden": 3.38046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19070623591542243,
+      "step": 16970
+    },
+    {
+      "epoch": 0.4245,
+      "grad_norm": 29.625,
+      "grad_norm_var": 2.3358723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3938,
+      "loss/crossentropy": 2.1281508088111876,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18748594932258128,
+      "step": 16980
+    },
+    {
+      "epoch": 0.42475,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.9358723958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4825,
+      "loss/crossentropy": 2.2469502836465836,
+      "loss/hidden": 3.2796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1956640187650919,
+      "step": 16990
+    },
+    {
+      "epoch": 0.425,
+      "grad_norm": 40.0,
+      "grad_norm_var": 7.366666666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4671,
+      "loss/crossentropy": 2.179307234287262,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1867943126708269,
+      "step": 17000
+    },
+    {
+      "epoch": 0.42525,
+      "grad_norm": 29.5,
+      "grad_norm_var": 6.927018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5113,
+      "loss/crossentropy": 2.101860895752907,
+      "loss/hidden": 3.35234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18607552852481604,
+      "step": 17010
+    },
+    {
+      "epoch": 0.4255,
+      "grad_norm": 30.125,
+      "grad_norm_var": 9.673372395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.434,
+      "loss/crossentropy": 2.1202177673578264,
+      "loss/hidden": 3.273046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18756014555692674,
+      "step": 17020
+    },
+    {
+      "epoch": 0.42575,
+      "grad_norm": 29.0,
+      "grad_norm_var": 8.856184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.1069195836782457,
+      "loss/hidden": 3.376953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19768658615648746,
+      "step": 17030
+    },
+    {
+      "epoch": 0.426,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.8353515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4817,
+      "loss/crossentropy": 2.191109284758568,
+      "loss/hidden": 3.362890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19470954779535532,
+      "step": 17040
+    },
+    {
+      "epoch": 0.42625,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.731184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.6624,
+      "loss/crossentropy": 2.1684040665626525,
+      "loss/hidden": 3.58046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.23237632848322393,
+      "step": 17050
+    },
+    {
+      "epoch": 0.4265,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.0947265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4689,
+      "loss/crossentropy": 2.1706847339868545,
+      "loss/hidden": 3.30234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18483849614858627,
+      "step": 17060
+    },
+    {
+      "epoch": 0.42675,
+      "grad_norm": 32.25,
+      "grad_norm_var": 1.8833333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4644,
+      "loss/crossentropy": 2.1269822865724564,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20061115603893995,
+      "step": 17070
+    },
+    {
+      "epoch": 0.427,
+      "grad_norm": 30.375,
+      "grad_norm_var": 14.857291666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4191,
+      "loss/crossentropy": 2.051969346404076,
+      "loss/hidden": 3.351953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19209920186549426,
+      "step": 17080
+    },
+    {
+      "epoch": 0.42725,
+      "grad_norm": 32.5,
+      "grad_norm_var": 14.401041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4766,
+      "loss/crossentropy": 2.1646567583084106,
+      "loss/hidden": 3.358203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1931615237146616,
+      "step": 17090
+    },
+    {
+      "epoch": 0.4275,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.0837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4649,
+      "loss/crossentropy": 2.1323696315288543,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18601270411163567,
+      "step": 17100
+    },
+    {
+      "epoch": 0.42775,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.4087890625,
+      "learning_rate": 0.0001,
+      "loss": 7.397,
+      "loss/crossentropy": 2.1776672542095183,
+      "loss/hidden": 3.483984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2124987743794918,
+      "step": 17110
+    },
+    {
+      "epoch": 0.428,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.8205729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4531,
+      "loss/crossentropy": 2.0366108894348143,
+      "loss/hidden": 3.485546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1920434921979904,
+      "step": 17120
+    },
+    {
+      "epoch": 0.42825,
+      "grad_norm": 29.625,
+      "grad_norm_var": 7.97890625,
+      "learning_rate": 0.0001,
+      "loss": 7.5251,
+      "loss/crossentropy": 2.2316421404480935,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1984385471791029,
+      "step": 17130
+    },
+    {
+      "epoch": 0.4285,
+      "grad_norm": 31.375,
+      "grad_norm_var": 43.22604166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.532,
+      "loss/crossentropy": 2.1584898799657823,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19015683736652136,
+      "step": 17140
+    },
+    {
+      "epoch": 0.42875,
+      "grad_norm": 34.75,
+      "grad_norm_var": 39.980143229166664,
+      "learning_rate": 0.0001,
+      "loss": 7.4923,
+      "loss/crossentropy": 2.0776416778564455,
+      "loss/hidden": 3.4015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19441335443407298,
+      "step": 17150
+    },
+    {
+      "epoch": 0.429,
+      "grad_norm": 29.375,
+      "grad_norm_var": 8.19375,
+      "learning_rate": 0.0001,
+      "loss": 7.4352,
+      "loss/crossentropy": 2.1459785103797913,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18559253066778184,
+      "step": 17160
+    },
+    {
+      "epoch": 0.42925,
+      "grad_norm": 33.5,
+      "grad_norm_var": 3.0900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4055,
+      "loss/crossentropy": 2.182371488213539,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2002473808825016,
+      "step": 17170
+    },
+    {
+      "epoch": 0.4295,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.212434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4465,
+      "loss/crossentropy": 2.1703537315130235,
+      "loss/hidden": 3.4,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2002835465595126,
+      "step": 17180
+    },
+    {
+      "epoch": 0.42975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.2900390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5187,
+      "loss/crossentropy": 2.1651883900165556,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19961190838366746,
+      "step": 17190
+    },
+    {
+      "epoch": 0.43,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.6395833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5114,
+      "loss/crossentropy": 2.0858636096119882,
+      "loss/hidden": 3.36875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19053794257342815,
+      "step": 17200
+    },
+    {
+      "epoch": 0.43025,
+      "grad_norm": 29.625,
+      "grad_norm_var": 3.9145833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.479,
+      "loss/crossentropy": 2.1603722527623175,
+      "loss/hidden": 3.30859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19217769615352154,
+      "step": 17210
+    },
+    {
+      "epoch": 0.4305,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.849739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3934,
+      "loss/crossentropy": 2.0350109085440637,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1894751563668251,
+      "step": 17220
+    },
+    {
+      "epoch": 0.43075,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.8285807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4516,
+      "loss/crossentropy": 2.0489366173744203,
+      "loss/hidden": 3.52109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1949023339897394,
+      "step": 17230
+    },
+    {
+      "epoch": 0.431,
+      "grad_norm": 36.5,
+      "grad_norm_var": 29.027083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4784,
+      "loss/crossentropy": 2.0739230304956435,
+      "loss/hidden": 3.380078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18924926947802306,
+      "step": 17240
+    },
+    {
+      "epoch": 0.43125,
+      "grad_norm": 33.75,
+      "grad_norm_var": 29.676497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4699,
+      "loss/crossentropy": 2.1982351303100587,
+      "loss/hidden": 3.37734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21114630959928035,
+      "step": 17250
+    },
+    {
+      "epoch": 0.4315,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.301822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5819,
+      "loss/crossentropy": 2.096822480857372,
+      "loss/hidden": 3.426171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2079928234219551,
+      "step": 17260
+    },
+    {
+      "epoch": 0.43175,
+      "grad_norm": 28.625,
+      "grad_norm_var": 8.306705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.593,
+      "loss/crossentropy": 2.115600660443306,
+      "loss/hidden": 3.432421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19750556647777556,
+      "step": 17270
+    },
+    {
+      "epoch": 0.432,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.0478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4366,
+      "loss/crossentropy": 2.1597479790449143,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19931532759219409,
+      "step": 17280
+    },
+    {
+      "epoch": 0.43225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.2518229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4325,
+      "loss/crossentropy": 2.026200716197491,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18456113785505296,
+      "step": 17290
+    },
+    {
+      "epoch": 0.4325,
+      "grad_norm": 31.375,
+      "grad_norm_var": 37.43118489583333,
+      "learning_rate": 0.0001,
+      "loss": 7.4414,
+      "loss/crossentropy": 2.149491161108017,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20337051767855882,
+      "step": 17300
+    },
+    {
+      "epoch": 0.43275,
+      "grad_norm": 34.5,
+      "grad_norm_var": 36.064518229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4859,
+      "loss/crossentropy": 2.1269571796059608,
+      "loss/hidden": 3.46484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2096902344375849,
+      "step": 17310
+    },
+    {
+      "epoch": 0.433,
+      "grad_norm": 31.875,
+      "grad_norm_var": 5.748958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4889,
+      "loss/crossentropy": 2.273072028160095,
+      "loss/hidden": 3.297265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18909230902791024,
+      "step": 17320
+    },
+    {
+      "epoch": 0.43325,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.46640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4258,
+      "loss/crossentropy": 2.1195196583867073,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19008973110467195,
+      "step": 17330
+    },
+    {
+      "epoch": 0.4335,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.076822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.1201743364334105,
+      "loss/hidden": 3.383203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21034939792007207,
+      "step": 17340
+    },
+    {
+      "epoch": 0.43375,
+      "grad_norm": 29.375,
+      "grad_norm_var": 4.154166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4493,
+      "loss/crossentropy": 2.193806689977646,
+      "loss/hidden": 3.305859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18929022550582886,
+      "step": 17350
+    },
+    {
+      "epoch": 0.434,
+      "grad_norm": 30.625,
+      "grad_norm_var": 0.9994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5285,
+      "loss/crossentropy": 2.20940330773592,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19673640970140696,
