File size: 27,960 Bytes
9defb95 |
1 |
{"GEM/web_nlg_en": {"0": {"PALM_prompt": {"bleu": 0.4712440224858371, "bleu_stderr": 0.041795896662364315, "rouge1_fmeasure": 0.11420475482093911, "rouge1_fmeasure_stderr": 0.0022183305831393195, "rouge1_precision": 0.08494278433659254, "rouge1_precision_stderr": 0.002682014199464667, "rouge1_recall": 0.3010315100774798, "rouge1_recall_stderr": 0.005434382926718196, "rouge2_fmeasure": 0.051776101353314175, "rouge2_fmeasure_stderr": 0.0012810666933093713, "rouge2_precision": 0.03708415721222578, "rouge2_precision_stderr": 0.0015055606328312114, "rouge2_recall": 0.14224175280721357, "rouge2_recall_stderr": 0.0032367826210838984, "rougeL_fmeasure": 0.10723944522759159, "rougeL_fmeasure_stderr": 0.002017538602140417, "rougeL_precision": 0.07986488274116094, "rougeL_precision_stderr": 0.0025364066921602967, "rougeL_recall": 0.2851084703862567, "rougeL_recall_stderr": 0.005103727035204417, "rougeLsum_fmeasure": 0.10754705129811067, "rougeLsum_fmeasure_stderr": 0.002066923006506901, "rougeLsum_precision": 0.08031203138023377, "rougeLsum_precision_stderr": 0.0025700547990470703, "rougeLsum_recall": 0.2832049885841092, "rougeLsum_recall_stderr": 0.00501983173550152}}, "1": {"PALM_prompt": {"bleu": 0.6352422858353469, "bleu_stderr": 0.04386319262846487, "rouge1_fmeasure": 0.17595828606649466, "rouge1_fmeasure_stderr": 0.0037103128590180736, "rouge1_precision": 0.14729082149250516, "rouge1_precision_stderr": 0.004315116801617204, "rouge1_recall": 0.34501701041772753, "rouge1_recall_stderr": 0.004883560454809294, "rouge2_fmeasure": 0.08797235452625987, "rouge2_fmeasure_stderr": 0.002563800160615811, "rouge2_precision": 0.0749912603946425, "rouge2_precision_stderr": 0.0029697645981284103, "rouge2_recall": 0.17507177387031794, "rouge2_recall_stderr": 0.0035680879509635034, "rougeL_fmeasure": 0.1587241385605771, "rougeL_fmeasure_stderr": 0.0031813191221144485, "rougeL_precision": 0.13087012462145003, "rougeL_precision_stderr": 0.0037359089786958557, "rougeL_recall": 0.3214469196295103, "rougeL_recall_stderr": 0.004518231473535444, "rougeLsum_fmeasure": 0.16209364614512511, "rougeLsum_fmeasure_stderr": 0.003268839028418146, "rougeLsum_precision": 0.13443162008077075, "rougeLsum_precision_stderr": 0.0038480943977239573, "rougeLsum_recall": 0.32447309188867546, "rougeLsum_recall_stderr": 0.0044995663315275155}}, "2": {"PALM_prompt": {"bleu": 0.9195810919624675, "bleu_stderr": 0.04315216107683015, "rouge1_fmeasure": 0.21485950562026987, "rouge1_fmeasure_stderr": 0.004330200790716057, "rouge1_precision": 0.19281853446958125, "rouge1_precision_stderr": 0.005348179657635172, "rouge1_recall": 0.3839239693259562, "rouge1_recall_stderr": 0.004896794421772872, "rouge2_fmeasure": 0.11382579286837109, "rouge2_fmeasure_stderr": 0.0029731216918058084, "rouge2_precision": 0.1059225294754287, "rouge2_precision_stderr": 0.0036916446521448806, "rouge2_recall": 0.20457409251843786, "rouge2_recall_stderr": 0.0037775743582373947, "rougeL_fmeasure": 0.1905777152487244, "rougeL_fmeasure_stderr": 0.0036266046372882984, "rougeL_precision": 0.16888883147055286, "rougeL_precision_stderr": 0.00457575327708451, "rougeL_recall": 0.35302151670878984, "rougeL_recall_stderr": 0.004443535122018173, "rougeLsum_fmeasure": 0.1966781681343204, "rougeLsum_fmeasure_stderr": 0.0037869984327305512, "rougeLsum_precision": 0.17526191475368208, "rougeLsum_precision_stderr": 0.004769580382692894, "rougeLsum_recall": 0.35984651407202956, "rougeLsum_recall_stderr": 