| [ |
| { |
| "loss": 0.6836, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.02, |
| "step": 1 |
| }, |
| { |
| "loss": 0.6947, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.04, |
| "step": 2 |
| }, |
| { |
| "loss": 0.7068, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.05, |
| "step": 3 |
| }, |
| { |
| "loss": 0.6936, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.07, |
| "step": 4 |
| }, |
| { |
| "loss": 0.6894, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.09, |
| "step": 5 |
| }, |
| { |
| "loss": 0.6947, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.11, |
| "step": 6 |
| }, |
| { |
| "loss": 0.7127, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.12, |
| "step": 7 |
| }, |
| { |
| "loss": 0.6564, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.14, |
| "step": 8 |
| }, |
| { |
| "loss": 0.7051, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.16, |
| "step": 9 |
| }, |
| { |
| "loss": 0.6898, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.18, |
| "step": 10 |
| }, |
| { |
| "loss": 0.6269, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.19, |
| "step": 11 |
| }, |
| { |
| "loss": 0.6373, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.21, |
| "step": 12 |
| }, |
| { |
| "loss": 0.6626, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.23, |
| "step": 13 |
| }, |
| { |
| "loss": 0.671, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.25, |
| "step": 14 |
| }, |
| { |
| "loss": 0.6446, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.26, |
| "step": 15 |
| }, |
| { |
| "loss": 0.6511, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.28, |
| "step": 16 |
| }, |
| { |
| "loss": 0.6861, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.3, |
| "step": 17 |
| }, |
| { |
| "loss": 0.6569, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.32, |
| "step": 18 |
| }, |
| { |
| "loss": 0.6328, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.33, |
| "step": 19 |
| }, |
| { |
| "loss": 0.6302, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.35, |
| "step": 20 |
| }, |
| { |
| "loss": 0.6311, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.37, |
| "step": 21 |
| }, |
| { |
| "loss": 0.6667, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.39, |
| "step": 22 |
| }, |
| { |
| "loss": 0.6476, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.4, |
| "step": 23 |
| }, |
| { |
| "loss": 0.6099, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.42, |
| "step": 24 |
| }, |
| { |
| "loss": 0.636, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.44, |
| "step": 25 |
| }, |
| { |
| "eval_code_loss": 0.6500930190086365, |
| "eval_code_score": -0.23084141314029694, |
| "eval_code_brier_score": 0.23084141314029694, |
| "eval_code_average_probability": 0.5324862003326416, |
| "eval_code_accuracy": 0.63, |
| "eval_code_probabilities": [ |
| 0.8981996178627014, |
| 0.9140064120292664, |
| 0.9086233377456665, |
| 0.4930489957332611, |
| 0.5311785340309143, |
| 0.49541160464286804, |
| 0.5163961052894592, |
| 0.5207482576370239, |
| 0.4978530704975128, |
| 0.4979601800441742, |
| 0.5097731351852417, |
| 0.5298001766204834, |
| 0.4631284475326538, |
| 0.4101221561431885, |
| 0.3820548951625824, |
| 0.5316497683525085, |
| 0.4804396629333496, |
| 0.5146058797836304, |
| 0.5150214433670044, |
| 0.4998961091041565, |
| 0.5088847875595093, |
| 0.5062065720558167, |
| 0.49155697226524353, |
| 0.5065352916717529, |
| 0.3692309260368347, |
| 0.3464958965778351, |
| 0.345247745513916, |
| 0.5082260370254517, |
| 0.4985589385032654, |
| 0.4908699691295624, |
| 0.5014219284057617, |
| 0.5078413486480713, |
| 0.5118864178657532, |
| 0.9250859618186951, |
| 0.9192001819610596, |
| 0.9183722734451294, |
| 0.5567901134490967, |
| 0.5061033368110657, |
| 0.5202013850212097, |
| 0.4856383204460144, |
| 0.46390581130981445, |
| 0.4597606658935547, |
| 0.5073948502540588, |
| 0.49284979701042175, |
| 0.49324461817741394, |
| 0.4991315007209778, |
| 0.5046825408935547, |
| 0.5030663013458252, |
| 0.4927866756916046, |
| 0.5034404993057251, |
| 0.6324691772460938, |
| 0.527839183807373, |
| 0.6027267575263977, |
| 0.5166205167770386, |
| 0.5005240440368652, |
| 0.5037304162979126, |
| 0.5007357001304626, |