+      "step": 17360
+    },
+    {
+      "epoch": 0.43425,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.61015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4179,
+      "loss/crossentropy": 2.0897106766700744,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18753900453448297,
+      "step": 17370
+    },
+    {
+      "epoch": 0.4345,
+      "grad_norm": 30.75,
+      "grad_norm_var": 3.6556640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4271,
+      "loss/crossentropy": 2.1975408419966698,
+      "loss/hidden": 3.299609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.180303574539721,
+      "step": 17380
+    },
+    {
+      "epoch": 0.43475,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.5830729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3879,
+      "loss/crossentropy": 2.1741068333387377,
+      "loss/hidden": 3.346875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197440049611032,
+      "step": 17390
+    },
+    {
+      "epoch": 0.435,
+      "grad_norm": 31.75,
+      "grad_norm_var": 2.22890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4275,
+      "loss/crossentropy": 2.132155954837799,
+      "loss/hidden": 3.447265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19416624326258897,
+      "step": 17400
+    },
+    {
+      "epoch": 0.43525,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.5764973958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3642,
+      "loss/crossentropy": 2.0881365656852724,
+      "loss/hidden": 3.31328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1851549668237567,
+      "step": 17410
+    },
+    {
+      "epoch": 0.4355,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.6479166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4346,
+      "loss/crossentropy": 2.213974291086197,
+      "loss/hidden": 3.215234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17357883155345916,
+      "step": 17420
+    },
+    {
+      "epoch": 0.43575,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.375,
+      "learning_rate": 0.0001,
+      "loss": 7.4587,
+      "loss/crossentropy": 2.150649666786194,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19460058882832526,
+      "step": 17430
+    },
+    {
+      "epoch": 0.436,
+      "grad_norm": 42.0,
+      "grad_norm_var": 10.516666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4428,
+      "loss/crossentropy": 2.1933241873979568,
+      "loss/hidden": 3.416796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19797156639397145,
+      "step": 17440
+    },
+    {
+      "epoch": 0.43625,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.1166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4753,
+      "loss/crossentropy": 2.005639246106148,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19176286831498146,
+      "step": 17450
+    },
+    {
+      "epoch": 0.4365,
+      "grad_norm": 32.25,
+      "grad_norm_var": 5.373893229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3735,
+      "loss/crossentropy": 2.1921694964170455,
+      "loss/hidden": 3.338671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1854123178869486,
+      "step": 17460
+    },
+    {
+      "epoch": 0.43675,
+      "grad_norm": 30.0,
+      "grad_norm_var": 5.0478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4629,
+      "loss/crossentropy": 2.062588620185852,
+      "loss/hidden": 3.409375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18035206217318772,
+      "step": 17470
+    },
+    {
+      "epoch": 0.437,
+      "grad_norm": 33.25,
+      "grad_norm_var": 2.0369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4723,
+      "loss/crossentropy": 2.1528229281306266,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18847114518284797,
+      "step": 17480
+    },
+    {
+      "epoch": 0.43725,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.984375,
+      "learning_rate": 0.0001,
+      "loss": 7.3924,
+      "loss/crossentropy": 2.225735864043236,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19506857246160508,
+      "step": 17490
+    },
+    {
+      "epoch": 0.4375,
+      "grad_norm": 32.0,
+      "grad_norm_var": 14.176041666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6873,
+      "loss/crossentropy": 2.1927039504051207,
+      "loss/hidden": 3.41484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21184294372797013,
+      "step": 17500
+    },
+    {
+      "epoch": 0.43775,
+      "grad_norm": 32.5,
+      "grad_norm_var": 12.9025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6167,
+      "loss/crossentropy": 2.1623361676931383,
+      "loss/hidden": 3.40859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1988366713747382,
+      "step": 17510
+    },
+    {
+      "epoch": 0.438,
+      "grad_norm": 5536481280.0,
+      "grad_norm_var": 1.915789037802619e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4769,
+      "loss/crossentropy": 2.0769626796245575,
+      "loss/hidden": 3.32890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1847306966781616,
+      "step": 17520
+    },
+    {
+      "epoch": 0.43825,
+      "grad_norm": 30.125,
+      "grad_norm_var": 1.9157890376411382e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4514,
+      "loss/crossentropy": 2.204550328850746,
+      "loss/hidden": 3.2953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18567237444221973,
+      "step": 17530
+    },
+    {
+      "epoch": 0.4385,
+      "grad_norm": 33.75,
+      "grad_norm_var": 4.995572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4338,
+      "loss/crossentropy": 2.16412447988987,
+      "loss/hidden": 3.333203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19459595140069724,
+      "step": 17540
+    },
+    {
+      "epoch": 0.43875,
+      "grad_norm": 35.0,
+      "grad_norm_var": 11.065625,
+      "learning_rate": 0.0001,
+      "loss": 7.4111,
+      "loss/crossentropy": 2.161545941233635,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18747595753520727,
+      "step": 17550
+    },
+    {
+      "epoch": 0.439,
+      "grad_norm": 30.875,
+      "grad_norm_var": 18.93515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4927,
+      "loss/crossentropy": 2.0142488092184068,
+      "loss/hidden": 3.441015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20353511590510606,
+      "step": 17560
+    },
+    {
+      "epoch": 0.43925,
+      "grad_norm": 31.75,
+      "grad_norm_var": 21.154166666666665,
+      "learning_rate": 0.0001,
+      "loss": 7.557,
+      "loss/crossentropy": 2.1331784069538116,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18347944878041744,
+      "step": 17570
+    },
+    {
+      "epoch": 0.4395,
+      "grad_norm": 36.5,
+      "grad_norm_var": 11.025455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.6167,
+      "loss/crossentropy": 2.2123819947242738,
+      "loss/hidden": 3.294140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20334599036723375,
+      "step": 17580
+    },
+    {
+      "epoch": 0.43975,
+      "grad_norm": 35.0,
+      "grad_norm_var": 26.2962890625,
+      "learning_rate": 0.0001,
+      "loss": 7.443,
+      "loss/crossentropy": 2.054215854406357,
+      "loss/hidden": 3.3171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1868831802159548,
+      "step": 17590
+    },
+    {
+      "epoch": 0.44,
+      "grad_norm": 29.5,
+      "grad_norm_var": 8.619791666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3659,
+      "loss/crossentropy": 2.1821652933955193,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1797631662338972,
+      "step": 17600
+    },
+    {
+      "epoch": 0.44025,
+      "grad_norm": 33.25,
+      "grad_norm_var": 6.815559895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4363,
+      "loss/crossentropy": 2.230892112851143,
+      "loss/hidden": 3.350390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19118015244603156,
+      "step": 17610
+    },
+    {
+      "epoch": 0.4405,
+      "grad_norm": 31.75,
+      "grad_norm_var": 3.7552083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.516,
+      "loss/crossentropy": 2.152178055047989,
+      "loss/hidden": 3.42578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2148242775350809,
+      "step": 17620
+    },
+    {
+      "epoch": 0.44075,
+      "grad_norm": 37.25,
+      "grad_norm_var": 9.076497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3909,
+      "loss/crossentropy": 2.1013472080230713,
+      "loss/hidden": 3.408203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1838392995297909,
+      "step": 17630
+    },
+    {
+      "epoch": 0.441,
+      "grad_norm": 35.75,
+      "grad_norm_var": 10.3291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4749,
+      "loss/crossentropy": 2.12664113342762,
+      "loss/hidden": 3.375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19979833103716374,
+      "step": 17640
+    },
+    {
+      "epoch": 0.44125,
+      "grad_norm": 34.0,
+      "grad_norm_var": 12.180989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.439,
+      "loss/crossentropy": 2.2174566209316255,
+      "loss/hidden": 3.33984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21133954562246798,
+      "step": 17650
+    },
+    {
+      "epoch": 0.4415,
+      "grad_norm": 35.0,
+      "grad_norm_var": 9.773958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4036,
+      "loss/crossentropy": 2.1589852809906005,
+      "loss/hidden": 3.43671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1996469773352146,
+      "step": 17660
+    },
+    {
+      "epoch": 0.44175,
+      "grad_norm": 30.125,
+      "grad_norm_var": 4.467122395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3101,
+      "loss/crossentropy": 2.1177646607160567,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20224390178918839,
+      "step": 17670
+    },
+    {
+      "epoch": 0.442,
+      "grad_norm": 33.25,
+      "grad_norm_var": 4.853580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4531,
+      "loss/crossentropy": 2.035849825292826,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935986878350377,
+      "step": 17680
+    },
+    {
+      "epoch": 0.44225,
+      "grad_norm": 31.25,
+      "grad_norm_var": 6.71015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3643,
+      "loss/crossentropy": 2.0254479318857195,
+      "loss/hidden": 3.429296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19959452152252197,
+      "step": 17690
+    },
+    {
+      "epoch": 0.4425,
+      "grad_norm": 30.25,
+      "grad_norm_var": 49.7041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.487,
+      "loss/crossentropy": 2.128339225053787,
+      "loss/hidden": 3.428515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19117913916707038,
+      "step": 17700
+    },
+    {
+      "epoch": 0.44275,
+      "grad_norm": 31.5,
+      "grad_norm_var": 68.46451822916667,
+      "learning_rate": 0.0001,
+      "loss": 7.4991,
+      "loss/crossentropy": 2.2027375906705857,
+      "loss/hidden": 3.288671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18826800361275672,
+      "step": 17710
+    },
+    {
+      "epoch": 0.443,
+      "grad_norm": 32.0,
+      "grad_norm_var": 39.52682291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4788,
+      "loss/crossentropy": 2.1423556417226792,
+      "loss/hidden": 3.36953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19592436235398053,
+      "step": 17720
+    },
+    {
+      "epoch": 0.44325,
+      "grad_norm": 32.75,
+      "grad_norm_var": 25.954166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.6508,
+      "loss/crossentropy": 2.1767683640122413,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20649159606546164,
+      "step": 17730
+    },
+    {
+      "epoch": 0.4435,
+      "grad_norm": 28.625,
+      "grad_norm_var": 12.77265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4281,
+      "loss/crossentropy": 2.1573469266295433,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2059380615130067,
+      "step": 17740
+    },
+    {
+      "epoch": 0.44375,
+      "grad_norm": 31.25,
+      "grad_norm_var": 12.7837890625,