0.004507478564774203}}, "3": {"PALM_prompt": {"bleu": 1.0349566025343735, "bleu_stderr": 0.03887674612290353, "rouge1_fmeasure": 0.2237082219543607, "rouge1_fmeasure_stderr": 0.004572034500678942, "rouge1_precision": 0.20438835211537978, "rouge1_precision_stderr": 0.005665805208271208, "rouge1_recall": 0.39198992418866246, "rouge1_recall_stderr": 0.004947498335222262, "rouge2_fmeasure": 0.1207306381621409, "rouge2_fmeasure_stderr": 0.003262380279344123, "rouge2_precision": 0.11365499071190853, "rouge2_precision_stderr": 0.00397252439813096, "rouge2_recall": 0.2111302605559281, "rouge2_recall_stderr": 0.003926393767113415, "rougeL_fmeasure": 0.196789710556852, "rougeL_fmeasure_stderr": 0.0038281461566836554, "rougeL_precision": 0.17756713475764335, "rougeL_precision_stderr": 0.004833439612735641, "rougeL_recall": 0.3583014136638203, "rougeL_recall_stderr": 0.004476979554028604, "rougeLsum_fmeasure": 0.20416935580581227, "rougeLsum_fmeasure_stderr": 0.004021347054930177, "rougeLsum_precision": 0.18556820599751273, "rougeLsum_precision_stderr": 0.00508629808674153, "rougeLsum_recall": 0.3664900622762122, "rougeLsum_recall_stderr": 0.004561176586485258}}, "4": {"PALM_prompt": {"bleu": 1.1673998473534046, "bleu_stderr": 0.09133948510400369, "rouge1_fmeasure": 0.23196338111553216, "rouge1_fmeasure_stderr": 0.004558457306896048, "rouge1_precision": 0.2125195111528658, "rouge1_precision_stderr": 0.005730453983133926, "rouge1_recall": 0.40637639113635166, "rouge1_recall_stderr": 0.004933922377880338, "rouge2_fmeasure": 0.12559736589051002, "rouge2_fmeasure_stderr": 0.003188817274548042, "rouge2_precision": 0.11875092253811664, "rouge2_precision_stderr": 0.0039714331213516595, "rouge2_recall": 0.22248575515884728, "rouge2_recall_stderr": 0.004014145150661029, "rougeL_fmeasure": 0.2038615706419507, "rougeL_fmeasure_stderr": 0.003812860468542949, "rougeL_precision": 0.1843607825179914, "rougeL_precision_stderr": 0.0048800492055134925, "rougeL_recall": 0.370665372766866, "rougeL_recall_stderr": 0.004464677380992271, "rougeLsum_fmeasure": 0.2125920854603434, "rougeLsum_fmeasure_stderr": 0.004058013873763605, "rougeLsum_precision": 0.19370515184964268, "rougeLsum_precision_stderr": 0.005184788234116766, "rougeLsum_recall": 0.3806829975264072, "rougeLsum_recall_stderr": 0.004583640895651089}}, "5": {"PALM_prompt": {"bleu": 1.194760741144373, "bleu_stderr": 0.09038400286323779, "rouge1_fmeasure": 0.24067238350377898, "rouge1_fmeasure_stderr": 0.004627635578763758, "rouge1_precision": 0.22388107728965712, "rouge1_precision_stderr": 0.0058375878448932725, "rouge1_recall": 0.41258053843820386, "rouge1_recall_stderr": 0.004808891702805189, "rouge2_fmeasure": 0.12988590207293185, "rouge2_fmeasure_stderr": 0.003239297680717399, "rouge2_precision": 0.12470779957706617, "rouge2_precision_stderr": 0.004011608962377894, "rouge2_recall": 0.22396081182592556, "rouge2_recall_stderr": 0.003932938864761309, "rougeL_fmeasure": 0.20930811896578502, "rougeL_fmeasure_stderr": 0.0038259333936250185, "rougeL_precision": 0.19195297464964095, "rougeL_precision_stderr": 0.004914089701123312, "rougeL_recall": 0.37374160359881364, "rougeL_recall_stderr": 0.00433847098671097, "rougeLsum_fmeasure": 0.21923266732565286, "rougeLsum_fmeasure_stderr": 0.004087600756703223, "rougeLsum_precision": 0.20275170487892208, "rougeLsum_precision_stderr": 0.005253109590732199, "rougeLsum_recall": 0.3844755575959121, "rougeLsum_recall_stderr": 0.004448663215506157}}}, "GEM/wiki_lingua_en": {"0": {"tldr_en": {"bleu": 