File size: 28,020 Bytes
29393c3 |
1 |
{"GEM/web_nlg_en": {"0": {"PALM_prompt": {"bleu": 0.35092030525579987, "bleu_stderr": 0.04513752019671635, "rouge1_fmeasure": 0.10668665131573762, "rouge1_fmeasure_stderr": 0.0019963304628015832, "rouge1_precision": 0.06981936667408743, "rouge1_precision_stderr": 0.001473328613173937, "rouge1_recall": 0.2898378467591775, "rouge1_recall_stderr": 0.004420681682062187, "rouge2_fmeasure": 0.05030332747327783, "rouge2_fmeasure_stderr": 0.001248625720436871, "rouge2_precision": 0.032790254442078545, "rouge2_precision_stderr": 0.0009004212723231938, "rouge2_recall": 0.14110910701776389, "rouge2_recall_stderr": 0.003077172133370463, "rougeL_fmeasure": 0.10309579270491318, "rougeL_fmeasure_stderr": 0.0018586300746194161, "rougeL_precision": 0.06724512679042116, "rougeL_precision_stderr": 0.0013469997217561611, "rougeL_recall": 0.2826819906334472, "rougeL_recall_stderr": 0.004340393008834924, "rougeLsum_fmeasure": 0.10235320219135983, "rougeLsum_fmeasure_stderr": 0.0018735247246252473, "rougeLsum_precision": 0.06690927840170163, "rougeLsum_precision_stderr": 0.0013776966657478994, "rougeLsum_recall": 0.27937547516216105, "rougeLsum_recall_stderr": 0.004232594249716667}}, "1": {"PALM_prompt": {"bleu": 0.4887135545339655, "bleu_stderr": 0.038207624916546064, "rouge1_fmeasure": 0.11005350518870848, "rouge1_fmeasure_stderr": 0.001863018182834003, "rouge1_precision": 0.07047823375611638, "rouge1_precision_stderr": 0.0013629185289183434, "rouge1_recall": 0.352113375646551, "rouge1_recall_stderr": 0.005212176176263606, "rouge2_fmeasure": 0.051085906931171744, "rouge2_fmeasure_stderr": 0.0011718163430073354, "rouge2_precision": 0.032678708300078345, "rouge2_precision_stderr": 0.0008350705039830835, "rouge2_recall": 0.17036081855757026, "rouge2_recall_stderr": 0.0036533260651395675, "rougeL_fmeasure": 0.10321588752678051, "rougeL_fmeasure_stderr": 0.001659748600170577, "rougeL_precision": 0.06602243378358179, "rougeL_precision_stderr": 0.0012047343606144839, "rougeL_recall": 0.3299025252012076, "rougeL_recall_stderr": 0.004711883556335918, "rougeLsum_fmeasure": 0.10490474504796687, "rougeLsum_fmeasure_stderr": 0.0017630576769512532, "rougeLsum_precision": 0.06724334284661879, "rougeLsum_precision_stderr": 0.0012905898446151749, "rougeLsum_recall": 0.3335950480627132, "rougeLsum_recall_stderr": 0.004795979080397771}}, "2": {"PALM_prompt": {"bleu": 0.4954577156236074, "bleu_stderr": 0.03940656960812993, "rouge1_fmeasure": 0.11261142270012729, "rouge1_fmeasure_stderr": 0.0017357579521792847, "rouge1_precision": 0.0714866607573638, "rouge1_precision_stderr": 0.0012477048923611861, "rouge1_recall": 0.366704464940973, "rouge1_recall_stderr": 0.005105137507748489, "rouge2_fmeasure": 0.052438296552281494, "rouge2_fmeasure_stderr": 0.0011033633061244077, "rouge2_precision": 0.03311033995538127, "rouge2_precision_stderr": 0.0007648297655042665, "rouge2_recall": 0.18155051913557482, "rouge2_recall_stderr": 0.00371730981090646, "rougeL_fmeasure": 0.10537631476024473, "rougeL_fmeasure_stderr": 0.001548032182806995, "rougeL_precision": 0.06687610955905898, "rougeL_precision_stderr": 0.0011120004143352184, "rougeL_recall": 0.3422071101265069, "rougeL_recall_stderr": 0.004593456998282074, "rougeLsum_fmeasure": 0.1071828571812226, "rougeLsum_fmeasure_stderr": 0.00163276079489414, "rougeLsum_precision": 0.06807059659117397, "rougeLsum_precision_stderr": 0.0011745555579416242, "rougeLsum_recall": 0.3484145276857689, "rougeLsum_recall_stderr": 