+      "learning_rate": 0.0001,
+      "loss": 7.402,
+      "loss/crossentropy": 2.121631011366844,
+      "loss/hidden": 3.3046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18308492042124272,
+      "step": 17750
+    },
+    {
+      "epoch": 0.444,
+      "grad_norm": 27.5,
+      "grad_norm_var": 11.298958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.359,
+      "loss/crossentropy": 2.1934143617749213,
+      "loss/hidden": 3.332421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18884631469845772,
+      "step": 17760
+    },
+    {
+      "epoch": 0.44425,
+      "grad_norm": 31.625,
+      "grad_norm_var": 10.630989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4511,
+      "loss/crossentropy": 2.2182701379060745,
+      "loss/hidden": 3.360546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20777408666908742,
+      "step": 17770
+    },
+    {
+      "epoch": 0.4445,
+      "grad_norm": 41.5,
+      "grad_norm_var": 15.221875,
+      "learning_rate": 0.0001,
+      "loss": 7.5054,
+      "loss/crossentropy": 2.15917833596468,
+      "loss/hidden": 3.4875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20331293437629938,
+      "step": 17780
+    },
+    {
+      "epoch": 0.44475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 28.752018229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4046,
+      "loss/crossentropy": 2.2085522964596747,
+      "loss/hidden": 3.4359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20968901123851538,
+      "step": 17790
+    },
+    {
+      "epoch": 0.445,
+      "grad_norm": 27.625,
+      "grad_norm_var": 16.3994140625,
+      "learning_rate": 0.0001,
+      "loss": 7.3779,
+      "loss/crossentropy": 2.088705601543188,
+      "loss/hidden": 3.264453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17547062635421753,
+      "step": 17800
+    },
+    {
+      "epoch": 0.44525,
+      "grad_norm": 33.5,
+      "grad_norm_var": 15.054622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.3363,
+      "loss/crossentropy": 2.0361307077109814,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17406579544767736,
+      "step": 17810
+    },
+    {
+      "epoch": 0.4455,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.33515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3829,
+      "loss/crossentropy": 2.2537280052900313,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1972663702443242,
+      "step": 17820
+    },
+    {
+      "epoch": 0.44575,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.854166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4551,
+      "loss/crossentropy": 2.1048080518841745,
+      "loss/hidden": 3.366796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20005077831447124,
+      "step": 17830
+    },
+    {
+      "epoch": 0.446,
+      "grad_norm": 35.75,
+      "grad_norm_var": 5.8572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4493,
+      "loss/crossentropy": 2.1436916798353196,
+      "loss/hidden": 3.343359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1947880506515503,
+      "step": 17840
+    },
+    {
+      "epoch": 0.44625,
+      "grad_norm": 32.0,
+      "grad_norm_var": 6.080989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4321,
+      "loss/crossentropy": 2.186265227198601,
+      "loss/hidden": 3.3328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18987231217324735,
+      "step": 17850
+    },
+    {
+      "epoch": 0.4465,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.404622395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4713,
+      "loss/crossentropy": 2.1789949417114256,
+      "loss/hidden": 3.226171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.16992315649986267,
+      "step": 17860
+    },
+    {
+      "epoch": 0.44675,
+      "grad_norm": 35.75,
+      "grad_norm_var": 4.260416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4499,
+      "loss/crossentropy": 2.1180329382419587,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20158529318869114,
+      "step": 17870
+    },
+    {
+      "epoch": 0.447,
+      "grad_norm": 32.75,
+      "grad_norm_var": 5.462434895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4072,
+      "loss/crossentropy": 2.1144249960780144,
+      "loss/hidden": 3.277734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17686952501535416,
+      "step": 17880
+    },
+    {
+      "epoch": 0.44725,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.7018229166666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4559,
+      "loss/crossentropy": 2.1003448024392126,
+      "loss/hidden": 3.473828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19702099692076444,
+      "step": 17890
+    },
+    {
+      "epoch": 0.4475,
+      "grad_norm": 33.25,
+      "grad_norm_var": 31.665625,
+      "learning_rate": 0.0001,
+      "loss": 7.5754,
+      "loss/crossentropy": 2.165144255757332,
+      "loss/hidden": 3.41640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20823087207973004,
+      "step": 17900
+    },
+    {
+      "epoch": 0.44775,
+      "grad_norm": 29.75,
+      "grad_norm_var": 27.611458333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4127,
+      "loss/crossentropy": 2.1247920095920563,
+      "loss/hidden": 3.34453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18894042856991292,
+      "step": 17910
+    },
+    {
+      "epoch": 0.448,
+      "grad_norm": 31.375,
+      "grad_norm_var": 16.0291015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5839,
+      "loss/crossentropy": 2.2031987488269804,
+      "loss/hidden": 3.4328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18946342077106237,
+      "step": 17920
+    },
+    {
+      "epoch": 0.44825,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.564518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4259,
+      "loss/crossentropy": 2.08001571893692,
+      "loss/hidden": 3.3375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18759096097201108,
+      "step": 17930
+    },
+    {
+      "epoch": 0.4485,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.5405598958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4475,
+      "loss/crossentropy": 2.091191267967224,
+      "loss/hidden": 3.35078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192183992639184,
+      "step": 17940
+    },
+    {
+      "epoch": 0.44875,
+      "grad_norm": 31.875,
+      "grad_norm_var": 6.664322916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.3588,
+      "loss/crossentropy": 2.1699838608503343,
+      "loss/hidden": 3.3390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1987156704068184,
+      "step": 17950
+    },
+    {
+      "epoch": 0.449,
+      "grad_norm": 31.0,
+      "grad_norm_var": 5.8509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.532,
+      "loss/crossentropy": 2.2819882690906526,
+      "loss/hidden": 3.2984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1946489542722702,
+      "step": 17960
+    },
+    {
+      "epoch": 0.44925,
+      "grad_norm": 35.25,
+      "grad_norm_var": 4.6931640625,
+      "learning_rate": 0.0001,
+      "loss": 7.5208,
+      "loss/crossentropy": 2.188323587179184,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20090275909751654,
+      "step": 17970
+    },
+    {
+      "epoch": 0.4495,
+      "grad_norm": 28.125,
+      "grad_norm_var": 4.03125,
+      "learning_rate": 0.0001,
+      "loss": 7.498,
+      "loss/crossentropy": 2.077894252538681,
+      "loss/hidden": 3.373828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18269995506852865,
+      "step": 17980
+    },
+    {
+      "epoch": 0.44975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 33.514322916666664,
+      "learning_rate": 0.0001,
+      "loss": 7.4175,
+      "loss/crossentropy": 2.0424763709306717,
+      "loss/hidden": 3.2703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1786050383001566,
+      "step": 17990
+    },
+    {
+      "epoch": 0.45,
+      "grad_norm": 35.5,
+      "grad_norm_var": 5.978580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4981,
+      "loss/crossentropy": 2.0257104620337487,
+      "loss/hidden": 3.411328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.203111675940454,
+      "step": 18000
+    },
+    {
+      "epoch": 0.45025,
+      "grad_norm": 27.125,
+      "grad_norm_var": 8.2103515625,
+      "learning_rate": 0.0001,
+      "loss": 7.3876,
+      "loss/crossentropy": 2.00185070335865,
+      "loss/hidden": 3.483984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.207043712772429,
+      "step": 18010
+    },
+    {
+      "epoch": 0.4505,
+      "grad_norm": 30.5,
+      "grad_norm_var": 7.8541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3589,
+      "loss/crossentropy": 2.163762652873993,
+      "loss/hidden": 3.31171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18520545940846206,
+      "step": 18020
+    },
+    {
+      "epoch": 0.45075,
+      "grad_norm": 31.375,
+      "grad_norm_var": 3.8343098958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.3585,
+      "loss/crossentropy": 2.0801485389471055,
+      "loss/hidden": 3.296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19670285172760488,
+      "step": 18030
+    },
+    {
+      "epoch": 0.451,
+      "grad_norm": 32.5,
+      "grad_norm_var": 4.377018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3916,
+      "loss/crossentropy": 2.063150378316641,
+      "loss/hidden": 3.336328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18180835004895926,
+      "step": 18040
+    },
+    {
+      "epoch": 0.45125,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.050455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4234,
+      "loss/crossentropy": 2.085699537396431,
+      "loss/hidden": 3.450390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18879668060690163,
+      "step": 18050
+    },
+    {
+      "epoch": 0.4515,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.415559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4605,
+      "loss/crossentropy": 2.197605139017105,
+      "loss/hidden": 3.324609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1884280290454626,
+      "step": 18060
+    },
+    {
+      "epoch": 0.45175,
+      "grad_norm": 35.75,
+      "grad_norm_var": 3.74765625,
+      "learning_rate": 0.0001,
+      "loss": 7.383,
+      "loss/crossentropy": 2.1910621374845505,
+      "loss/hidden": 3.33125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18678564615547658,
+      "step": 18070
+    },
+    {
+      "epoch": 0.452,
+      "grad_norm": 31.0,
+      "grad_norm_var": 2.5136418856872443e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.4528,
+      "loss/crossentropy": 2.188386806845665,
+      "loss/hidden": 3.6078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18787388410419226,
+      "step": 18080
+    },
+    {
+      "epoch": 0.45225,
+      "grad_norm": 34.5,
+      "grad_norm_var": 5.255989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5167,
+      "loss/crossentropy": 2.086820271611214,
+      "loss/hidden": 3.521484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21262760646641254,
+      "step": 18090
+    },
+    {
+      "epoch": 0.4525,
+      "grad_norm": 30.25,
+      "grad_norm_var": 4.459309895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4389,
+      "loss/crossentropy": 2.156789070367813,
+      "loss/hidden": 3.46171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2072874901816249,
+      "step": 18100
+    },
+    {
+      "epoch": 0.45275,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.32265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4645,
+      "loss/crossentropy": 2.280264839529991,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18678643852472304,
+      "step": 18110
+    },
+    {
+      "epoch": 0.453,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.63515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5154,
+      "loss/crossentropy": 2.167169563472271,
+      "loss/hidden": 3.316015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19211388044059277,
+      "step": 18120
+    },
+    {
+      "epoch": 0.45325,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.9061848958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.564,