2.170472303270534, "bleu_stderr": 0.09245843828343826, "rouge1_fmeasure": 0.10762579083182928, "rouge1_fmeasure_stderr": 0.0025835076322910672, "rouge1_precision": 0.11259234653795358, "rouge1_precision_stderr": 0.0032411862971363117, "rouge1_recall": 0.14340893843595093, "rouge1_recall_stderr": 0.0034694907773717048, "rouge2_fmeasure": 0.02778920003491775, "rouge2_fmeasure_stderr": 0.000982242456017776, "rouge2_precision": 0.02582999428983234, "rouge2_precision_stderr": 0.0010000072112516562, "rouge2_recall": 0.03784241933814092, "rouge2_recall_stderr": 0.001450453403192633, "rougeL_fmeasure": 0.08192620888873649, "rougeL_fmeasure_stderr": 0.0019339948879164794, "rougeL_precision": 0.08841484558766141, "rougeL_precision_stderr": 0.0027853934411114305, "rougeL_recall": 0.1114317644839379, "rougeL_recall_stderr": 0.0027453844256091257, "rougeLsum_fmeasure": 0.10120163014874665, "rougeLsum_fmeasure_stderr": 0.002435237757238233, "rougeLsum_precision": 0.1068082984161991, "rougeLsum_precision_stderr": 0.0031424461382987463, "rougeLsum_recall": 0.13490305217744453, "rougeLsum_recall_stderr": 0.0032803950524800847}}, "1": {"tldr_en": {"bleu": 3.020335174960559, "bleu_stderr": 0.06787715772935045, "rouge1_fmeasure": 0.18153788588353004, "rouge1_fmeasure_stderr": 0.002424482153914544, "rouge1_precision": 0.19504169763237164, "rouge1_precision_stderr": 0.003161655570483573, "rouge1_recall": 0.2268980166797919, "rouge1_recall_stderr": 0.00336143207911689, "rouge2_fmeasure": 0.044909001327117046, "rouge2_fmeasure_stderr": 0.0011943824375694858, "rouge2_precision": 0.049740220187040864, "rouge2_precision_stderr": 0.001561722025312558, "rouge2_recall": 0.05782059703115125, "rouge2_recall_stderr": 0.0016860297998920056, "rougeL_fmeasure": 0.1341843957584949, "rougeL_fmeasure_stderr": 0.0017857657209747223, "rougeL_precision": 0.14594590659077963, "rougeL_precision_stderr": 0.002476347943750222, "rougeL_recall": 0.1691189101402509, "rougeL_recall_stderr": 0.002588936970621776, "rougeLsum_fmeasure": 0.16951014920125493, "rougeLsum_fmeasure_stderr": 0.0022640636265336893, "rougeLsum_precision": 0.18281924229155466, "rougeLsum_precision_stderr": 0.0029977938731759806, "rougeLsum_recall": 0.21164415433278383, "rougeLsum_recall_stderr": 0.0031361900114708872}}, "2": {"tldr_en": {"bleu": 4.287770006656117, "bleu_stderr": 0.11835519853257852, "rouge1_fmeasure": 0.24048738248155943, "rouge1_fmeasure_stderr": 0.0022924559211728, "rouge1_precision": 0.26704682686051745, "rouge1_precision_stderr": 0.003346061664825914, "rouge1_recall": 0.28723900360273236, "rouge1_recall_stderr": 0.0030648378055716376, "rouge2_fmeasure": 0.0668489557832933, "rouge2_fmeasure_stderr": 0.0013397924519285302, "rouge2_precision": 0.07760883007326022, "rouge2_precision_stderr": 0.0018992689472675252, "rouge2_recall": 0.07930488429722847, "rouge2_recall_stderr": 0.0016958561414342646, "rougeL_fmeasure": 0.17873946616791977, "rougeL_fmeasure_stderr": 0.0017622551083101064, "rougeL_precision": 0.2005421334534283, "rougeL_precision_stderr": 0.00272610264193461, "rougeL_recall": 0.2147121574085488, "rougeL_recall_stderr": 0.0023921943267692113, "rougeLsum_fmeasure": 0.22548796825222664, "rougeLsum_fmeasure_stderr": 0.002166294099617983, "rougeLsum_precision": 0.25120585518065286, "rougeLsum_precision_stderr": 0.0032054522314792226, "rougeLsum_recall": 0.2689555721836676, "rougeLsum_recall_stderr": 0.002872035212266304}}, "3": {"tldr_en": {"bleu": 