| 0.5299914479255676, |
| 0.49253302812576294, |
| 0.5389838814735413, |
| 0.7237929105758667, |
| 0.722010612487793, |
| 0.7248772978782654, |
| 0.5261573791503906, |
| 0.5009329319000244, |
| 0.5066245198249817, |
| 0.5175695419311523, |
| 0.48236575722694397, |
| 0.4813913404941559, |
| 0.4951794743537903, |
| 0.5201015472412109, |
| 0.5013536810874939, |
| 0.5012617111206055, |
| 0.5066967606544495, |
| 0.5075401067733765, |
| 0.5807880759239197, |
| 0.5756419897079468, |
| 0.5374066829681396, |
| 0.5401325821876526, |
| 0.4964595437049866, |
| 0.48420250415802, |
| 0.5349816083908081, |
| 0.4993498623371124, |
| 0.5083227157592773, |
| 0.4809667766094208, |
| 0.5351281762123108, |
| 0.45967379212379456, |
| 0.5043829679489136, |
| 0.4973975121974945, |
| 0.5018096566200256, |
| 0.4930133819580078, |
| 0.5008613467216492, |
| 0.5038902163505554, |
| 0.5041508078575134, |
| 0.4986605942249298, |
| 0.5054576992988586, |
| 0.4900255799293518, |
| 0.4928497076034546, |
| 0.5077279806137085, |
| 0.49679625034332275 |
| ], |
| "eval_code_runtime": 36.4656, |
| "eval_code_samples_per_second": 2.742, |
| "eval_code_steps_per_second": 0.055, |
| "epoch": 0.44, |
| "step": 25 |
| }, |
| { |
| "eval_counterfactual_python_loss": 0.6674560308456421, |
| "eval_counterfactual_python_score": -0.2368319183588028, |
| "eval_counterfactual_python_brier_score": 0.2368319183588028, |
| "eval_counterfactual_python_average_probability": 0.5182108879089355, |
| "eval_counterfactual_python_accuracy": 0.6, |
| "eval_counterfactual_python_probabilities": [ |
| 0.506840705871582, |
| 0.48089414834976196, |
| 0.497524231672287, |
| 0.5044610500335693, |
| 0.4899257719516754, |
| 0.5029208660125732, |
| 0.5528867840766907, |
| 0.6020012497901917, |
| 0.5903577208518982, |
| 0.5012052059173584, |
| 0.5010500550270081, |
| 0.5014556050300598, |
| 0.5129737257957458, |
| 0.5166401863098145, |
| 0.504987359046936, |
| 0.48754236102104187, |
| 0.4921533763408661, |
| 0.5110666751861572, |
| 0.5212812423706055, |
| 0.505253791809082, |
| 0.5044047236442566, |
| 0.5446178913116455, |
| 0.5504598021507263, |
| 0.5181146860122681, |
| 0.4997810125350952, |
| 0.5025537610054016, |
| 0.49983397126197815, |
| 0.5592316389083862, |
| 0.5561607480049133, |
| 0.49701419472694397, |
| 0.5067521929740906, |
| 0.5109471082687378, |
| 0.49662530422210693, |
| 0.4926052391529083, |
| 0.5500054955482483, |
| 0.5147995352745056, |
| 0.5712809562683105, |
| 0.5422273874282837, |
| 0.5272724032402039, |
| 0.5182323455810547, |
| 0.504339337348938, |
| 0.524910569190979, |
| 0.5017884373664856, |
| 0.49858200550079346, |
| 0.5003425478935242, |
| 0.49639299511909485, |
| 0.5009803175926208, |
| 0.506009042263031, |
| 0.5095840096473694, |
| 0.49511101841926575, |
| 0.49724477529525757, |
| 0.5003291964530945, |
| 0.49959462881088257, |
| 0.5014533996582031, |
| 0.5121566653251648, |
| 0.4868036210536957, |
| 0.4927564561367035, |
| 0.49987566471099854, |
| 0.48913922905921936, |
| 0.4933817982673645, |
| 0.5025197267532349, |
| 0.4877781271934509, |
| 0.4988517463207245, |
| 0.48360976576805115, |
| 0.47294488549232483, |
| 0.44178009033203125, |
| 0.5168694853782654, |
| 0.5422465801239014, |
| 0.5201556086540222, |
| 0.5059149861335754, |
| 0.49637845158576965, |
| 0.518649697303772, |
| 0.4840700328350067, |
| 0.4878598749637604, |
| 0.4397459328174591, |
| 0.5004118084907532, |
| 0.4982900023460388, |
| 0.49845632910728455, |
| 0.4866858422756195, |
| 0.5425868630409241, |
| 0.5578758120536804, |
| 0.5089701414108276, |
| 0.4936344027519226, |
| 0.49497488141059875, |
| 0.4820954203605652, |
| 0.4383483827114105, |
| 0.4935499131679535, |
| 0.8830961585044861, |
| 0.8793171048164368, |
| 0.8719192743301392, |
| 0.44867223501205444, |
| 0.4727223813533783, |
| 0.5310845971107483, |
| 0.4810177981853485, |
| 0.5083786249160767, |
| 0.5173466205596924, |
| 0.5266315340995789, |
| 0.513174831867218, |
| 0.5350879430770874, |
| 0.5002694725990295 |
| ], |
| "eval_counterfactual_python_runtime": 42.2718, |
| "eval_counterfactual_python_samples_per_second": 2.366, |
| "eval_counterfactual_python_steps_per_second": 0.047, |
| "epoch": 0.44, |
| "step": 25 |
| }, |
| { |
| "loss": 0.6523, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.46, |