+      "loss/crossentropy": 2.070133328437805,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18248395789414645,
+      "step": 18130
+    },
+    {
+      "epoch": 0.4535,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.637955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4519,
+      "loss/crossentropy": 2.068394711613655,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19283174835145472,
+      "step": 18140
+    },
+    {
+      "epoch": 0.45375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 1.7754557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4121,
+      "loss/crossentropy": 2.0715004056692123,
+      "loss/hidden": 3.464453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21265985630452633,
+      "step": 18150
+    },
+    {
+      "epoch": 0.454,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.4551432291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5993,
+      "loss/crossentropy": 2.201769821345806,
+      "loss/hidden": 3.378125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19506389424204826,
+      "step": 18160
+    },
+    {
+      "epoch": 0.45425,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.2499348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4775,
+      "loss/crossentropy": 2.22137137055397,
+      "loss/hidden": 3.260546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18893426284193993,
+      "step": 18170
+    },
+    {
+      "epoch": 0.4545,
+      "grad_norm": 31.25,
+      "grad_norm_var": 27.433268229166668,
+      "learning_rate": 0.0001,
+      "loss": 7.5006,
+      "loss/crossentropy": 2.194701671600342,
+      "loss/hidden": 3.4265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19335605893284083,
+      "step": 18180
+    },
+    {
+      "epoch": 0.45475,
+      "grad_norm": 27.375,
+      "grad_norm_var": 33.23587239583333,
+      "learning_rate": 0.0001,
+      "loss": 7.5273,
+      "loss/crossentropy": 2.224129018187523,
+      "loss/hidden": 3.42734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21178614553064107,
+      "step": 18190
+    },
+    {
+      "epoch": 0.455,
+      "grad_norm": 31.125,
+      "grad_norm_var": 9.046875,
+      "learning_rate": 0.0001,
+      "loss": 7.4934,
+      "loss/crossentropy": 2.168444353342056,
+      "loss/hidden": 3.369921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18636115789413452,
+      "step": 18200
+    },
+    {
+      "epoch": 0.45525,
+      "grad_norm": 28.25,
+      "grad_norm_var": 9.622330729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.5636,
+      "loss/crossentropy": 2.1226459950208665,
+      "loss/hidden": 3.462109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19698166605085135,
+      "step": 18210
+    },
+    {
+      "epoch": 0.4555,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.7848307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4006,
+      "loss/crossentropy": 2.248943442106247,
+      "loss/hidden": 3.337890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1881408639252186,
+      "step": 18220
+    },
+    {
+      "epoch": 0.45575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.893684895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4471,
+      "loss/crossentropy": 2.150214585661888,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2005108630284667,
+      "step": 18230
+    },
+    {
+      "epoch": 0.456,
+      "grad_norm": 30.875,
+      "grad_norm_var": 5.280989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4975,
+      "loss/crossentropy": 2.149556818604469,
+      "loss/hidden": 3.3828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19575091004371642,
+      "step": 18240
+    },
+    {
+      "epoch": 0.45625,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.73125,
+      "learning_rate": 0.0001,
+      "loss": 7.4075,
+      "loss/crossentropy": 2.0766668774187567,
+      "loss/hidden": 3.3984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20460515054874123,
+      "step": 18250
+    },
+    {
+      "epoch": 0.4565,
+      "grad_norm": 37.0,
+      "grad_norm_var": 4.530989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3759,
+      "loss/crossentropy": 2.04805474281311,
+      "loss/hidden": 3.384765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18103201519697903,
+      "step": 18260
+    },
+    {
+      "epoch": 0.45675,
+      "grad_norm": 35.25,
+      "grad_norm_var": 70.37265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5674,
+      "loss/crossentropy": 2.1427599877119063,
+      "loss/hidden": 3.4375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2034947697073221,
+      "step": 18270
+    },
+    {
+      "epoch": 0.457,
+      "grad_norm": 29.25,
+      "grad_norm_var": 7.63515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4797,
+      "loss/crossentropy": 2.074523624777794,
+      "loss/hidden": 3.405078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19340968001633882,
+      "step": 18280
+    },
+    {
+      "epoch": 0.45725,
+      "grad_norm": 31.625,
+      "grad_norm_var": 4.620247395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4549,
+      "loss/crossentropy": 2.224767416715622,
+      "loss/hidden": 3.319140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18859731387346984,
+      "step": 18290
+    },
+    {
+      "epoch": 0.4575,
+      "grad_norm": 32.75,
+      "grad_norm_var": 16.937955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4625,
+      "loss/crossentropy": 2.1193591982126234,
+      "loss/hidden": 3.419921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2029947843402624,
+      "step": 18300
+    },
+    {
+      "epoch": 0.45775,
+      "grad_norm": 39.5,
+      "grad_norm_var": 18.048893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4101,
+      "loss/crossentropy": 2.1422934621572494,
+      "loss/hidden": 3.354296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19083867389708759,
+      "step": 18310
+    },
+    {
+      "epoch": 0.458,
+      "grad_norm": 28.125,
+      "grad_norm_var": 8.314322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4655,
+      "loss/crossentropy": 2.0613017052412035,
+      "loss/hidden": 3.433203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2031024197116494,
+      "step": 18320
+    },
+    {
+      "epoch": 0.45825,
+      "grad_norm": 40.5,
+      "grad_norm_var": 15.055989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3438,
+      "loss/crossentropy": 2.208176797628403,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1937787916511297,
+      "step": 18330
+    },
+    {
+      "epoch": 0.4585,
+      "grad_norm": 29.75,
+      "grad_norm_var": 12.668489583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4472,
+      "loss/crossentropy": 2.1267407298088075,
+      "loss/hidden": 3.444140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19318134970963002,
+      "step": 18340
+    },
+    {
+      "epoch": 0.45875,
+      "grad_norm": 29.375,
+      "grad_norm_var": 10.7384765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3673,
+      "loss/crossentropy": 2.0367541924118995,
+      "loss/hidden": 3.29375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1756553279235959,
+      "step": 18350
+    },
+    {
+      "epoch": 0.459,
+      "grad_norm": 33.5,
+      "grad_norm_var": 11.2697265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5204,
+      "loss/crossentropy": 2.187627172470093,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.192278266325593,
+      "step": 18360
+    },
+    {
+      "epoch": 0.45925,
+      "grad_norm": 35.0,
+      "grad_norm_var": 49.330989583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3755,
+      "loss/crossentropy": 2.0965137451887133,
+      "loss/hidden": 3.3796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17834257557988167,
+      "step": 18370
+    },
+    {
+      "epoch": 0.4595,
+      "grad_norm": 31.625,
+      "grad_norm_var": 14.001822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.454,
+      "loss/crossentropy": 2.1809972152113914,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19421270694583653,
+      "step": 18380
+    },
+    {
+      "epoch": 0.45975,
+      "grad_norm": 29.375,
+      "grad_norm_var": 12.6072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4537,
+      "loss/crossentropy": 2.0333224818110467,
+      "loss/hidden": 3.307421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18099675104022026,
+      "step": 18390
+    },
+    {
+      "epoch": 0.46,
+      "grad_norm": 31.875,
+      "grad_norm_var": 7.2728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5258,
+      "loss/crossentropy": 2.0772110372781754,
+      "loss/hidden": 3.532421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2270737275481224,
+      "step": 18400
+    },
+    {
+      "epoch": 0.46025,
+      "grad_norm": 28.75,
+      "grad_norm_var": 5.723893229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.3846,
+      "loss/crossentropy": 2.092162123322487,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2126236293464899,
+      "step": 18410
+    },
+    {
+      "epoch": 0.4605,
+      "grad_norm": 30.5,
+      "grad_norm_var": 3.8921223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4763,
+      "loss/crossentropy": 2.118947532773018,
+      "loss/hidden": 3.405859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20705053452402353,
+      "step": 18420
+    },
+    {
+      "epoch": 0.46075,
+      "grad_norm": 30.5,
+      "grad_norm_var": 5.431184895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3772,
+      "loss/crossentropy": 2.150417110323906,
+      "loss/hidden": 3.244140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18132503498345615,
+      "step": 18430
+    },
+    {
+      "epoch": 0.461,
+      "grad_norm": 31.125,
+      "grad_norm_var": 5.370572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4958,
+      "loss/crossentropy": 2.0525425523519516,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.178656730055809,
+      "step": 18440
+    },
+    {
+      "epoch": 0.46125,
+      "grad_norm": 34.75,
+      "grad_norm_var": 4.55390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4681,
+      "loss/crossentropy": 2.1539366483688354,
+      "loss/hidden": 3.40234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20078267380595208,
+      "step": 18450
+    },
+    {
+      "epoch": 0.4615,
+      "grad_norm": 28.625,
+      "grad_norm_var": 6.175,
+      "learning_rate": 0.0001,
+      "loss": 7.4644,
+      "loss/crossentropy": 2.238241583108902,
+      "loss/hidden": 3.36640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2120617249980569,
+      "step": 18460
+    },
+    {
+      "epoch": 0.46175,
+      "grad_norm": 34.0,
+      "grad_norm_var": 6.612955729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.452,
+      "loss/crossentropy": 2.061172368377447,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1729869638569653,
+      "step": 18470
+    },
+    {
+      "epoch": 0.462,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.0072265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4506,
+      "loss/crossentropy": 2.1220480680465696,
+      "loss/hidden": 3.4203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20648412220180035,
+      "step": 18480
+    },
+    {
+      "epoch": 0.46225,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.564322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5257,
+      "loss/crossentropy": 2.1619621008634566,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22688564024865626,
+      "step": 18490
+    },
+    {
+      "epoch": 0.4625,
+      "grad_norm": 32.0,
+      "grad_norm_var": 5.120833333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3694,
+      "loss/crossentropy": 2.264406609535217,
+      "loss/hidden": 3.298046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18825284875929355,
+      "step": 18500
+    },
+    {
+      "epoch": 0.46275,
+      "grad_norm": 28.125,
+      "grad_norm_var": 5.265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5201,
+      "loss/crossentropy": 2.2461962729692457,