3.8253545690458486, "bleu_stderr": 0.11601169899524703, "rouge1_fmeasure": 0.21120410580427312, "rouge1_fmeasure_stderr": 0.002645585016214612, "rouge1_precision": 0.24693572848760356, "rouge1_precision_stderr": 0.0036808375259794837, "rouge1_recall": 0.24354376203183603, "rouge1_recall_stderr": 0.0034150456640924024, "rouge2_fmeasure": 0.05941725493999427, "rouge2_fmeasure_stderr": 0.0013446993130627589, "rouge2_precision": 0.07085834748957563, "rouge2_precision_stderr": 0.001874517454430376, "rouge2_recall": 0.0691245743627346, "rouge2_recall_stderr": 0.0017265463033494647, "rougeL_fmeasure": 0.1574847757182755, "rougeL_fmeasure_stderr": 0.0020235834993420653, "rougeL_precision": 0.18619149092636542, "rougeL_precision_stderr": 0.0029566758253007693, "rougeL_recall": 0.18275407900501345, "rougeL_recall_stderr": 0.002678372152846199, "rougeLsum_fmeasure": 0.19767422939876575, "rougeLsum_fmeasure_stderr": 0.00249381697530238, "rougeLsum_precision": 0.23179629616820094, "rougeLsum_precision_stderr": 0.003517652101595812, "rougeLsum_recall": 0.22762811767303537, "rougeLsum_recall_stderr": 0.0031986554975203343}}, "4": {"tldr_en": {"bleu": 0.2364078795997809, "bleu_stderr": 0.024391601699964672, "rouge1_fmeasure": 0.069564585739879, "rouge1_fmeasure_stderr": 0.002354877693129134, "rouge1_precision": 0.08491288146025691, "rouge1_precision_stderr": 0.003092982485525366, "rouge1_recall": 0.08011242516269225, "rouge1_recall_stderr": 0.0028916714775573983, "rouge2_fmeasure": 0.02020024245182335, "rouge2_fmeasure_stderr": 0.0009720348342506413, "rouge2_precision": 0.02565888234744232, "rouge2_precision_stderr": 0.0014806953806721635, "rouge2_recall": 0.023735577183813514, "rouge2_recall_stderr": 0.0012482652822662537, "rougeL_fmeasure": 0.0532771556440879, "rougeL_fmeasure_stderr": 0.001820427400127109, "rougeL_precision": 0.06624573735046896, "rougeL_precision_stderr": 0.0025139925050591855, "rougeL_recall": 0.06166301644336901, "rougeL_recall_stderr": 0.002275982977075628, "rougeLsum_fmeasure": 0.06493333871562736, "rougeLsum_fmeasure_stderr": 0.0022026091321757347, "rougeLsum_precision": 0.07976639037062659, "rougeLsum_precision_stderr": 0.0029385774245711256, "rougeLsum_recall": 0.07471788462506215, "rougeLsum_recall_stderr": 0.002705709953549525}}, "5": {"tldr_en": {"bleu": 2.099893037305623e-12, "bleu_stderr": 2.420929985001087e-11, "rouge1_fmeasure": 0.010622171790435694, "rouge1_fmeasure_stderr": 0.0010302186253132802, "rouge1_precision": 0.015915821872377674, "rouge1_precision_stderr": 0.0016702869678876702, "rouge1_recall": 0.011688601517501347, "rouge1_recall_stderr": 0.0011952240401007063, "rouge2_fmeasure": 0.0035013576004744333, "rouge2_fmeasure_stderr": 0.00047232485179348453, "rouge2_precision": 0.005627833349821597, "rouge2_precision_stderr": 0.0009009236122981561, "rouge2_recall": 0.0037129642663769444, "rouge2_recall_stderr": 0.0005184603335382291, "rougeL_fmeasure": 0.008452079824288578, "rougeL_fmeasure_stderr": 0.0008242289394208465, "rougeL_precision": 0.013277596011797644, "rougeL_precision_stderr": 0.0014699518837369332, "rougeL_recall": 0.009310479193227725, "rougeL_recall_stderr": 0.0009665047884074748, "rougeLsum_fmeasure": 0.009996482786372373, "rougeLsum_fmeasure_stderr": 0.0009674414655995565, "rougeLsum_precision": 0.015271784138576902, "rougeLsum_precision_stderr": 0.001626300881272708, "rougeLsum_recall": 0.010986042036294772, "rougeLsum_recall_stderr": 0.0011291877819487723}}}, "e2e_nlg_cleaned": {"0": {"generate_text_restaurant": {"bleu": 