| "step": 26 |
| }, |
| { |
| "loss": 0.6145, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.47, |
| "step": 27 |
| }, |
| { |
| "loss": 0.6282, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.49, |
| "step": 28 |
| }, |
| { |
| "loss": 0.5889, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.51, |
| "step": 29 |
| }, |
| { |
| "loss": 0.6173, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.53, |
| "step": 30 |
| }, |
| { |
| "loss": 0.6474, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.54, |
| "step": 31 |
| }, |
| { |
| "loss": 0.5995, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.56, |
| "step": 32 |
| }, |
| { |
| "loss": 0.5861, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.58, |
| "step": 33 |
| }, |
| { |
| "loss": 0.5736, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.6, |
| "step": 34 |
| }, |
| { |
| "loss": 0.6119, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.61, |
| "step": 35 |
| }, |
| { |
| "loss": 0.6247, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.63, |
| "step": 36 |
| }, |
| { |
| "loss": 0.6276, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.65, |
| "step": 37 |
| }, |
| { |
| "loss": 0.6098, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.67, |
| "step": 38 |
| }, |
| { |
| "loss": 0.6301, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.68, |
| "step": 39 |
| }, |
| { |
| "loss": 0.6021, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.7, |
| "step": 40 |
| }, |
| { |
| "loss": 0.555, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.72, |
| "step": 41 |
| }, |
| { |
| "loss": 0.5695, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.74, |
| "step": 42 |
| }, |
| { |
| "loss": 0.5669, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.75, |
| "step": 43 |
| }, |
| { |
| "loss": 0.5473, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.77, |
| "step": 44 |
| }, |
| { |
| "loss": 0.5369, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.79, |
| "step": 45 |
| }, |
| { |
| "loss": 0.5351, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.81, |
| "step": 46 |
| }, |
| { |
| "loss": 0.4989, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.82, |
| "step": 47 |
| }, |
| { |
| "loss": 0.4687, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.84, |
| "step": 48 |
| }, |
| { |
| "loss": 0.4431, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.86, |
| "step": 49 |
| }, |
| { |
| "loss": 0.475, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.88, |
| "step": 50 |
| }, |
| { |
| "eval_code_loss": 0.6905087232589722, |
| "eval_code_score": -0.21665707230567932, |
| "eval_code_brier_score": 0.21665707230567932, |
| "eval_code_average_probability": 0.576951265335083, |
| "eval_code_accuracy": 0.68, |
| "eval_code_probabilities": [ |
| 0.8747886419296265, |
| 0.897834300994873, |
| 0.9301072955131531, |
| 0.5637628436088562, |
| 0.6103933453559875, |
| 0.5329621434211731, |
| 0.661848783493042, |
| 0.8006942868232727, |
| 0.5047258138656616, |
| 0.5474675297737122, |
| 0.5803297162055969, |
| 0.6606214642524719, |
| 0.6511579155921936, |
| 0.5316810011863708, |
| 0.5710970163345337, |
| 0.7186195850372314, |
| 0.40853649377822876, |
| 0.6621276140213013, |
| 0.7090564370155334, |
| 0.686518132686615, |
| 0.6983659863471985, |
| 0.3943677544593811, |
| 0.7656291127204895, |
| 0.7813704609870911, |
| 0.01187042985111475, |
| 0.008863707073032856, |
| 0.007746896706521511, |
| 0.5090981721878052, |
| 0.5062975287437439, |
| 0.45505571365356445, |
| 0.4951728880405426, |
| 0.5581167340278625, |
| 0.5970812439918518, |
| 0.9008167386054993, |
| 0.9059542417526245, |
| 0.9084323644638062, |
| 0.814932644367218, |
| 0.5395833849906921, |
| 0.6921597719192505, |
| 0.45764750242233276, |
| 0.3827315866947174, |
| 0.5150556564331055, |
| 0.5219369530677795, |
| 0.4390372037887573, |
| 0.4954327344894409, |
| 0.47707119584083557, |
| 0.5074189305305481, |
| 0.5035202503204346, |
| 0.5605807304382324, |
| 0.5175653696060181, |
| 0.9902499318122864, |
| 0.5175576210021973, |
| 0.8499017953872681, |
| 0.5307424068450928, |
| 0.5028649568557739, |
| 0.515737771987915, |
| 0.4969106614589691, |
| 0.2922719717025757, |