+      "loss/hidden": 3.420703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21331474296748637,
+      "step": 18510
+    },
+    {
+      "epoch": 0.463,
+      "grad_norm": 30.75,
+      "grad_norm_var": 4.0728515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4195,
+      "loss/crossentropy": 1.9451055109500885,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18808368891477584,
+      "step": 18520
+    },
+    {
+      "epoch": 0.46325,
+      "grad_norm": 31.0,
+      "grad_norm_var": 3.0927083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4772,
+      "loss/crossentropy": 2.110605201125145,
+      "loss/hidden": 3.4671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19988117516040801,
+      "step": 18530
+    },
+    {
+      "epoch": 0.4635,
+      "grad_norm": 30.25,
+      "grad_norm_var": 1.9395182291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4795,
+      "loss/crossentropy": 1.9315606564283372,
+      "loss/hidden": 3.428515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17853899393230677,
+      "step": 18540
+    },
+    {
+      "epoch": 0.46375,
+      "grad_norm": 34.5,
+      "grad_norm_var": 26.724934895833332,
+      "learning_rate": 0.0001,
+      "loss": 7.5498,
+      "loss/crossentropy": 2.1117912903428078,
+      "loss/hidden": 3.54453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2193370433524251,
+      "step": 18550
+    },
+    {
+      "epoch": 0.464,
+      "grad_norm": 29.25,
+      "grad_norm_var": 27.626822916666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4152,
+      "loss/crossentropy": 2.057279494404793,
+      "loss/hidden": 3.337109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18574363086372614,
+      "step": 18560
+    },
+    {
+      "epoch": 0.46425,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.3452473958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4341,
+      "loss/crossentropy": 2.0958516538143157,
+      "loss/hidden": 3.36015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18902888633310794,
+      "step": 18570
+    },
+    {
+      "epoch": 0.4645,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.6830729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5493,
+      "loss/crossentropy": 2.177600572258234,
+      "loss/hidden": 3.457421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2118171840906143,
+      "step": 18580
+    },
+    {
+      "epoch": 0.46475,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.475455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4925,
+      "loss/crossentropy": 2.087388700246811,
+      "loss/hidden": 3.41953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19477690532803535,
+      "step": 18590
+    },
+    {
+      "epoch": 0.465,
+      "grad_norm": 30.75,
+      "grad_norm_var": 10.176822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5546,
+      "loss/crossentropy": 2.076089936494827,
+      "loss/hidden": 3.38203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18885728288441897,
+      "step": 18600
+    },
+    {
+      "epoch": 0.46525,
+      "grad_norm": 30.0,
+      "grad_norm_var": 1.8504557291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4298,
+      "loss/crossentropy": 2.2245807096362116,
+      "loss/hidden": 3.455859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2102813944220543,
+      "step": 18610
+    },
+    {
+      "epoch": 0.4655,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.06015625,
+      "learning_rate": 0.0001,
+      "loss": 7.5033,
+      "loss/crossentropy": 2.037870594859123,
+      "loss/hidden": 3.42421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19886170476675033,
+      "step": 18620
+    },
+    {
+      "epoch": 0.46575,
+      "grad_norm": 29.25,
+      "grad_norm_var": 4.106705729166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4177,
+      "loss/crossentropy": 2.1507607758045197,
+      "loss/hidden": 3.378515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20115135256201028,
+      "step": 18630
+    },
+    {
+      "epoch": 0.466,
+      "grad_norm": 29.25,
+      "grad_norm_var": 3.1455729166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4323,
+      "loss/crossentropy": 2.208960172533989,
+      "loss/hidden": 3.425,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21820493452250958,
+      "step": 18640
+    },
+    {
+      "epoch": 0.46625,
+      "grad_norm": 29.875,
+      "grad_norm_var": 12.363541666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5667,
+      "loss/crossentropy": 2.062282791733742,
+      "loss/hidden": 3.4140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2069956874474883,
+      "step": 18650
+    },
+    {
+      "epoch": 0.4665,
+      "grad_norm": 31.5,
+      "grad_norm_var": 3.501822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4799,
+      "loss/crossentropy": 2.155339244008064,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20734498463571072,
+      "step": 18660
+    },
+    {
+      "epoch": 0.46675,
+      "grad_norm": 35.5,
+      "grad_norm_var": 10.81875,
+      "learning_rate": 0.0001,
+      "loss": 7.4112,
+      "loss/crossentropy": 2.116558775305748,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.22728623170405626,
+      "step": 18670
+    },
+    {
+      "epoch": 0.467,
+      "grad_norm": 48.0,
+      "grad_norm_var": 112.26555989583333,
+      "learning_rate": 0.0001,
+      "loss": 7.5819,
+      "loss/crossentropy": 2.1996765822172164,
+      "loss/hidden": 3.391796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20961329583078622,
+      "step": 18680
+    },
+    {
+      "epoch": 0.46725,
+      "grad_norm": 29.625,
+      "grad_norm_var": 121.28932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4733,
+      "loss/crossentropy": 2.178187218308449,
+      "loss/hidden": 3.476953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20361957363784314,
+      "step": 18690
+    },
+    {
+      "epoch": 0.4675,
+      "grad_norm": 29.375,
+      "grad_norm_var": 8.421809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4133,
+      "loss/crossentropy": 2.3110476464033125,
+      "loss/hidden": 3.28359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19236205983906984,
+      "step": 18700
+    },
+    {
+      "epoch": 0.46775,
+      "grad_norm": 40.0,
+      "grad_norm_var": 2.408370244508014e+18,
+      "learning_rate": 0.0001,
+      "loss": 7.5826,
+      "loss/crossentropy": 1.9763169527053832,
+      "loss/hidden": 3.5734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19957512021064758,
+      "step": 18710
+    },
+    {
+      "epoch": 0.468,
+      "grad_norm": 33.25,
+      "grad_norm_var": 7.312239583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5,
+      "loss/crossentropy": 2.1549572572112083,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.197417950630188,
+      "step": 18720
+    },
+    {
+      "epoch": 0.46825,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.8677083333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4042,
+      "loss/crossentropy": 2.168655255436897,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19954893179237843,
+      "step": 18730
+    },
+    {
+      "epoch": 0.4685,
+      "grad_norm": 31.5,
+      "grad_norm_var": 2.6478515625,
+      "learning_rate": 0.0001,
+      "loss": 7.5232,
+      "loss/crossentropy": 2.11354903280735,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1892693081870675,
+      "step": 18740
+    },
+    {
+      "epoch": 0.46875,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.4692057291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4898,
+      "loss/crossentropy": 2.1568462401628494,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20531034972518683,
+      "step": 18750
+    },
+    {
+      "epoch": 0.469,
+      "grad_norm": 32.0,
+      "grad_norm_var": 2.2113932291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.6474,
+      "loss/crossentropy": 2.1167890459299086,
+      "loss/hidden": 3.372265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18697490729391575,
+      "step": 18760
+    },
+    {
+      "epoch": 0.46925,
+      "grad_norm": 28.625,
+      "grad_norm_var": 2.9041015625,
+      "learning_rate": 0.0001,
+      "loss": 7.3638,
+      "loss/crossentropy": 2.065881370007992,
+      "loss/hidden": 3.2921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18317838590592145,
+      "step": 18770
+    },
+    {
+      "epoch": 0.4695,
+      "grad_norm": 30.375,
+      "grad_norm_var": 3.733268229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4584,
+      "loss/crossentropy": 2.2059927970170974,
+      "loss/hidden": 3.390234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19830067679286004,
+      "step": 18780
+    },
+    {
+      "epoch": 0.46975,
+      "grad_norm": 29.5,
+      "grad_norm_var": 3.7934895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4786,
+      "loss/crossentropy": 2.0999852418899536,
+      "loss/hidden": 3.26484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1729890163987875,
+      "step": 18790
+    },
+    {
+      "epoch": 0.47,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.5400390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4596,
+      "loss/crossentropy": 2.015045887231827,
+      "loss/hidden": 3.280859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1711340345442295,
+      "step": 18800
+    },
+    {
+      "epoch": 0.47025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.2223307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.4839,
+      "loss/crossentropy": 2.0983877390623094,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18115116618573665,
+      "step": 18810
+    },
+    {
+      "epoch": 0.4705,
+      "grad_norm": 30.0,
+      "grad_norm_var": 78.5822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5142,
+      "loss/crossentropy": 2.122063784301281,
+      "loss/hidden": 3.3453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18946684896945953,
+      "step": 18820
+    },
+    {
+      "epoch": 0.47075,
+      "grad_norm": 28.375,
+      "grad_norm_var": 97.58639322916666,
+      "learning_rate": 0.0001,
+      "loss": 7.6061,
+      "loss/crossentropy": 2.237548753619194,
+      "loss/hidden": 3.2828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19166598822921516,
+      "step": 18830
+    },
+    {
+      "epoch": 0.471,
+      "grad_norm": 30.25,
+      "grad_norm_var": 9.527083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4705,
+      "loss/crossentropy": 2.081946098059416,
+      "loss/hidden": 3.38515625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18989304369315504,
+      "step": 18840
+    },
+    {
+      "epoch": 0.47125,
+      "grad_norm": 31.125,
+      "grad_norm_var": 8.614583333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.3761,
+      "loss/crossentropy": 2.2694112062454224,
+      "loss/hidden": 3.291015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1898673100396991,
+      "step": 18850
+    },
+    {
+      "epoch": 0.4715,
+      "grad_norm": 35.0,
+      "grad_norm_var": 3.9619140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5445,
+      "loss/crossentropy": 2.1576954081654547,
+      "loss/hidden": 3.423046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19350357558578252,
+      "step": 18860
+    },
+    {
+      "epoch": 0.47175,
+      "grad_norm": 33.5,
+      "grad_norm_var": 13.674934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.498,
+      "loss/crossentropy": 2.0821513146162034,
+      "loss/hidden": 3.458984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19236762393265963,
+      "step": 18870
+    },
+    {
+      "epoch": 0.472,
+      "grad_norm": 30.0,
+      "grad_norm_var": 14.720572916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.44,
+      "loss/crossentropy": 2.097795248031616,
+      "loss/hidden": 3.4109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1873439846560359,
+      "step": 18880
+    },
+    {
+      "epoch": 0.47225,
+      "grad_norm": 50.5,
+      "grad_norm_var": 28.486458333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5299,
+      "loss/crossentropy": 2.1926951453089716,
+      "loss/hidden": 3.483203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19845539480447769,