2.235702101845383, "bleu_stderr": 0.14717774293969751, "rouge1_fmeasure": 0.18121867050334917, "rouge1_fmeasure_stderr": 0.0018474922816842606, "rouge1_precision": 0.7622052797070503, "rouge1_precision_stderr": 0.006647669795578922, "rouge1_recall": 0.1532954300944369, "rouge1_recall_stderr": 0.0031267049368978316, "rouge2_fmeasure": 0.08752141617950068, "rouge2_fmeasure_stderr": 0.0010542837911429734, "rouge2_precision": 0.7122733552799622, "rouge2_precision_stderr": 0.007806658717987775, "rouge2_recall": 0.07189313722539607, "rouge2_recall_stderr": 0.0017250469433453682, "rougeL_fmeasure": 0.1698326390033901, "rougeL_fmeasure_stderr": 0.0014973431260616968, "rougeL_precision": 0.7500255707124592, "rougeL_precision_stderr": 0.006889071529543362, "rougeL_recall": 0.13882842661597, "rougeL_recall_stderr": 0.0025055451424711027, "rougeLsum_fmeasure": 0.17261995880897366, "rougeLsum_fmeasure_stderr": 0.0016987141071618745, "rougeLsum_precision": 0.752191135927755, "rougeLsum_precision_stderr": 0.00685154752320934, "rougeLsum_recall": 0.14346581040848494, "rougeLsum_recall_stderr": 0.002891637286051436}}, "1": {"generate_text_restaurant": {"bleu": 12.507959373818647, "bleu_stderr": 0.12774308491764494, "rouge1_fmeasure": 0.48125855004986695, "rouge1_fmeasure_stderr": 0.0023747652067436864, "rouge1_precision": 0.5901037503482837, "rouge1_precision_stderr": 0.0032159724501758416, "rouge1_recall": 0.4453954075758376, "rouge1_recall_stderr": 0.0030647232677422412, "rouge2_fmeasure": 0.23034576083422903, "rouge2_fmeasure_stderr": 0.0020975083120057516, "rouge2_precision": 0.2860806077756621, "rouge2_precision_stderr": 0.0027716337523015977, "rouge2_recall": 0.21297260057374331, "rouge2_recall_stderr": 0.002224122053216537, "rougeL_fmeasure": 0.34846752107389445, "rougeL_fmeasure_stderr": 0.0021243881165946155, "rougeL_precision": 0.4308357021587011, "rougeL_precision_stderr": 0.003009420690844223, "rougeL_recall": 0.3212543906087175, "rougeL_recall_stderr": 0.002480078078958591, "rougeLsum_fmeasure": 0.3931631455827191, "rougeLsum_fmeasure_stderr": 0.0023981333184838593, "rougeLsum_precision": 0.4830396285828985, "rougeLsum_precision_stderr": 0.003213107569523447, "rougeLsum_recall": 0.3634977467417529, "rougeLsum_recall_stderr": 0.002819214726631758}}, "2": {"generate_text_restaurant": {"bleu": 15.158653678134238, "bleu_stderr": 0.11217626725691181, "rouge1_fmeasure": 0.5131292320126455, "rouge1_fmeasure_stderr": 0.002253031761361406, "rouge1_precision": 0.6066145287034505, "rouge1_precision_stderr": 0.003147539042422012, "rouge1_recall": 0.48239587619310037, "rouge1_recall_stderr": 0.0029532891322940407, "rouge2_fmeasure": 0.2578530121604481, "rouge2_fmeasure_stderr": 0.0021628607938910917, "rouge2_precision": 0.3089787427206784, "rouge2_precision_stderr": 0.0028203707876968153, "rouge2_recall": 0.24188957715036025, "rouge2_recall_stderr": 0.002292974031208558, "rougeL_fmeasure": 0.37525966202827027, "rougeL_fmeasure_stderr": 0.0021725307399361672, "rougeL_precision": 0.44560511468454206, "rougeL_precision_stderr": 0.002995038040318342, "rougeL_recall": 0.35193582043955735, "rougeL_recall_stderr": 0.0025196283655795136, "rougeLsum_fmeasure": 0.4285131129511556, "rougeLsum_fmeasure_stderr": 0.0024045780604980463, "rougeLsum_precision": 0.5064590234844139, "rougeLsum_precision_stderr": 0.003190583549175626, "rougeLsum_recall": 0.4026845079975637, "rougeLsum_recall_stderr": 0.002836627735824717}}, "3": {"generate_text_restaurant": {"bleu": 