| 0.44801953434944153, |
| 0.32066962122917175, |
| 0.9447461366653442, |
| 0.9428120255470276, |
| 0.9458075165748596, |
| 0.7382722496986389, |
| 0.5110031962394714, |
| 0.5486981868743896, |
| 0.4352004826068878, |
| 0.41556352376937866, |
| 0.4588325619697571, |
| 0.5106350779533386, |
| 0.5079880952835083, |
| 0.5065162181854248, |
| 0.5341932773590088, |
| 0.5115707516670227, |
| 0.49947670102119446, |
| 0.9284617304801941, |
| 0.510497510433197, |
| 0.6312513947486877, |
| 0.9168699979782104, |
| 0.9026933312416077, |
| 0.8503899574279785, |
| 0.5394543409347534, |
| 0.4984402060508728, |
| 0.5079206228256226, |
| 0.29503393173217773, |
| 0.7989310026168823, |
| 0.3873499035835266, |
| 0.552551805973053, |
| 0.49572017788887024, |
| 0.5294312834739685, |
| 0.44464293122291565, |
| 0.5211971998214722, |
| 0.46848106384277344, |
| 0.5291138887405396, |
| 0.4984701871871948, |
| 0.4968411326408386, |
| 0.46553924679756165, |
| 0.46003156900405884, |
| 0.493184357881546, |
| 0.463138610124588 |
| ], |
| "eval_code_runtime": 36.5772, |
| "eval_code_samples_per_second": 2.734, |
| "eval_code_steps_per_second": 0.055, |
| "epoch": 0.88, |
| "step": 50 |
| }, |
| { |
| "eval_counterfactual_python_loss": 0.6410097479820251, |
| "eval_counterfactual_python_score": -0.222365602850914, |
| "eval_counterfactual_python_brier_score": 0.222365602850914, |
| "eval_counterfactual_python_average_probability": 0.5518541932106018, |
| "eval_counterfactual_python_accuracy": 0.66, |
| "eval_counterfactual_python_probabilities": [ |
| 0.5386440753936768, |
| 0.49819016456604004, |
| 0.4844203591346741, |
| 0.5327919125556946, |
| 0.4123155176639557, |
| 0.5146786570549011, |
| 0.44351926445961, |
| 0.4285781681537628, |
| 0.41586360335350037, |
| 0.5060333609580994, |
| 0.5129002928733826, |
| 0.5073186755180359, |
| 0.5133093595504761, |
| 0.30924591422080994, |
| 0.5565316081047058, |
| 0.41158172488212585, |
| 0.42925503849983215, |
| 0.45506441593170166, |
| 0.5124459862709045, |
| 0.5084977746009827, |
| 0.5260626673698425, |
| 0.9784466624259949, |
| 0.977891206741333, |
| 0.6758217811584473, |
| 0.4996006488800049, |
| 0.502190351486206, |
| 0.49963271617889404, |
| 0.24913643300533295, |
| 0.24497272074222565, |
| 0.15005168318748474, |
| 0.5284581780433655, |
| 0.4637279510498047, |
| 0.46632638573646545, |
| 0.49904829263687134, |
| 0.5982093811035156, |
| 0.5048952102661133, |
| 0.7960643768310547, |
| 0.6252861022949219, |
| 0.6371491551399231, |
| 0.6120105385780334, |
| 0.44951215386390686, |
| 0.30023863911628723, |
| 0.5029293298721313, |
| 0.49918922781944275, |
| 0.5002826452255249, |
| 0.48498842120170593, |
| 0.49545082449913025, |
| 0.5000977516174316, |
| 0.5497520565986633, |
| 0.4658561646938324, |
| 0.47568437457084656, |
| 0.5000590682029724, |
| 0.4994635581970215, |
| 0.501465380191803, |
| 0.5806223154067993, |
| 0.5407394766807556, |
| 0.548596203327179, |
| 0.49953749775886536, |
| 0.5177533626556396, |
| 0.5937445163726807, |
| 0.51665860414505, |
| 0.5505573749542236, |
| 0.5108556151390076, |
| 0.48964056372642517, |
| 0.6833782196044922, |
| 0.7741197943687439, |
| 0.5394606590270996, |
| 0.6131158471107483, |
| 0.5644785165786743, |
| 0.601515531539917, |
| 0.4582086503505707, |
| 0.5712682604789734, |
| 0.6359525918960571, |
| 0.740151047706604, |
| 0.453473836183548, |
| 0.5005698204040527, |
| 0.49849438667297363, |
| 0.4981241226196289, |
| 0.5342211127281189, |
| 0.7014966607093811, |
| 0.8450021743774414, |
| 0.519537091255188, |
| 0.5990186333656311, |
| 0.7578335404396057, |
| 0.5165602564811707, |
| 0.46256765723228455, |
| 0.5105525851249695, |
| 0.9811468124389648, |
| 0.9849089980125427, |
| 0.9816316366195679, |
| 0.6144223809242249, |
| 0.4611106216907501, |
| 0.8631046414375305, |
| 0.5419878363609314, |
| 0.519582211971283, |
| 0.6103906035423279, |
| 0.667199969291687, |
| 0.5551198720932007, |
| 0.669242262840271, |
| 0.5566273331642151 |
| ], |
| "eval_counterfactual_python_runtime": 42.2853, |
| "eval_counterfactual_python_samples_per_second": 2.365, |
| "eval_counterfactual_python_steps_per_second": 0.047, |
| "epoch": 0.88, |
| "step": 50 |
| }, |
| { |
| "loss": 0.4995, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.89, |