0.004752677986072459}}, "3": {"PALM_prompt": {"bleu": 0.5376401397749783, "bleu_stderr": 0.042953510015901336, "rouge1_fmeasure": 0.11391880759717438, "rouge1_fmeasure_stderr": 0.0017214645061344748, "rouge1_precision": 0.07213772505487824, "rouge1_precision_stderr": 0.0012404738886852833, "rouge1_recall": 0.3776942956991359, "rouge1_recall_stderr": 0.00518895191365285, "rouge2_fmeasure": 0.053701976919488074, "rouge2_fmeasure_stderr": 0.0011062693098051102, "rouge2_precision": 0.033831854491709834, "rouge2_precision_stderr": 0.000767112875303506, "rouge2_recall": 0.18902052205137462, "rouge2_recall_stderr": 0.003745503144139708, "rougeL_fmeasure": 0.10636377632235701, "rougeL_fmeasure_stderr": 0.0015396374929510118, "rougeL_precision": 0.0673760809566409, "rougeL_precision_stderr": 0.0011092327236371367, "rougeL_recall": 0.3510488403621821, "rougeL_recall_stderr": 0.0046683703819187594, "rougeLsum_fmeasure": 0.10848385182183909, "rougeLsum_fmeasure_stderr": 0.0016325541592215193, "rougeLsum_precision": 0.06876714809423898, "rougeLsum_precision_stderr": 0.001180603104206542, "rougeLsum_recall": 0.3584958814094978, "rougeLsum_recall_stderr": 0.004825789898873508}}, "4": {"PALM_prompt": {"bleu": 0.5162458189892425, "bleu_stderr": 0.04567561479003356, "rouge1_fmeasure": 0.11538388270081333, "rouge1_fmeasure_stderr": 0.0017082303946239072, "rouge1_precision": 0.07313568376833962, "rouge1_precision_stderr": 0.0012433254028699097, "rouge1_recall": 0.3808222176442103, "rouge1_recall_stderr": 0.005017313267908712, "rouge2_fmeasure": 0.05408499700220169, "rouge2_fmeasure_stderr": 0.0010803111391284627, "rouge2_precision": 0.03409652402405422, "rouge2_precision_stderr": 0.0007575603665210839, "rouge2_recall": 0.1913005341448632, "rouge2_recall_stderr": 0.003690388734385289, "rougeL_fmeasure": 0.10721312824281004, "rougeL_fmeasure_stderr": 0.0015059136603345965, "rougeL_precision": 0.06801047825803394, "rougeL_precision_stderr": 0.0011036108321794546, "rougeL_recall": 0.35323488966561345, "rougeL_recall_stderr": 0.004471460082976139, "rougeLsum_fmeasure": 0.10984711639426804, "rougeLsum_fmeasure_stderr": 0.0016152114960555307, "rougeLsum_precision": 0.06970710378600671, "rougeLsum_precision_stderr": 0.0011816060505639649, "rougeLsum_recall": 0.3612576150943837, "rougeLsum_recall_stderr": 0.004665280252810779}}, "5": {"PALM_prompt": {"bleu": 0.5915390188723657, "bleu_stderr": 0.044999471378700744, "rouge1_fmeasure": 0.1156135546890491, "rouge1_fmeasure_stderr": 0.0016639879184558524, "rouge1_precision": 0.07308133165879348, "rouge1_precision_stderr": 0.0012280730859575018, "rouge1_recall": 0.39313349727864033, "rouge1_recall_stderr": 0.005208530742307819, "rouge2_fmeasure": 0.054252618076757034, "rouge2_fmeasure_stderr": 0.0010618703958199596, "rouge2_precision": 0.03409018556609139, "rouge2_precision_stderr": 0.000749916416836098, "rouge2_recall": 0.19987682746116572, "rouge2_recall_stderr": 0.003867392245272622, "rougeL_fmeasure": 0.10635472804850638, "rougeL_fmeasure_stderr": 0.001478213765313952, "rougeL_precision": 0.06736116168446518, "rougeL_precision_stderr": 0.0011057761206049632, "rougeL_recall": 0.3596279768743875, "rougeL_recall_stderr": 0.004551765129386944, "rougeLsum_fmeasure": 0.10968223642629092, "rougeLsum_fmeasure_stderr": 0.0015751819563328757, "rougeLsum_precision": 0.06946965842837552, "rougeLsum_precision_stderr": 0.0011735301458215054, "rougeLsum_recall": 0.37102628840990604, "rougeLsum_recall_stderr": 