15.759812537730546, "bleu_stderr": 0.16130432495289687, "rouge1_fmeasure": 0.5204784544081206, "rouge1_fmeasure_stderr": 0.002251275712244091, "rouge1_precision": 0.6077270325813979, "rouge1_precision_stderr": 0.0031150320109689183, "rouge1_recall": 0.49124527491366843, "rouge1_recall_stderr": 0.0029113142559981263, "rouge2_fmeasure": 0.26545692823299993, "rouge2_fmeasure_stderr": 0.0021835553197599713, "rouge2_precision": 0.31332017254120853, "rouge2_precision_stderr": 0.0027772905192189512, "rouge2_recall": 0.2502643652420692, "rouge2_recall_stderr": 0.0023315173133087275, "rougeL_fmeasure": 0.3810216920994399, "rougeL_fmeasure_stderr": 0.002194042986121944, "rougeL_precision": 0.4459205694292829, "rougeL_precision_stderr": 0.0029361567575171836, "rougeL_recall": 0.3592970906286205, "rougeL_recall_stderr": 0.0025375120589350396, "rougeLsum_fmeasure": 0.4357908055166842, "rougeLsum_fmeasure_stderr": 0.0024150704709467426, "rougeLsum_precision": 0.5082193147554575, "rougeLsum_precision_stderr": 0.003142597202698407, "rougeLsum_recall": 0.41160458845561226, "rougeLsum_recall_stderr": 0.002840955053130549}}, "4": {"generate_text_restaurant": {"bleu": 16.134799256710256, "bleu_stderr": 0.1070119918698532, "rouge1_fmeasure": 0.523909375920506, "rouge1_fmeasure_stderr": 0.0022880595790812155, "rouge1_precision": 0.6054119860771526, "rouge1_precision_stderr": 0.0031401965578103136, "rouge1_recall": 0.4946565816842518, "rouge1_recall_stderr": 0.0028364911596105765, "rouge2_fmeasure": 0.26757332273287654, "rouge2_fmeasure_stderr": 0.0022434333099258102, "rouge2_precision": 0.3124961122227526, "rouge2_precision_stderr": 0.002808954903697186, "rouge2_recall": 0.25220130152340275, "rouge2_recall_stderr": 0.0023493523906758557, "rougeL_fmeasure": 0.382175935620986, "rougeL_fmeasure_stderr": 0.002236237383536334, "rougeL_precision": 0.44233843950127705, "rougeL_precision_stderr": 0.0029306698549013154, "rougeL_recall": 0.3607369273273119, "rougeL_recall_stderr": 0.002523691933961996, "rougeLsum_fmeasure": 0.4381722454639816, "rougeLsum_fmeasure_stderr": 0.0024443106668905357, "rougeLsum_precision": 0.5057660051353207, "rougeLsum_precision_stderr": 0.003153645376968123, "rougeLsum_recall": 0.4139501510871044, "rougeLsum_recall_stderr": 0.002788136076678763}}, "5": {"generate_text_restaurant": {"bleu": 16.107165763126492, "bleu_stderr": 0.1725403495293288, "rouge1_fmeasure": 0.5240001141378704, "rouge1_fmeasure_stderr": 0.0022053636458810423, "rouge1_precision": 0.6037882438302521, "rouge1_precision_stderr": 0.00307437348704947, "rouge1_recall": 0.49496382545882733, "rouge1_recall_stderr": 0.0027753751046680057, "rouge2_fmeasure": 0.26817812765923393, "rouge2_fmeasure_stderr": 0.002197626024453061, "rouge2_precision": 0.31296117294303094, "rouge2_precision_stderr": 0.0027976539566726744, "rouge2_recall": 0.25262525126987617, "rouge2_recall_stderr": 0.0023072545439166217, "rougeL_fmeasure": 0.3843695820781079, "rougeL_fmeasure_stderr": 0.0022098458849370652, "rougeL_precision": 0.44434419215812615, "rougeL_precision_stderr": 0.002949180064465491, "rougeL_recall": 0.3625276939571968, "rougeL_recall_stderr": 0.0024933568400759844, "rougeLsum_fmeasure": 0.4393845856350276, "rougeLsum_fmeasure_stderr": 0.002395852175826684, "rougeLsum_precision": 0.5068534628215157, "rougeLsum_precision_stderr": 0.0031588710330076797, "rougeLsum_recall": 0.41448240798432473, "rougeLsum_recall_stderr": 0.0027203178472831524}}}, "gem_xsum": {"0": {"article_DOC_summary": {"bleu": 