| "step": 51 |
| }, |
| { |
| "loss": 0.5482, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.91, |
| "step": 52 |
| }, |
| { |
| "loss": 0.6127, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.93, |
| "step": 53 |
| }, |
| { |
| "loss": 0.5, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.95, |
| "step": 54 |
| }, |
| { |
| "loss": 0.5382, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.96, |
| "step": 55 |
| }, |
| { |
| "loss": 0.5227, |
| "learning_rate": 7.2e-05, |
| "epoch": 0.98, |
| "step": 56 |
| }, |
| { |
| "loss": 0.6282, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.0, |
| "step": 57 |
| }, |
| { |
| "loss": 0.4611, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.02, |
| "step": 58 |
| }, |
| { |
| "loss": 0.4296, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.04, |
| "step": 59 |
| }, |
| { |
| "loss": 0.3265, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.05, |
| "step": 60 |
| }, |
| { |
| "loss": 0.3327, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.07, |
| "step": 61 |
| }, |
| { |
| "loss": 0.5471, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.09, |
| "step": 62 |
| }, |
| { |
| "loss": 0.3814, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.11, |
| "step": 63 |
| }, |
| { |
| "loss": 0.3941, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.12, |
| "step": 64 |
| }, |
| { |
| "loss": 0.3218, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.14, |
| "step": 65 |
| }, |
| { |
| "loss": 0.3258, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.16, |
| "step": 66 |
| }, |
| { |
| "loss": 0.5251, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.18, |
| "step": 67 |
| }, |
| { |
| "loss": 0.469, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.19, |
| "step": 68 |
| }, |
| { |
| "loss": 0.3852, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.21, |
| "step": 69 |
| }, |
| { |
| "loss": 0.3409, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.23, |
| "step": 70 |
| }, |
| { |
| "loss": 0.4473, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.25, |
| "step": 71 |
| }, |
| { |
| "loss": 0.2945, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.26, |
| "step": 72 |
| }, |
| { |
| "loss": 0.3608, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.28, |
| "step": 73 |
| }, |
| { |
| "loss": 0.4042, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.3, |
| "step": 74 |
| }, |
| { |
| "loss": 0.3802, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.32, |
| "step": 75 |
| }, |
| { |
| "eval_code_loss": 0.4694095849990845, |
| "eval_code_score": -0.16457700729370117, |
| "eval_code_brier_score": 0.16457700729370117, |
| "eval_code_average_probability": 0.6747273802757263, |
| "eval_code_accuracy": 0.77, |
| "eval_code_probabilities": [ |
| 0.9998944997787476, |
| 0.9999444484710693, |
| 0.9999792575836182, |
| 0.7207562327384949, |
| 0.9202108383178711, |
| 0.5724186301231384, |
| 0.8467735648155212, |
| 0.8126452565193176, |
| 0.44478896260261536, |
| 0.7643183469772339, |
| 0.7819174528121948, |
| 0.9309266805648804, |
| 0.9411635994911194, |
| 0.8424233198165894, |
| 0.9049078226089478, |
| 0.952031672000885, |
| 0.3711952567100525, |
| 0.9328634738922119, |
| 0.9677587151527405, |
| 0.9494214057922363, |
| 0.9498708844184875, |
| 0.30629244446754456, |
| 0.8959491848945618, |
| 0.9032987952232361, |
| 0.9084609150886536, |
| 0.8893307447433472, |
| 0.8730643391609192, |
| 0.5209372043609619, |
| 0.5036833882331848, |
| 0.4168521463871002, |
| 0.5004851818084717, |
| 0.6079723834991455, |
| 0.5459354519844055, |
| 0.9346232414245605, |
| 0.9303449988365173, |
| 0.9605541825294495, |
| 0.9573792815208435, |
| 0.6248452067375183, |
| 0.9251506328582764, |
| 0.4677916467189789, |
| 0.37387004494667053, |
| 0.42620334029197693, |
| 0.600398063659668, |
| 0.48431196808815, |
| 0.7895669341087341, |
| 0.4504677951335907, |
| 0.5762742757797241, |
| 0.6442916989326477, |
| 0.8840709328651428, |
| 0.7550206780433655, |
| 0.9939717650413513, |
| 0.6135303378105164, |
| 0.6339468359947205, |
| 0.5852696895599365, |
| 0.602887749671936, |
| 0.5735942125320435, |
| 0.4814770221710205, |
| 0.9196935892105103, |
| 0.1881166696548462, |
| 0.8902093172073364, |
| 0.977299153804779, |