0.004786162392169987}}}, "GEM/wiki_lingua_en": {"0": {"tldr_en": {"bleu": 1.2222578150102583, "bleu_stderr": 0.06328685065190241, "rouge1_fmeasure": 0.1471218733322546, "rouge1_fmeasure_stderr": 0.0019086674422868044, "rouge1_precision": 0.12728317987175145, "rouge1_precision_stderr": 0.0019543197768273073, "rouge1_recall": 0.21340145861239856, "rouge1_recall_stderr": 0.0027771001830022594, "rouge2_fmeasure": 0.027536298312222936, "rouge2_fmeasure_stderr": 0.0007720385992298698, "rouge2_precision": 0.023495678323026473, "rouge2_precision_stderr": 0.0007149798061630863, "rouge2_recall": 0.04189401531194056, "rouge2_recall_stderr": 0.0013277090457311043, "rougeL_fmeasure": 0.11779028020448569, "rougeL_fmeasure_stderr": 0.001400602568948609, "rougeL_precision": 0.10103719194551143, "rougeL_precision_stderr": 0.0014559595916573205, "rougeL_recall": 0.1749892611099076, "rougeL_recall_stderr": 0.0022848816135267102, "rougeLsum_fmeasure": 0.13579115205056982, "rougeLsum_fmeasure_stderr": 0.0017452654195332434, "rougeLsum_precision": 0.1174405809942993, "rougeLsum_precision_stderr": 0.001802707885416322, "rougeLsum_recall": 0.19766216246005014, "rougeLsum_recall_stderr": 0.00258101670904122}}, "1": {"tldr_en": {"bleu": 2.391147100552601, "bleu_stderr": 0.04006224158757559, "rouge1_fmeasure": 0.19732650210441122, "rouge1_fmeasure_stderr": 0.0020510877426877347, "rouge1_precision": 0.1736334455823323, "rouge1_precision_stderr": 0.0022976987647117656, "rouge1_recall": 0.28354394939217903, "rouge1_recall_stderr": 0.0029641856324906304, "rouge2_fmeasure": 0.04692285315751292, "rouge2_fmeasure_stderr": 0.0010196597076878576, "rouge2_precision": 0.04141351020002613, "rouge2_precision_stderr": 0.001065936730790372, "rouge2_recall": 0.0699662145493009, "rouge2_recall_stderr": 0.0016954247431350791, "rougeL_fmeasure": 0.14379398280935196, "rougeL_fmeasure_stderr": 0.001412248102591144, "rougeL_precision": 0.12577530969575892, "rougeL_precision_stderr": 0.0016573447824456509, "rougeL_recall": 0.2115227804388362, "rougeL_recall_stderr": 0.0023043951894502297, "rougeLsum_fmeasure": 0.18431976378633813, "rougeLsum_fmeasure_stderr": 0.001921122990992563, "rougeLsum_precision": 0.16211598975071775, "rougeLsum_precision_stderr": 0.002164725932714466, "rougeLsum_recall": 0.2656809680060299, "rougeLsum_recall_stderr": 0.002812085736561483}}, "2": {"tldr_en": {"bleu": 2.7171126284576332, "bleu_stderr": 0.04815677557686793, "rouge1_fmeasure": 0.2084266165739667, "rouge1_fmeasure_stderr": 0.001999753547385771, "rouge1_precision": 0.19459311097597148, "rouge1_precision_stderr": 0.002498974869930833, "rouge1_recall": 0.28833446397621476, "rouge1_recall_stderr": 0.002863676862389607, "rouge2_fmeasure": 0.05192650547465838, "rouge2_fmeasure_stderr": 0.0010701161009785201, "rouge2_precision": 0.050412113693174034, "rouge2_precision_stderr": 0.001351379848487111, "rouge2_recall": 0.07329394037381169, "rouge2_recall_stderr": 0.0016677245426515634, "rougeL_fmeasure": 0.15280735738097753, "rougeL_fmeasure_stderr": 0.0014043929947908011, "rougeL_precision": 0.1431722098886608, "rougeL_precision_stderr": 0.001914919098242861, "rougeL_recall": 0.2153296115668877, "rougeL_recall_stderr": 0.0022634798460424333, "rougeLsum_fmeasure": 0.1956803458833979, "rougeLsum_fmeasure_stderr": 0.001874943202214558, "rougeLsum_precision": 0.1826592257718392, "rougeLsum_precision_stderr": 0.0023613926909123446, "rougeLsum_recall": 0.27102904289356916, "rougeLsum_recall_stderr": 0.002697214827427601}}, "3": {"tldr_en": {"bleu": 