+      "step": 18890
+    },
+    {
+      "epoch": 0.4725,
+      "grad_norm": 28.0,
+      "grad_norm_var": 27.077018229166665,
+      "learning_rate": 0.0001,
+      "loss": 7.4255,
+      "loss/crossentropy": 2.1055196806788445,
+      "loss/hidden": 3.437109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1981551146134734,
+      "step": 18900
+    },
+    {
+      "epoch": 0.47275,
+      "grad_norm": 28.375,
+      "grad_norm_var": 2.6311848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4186,
+      "loss/crossentropy": 2.1644329100847246,
+      "loss/hidden": 3.3859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1863600155338645,
+      "step": 18910
+    },
+    {
+      "epoch": 0.473,
+      "grad_norm": 37.25,
+      "grad_norm_var": 5.633333333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5188,
+      "loss/crossentropy": 2.111950933933258,
+      "loss/hidden": 3.329296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19227960985153913,
+      "step": 18920
+    },
+    {
+      "epoch": 0.47325,
+      "grad_norm": 30.5,
+      "grad_norm_var": 4.974934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4124,
+      "loss/crossentropy": 2.0366090714931486,
+      "loss/hidden": 3.446484375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20001390851102768,
+      "step": 18930
+    },
+    {
+      "epoch": 0.4735,
+      "grad_norm": 28.5,
+      "grad_norm_var": 4.0509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4834,
+      "loss/crossentropy": 2.136207638680935,
+      "loss/hidden": 3.386328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18433616552501916,
+      "step": 18940
+    },
+    {
+      "epoch": 0.47375,
+      "grad_norm": 33.0,
+      "grad_norm_var": 2.98515625,
+      "learning_rate": 0.0001,
+      "loss": 7.6108,
+      "loss/crossentropy": 2.1925762712955477,
+      "loss/hidden": 3.492578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2353795062750578,
+      "step": 18950
+    },
+    {
+      "epoch": 0.474,
+      "grad_norm": 32.75,
+      "grad_norm_var": 3.55390625,
+      "learning_rate": 0.0001,
+      "loss": 7.6327,
+      "loss/crossentropy": 2.1619835913181307,
+      "loss/hidden": 3.4234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20424192007631065,
+      "step": 18960
+    },
+    {
+      "epoch": 0.47425,
+      "grad_norm": 32.5,
+      "grad_norm_var": 2.7983723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.538,
+      "loss/crossentropy": 2.1776172876358033,
+      "loss/hidden": 3.4296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20042523313313723,
+      "step": 18970
+    },
+    {
+      "epoch": 0.4745,
+      "grad_norm": 30.625,
+      "grad_norm_var": 4.616080729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5789,
+      "loss/crossentropy": 2.124316268414259,
+      "loss/hidden": 3.326953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18864221088588237,
+      "step": 18980
+    },
+    {
+      "epoch": 0.47475,
+      "grad_norm": 31.875,
+      "grad_norm_var": 2.0697916666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.574,
+      "loss/crossentropy": 2.073023219406605,
+      "loss/hidden": 3.45078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2027593031525612,
+      "step": 18990
+    },
+    {
+      "epoch": 0.475,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.664322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4101,
+      "loss/crossentropy": 2.0187594324350355,
+      "loss/hidden": 3.48671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21064169742166997,
+      "step": 19000
+    },
+    {
+      "epoch": 0.47525,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.8197265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4446,
+      "loss/crossentropy": 2.2003646433353423,
+      "loss/hidden": 3.28359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18558428715914488,
+      "step": 19010
+    },
+    {
+      "epoch": 0.4755,
+      "grad_norm": 33.5,
+      "grad_norm_var": 85.84348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5855,
+      "loss/crossentropy": 2.2160910099744795,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20706005822867155,
+      "step": 19020
+    },
+    {
+      "epoch": 0.47575,
+      "grad_norm": 27.875,
+      "grad_norm_var": 6.8462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4416,
+      "loss/crossentropy": 2.080396056175232,
+      "loss/hidden": 3.3640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18421638533473014,
+      "step": 19030
+    },
+    {
+      "epoch": 0.476,
+      "grad_norm": 30.0,
+      "grad_norm_var": 2.349739583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4877,
+      "loss/crossentropy": 2.0878879494965075,
+      "loss/hidden": 3.530078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20337579548358917,
+      "step": 19040
+    },
+    {
+      "epoch": 0.47625,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.975455729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3789,
+      "loss/crossentropy": 2.0772477351129055,
+      "loss/hidden": 3.3125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1744394849985838,
+      "step": 19050
+    },
+    {
+      "epoch": 0.4765,
+      "grad_norm": 30.25,
+      "grad_norm_var": 24.359375,
+      "learning_rate": 0.0001,
+      "loss": 7.4862,
+      "loss/crossentropy": 2.100569760799408,
+      "loss/hidden": 3.410546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19498176760971547,
+      "step": 19060
+    },
+    {
+      "epoch": 0.47675,
+      "grad_norm": 33.5,
+      "grad_norm_var": 23.444791666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4441,
+      "loss/crossentropy": 2.088040843605995,
+      "loss/hidden": 3.4046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1772982817143202,
+      "step": 19070
+    },
+    {
+      "epoch": 0.477,
+      "grad_norm": 29.625,
+      "grad_norm_var": 4.158072916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4619,
+      "loss/crossentropy": 2.176638293266296,
+      "loss/hidden": 3.430859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19879398066550494,
+      "step": 19080
+    },
+    {
+      "epoch": 0.47725,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.5978515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4313,
+      "loss/crossentropy": 2.091726778447628,
+      "loss/hidden": 3.393359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18576510678976774,
+      "step": 19090
+    },
+    {
+      "epoch": 0.4775,
+      "grad_norm": 31.5,
+      "grad_norm_var": 5.673372395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4909,
+      "loss/crossentropy": 2.235911877453327,
+      "loss/hidden": 3.3921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21205585449934006,
+      "step": 19100
+    },
+    {
+      "epoch": 0.47775,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.5973307291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.394,
+      "loss/crossentropy": 2.2365425676107407,
+      "loss/hidden": 3.309765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19166039237752558,
+      "step": 19110
+    },
+    {
+      "epoch": 0.478,
+      "grad_norm": 31.25,
+      "grad_norm_var": 3.24375,
+      "learning_rate": 0.0001,
+      "loss": 7.5406,
+      "loss/crossentropy": 2.108828788995743,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18494639433920385,
+      "step": 19120
+    },
+    {
+      "epoch": 0.47825,
+      "grad_norm": 29.25,
+      "grad_norm_var": 6.721809895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.5922,
+      "loss/crossentropy": 2.148023870587349,
+      "loss/hidden": 3.448828125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19042143411934376,
+      "step": 19130
+    },
+    {
+      "epoch": 0.4785,
+      "grad_norm": 28.5,
+      "grad_norm_var": 2.0660807291666665,
+      "learning_rate": 0.0001,
+      "loss": 7.3712,
+      "loss/crossentropy": 2.1488536432385446,
+      "loss/hidden": 3.3546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19772066082805395,
+      "step": 19140
+    },
+    {
+      "epoch": 0.47875,
+      "grad_norm": 32.5,
+      "grad_norm_var": 3.3541015625,
+      "learning_rate": 0.0001,
+      "loss": 7.542,
+      "loss/crossentropy": 2.0818497955799105,
+      "loss/hidden": 3.45703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20143606336787342,
+      "step": 19150
+    },
+    {
+      "epoch": 0.479,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.6796223958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5226,
+      "loss/crossentropy": 2.213534101843834,
+      "loss/hidden": 3.39375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20065031982958317,
+      "step": 19160
+    },
+    {
+      "epoch": 0.47925,
+      "grad_norm": 29.625,
+      "grad_norm_var": 5.286393229166666,
+      "learning_rate": 0.0001,
+      "loss": 7.4779,
+      "loss/crossentropy": 2.094844227284193,
+      "loss/hidden": 3.417578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19940570322796702,
+      "step": 19170
+    },
+    {
+      "epoch": 0.4795,
+      "grad_norm": 29.5,
+      "grad_norm_var": 2.856705729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4351,
+      "loss/crossentropy": 2.0740013778209687,
+      "loss/hidden": 3.31640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1735646616667509,
+      "step": 19180
+    },
+    {
+      "epoch": 0.47975,
+      "grad_norm": 31.125,
+      "grad_norm_var": 4.551497395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3991,
+      "loss/crossentropy": 2.1303412437438967,
+      "loss/hidden": 3.367578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20425421688705683,
+      "step": 19190
+    },
+    {
+      "epoch": 0.48,
+      "grad_norm": 28.25,
+      "grad_norm_var": 4.960416666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4677,
+      "loss/crossentropy": 2.211458620429039,
+      "loss/hidden": 3.36171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20086632966995238,
+      "step": 19200
+    },
+    {
+      "epoch": 0.48025,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.3056640625,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.098607787489891,
+      "loss/hidden": 3.3359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18264676369726657,
+      "step": 19210
+    },
+    {
+      "epoch": 0.4805,
+      "grad_norm": 30.375,
+      "grad_norm_var": 2.2363932291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4141,
+      "loss/crossentropy": 2.1743356212973595,
+      "loss/hidden": 3.23203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18333771340548993,
+      "step": 19220
+    },
+    {
+      "epoch": 0.48075,
+      "grad_norm": 29.625,
+      "grad_norm_var": 79.7166015625,
+      "learning_rate": 0.0001,
+      "loss": 7.4269,
+      "loss/crossentropy": 2.269570749998093,
+      "loss/hidden": 3.339453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1983551263809204,
+      "step": 19230
+    },
+    {
+      "epoch": 0.481,
+      "grad_norm": 30.25,
+      "grad_norm_var": 3.2462890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4745,
+      "loss/crossentropy": 2.0477301344275474,
+      "loss/hidden": 3.256640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18232697807252407,
+      "step": 19240
+    },
+    {
+      "epoch": 0.48125,
+      "grad_norm": 34.0,
+      "grad_norm_var": 4.055989583333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4972,
+      "loss/crossentropy": 2.0904338479042055,
+      "loss/hidden": 3.39765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19662635941058398,
+      "step": 19250
+    },
+    {
+      "epoch": 0.4815,
+      "grad_norm": 33.75,
+      "grad_norm_var": 23.349739583333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5738,
+      "loss/crossentropy": 2.146761494874954,
+      "loss/hidden": 3.375390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2060481060296297,
+      "step": 19260
+    },
+    {
+      "epoch": 0.48175,
+      "grad_norm": 30.75,
+      "grad_norm_var": 25.313997395833333,
+      "learning_rate": 0.0001,