| 0.9783608913421631, |
| 0.9801336526870728, |
| 0.9884994626045227, |
| 0.5481224656105042, |
| 0.5878416299819946, |
| 0.2562265396118164, |
| 0.10622240602970123, |
| 0.1097337156534195, |
| 0.5593991875648499, |
| 0.5126798152923584, |
| 0.4901875853538513, |
| 0.539310097694397, |
| 0.5853370428085327, |
| 0.5666263103485107, |
| 0.9994970560073853, |
| 0.7907701730728149, |
| 0.9174056649208069, |
| 0.9986317753791809, |
| 0.9991476535797119, |
| 0.9870390295982361, |
| 0.6845882534980774, |
| 0.5273832082748413, |
| 0.5825375914573669, |
| 0.05030998960137367, |
| 0.7106683254241943, |
| 0.376408189535141, |
| 0.6357782483100891, |
| 0.5115633606910706, |
| 0.5071816444396973, |
| 0.39829522371292114, |
| 0.535811722278595, |
| 0.3911305069923401, |
| 0.6767864227294922, |
| 0.50504070520401, |
| 0.43804654479026794, |
| 0.5036672949790955, |
| 0.4026598036289215, |
| 0.35143136978149414, |
| 0.4586479961872101 |
| ], |
| "eval_code_runtime": 36.5352, |
| "eval_code_samples_per_second": 2.737, |
| "eval_code_steps_per_second": 0.055, |
| "epoch": 1.32, |
| "step": 75 |
| }, |
| { |
| "eval_counterfactual_python_loss": 0.6137931942939758, |
| "eval_counterfactual_python_score": -0.20746655762195587, |
| "eval_counterfactual_python_brier_score": 0.20746655762195587, |
| "eval_counterfactual_python_average_probability": 0.5891837477684021, |
| "eval_counterfactual_python_accuracy": 0.63, |
| "eval_counterfactual_python_probabilities": [ |
| 0.5506163239479065, |
| 0.42313048243522644, |
| 0.4198402762413025, |
| 0.5754298567771912, |
| 0.38630211353302, |
| 0.47240203619003296, |
| 0.7205502986907959, |
| 0.6168351769447327, |
| 0.5663421750068665, |
| 0.45957374572753906, |
| 0.5441972017288208, |
| 0.4583946764469147, |
| 0.5182128548622131, |
| 0.13082706928253174, |
| 0.6158074140548706, |
| 0.30370235443115234, |
| 0.30528637766838074, |
| 0.36947813630104065, |
| 0.5232859253883362, |
| 0.5014183521270752, |
| 0.49940726161003113, |
| 0.6633936166763306, |
| 0.8305829763412476, |
| 0.8158023953437805, |
| 0.4996267259120941, |
| 0.5016799569129944, |
| 0.49929139018058777, |
| 0.37765786051750183, |
| 0.3638369143009186, |
| 0.2675539255142212, |
| 0.39653363823890686, |
| 0.5094757676124573, |
| 0.4743023216724396, |
| 0.6519900560379028, |
| 0.726960301399231, |
| 0.4309100806713104, |
| 0.9850045442581177, |
| 0.954617977142334, |
| 0.9593468904495239, |
| 0.6046375036239624, |
| 0.4302014708518982, |
| 0.3352057933807373, |
| 0.5038560628890991, |
| 0.49972790479660034, |
| 0.5001466870307922, |
| 0.42995354533195496, |
| 0.47737735509872437, |
| 0.45605820417404175, |
| 0.6769935488700867, |
| 0.4527358412742615, |
| 0.4564270079135895, |
| 0.49967876076698303, |
| 0.49923568964004517, |
| 0.5014398694038391, |
| 0.61121666431427, |
| 0.5898076295852661, |
| 0.6469179391860962, |
| 0.5191650390625, |
| 0.43958330154418945, |
| 0.513123631477356, |
| 0.6665530800819397, |
| 0.7978847622871399, |
| 0.5442570447921753, |
| 0.36048850417137146, |
| 0.5295901894569397, |
| 0.9162126779556274, |
| 0.6021609902381897, |
| 0.7771204710006714, |
| 0.7251576781272888, |
| 0.6523131728172302, |
| 0.46528783440589905, |
| 0.6797440648078918, |
| 0.5512087941169739, |
| 0.26883381605148315, |
| 0.2967521548271179, |
| 0.5009620189666748, |
| 0.4986320734024048, |
| 0.49721115827560425, |
| 0.6624665856361389, |
| 0.9735934734344482, |
| 0.9868214726448059, |
| 0.5706613063812256, |
| 0.7753136157989502, |
| 0.9256560206413269, |
| 0.5435609221458435, |
| 0.5188571214675903, |
| 0.554946780204773, |
| 0.999994158744812, |
| 0.9999972581863403, |
| 0.9999933242797852, |
| 0.8329089283943176, |
| 0.4099949896335602, |
| 0.9683297872543335, |
| 0.8267971873283386, |
| 0.504895031452179, |
| 0.6182412505149841, |
| 0.9561484456062317, |
| 0.8357421159744263, |
| 0.9478693604469299, |
| 0.6621176600456238 |
| ], |
| "eval_counterfactual_python_runtime": 42.2604, |
| "eval_counterfactual_python_samples_per_second": 2.366, |
| "eval_counterfactual_python_steps_per_second": 0.047, |
| "epoch": 1.32, |
| "step": 75 |
| }, |
| { |
| "loss": 0.3477, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.33, |
| "step": 76 |
| }, |
| { |