2.7422220260434083, "bleu_stderr": 0.07994048977227425, "rouge1_fmeasure": 0.17822188610923453, "rouge1_fmeasure_stderr": 0.002313579710983267, "rouge1_precision": 0.1833771887413479, "rouge1_precision_stderr": 0.0030737417465454956, "rouge1_recall": 0.24146816828202433, "rouge1_recall_stderr": 0.0033746604530580163, "rouge2_fmeasure": 0.045243645176363284, "rouge2_fmeasure_stderr": 0.001053472542589495, "rouge2_precision": 0.04870214637199058, "rouge2_precision_stderr": 0.0015294386752092305, "rouge2_recall": 0.06300025230653579, "rouge2_recall_stderr": 0.001665082880499784, "rougeL_fmeasure": 0.131733000955773, "rougeL_fmeasure_stderr": 0.0016654922516170963, "rougeL_precision": 0.13827180021901367, "rougeL_precision_stderr": 0.0024999501979138853, "rougeL_recall": 0.18075620846507245, "rougeL_recall_stderr": 0.002606651906742854, "rougeLsum_fmeasure": 0.16787441794691876, "rougeLsum_fmeasure_stderr": 0.002180132432051263, "rougeLsum_precision": 0.17302447151119815, "rougeLsum_precision_stderr": 0.0029298205686312893, "rougeLsum_recall": 0.2277850010309598, "rougeLsum_recall_stderr": 0.003197183070245692}}, "4": {"tldr_en": {"bleu": 0.5925437754263059, "bleu_stderr": 0.04242065704358776, "rouge1_fmeasure": 0.05723884996198824, "rouge1_fmeasure_stderr": 0.0019771316422231024, "rouge1_precision": 0.06330716830934252, "rouge1_precision_stderr": 0.0025709225823679588, "rouge1_recall": 0.07845690859829274, "rouge1_recall_stderr": 0.002822969061122126, "rouge2_fmeasure": 0.014905869113824048, "rouge2_fmeasure_stderr": 0.0007403899631872741, "rouge2_precision": 0.017858459821703032, "rouge2_precision_stderr": 0.0012094921987221639, "rouge2_recall": 0.021543413757512866, "rouge2_recall_stderr": 0.0011877190270122848, "rougeL_fmeasure": 0.043631603403088595, "rougeL_fmeasure_stderr": 0.001500131477167622, "rougeL_precision": 0.0491984487986743, "rougeL_precision_stderr": 0.0020685970258627613, "rougeL_recall": 0.06047877336706885, "rougeL_recall_stderr": 0.002215110058371633, "rougeLsum_fmeasure": 0.05373718833644186, "rougeLsum_fmeasure_stderr": 0.0018538420089590803, "rougeLsum_precision": 0.0596627080267218, "rougeLsum_precision_stderr": 0.0024418625671780636, "rougeLsum_recall": 0.07373240073484355, "rougeLsum_recall_stderr": 0.0026557359756667646}}, "5": {"tldr_en": {"bleu": 4.93737386207863e-07, "bleu_stderr": 1.1103847235234578e-06, "rouge1_fmeasure": 0.008992771424476102, "rouge1_fmeasure_stderr": 0.0008967572939909481, "rouge1_precision": 0.01062522482426178, "rouge1_precision_stderr": 0.0012092617892710666, "rouge1_recall": 0.01185222370379365, "rouge1_recall_stderr": 0.0012020911444765917, "rouge2_fmeasure": 0.0023077580857873777, "rouge2_fmeasure_stderr": 0.0003193569948458837, "rouge2_precision": 0.002927154149860958, "rouge2_precision_stderr": 0.000550773136343612, "rouge2_recall": 0.002996639644150718, "rouge2_recall_stderr": 0.0004061218977723996, "rougeL_fmeasure": 0.006734668579961652, "rougeL_fmeasure_stderr": 0.0006654911916185634, "rougeL_precision": 0.00826237840691835, "rougeL_precision_stderr": 0.000987476405544466, "rougeL_recall": 0.008978430311958394, "rougeL_recall_stderr": 0.0009171739341734633, "rougeLsum_fmeasure": 0.008370404629808693, "rougeLsum_fmeasure_stderr": 0.000834783139372695, "rougeLsum_precision": 0.009989825480627754, "rougeLsum_precision_stderr": 0.001152294422344019, "rougeLsum_recall": 0.011036254899513882, "rougeLsum_recall_stderr": 0.0011187636476075458}}}, "e2e_nlg_cleaned": {"0": {"generate_text_restaurant": {"bleu": 