2.1587870129118882, "bleu_stderr": 0.09029922216218629, "rouge1_fmeasure": 0.2131475561803473, "rouge1_fmeasure_stderr": 0.002809398976147748, "rouge1_precision": 0.1554932484923166, "rouge1_precision_stderr": 0.002221691062513405, "rouge1_recall": 0.35841398237522276, "rouge1_recall_stderr": 0.004880460975300627, "rouge2_fmeasure": 0.05062634534913608, "rouge2_fmeasure_stderr": 0.001690768330530915, "rouge2_precision": 0.036146559262545196, "rouge2_precision_stderr": 0.0012086761533876766, "rouge2_recall": 0.08868762877243597, "rouge2_recall_stderr": 0.0030835154047321074, "rougeL_fmeasure": 0.15595658881556435, "rougeL_fmeasure_stderr": 0.002127497880378741, "rougeL_precision": 0.11383026012975615, "rougeL_precision_stderr": 0.001748646718497343, "rougeL_recall": 0.2638855275835188, "rougeL_recall_stderr": 0.0038587410761858755, "rougeLsum_fmeasure": 0.16967238254142056, "rougeLsum_fmeasure_stderr": 0.00237624607133809, "rougeLsum_precision": 0.12373439304630478, "rougeLsum_precision_stderr": 0.0019078089158618033, "rougeLsum_recall": 0.2868552037024372, "rougeLsum_recall_stderr": 0.004272131408381669}}, "1": {"article_DOC_summary": {"bleu": 1.9813214649410344, "bleu_stderr": 0.06605541875330423, "rouge1_fmeasure": 0.20475174329785, "rouge1_fmeasure_stderr": 0.0030292857270426262, "rouge1_precision": 0.17602273459783957, "rouge1_precision_stderr": 0.0034912567740095035, "rouge1_recall": 0.2978672198691626, "rouge1_recall_stderr": 0.004325712043489534, "rouge2_fmeasure": 0.046694043648379924, "rouge2_fmeasure_stderr": 0.001891957305961457, "rouge2_precision": 0.04044850363752258, "rouge2_precision_stderr": 0.0019137570100629107, "rouge2_recall": 0.06884979773351274, "rouge2_recall_stderr": 0.002724070824860895, "rougeL_fmeasure": 0.1593397594005996, "rougeL_fmeasure_stderr": 0.0024118116760541847, "rougeL_precision": 0.13694272638916472, "rougeL_precision_stderr": 0.0028412360369673493, "rougeL_recall": 0.2334897227634632, "rougeL_recall_stderr": 0.003556590427451311, "rougeLsum_fmeasure": 0.15946764140405875, "rougeLsum_fmeasure_stderr": 0.0025141424269190023, "rougeLsum_precision": 0.137142234387202, "rougeLsum_precision_stderr": 0.0029073893715640745, "rougeLsum_recall": 0.23357550407307784, "rougeLsum_recall_stderr": 0.003728086289055363}}, "2": {"article_DOC_summary": {"bleu": 2.568315279137756, "bleu_stderr": 0.13888090668923836, "rouge1_fmeasure": 0.2303304665083427, "rouge1_fmeasure_stderr": 0.003468001334593542, "rouge1_precision": 0.2229374260378376, "rouge1_precision_stderr": 0.004280928411283304, "rouge1_recall": 0.28304069580526636, "rouge1_recall_stderr": 0.003972947611177992, "rouge2_fmeasure": 0.05750919277461529, "rouge2_fmeasure_stderr": 0.002337201303458201, "rouge2_precision": 0.057639208977718646, "rouge2_precision_stderr": 0.002614759341604805, "rouge2_recall": 0.06850665995022794, "rouge2_recall_stderr": 0.0026514610571335023, "rougeL_fmeasure": 0.17795352717964225, "rougeL_fmeasure_stderr": 0.0028890194692426686, "rougeL_precision": 0.17219483167439611, "rougeL_precision_stderr": 0.003538412807763811, "rougeL_recall": 0.21973513766849492, "rougeL_recall_stderr": 0.0033149507650464807, "rougeLsum_fmeasure": 0.17909978174756772, "rougeLsum_fmeasure_stderr": 0.002921283919599673, "rougeLsum_precision": 0.17313600638730706, "rougeLsum_precision_stderr": 0.0035630889990072444, "rougeLsum_recall": 0.2218290481310322, "rougeLsum_recall_stderr": 0.0034301100563963126}}, "3": {"article_DOC_summary": {"bleu": 