| "loss": 0.3441, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.35, |
| "step": 77 |
| }, |
| { |
| "loss": 0.2051, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.37, |
| "step": 78 |
| }, |
| { |
| "loss": 0.3386, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.39, |
| "step": 79 |
| }, |
| { |
| "loss": 0.3102, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.4, |
| "step": 80 |
| }, |
| { |
| "loss": 0.2749, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.42, |
| "step": 81 |
| }, |
| { |
| "loss": 0.393, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.44, |
| "step": 82 |
| }, |
| { |
| "loss": 0.3211, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.46, |
| "step": 83 |
| }, |
| { |
| "loss": 0.3656, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.47, |
| "step": 84 |
| }, |
| { |
| "loss": 0.2527, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.49, |
| "step": 85 |
| }, |
| { |
| "loss": 0.3756, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.51, |
| "step": 86 |
| }, |
| { |
| "loss": 0.5116, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.53, |
| "step": 87 |
| }, |
| { |
| "loss": 0.3191, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.54, |
| "step": 88 |
| }, |
| { |
| "loss": 0.2474, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.56, |
| "step": 89 |
| }, |
| { |
| "loss": 0.3117, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.58, |
| "step": 90 |
| }, |
| { |
| "loss": 0.2992, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.6, |
| "step": 91 |
| }, |
| { |
| "loss": 0.158, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.61, |
| "step": 92 |
| }, |
| { |
| "loss": 0.2684, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.63, |
| "step": 93 |
| }, |
| { |
| "loss": 0.5338, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.65, |
| "step": 94 |
| }, |
| { |
| "loss": 0.3467, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.67, |
| "step": 95 |
| }, |
| { |
| "loss": 0.2647, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.68, |
| "step": 96 |
| }, |
| { |
| "loss": 0.2894, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.7, |
| "step": 97 |
| }, |
| { |
| "loss": 0.2537, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.72, |
| "step": 98 |
| }, |
| { |
| "loss": 0.4454, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.74, |
| "step": 99 |
| }, |
| { |
| "loss": 0.2264, |
| "learning_rate": 7.2e-05, |
| "epoch": 1.75, |
| "step": 100 |
| }, |
| { |
| "eval_code_loss": 0.562386691570282, |
| "eval_code_score": -0.1894664168357849, |
| "eval_code_brier_score": 0.1894664168357849, |
| "eval_code_average_probability": 0.6813782453536987, |
| "eval_code_accuracy": 0.75, |
| "eval_code_probabilities": [ |
| 0.99964439868927, |
| 0.9998418092727661, |
| 0.9999876022338867, |
| 0.8987520337104797, |
| 0.9715626835823059, |
| 0.8736424446105957, |
| 0.992048442363739, |
| 0.9987469911575317, |
| 0.5219283103942871, |
| 0.9212198853492737, |
| 0.9132972955703735, |
| 0.975989818572998, |
| 0.9921802282333374, |
| 0.9395624399185181, |
| 0.9364026784896851, |
| 0.9374613165855408, |
| 0.27392876148223877, |
| 0.916027307510376, |
| 0.9856477975845337, |
| 0.9577564001083374, |
| 0.9659378528594971, |
| 0.44796139001846313, |
| 0.9555050134658813, |
| 0.9627187848091125, |
| 0.49647897481918335, |
| 0.2568376362323761, |
| 0.1927042156457901, |
| 0.561137318611145, |
| 0.5502861738204956, |
| 0.3930111825466156, |
| 0.46675917506217957, |
| 0.7520542144775391, |
| 0.3132498860359192, |
| 0.1046283170580864, |
| 0.12346971035003662, |
| 0.16289445757865906, |
| 0.998822033405304, |
| 0.8946499228477478, |
| 0.9992140531539917, |
| 0.5172696709632874, |
| 0.1285436749458313, |
| 0.5089917182922363, |
| 0.8845463991165161, |
| 0.7489760518074036, |
| 0.8961533308029175, |
| 0.41095906496047974, |
| 0.733329713344574, |
| 0.6693552136421204, |
| 0.9867847561836243, |
| 0.8965162038803101, |
| 0.9999988079071045, |
| 0.9335614442825317, |
| 0.9917327165603638, |
| 0.7795218825340271, |
| 0.673981785774231, |
| 0.5716884136199951, |
| 0.44703209400177, |
| 0.990574300289154, |
| 0.035506464540958405, |
| 0.976534366607666, |
| 0.9999401569366455, |
| 0.9999487400054932, |
| 0.9999538660049438, |