3.5763061428514815, "bleu_stderr": 0.08841895367881542, "rouge1_fmeasure": 0.18522377860735378, "rouge1_fmeasure_stderr": 0.001793031425512448, "rouge1_precision": 0.13693099764528335, "rouge1_precision_stderr": 0.0015356004169807865, "rouge1_recall": 0.30589803506201596, "rouge1_recall_stderr": 0.002569064963211563, "rouge2_fmeasure": 0.0750677147834178, "rouge2_fmeasure_stderr": 0.001192303759950116, "rouge2_precision": 0.05608296665902916, "rouge2_precision_stderr": 0.001060781879450204, "rouge2_recall": 0.12405447717993638, "rouge2_recall_stderr": 0.0019165396671815943, "rougeL_fmeasure": 0.1637374521643533, "rougeL_fmeasure_stderr": 0.0014672700966103522, "rougeL_precision": 0.12077559035745344, "rougeL_precision_stderr": 0.0012933898621559714, "rougeL_recall": 0.2725557899490075, "rougeL_recall_stderr": 0.0021592023164069916, "rougeLsum_fmeasure": 0.16156525416128376, "rougeLsum_fmeasure_stderr": 0.0016584383409070493, "rougeLsum_precision": 0.11941400647321004, "rougeLsum_precision_stderr": 0.0014208570924380114, "rougeLsum_recall": 0.2675565021971549, "rougeLsum_recall_stderr": 0.0024460229189680735}}, "1": {"generate_text_restaurant": {"bleu": 7.889074185047009, "bleu_stderr": 0.12895617864019385, "rouge1_fmeasure": 0.3819329232968168, "rouge1_fmeasure_stderr": 0.002058924047819458, "rouge1_precision": 0.37345813482947465, "rouge1_precision_stderr": 0.00235563045864244, "rouge1_recall": 0.425568058263057, "rouge1_recall_stderr": 0.0027502764404919496, "rouge2_fmeasure": 0.14986852530881986, "rouge2_fmeasure_stderr": 0.001646906448593975, "rouge2_precision": 0.1457310116329793, "rouge2_precision_stderr": 0.0016791811257879754, "rouge2_recall": 0.1691753347651205, "rouge2_recall_stderr": 0.0020338555184736705, "rougeL_fmeasure": 0.26455232520366323, "rougeL_fmeasure_stderr": 0.0016261234117796919, "rougeL_precision": 0.25802392248926737, "rougeL_precision_stderr": 0.0017801562545908425, "rougeL_recall": 0.2964926298625058, "rougeL_recall_stderr": 0.0022135158513236245, "rougeLsum_fmeasure": 0.31785134565072753, "rougeLsum_fmeasure_stderr": 0.0019390533889751562, "rougeLsum_precision": 0.31083590587813814, "rougeLsum_precision_stderr": 0.0021555685632575444, "rougeLsum_recall": 0.3542250176651235, "rougeLsum_recall_stderr": 0.002527520077331439}}, "2": {"generate_text_restaurant": {"bleu": 9.430394273853375, "bleu_stderr": 0.1847232705243506, "rouge1_fmeasure": 0.418885586013751, "rouge1_fmeasure_stderr": 0.002007463993097918, "rouge1_precision": 0.40831216737868514, "rouge1_precision_stderr": 0.0023142524211960777, "rouge1_recall": 0.4661123939279311, "rouge1_recall_stderr": 0.002837142860791685, "rouge2_fmeasure": 0.181238613552319, "rouge2_fmeasure_stderr": 0.0017018054740245462, "rouge2_precision": 0.17588735757182558, "rouge2_precision_stderr": 0.0017565431329739812, "rouge2_recall": 0.20469246454092782, "rouge2_recall_stderr": 0.0021698678454790243, "rougeL_fmeasure": 0.2955881450808913, "rougeL_fmeasure_stderr": 0.0016512442074684325, "rougeL_precision": 0.28785435905822937, "rougeL_precision_stderr": 0.001838982518273614, "rougeL_recall": 0.3302177769679092, "rougeL_recall_stderr": 0.0023254365486443333, "rougeLsum_fmeasure": 0.34857096921066205, "rougeLsum_fmeasure_stderr": 0.0019611752925251216, "rougeLsum_precision": 0.3397731417682527, "rougeLsum_precision_stderr": 0.0021769796707002188, "rougeLsum_recall": 0.3880429438073185, "rougeLsum_recall_stderr": 0.002651222748349836}}, "3": {"generate_text_restaurant": {"bleu": 