+      "loss": 7.436,
+      "loss/crossentropy": 2.136366602778435,
+      "loss/hidden": 3.45390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18981903176754714,
+      "step": 19270
+    },
+    {
+      "epoch": 0.482,
+      "grad_norm": 30.875,
+      "grad_norm_var": 10.1134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.5383,
+      "loss/crossentropy": 2.155537909269333,
+      "loss/hidden": 3.356640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19545716531574725,
+      "step": 19280
+    },
+    {
+      "epoch": 0.48225,
+      "grad_norm": 35.0,
+      "grad_norm_var": 22.241666666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4286,
+      "loss/crossentropy": 2.2027793377637863,
+      "loss/hidden": 3.33203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18559762574732303,
+      "step": 19290
+    },
+    {
+      "epoch": 0.4825,
+      "grad_norm": 29.0,
+      "grad_norm_var": 22.192122395833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4078,
+      "loss/crossentropy": 2.146567094326019,
+      "loss/hidden": 3.412109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19877713918685913,
+      "step": 19300
+    },
+    {
+      "epoch": 0.48275,
+      "grad_norm": 28.625,
+      "grad_norm_var": 10.4509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4685,
+      "loss/crossentropy": 2.1408349931240083,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20110496152192353,
+      "step": 19310
+    },
+    {
+      "epoch": 0.483,
+      "grad_norm": 27.625,
+      "grad_norm_var": 38.0509765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4049,
+      "loss/crossentropy": 2.0760223269462585,
+      "loss/hidden": 3.401953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1911366345360875,
+      "step": 19320
+    },
+    {
+      "epoch": 0.48325,
+      "grad_norm": 29.125,
+      "grad_norm_var": 6.47265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5924,
+      "loss/crossentropy": 2.0804271429777144,
+      "loss/hidden": 3.48046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20118813700973986,
+      "step": 19330
+    },
+    {
+      "epoch": 0.4835,
+      "grad_norm": 31.375,
+      "grad_norm_var": 5.389322916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4406,
+      "loss/crossentropy": 2.1793388813734054,
+      "loss/hidden": 3.39609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19482165575027466,
+      "step": 19340
+    },
+    {
+      "epoch": 0.48375,
+      "grad_norm": 28.375,
+      "grad_norm_var": 1.6296223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.3994,
+      "loss/crossentropy": 2.037779450416565,
+      "loss/hidden": 3.459375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19606408532708883,
+      "step": 19350
+    },
+    {
+      "epoch": 0.484,
+      "grad_norm": 29.75,
+      "grad_norm_var": 2.1025390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5035,
+      "loss/crossentropy": 2.108065330982208,
+      "loss/hidden": 3.300390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1907705569639802,
+      "step": 19360
+    },
+    {
+      "epoch": 0.48425,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.8333333333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5038,
+      "loss/crossentropy": 2.17846602499485,
+      "loss/hidden": 3.3265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19125208593904972,
+      "step": 19370
+    },
+    {
+      "epoch": 0.4845,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.4442057291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5134,
+      "loss/crossentropy": 2.139002203941345,
+      "loss/hidden": 3.43203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20099837705492973,
+      "step": 19380
+    },
+    {
+      "epoch": 0.48475,
+      "grad_norm": 30.5,
+      "grad_norm_var": 2.810416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5401,
+      "loss/crossentropy": 2.199608436226845,
+      "loss/hidden": 3.35546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20445926804095507,
+      "step": 19390
+    },
+    {
+      "epoch": 0.485,
+      "grad_norm": 31.875,
+      "grad_norm_var": 3.9160807291666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5383,
+      "loss/crossentropy": 2.12961964905262,
+      "loss/hidden": 3.445703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20509012490510942,
+      "step": 19400
+    },
+    {
+      "epoch": 0.48525,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.0052083333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.507,
+      "loss/crossentropy": 2.1178236939013004,
+      "loss/hidden": 3.406640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2015313227660954,
+      "step": 19410
+    },
+    {
+      "epoch": 0.4855,
+      "grad_norm": 31.25,
+      "grad_norm_var": 4.2712890625,
+      "learning_rate": 0.0001,
+      "loss": 7.4174,
+      "loss/crossentropy": 2.243345336616039,
+      "loss/hidden": 3.415234375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19893982484936715,
+      "step": 19420
+    },
+    {
+      "epoch": 0.48575,
+      "grad_norm": 28.875,
+      "grad_norm_var": 4.418684895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.4455,
+      "loss/crossentropy": 2.2348459392786024,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19131067730486392,
+      "step": 19430
+    },
+    {
+      "epoch": 0.486,
+      "grad_norm": 31.5,
+      "grad_norm_var": 1.3369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4526,
+      "loss/crossentropy": 2.0366022780537607,
+      "loss/hidden": 3.37109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20100958244875072,
+      "step": 19440
+    },
+    {
+      "epoch": 0.48625,
+      "grad_norm": 29.25,
+      "grad_norm_var": 1.0559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5342,
+      "loss/crossentropy": 2.1607364803552627,
+      "loss/hidden": 3.4796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1935076829046011,
+      "step": 19450
+    },
+    {
+      "epoch": 0.4865,
+      "grad_norm": 29.625,
+      "grad_norm_var": 1.4410807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5696,
+      "loss/crossentropy": 2.075075288116932,
+      "loss/hidden": 3.396875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20319369733333587,
+      "step": 19460
+    },
+    {
+      "epoch": 0.48675,
+      "grad_norm": 29.5,
+      "grad_norm_var": 1.5134765625,
+      "learning_rate": 0.0001,
+      "loss": 7.3679,
+      "loss/crossentropy": 2.1516528606414793,
+      "loss/hidden": 3.330078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18369753509759904,
+      "step": 19470
+    },
+    {
+      "epoch": 0.487,
+      "grad_norm": 30.25,
+      "grad_norm_var": 10.126822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5018,
+      "loss/crossentropy": 2.059604635834694,
+      "loss/hidden": 3.476171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19641294488683342,
+      "step": 19480
+    },
+    {
+      "epoch": 0.48725,
+      "grad_norm": 30.625,
+      "grad_norm_var": 17.711458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4581,
+      "loss/crossentropy": 2.0580811988562346,
+      "loss/hidden": 3.34765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1850988393649459,
+      "step": 19490
+    },
+    {
+      "epoch": 0.4875,
+      "grad_norm": 31.25,
+      "grad_norm_var": 18.8494140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4531,
+      "loss/crossentropy": 2.213153839111328,
+      "loss/hidden": 3.34375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1889218719676137,
+      "step": 19500
+    },
+    {
+      "epoch": 0.48775,
+      "grad_norm": 29.0,
+      "grad_norm_var": 11.201041666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5007,
+      "loss/crossentropy": 2.1869974330067636,
+      "loss/hidden": 3.50546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21722153890877963,
+      "step": 19510
+    },
+    {
+      "epoch": 0.488,
+      "grad_norm": 28.625,
+      "grad_norm_var": 11.2244140625,
+      "learning_rate": 0.0001,
+      "loss": 7.5173,
+      "loss/crossentropy": 2.3184584975242615,
+      "loss/hidden": 3.28203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19578878507018088,
+      "step": 19520
+    },
+    {
+      "epoch": 0.48825,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.477083333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4547,
+      "loss/crossentropy": 2.2146874606609344,
+      "loss/hidden": 3.3203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18467055093497037,
+      "step": 19530
+    },
+    {
+      "epoch": 0.4885,
+      "grad_norm": 34.0,
+      "grad_norm_var": 1.90390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5217,
+      "loss/crossentropy": 2.121046042442322,
+      "loss/hidden": 3.37890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19369308706372976,
+      "step": 19540
+    },
+    {
+      "epoch": 0.48875,
+      "grad_norm": 30.75,
+      "grad_norm_var": 2.3322265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3847,
+      "loss/crossentropy": 2.03256713449955,
+      "loss/hidden": 3.44453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18962962226942182,
+      "step": 19550
+    },
+    {
+      "epoch": 0.489,
+      "grad_norm": 28.625,
+      "grad_norm_var": 1.7330729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4378,
+      "loss/crossentropy": 2.0885259807109833,
+      "loss/hidden": 3.368359375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941137770190835,
+      "step": 19560
+    },
+    {
+      "epoch": 0.48925,
+      "grad_norm": 30.25,
+      "grad_norm_var": 2.25625,
+      "learning_rate": 0.0001,
+      "loss": 7.4224,
+      "loss/crossentropy": 2.0018733590841293,
+      "loss/hidden": 3.4703125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18955971095710994,
+      "step": 19570
+    },
+    {
+      "epoch": 0.4895,
+      "grad_norm": 32.5,
+      "grad_norm_var": 1.94375,
+      "learning_rate": 0.0001,
+      "loss": 7.4489,
+      "loss/crossentropy": 2.136213929951191,
+      "loss/hidden": 3.366015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2040450617671013,
+      "step": 19580
+    },
+    {
+      "epoch": 0.48975,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.5559895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4945,
+      "loss/crossentropy": 2.049376104027033,
+      "loss/hidden": 3.481640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941303763538599,
+      "step": 19590
+    },
+    {
+      "epoch": 0.49,
+      "grad_norm": 32.25,
+      "grad_norm_var": 2.3643229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4543,
+      "loss/crossentropy": 2.189481696486473,
+      "loss/hidden": 3.351171875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19288850836455823,
+      "step": 19600
+    },
+    {
+      "epoch": 0.49025,
+      "grad_norm": 31.125,
+      "grad_norm_var": 1.6760416666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.3698,
+      "loss/crossentropy": 2.0647163964807986,
+      "loss/hidden": 3.398046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1837151089683175,
+      "step": 19610
+    },
+    {
+      "epoch": 0.4905,
+      "grad_norm": 31.25,
+      "grad_norm_var": 1.3643229166666666,
+      "learning_rate": 0.0001,
+      "loss": 7.4861,
+      "loss/crossentropy": 2.116924060881138,
+      "loss/hidden": 3.383984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19573113061487674,
+      "step": 19620
+    },
+    {
+      "epoch": 0.49075,
+      "grad_norm": 33.75,
+      "grad_norm_var": 2.9848307291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4356,
+      "loss/crossentropy": 2.0818495750427246,
+      "loss/hidden": 3.359765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19164716321974992,
+      "step": 19630
+    },
+    {
+      "epoch": 0.491,
+      "grad_norm": 29.875,
+      "grad_norm_var": 2.0369140625,
+      "learning_rate": 0.0001,
+      "loss": 7.4685,
+      "loss/crossentropy": 2.0525588363409044,
+      "loss/hidden": 3.428125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19741135966032744,