| 0.9866085648536682, |
| 0.610414445400238, |
| 0.4811190962791443, |
| 0.10195612162351608, |
| 0.022535108029842377, |
| 0.03303113952279091, |
| 0.6767235994338989, |
| 0.513542890548706, |
| 0.5040165781974792, |
| 0.5857304930686951, |
| 0.5917763113975525, |
| 0.6388770341873169, |
| 0.999974250793457, |
| 0.6209684610366821, |
| 0.9876555800437927, |
| 0.9999986886978149, |
| 0.9999998807907104, |
| 0.9999920129776001, |
| 0.8514988422393799, |
| 0.5619485378265381, |
| 0.6238705515861511, |
| 0.08037617802619934, |
| 0.9763862490653992, |
| 0.21729014813899994, |
| 0.7570930123329163, |
| 0.6862062215805054, |
| 0.6152291893959045, |
| 0.28896594047546387, |
| 0.5001997947692871, |
| 0.3280743360519409, |
| 0.6856067180633545, |
| 0.5078412890434265, |
| 0.4755324721336365, |
| 0.6357707977294922, |
| 0.4387938678264618, |
| 0.530180811882019, |
| 0.6366823315620422 |
| ], |
| "eval_code_runtime": 36.6096, |
| "eval_code_samples_per_second": 2.732, |
| "eval_code_steps_per_second": 0.055, |
| "epoch": 1.75, |
| "step": 100 |
| }, |
| { |
| "eval_counterfactual_python_loss": 0.5543044209480286, |
| "eval_counterfactual_python_score": -0.17998556792736053, |
| "eval_counterfactual_python_brier_score": 0.17998556792736053, |
| "eval_counterfactual_python_average_probability": 0.641478955745697, |
| "eval_counterfactual_python_accuracy": 0.68, |
| "eval_counterfactual_python_probabilities": [ |
| 0.5931702852249146, |
| 0.36623018980026245, |
| 0.4083799719810486, |
| 0.7441582679748535, |
| 0.3787698447704315, |
| 0.4585961401462555, |
| 0.9636162519454956, |
| 0.8045032620429993, |
| 0.7646597027778625, |
| 0.4695722460746765, |
| 0.6623252630233765, |
| 0.5570592880249023, |
| 0.4867101013660431, |
| 0.09203716367483139, |
| 0.659835934638977, |
| 0.38409557938575745, |
| 0.3676278293132782, |
| 0.4404756426811218, |
| 0.5268372893333435, |
| 0.5405157804489136, |
| 0.545403003692627, |
| 0.7553650140762329, |
| 0.917907178401947, |
| 0.9101668000221252, |
| 0.4998391270637512, |
| 0.500482439994812, |
| 0.4992232620716095, |
| 0.36769258975982666, |
| 0.3600624203681946, |
| 0.10161764174699783, |
| 0.27137765288352966, |
| 0.5764327645301819, |
| 0.5460357069969177, |
| 0.6896645426750183, |
| 0.5904480814933777, |
| 0.43029138445854187, |
| 0.9749554991722107, |
| 0.9570443034172058, |
| 0.9645709991455078, |
| 0.8437580466270447, |
| 0.3237834870815277, |
| 0.4067584276199341, |
| 0.505946695804596, |
| 0.5007390975952148, |
| 0.5010091066360474, |
| 0.40068086981773376, |
| 0.46205294132232666, |
| 0.43641653656959534, |
| 0.8441718220710754, |
| 0.4741505980491638, |
| 0.36617419123649597, |
| 0.49912282824516296, |
| 0.49875274300575256, |
| 0.5011247396469116, |
| 0.6660105586051941, |
| 0.6525660157203674, |
| 0.7143298387527466, |
| 0.5283794403076172, |
| 0.5694890022277832, |
| 0.7942121624946594, |
| 0.9201071858406067, |
| 0.9259344935417175, |
| 0.5873693227767944, |
| 0.24995583295822144, |
| 0.7780645489692688, |
| 0.9872015714645386, |
| 0.7106945514678955, |
| 0.9213557839393616, |
| 0.8370577692985535, |
| 0.816162645816803, |
| 0.385659784078598, |
| 0.815650224685669, |
| 0.9024463295936584, |
| 0.40921616554260254, |
| 0.7233693599700928, |
| 0.5013810992240906, |
| 0.4987756907939911, |
| 0.4964328408241272, |
| 0.7962855100631714, |
| 0.9599968791007996, |
| 0.967514157295227, |
| 0.5222499966621399, |
| 0.8373121619224548, |
| 0.926533579826355, |
| 0.6371393203735352, |
| 0.5046852827072144, |
| 0.6500534415245056, |
| 0.9999998807907104, |
| 1.0, |
| 0.9999998807907104, |
| 0.8575750589370728, |
| 0.44756925106048584, |
| 0.9889571070671082, |
| 0.9346925616264343, |
| 0.5233546495437622, |
| 0.7729074954986572, |
| 0.9984490871429443, |
| 0.9480088353157043, |
| 0.9992019534111023, |
| 0.7911841869354248 |
| ], |
| "eval_counterfactual_python_runtime": 42.2823, |
| "eval_counterfactual_python_samples_per_second": 2.365, |
| "eval_counterfactual_python_steps_per_second": 0.047, |
| "epoch": 1.75, |
| "step": 100 |
| }, |
| { |
| "train_runtime": 3769.2128, |
| "train_samples_per_second": 0.849, |
| "train_steps_per_second": 0.027, |
| "total_flos": 0.0, |
| "train_loss": 0.5000894145667553, |
| "epoch": 1.75, |
| "step": 100 |
| } |
| ] |