10.113902500867782, "bleu_stderr": 0.11723679605888841, "rouge1_fmeasure": 0.42494100808108426, "rouge1_fmeasure_stderr": 0.0019848574037000756, "rouge1_precision": 0.4113193603539492, "rouge1_precision_stderr": 0.002326696320360366, "rouge1_recall": 0.4762578661058886, "rouge1_recall_stderr": 0.002803809063258203, "rouge2_fmeasure": 0.18974320403636494, "rouge2_fmeasure_stderr": 0.001728075653184395, "rouge2_precision": 0.18289494079057925, "rouge2_precision_stderr": 0.0017885733052731077, "rouge2_recall": 0.21585034907240166, "rouge2_recall_stderr": 0.0022144084445730352, "rougeL_fmeasure": 0.3038282693785727, "rougeL_fmeasure_stderr": 0.0016658099172296667, "rougeL_precision": 0.2943005681132957, "rougeL_precision_stderr": 0.0019138188823571657, "rougeL_recall": 0.341537961684278, "rougeL_recall_stderr": 0.0023341037495063265, "rougeLsum_fmeasure": 0.35595547534863364, "rougeLsum_fmeasure_stderr": 0.0019811017835625216, "rougeLsum_precision": 0.3448175191058933, "rougeLsum_precision_stderr": 0.0022377191876361987, "rougeLsum_recall": 0.3990338396086915, "rougeLsum_recall_stderr": 0.00266713283320383}}, "4": {"generate_text_restaurant": {"bleu": 10.433235636684694, "bleu_stderr": 0.16559538387279382, "rouge1_fmeasure": 0.4276906087969391, "rouge1_fmeasure_stderr": 0.0019888873745294716, "rouge1_precision": 0.41299825090051584, "rouge1_precision_stderr": 0.0023368767157547292, "rouge1_recall": 0.47962543535821944, "rouge1_recall_stderr": 0.002770194421824991, "rouge2_fmeasure": 0.19376219920647275, "rouge2_fmeasure_stderr": 0.0017880215942833856, "rouge2_precision": 0.18652386544213564, "rouge2_precision_stderr": 0.001859671426693175, "rouge2_recall": 0.22057943475581543, "rouge2_recall_stderr": 0.002268269935834746, "rougeL_fmeasure": 0.30693273428518913, "rougeL_fmeasure_stderr": 0.0017093495953663948, "rougeL_precision": 0.29632873268599635, "rougeL_precision_stderr": 0.0019349036829606914, "rougeL_recall": 0.3453892484465583, "rougeL_recall_stderr": 0.0023533437638236336, "rougeLsum_fmeasure": 0.3575618775034696, "rougeLsum_fmeasure_stderr": 0.0020284523859768924, "rougeLsum_precision": 0.3452141125875863, "rougeLsum_precision_stderr": 0.0022633006759387865, "rougeLsum_recall": 0.4013412942257543, "rougeLsum_recall_stderr": 0.002691394699368152}}, "5": {"generate_text_restaurant": {"bleu": 10.524222254347402, "bleu_stderr": 0.1570733014354235, "rouge1_fmeasure": 0.4297336508488616, "rouge1_fmeasure_stderr": 0.0019877776713711744, "rouge1_precision": 0.41338865705266253, "rouge1_precision_stderr": 0.0023151584252829207, "rouge1_recall": 0.48229437905400363, "rouge1_recall_stderr": 0.002780755795312751, "rouge2_fmeasure": 0.19587324037834244, "rouge2_fmeasure_stderr": 0.0017783986404322938, "rouge2_precision": 0.18750452947717144, "rouge2_precision_stderr": 0.0017854738914506623, "rouge2_recall": 0.22300380292777322, "rouge2_recall_stderr": 0.002292337991373791, "rougeL_fmeasure": 0.3112157685715752, "rougeL_fmeasure_stderr": 0.0017222919884681302, "rougeL_precision": 0.29901837534732933, "rougeL_precision_stderr": 0.0018933521587321512, "rougeL_recall": 0.3505715839455596, "rougeL_recall_stderr": 0.0023947239138506207, "rougeLsum_fmeasure": 0.3605533829938833, "rougeLsum_fmeasure_stderr": 0.0019926332071944987, "rougeLsum_precision": 0.346626530434538, "rougeLsum_precision_stderr": 0.0021961576508456597, "rougeLsum_recall": 0.40525157315277366, "rougeLsum_recall_stderr": 0.002694650983597002}}}, "gem_xsum": {"0": {"article_DOC_summary": {"bleu": 