2.911154273868146, "bleu_stderr": 0.15584859367239043, "rouge1_fmeasure": 0.22535248615052844, "rouge1_fmeasure_stderr": 0.003791735783514838, "rouge1_precision": 0.22724912362863622, "rouge1_precision_stderr": 0.004617605872422519, "rouge1_recall": 0.26380458901019693, "rouge1_recall_stderr": 0.004323907531182028, "rouge2_fmeasure": 0.05815173892712961, "rouge2_fmeasure_stderr": 0.002414756084543016, "rouge2_precision": 0.06038128094861811, "rouge2_precision_stderr": 0.002741435955597867, "rouge2_recall": 0.06597238885255258, "rouge2_recall_stderr": 0.0026568422218367894, "rougeL_fmeasure": 0.17298927032856967, "rougeL_fmeasure_stderr": 0.0031337802161657533, "rougeL_precision": 0.17388874536583043, "rougeL_precision_stderr": 0.0037665824275065446, "rougeL_recall": 0.2036906299119507, "rougeL_recall_stderr": 0.0035595683530041146, "rougeLsum_fmeasure": 0.17531050854597707, "rougeLsum_fmeasure_stderr": 0.003177216988533208, "rougeLsum_precision": 0.17573844022643165, "rougeLsum_precision_stderr": 0.003788816404981867, "rougeLsum_recall": 0.20748309723587363, "rougeLsum_recall_stderr": 0.0037272616939512094}}, "4": {"article_DOC_summary": {"bleu": 0.3728774440118906, "bleu_stderr": 0.09579003392790254, "rouge1_fmeasure": 0.058891655759634634, "rouge1_fmeasure_stderr": 0.0034788169061019582, "rouge1_precision": 0.06384997695789031, "rouge1_precision_stderr": 0.004081417327764998, "rouge1_recall": 0.06700890326086541, "rouge1_recall_stderr": 0.00398883760575528, "rouge2_fmeasure": 0.014176278266192046, "rouge2_fmeasure_stderr": 0.0013476830056293153, "rouge2_precision": 0.01497412391522327, "rouge2_precision_stderr": 0.0015535957922032293, "rouge2_recall": 0.016447549994388213, "rouge2_recall_stderr": 0.0015411299403550512, "rougeL_fmeasure": 0.045492505164083784, "rougeL_fmeasure_stderr": 0.00272525124648742, "rougeL_precision": 0.04936509536592748, "rougeL_precision_stderr": 0.003202638138772514, "rougeL_recall": 0.05218241102055928, "rougeL_recall_stderr": 0.003145249972269267, "rougeLsum_fmeasure": 0.04608216819829664, "rougeLsum_fmeasure_stderr": 0.0027590740017696767, "rougeLsum_precision": 0.049919934471052654, "rougeLsum_precision_stderr": 0.003226500601808736, "rougeLsum_recall": 0.05295420506561277, "rougeLsum_recall_stderr": 0.0032077224941783403}}, "5": {"article_DOC_summary": {"bleu": 4.5775755670070436e-24, "bleu_stderr": 1.165194982722847e-20, "rouge1_fmeasure": 0.003209685928721461, "rouge1_fmeasure_stderr": 0.0009321057176689659, "rouge1_precision": 0.003245026115016397, "rouge1_precision_stderr": 0.0010018752109838872, "rouge1_recall": 0.003816108316949039, "rouge1_recall_stderr": 0.0010591733379027508, "rouge2_fmeasure": 0.0005731642694331555, "rouge2_fmeasure_stderr": 0.00023579174692404015, "rouge2_precision": 0.0006149053716742533, "rouge2_precision_stderr": 0.0002603220643343989, "rouge2_recall": 0.0005865993976426828, "rouge2_recall_stderr": 0.00023171003012774805, "rougeL_fmeasure": 0.002272633054035741, "rougeL_fmeasure_stderr": 0.0006656958601438186, "rougeL_precision": 0.0022714483278982152, "rougeL_precision_stderr": 0.0007108947692843597, "rougeL_recall": 0.0028112298473500735, "rougeL_recall_stderr": 0.0008069199908163127, "rougeLsum_fmeasure": 0.0023210463979317724, "rougeLsum_fmeasure_stderr": 0.0006849794862388629, "rougeLsum_precision": 0.002332315692638469, "rougeLsum_precision_stderr": 0.000733094697158169, "rougeLsum_recall": 0.0027357899134195732, "rougeLsum_recall_stderr": 0.0007566942257832804}}}} |