+      "step": 19640
+    },
+    {
+      "epoch": 0.49125,
+      "grad_norm": 29.125,
+      "grad_norm_var": 3.5171223958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4433,
+      "loss/crossentropy": 2.167221999168396,
+      "loss/hidden": 3.384375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19002752266824247,
+      "step": 19650
+    },
+    {
+      "epoch": 0.4915,
+      "grad_norm": 35.5,
+      "grad_norm_var": 4.730208333333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5538,
+      "loss/crossentropy": 2.125367084145546,
+      "loss/hidden": 3.456640625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2023206612095237,
+      "step": 19660
+    },
+    {
+      "epoch": 0.49175,
+      "grad_norm": 30.0,
+      "grad_norm_var": 4.536458333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5214,
+      "loss/crossentropy": 2.2167422816157343,
+      "loss/hidden": 3.454296875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19829270094633103,
+      "step": 19670
+    },
+    {
+      "epoch": 0.492,
+      "grad_norm": 31.625,
+      "grad_norm_var": 2.948958333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4018,
+      "loss/crossentropy": 2.16324964761734,
+      "loss/hidden": 3.335546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1941527757793665,
+      "step": 19680
+    },
+    {
+      "epoch": 0.49225,
+      "grad_norm": 29.875,
+      "grad_norm_var": 3.678580729166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4845,
+      "loss/crossentropy": 2.104529711604118,
+      "loss/hidden": 3.26796875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18227193616330623,
+      "step": 19690
+    },
+    {
+      "epoch": 0.4925,
+      "grad_norm": 29.875,
+      "grad_norm_var": 1.3080729166666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4912,
+      "loss/crossentropy": 2.0313437558710574,
+      "loss/hidden": 3.387890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19764493741095066,
+      "step": 19700
+    },
+    {
+      "epoch": 0.49275,
+      "grad_norm": 32.0,
+      "grad_norm_var": 1.9884765625,
+      "learning_rate": 0.0001,
+      "loss": 7.4652,
+      "loss/crossentropy": 2.1517415940761566,
+      "loss/hidden": 3.342578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19597405511885882,
+      "step": 19710
+    },
+    {
+      "epoch": 0.493,
+      "grad_norm": 31.375,
+      "grad_norm_var": 2.220833333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5347,
+      "loss/crossentropy": 2.1508257031440734,
+      "loss/hidden": 3.4734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19206973258405924,
+      "step": 19720
+    },
+    {
+      "epoch": 0.49325,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.0479166666666666,
+      "learning_rate": 0.0001,
+      "loss": 7.5137,
+      "loss/crossentropy": 2.2038475602865217,
+      "loss/hidden": 3.3609375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19588778279721736,
+      "step": 19730
+    },
+    {
+      "epoch": 0.4935,
+      "grad_norm": 28.875,
+      "grad_norm_var": 3.449934895833333,
+      "learning_rate": 0.0001,
+      "loss": 7.4967,
+      "loss/crossentropy": 2.0373061880469323,
+      "loss/hidden": 3.508203125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20586073007434608,
+      "step": 19740
+    },
+    {
+      "epoch": 0.49375,
+      "grad_norm": 31.0,
+      "grad_norm_var": 4.093489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5691,
+      "loss/crossentropy": 2.1521503664553165,
+      "loss/hidden": 3.29921875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18537704246118664,
+      "step": 19750
+    },
+    {
+      "epoch": 0.494,
+      "grad_norm": 29.75,
+      "grad_norm_var": 1.8858723958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4954,
+      "loss/crossentropy": 2.0740778051316737,
+      "loss/hidden": 3.4859375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20432667913846672,
+      "step": 19760
+    },
+    {
+      "epoch": 0.49425,
+      "grad_norm": 34.25,
+      "grad_norm_var": 2.8061848958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5234,
+      "loss/crossentropy": 2.1025756657123567,
+      "loss/hidden": 3.31015625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1960667993873358,
+      "step": 19770
+    },
+    {
+      "epoch": 0.4945,
+      "grad_norm": 34.0,
+      "grad_norm_var": 3.1733723958333333,
+      "learning_rate": 0.0001,
+      "loss": 7.416,
+      "loss/crossentropy": 2.0789266645908357,
+      "loss/hidden": 3.363671875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.17625574823468923,
+      "step": 19780
+    },
+    {
+      "epoch": 0.49475,
+      "grad_norm": 32.0,
+      "grad_norm_var": 3.1150390625,
+      "learning_rate": 0.0001,
+      "loss": 7.5526,
+      "loss/crossentropy": 2.1197325706481935,
+      "loss/hidden": 3.397265625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19719154592603444,
+      "step": 19790
+    },
+    {
+      "epoch": 0.495,
+      "grad_norm": 28.75,
+      "grad_norm_var": 1.7999348958333334,
+      "learning_rate": 0.0001,
+      "loss": 7.4495,
+      "loss/crossentropy": 2.1562264621257783,
+      "loss/hidden": 3.436328125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2082188345491886,
+      "step": 19800
+    },
+    {
+      "epoch": 0.49525,
+      "grad_norm": 30.125,
+      "grad_norm_var": 2.2561848958333335,
+      "learning_rate": 0.0001,
+      "loss": 7.5052,
+      "loss/crossentropy": 2.1419773191213607,
+      "loss/hidden": 3.4546875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19062785096466542,
+      "step": 19810
+    },
+    {
+      "epoch": 0.4955,
+      "grad_norm": 28.125,
+      "grad_norm_var": 2.42265625,
+      "learning_rate": 0.0001,
+      "loss": 7.3991,
+      "loss/crossentropy": 2.144966857135296,
+      "loss/hidden": 3.323046875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19392486391589045,
+      "step": 19820
+    },
+    {
+      "epoch": 0.49575,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.3354166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.5459,
+      "loss/crossentropy": 2.317871165275574,
+      "loss/hidden": 3.37578125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20522685311734676,
+      "step": 19830
+    },
+    {
+      "epoch": 0.496,
+      "grad_norm": 35.25,
+      "grad_norm_var": 8.90625,
+      "learning_rate": 0.0001,
+      "loss": 7.458,
+      "loss/crossentropy": 2.106313969194889,
+      "loss/hidden": 3.3953125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18565387930721045,
+      "step": 19840
+    },
+    {
+      "epoch": 0.49625,
+      "grad_norm": 32.25,
+      "grad_norm_var": 10.2275390625,
+      "learning_rate": 0.0001,
+      "loss": 7.4956,
+      "loss/crossentropy": 2.1663008645176887,
+      "loss/hidden": 3.333984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19335117507725955,
+      "step": 19850
+    },
+    {
+      "epoch": 0.4965,
+      "grad_norm": 29.875,
+      "grad_norm_var": 0.6684895833333333,
+      "learning_rate": 0.0001,
+      "loss": 7.4191,
+      "loss/crossentropy": 2.2045235991477967,
+      "loss/hidden": 3.314453125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18433119654655455,
+      "step": 19860
+    },
+    {
+      "epoch": 0.49675,
+      "grad_norm": 31.75,
+      "grad_norm_var": 6.604166666666667,
+      "learning_rate": 0.0001,
+      "loss": 7.614,
+      "loss/crossentropy": 2.102213513851166,
+      "loss/hidden": 3.437890625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19189696963876485,
+      "step": 19870
+    },
+    {
+      "epoch": 0.497,
+      "grad_norm": 29.75,
+      "grad_norm_var": 3.9853515625,
+      "learning_rate": 0.0001,
+      "loss": 7.4145,
+      "loss/crossentropy": 2.1133465990424156,
+      "loss/hidden": 3.408984375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19469942338764668,
+      "step": 19880
+    },
+    {
+      "epoch": 0.49725,
+      "grad_norm": 31.75,
+      "grad_norm_var": 5.577018229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.3473,
+      "loss/crossentropy": 2.1288680538535116,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18879411723464729,
+      "step": 19890
+    },
+    {
+      "epoch": 0.4975,
+      "grad_norm": 31.875,
+      "grad_norm_var": 4.814518229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.5347,
+      "loss/crossentropy": 2.125280204415321,
+      "loss/hidden": 3.325390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1846629874780774,
+      "step": 19900
+    },
+    {
+      "epoch": 0.49775,
+      "grad_norm": 35.25,
+      "grad_norm_var": 20.151822916666667,
+      "learning_rate": 0.0001,
+      "loss": 7.4436,
+      "loss/crossentropy": 2.0771165072917936,
+      "loss/hidden": 3.419140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18763564769178628,
+      "step": 19910
+    },
+    {
+      "epoch": 0.498,
+      "grad_norm": 29.75,
+      "grad_norm_var": 21.658333333333335,
+      "learning_rate": 0.0001,
+      "loss": 7.4811,
+      "loss/crossentropy": 1.9571763649582863,
+      "loss/hidden": 3.377734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.1826960850507021,
+      "step": 19920
+    },
+    {
+      "epoch": 0.49825,
+      "grad_norm": 31.5,
+      "grad_norm_var": 5.590559895833334,
+      "learning_rate": 0.0001,
+      "loss": 7.3641,
+      "loss/crossentropy": 2.099938778579235,
+      "loss/hidden": 3.402734375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19888236112892627,
+      "step": 19930
+    },
+    {
+      "epoch": 0.4985,
+      "grad_norm": 30.375,
+      "grad_norm_var": 1.5239583333333333,
+      "learning_rate": 0.0001,
+      "loss": 7.5621,
+      "loss/crossentropy": 2.146530794352293,
+      "loss/hidden": 3.400390625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19427433758974075,
+      "step": 19940
+    },
+    {
+      "epoch": 0.49875,
+      "grad_norm": 30.875,
+      "grad_norm_var": 1.5035807291666667,
+      "learning_rate": 0.0001,
+      "loss": 7.483,
+      "loss/crossentropy": 2.141235402226448,
+      "loss/hidden": 3.455078125,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.21379205845296384,
+      "step": 19950
+    },
+    {
+      "epoch": 0.499,
+      "grad_norm": 30.0,
+      "grad_norm_var": 3.6572265625,
+      "learning_rate": 0.0001,
+      "loss": 7.4701,
+      "loss/crossentropy": 2.2140693843364714,
+      "loss/hidden": 3.35,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18884072303771973,
+      "step": 19960
+    },
+    {
+      "epoch": 0.49925,
+      "grad_norm": 34.25,
+      "grad_norm_var": 4.355143229166667,
+      "learning_rate": 0.0001,
+      "loss": 7.4807,
+      "loss/crossentropy": 2.106860537827015,
+      "loss/hidden": 3.409765625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.20601125992834568,
+      "step": 19970
+    },
+    {
+      "epoch": 0.4995,
+      "grad_norm": 30.25,
+      "grad_norm_var": 11.820572916666666,
+      "learning_rate": 0.0001,
+      "loss": 7.503,
+      "loss/crossentropy": 2.1892090171575544,
+      "loss/hidden": 3.34140625,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.18834524974226952,
+      "step": 19980
+    },
+    {
+      "epoch": 0.49975,
+      "grad_norm": 29.75,
+      "grad_norm_var": 4.793489583333334,
+      "learning_rate": 0.0001,
+      "loss": 7.5279,
+      "loss/crossentropy": 2.103451582789421,
+      "loss/hidden": 3.4421875,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.2049520380795002,
+      "step": 19990
+    },
+    {
+      "epoch": 0.5,
+      "grad_norm": 32.25,
+      "grad_norm_var": 28.2822265625,
+      "learning_rate": 0.0001,
+      "loss": 7.5062,
+      "loss/crossentropy": 1.972146065533161,
+      "loss/hidden": 3.387109375,
+      "loss/jsd": 0.0,
+      "loss/logits": 0.19209758015349507,
+      "step": 20000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 40000,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 9223372036854775807,
+  "save_steps": 5000,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 5.715020064017613e+19,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}