1.6443865340332708, "bleu_stderr": 0.06448573771008022, "rouge1_fmeasure": 0.20894728461895484, "rouge1_fmeasure_stderr": 0.002676706164139504, "rouge1_precision": 0.15977268662525132, "rouge1_precision_stderr": 0.002498597687577891, "rouge1_recall": 0.3439802391671489, "rouge1_recall_stderr": 0.00444649707695747, "rouge2_fmeasure": 0.04399744295623738, "rouge2_fmeasure_stderr": 0.0015740981350718616, "rouge2_precision": 0.03339906984759204, "rouge2_precision_stderr": 0.001300434810680209, "rouge2_recall": 0.07442330924142955, "rouge2_recall_stderr": 0.002682978761295735, "rougeL_fmeasure": 0.1503650560043469, "rougeL_fmeasure_stderr": 0.0019556473810402567, "rougeL_precision": 0.11534576817816335, "rougeL_precision_stderr": 0.0018893138267035729, "rougeL_recall": 0.24796063756999268, "rougeL_recall_stderr": 0.003283686560553736, "rougeLsum_fmeasure": 0.1619728615486753, "rougeLsum_fmeasure_stderr": 0.0021854057188140374, "rougeLsum_precision": 0.12356200850631334, "rougeLsum_precision_stderr": 0.001992934285663947, "rougeLsum_recall": 0.26845293758700167, "rougeLsum_recall_stderr": 0.0037961745448933685}}, "1": {"article_DOC_summary": {"bleu": 1.179463416983736, "bleu_stderr": 0.049322464873007926, "rouge1_fmeasure": 0.17789279128849939, "rouge1_fmeasure_stderr": 0.002409867096248393, "rouge1_precision": 0.1283487315327207, "rouge1_precision_stderr": 0.0018698659023036683, "rouge1_recall": 0.30677115841434355, "rouge1_recall_stderr": 0.004142945172247196, "rouge2_fmeasure": 0.03148436075040743, "rouge2_fmeasure_stderr": 0.0012999151791325903, "rouge2_precision": 0.02229153527620068, "rouge2_precision_stderr": 0.0009290255555238756, "rouge2_recall": 0.05667515961237042, "rouge2_recall_stderr": 0.0024282370591926813, "rougeL_fmeasure": 0.13571854568454508, "rougeL_fmeasure_stderr": 0.0017897390119063444, "rougeL_precision": 0.09766150074500464, "rougeL_precision_stderr": 0.0013689556817164002, "rougeL_recall": 0.2359307623025628, "rougeL_recall_stderr": 0.0032602221353705487, "rougeLsum_fmeasure": 0.1391691564025729, "rougeLsum_fmeasure_stderr": 0.0019802562047267817, "rougeLsum_precision": 0.10010552349835154, "rougeLsum_precision_stderr": 0.00150075160578187, "rougeLsum_recall": 0.24214462905923226, "rougeLsum_recall_stderr": 0.00358086419270699}}, "2": {"article_DOC_summary": {"bleu": 1.171605281185158, "bleu_stderr": 0.09705805598592233, "rouge1_fmeasure": 0.16607212501588745, "rouge1_fmeasure_stderr": 0.002465797526765075, "rouge1_precision": 0.12026259340246569, "rouge1_precision_stderr": 0.001959615202442428, "rouge1_recall": 0.28539768882214345, "rouge1_recall_stderr": 0.0042054437997353904, "rouge2_fmeasure": 0.031093372281268725, "rouge2_fmeasure_stderr": 0.001315103452152175, "rouge2_precision": 0.022171538116183456, "rouge2_precision_stderr": 0.0009481516225085209, "rouge2_recall": 0.05565794662443053, "rouge2_recall_stderr": 0.0024464114450631824, "rougeL_fmeasure": 0.13412112184120395, "rougeL_fmeasure_stderr": 0.001883651866644435, "rougeL_precision": 0.09678284723300769, "rougeL_precision_stderr": 0.0014596259783767793, "rougeL_recall": 0.2322804210017996, "rougeL_recall_stderr": 0.0034048283128554914, "rougeLsum_fmeasure": 0.12819178525497177, "rougeLsum_fmeasure_stderr": 0.001989564938897622, "rougeLsum_precision": 0.09254890915149624, "rougeLsum_precision_stderr": 0.0015482197742084348, "rougeLsum_recall": 0.2221642225599103, "rougeLsum_recall_stderr": 0.0035391189699574043}}, "3": {"article_DOC_summary": {"bleu": 1.2421369448331907, "bleu_stderr": 0.09404310747626213, "rouge1_fmeasure": 0.16125775260746725, "rouge1_fmeasure_stderr": 0.0024929400149637957, "rouge1_precision": 0.11937762777694022, "rouge1_precision_stderr": 0.0020785359403546226, "rouge1_recall": 0.27229911353559627, "rouge1_recall_stderr": 0.004290006567245677, "rouge2_fmeasure": 0.03135863836588418, "rouge2_fmeasure_stderr": 0.0012490119416798898, "rouge2_precision": 0.022594629033534534, "rouge2_precision_stderr": 0.0009043350501605985, "rouge2_recall": 0.05460155665840838, "rouge2_recall_stderr": 0.0022724335465863234, "rougeL_fmeasure": 0.13340144911682686, "rougeL_fmeasure_stderr": 0.001985882478626079, "rougeL_precision": 0.09874208631971298, "rougeL_precision_stderr": 0.001712176793787427, "rougeL_recall": 0.2263072791819812, "rougeL_recall_stderr": 0.0035183479716809104, "rougeLsum_fmeasure": 0.1244865414349181, "rougeLsum_fmeasure_stderr": 0.0020109053608498664, "rougeLsum_precision": 0.09228468276563344, "rougeLsum_precision_stderr": 0.0017353557651595074, "rougeLsum_recall": 0.2115269820097042, "rougeLsum_recall_stderr": 0.003570966623025798}}, "4": {"article_DOC_summary": {"bleu": 0.7507897873071828, "bleu_stderr": 0.09022391990160196, "rouge1_fmeasure": 0.04587050315380673, "rouge1_fmeasure_stderr": 0.002682921061867743, "rouge1_precision": 0.0396817017668287, "rouge1_precision_stderr": 0.0025276425448518785, "rouge1_recall": 0.07000574538399217, "rouge1_recall_stderr": 0.0041107759713876766, "rouge2_fmeasure": 0.009627528506364348, "rouge2_fmeasure_stderr": 0.0009541396465140421, "rouge2_precision": 0.007710704081296013, "rouge2_precision_stderr": 0.0008327806241196672, "rouge2_recall": 0.015032913904949322, "rouge2_recall_stderr": 0.0014695428313471306, "rougeL_fmeasure": 0.03768209350615277, "rougeL_fmeasure_stderr": 0.002167464042953999, "rougeL_precision": 0.03284394634659473, "rougeL_precision_stderr": 0.002105616405808581, "rougeL_recall": 0.05787260392319605, "rougeL_recall_stderr": 0.0033756839248268904, "rougeLsum_fmeasure": 0.03690509099375681, "rougeLsum_fmeasure_stderr": 0.002174042106201563, "rougeLsum_precision": 0.032354994329316655, "rougeLsum_precision_stderr": 0.002122514792763723, "rougeLsum_recall": 0.05653870801404548, "rougeLsum_recall_stderr": 0.0033699559531707246}}, "5": {"article_DOC_summary": {"bleu": 8.215991966026401e-36, "bleu_stderr": 2.035943344842957e-31, "rouge1_fmeasure": 0.0025391991198481386, "rouge1_fmeasure_stderr": 0.0007178274187167163, "rouge1_precision": 0.0029852282009972296, "rouge1_precision_stderr": 0.000878002695182032, "rouge1_recall": 0.002256034533050118, "rouge1_recall_stderr": 0.0006291747215958679, "rouge2_fmeasure": 0.0004511007812344588, "rouge2_fmeasure_stderr": 0.00018994877121643913, "rouge2_precision": 0.0005539957329635465, "rouge2_precision_stderr": 0.0002384720401159147, "rouge2_recall": 0.00039131204225543846, "rouge2_recall_stderr": 0.00016680843472868898, "rougeL_fmeasure": 0.0019230606581652292, "rougeL_fmeasure_stderr": 0.0005609337748918923, "rougeL_precision": 0.00225272205418534, "rougeL_precision_stderr": 0.0006884647595916458, "rougeL_recall": 0.001717413213020426, "rougeL_recall_stderr": 0.0004941575994563838, "rougeLsum_fmeasure": 0.0020240460078641293, "rougeLsum_fmeasure_stderr": 0.0005911504360672772, "rougeLsum_precision": 0.0023567731085693583, "rougeLsum_precision_stderr": 0.0007147557896729491, "rougeLsum_recall": 0.0018155084831467728, "rougeLsum_recall_stderr": 0.0005263836320326491}}}} |