File size: 28,111 Bytes
2aef930 |
1 |
{"GEM/web_nlg_en": {"0": {"PALM_prompt": {"bleu": 0.26787629279095865, "bleu_stderr": 0.02421388470865588, "rouge1_fmeasure": 0.07916714595011602, "rouge1_fmeasure_stderr": 0.0016165914559961, "rouge1_precision": 0.053140729220499526, "rouge1_precision_stderr": 0.001447561888642144, "rouge1_recall": 0.2650539620149166, "rouge1_recall_stderr": 0.004611089062355459, "rouge2_fmeasure": 0.03206245161358705, "rouge2_fmeasure_stderr": 0.0009944025304645106, "rouge2_precision": 0.021704917896300822, "rouge2_precision_stderr": 0.0009077529999546599, "rouge2_recall": 0.10624239698539174, "rouge2_recall_stderr": 0.0029454877090427634, "rougeL_fmeasure": 0.07256452976899888, "rougeL_fmeasure_stderr": 0.0014455163882477753, "rougeL_precision": 0.04859565080529351, "rougeL_precision_stderr": 0.0013030805071960487, "rougeL_recall": 0.24447365028169235, "rougeL_recall_stderr": 0.0042597620469032324, "rougeLsum_fmeasure": 0.07317950398891415, "rougeLsum_fmeasure_stderr": 0.0014738730467798152, "rougeLsum_precision": 0.04911887572948648, "rougeLsum_precision_stderr": 0.0013308973848181627, "rougeLsum_recall": 0.24529807807313986, "rougeLsum_recall_stderr": 0.004231057274690599}}, "1": {"PALM_prompt": {"bleu": 0.27346154994880517, "bleu_stderr": 0.02188738307683954, "rouge1_fmeasure": 0.07424112149389747, "rouge1_fmeasure_stderr": 0.0013991731147740165, "rouge1_precision": 0.047626171506349936, "rouge1_precision_stderr": 0.0011352906867889128, "rouge1_recall": 0.26035399188160574, "rouge1_recall_stderr": 0.004602505397267029, "rouge2_fmeasure": 0.028108438388606747, "rouge2_fmeasure_stderr": 0.0008209525600219419, "rouge2_precision": 0.01795733317688481, "rouge2_precision_stderr": 0.0006207760872719163, "rouge2_recall": 0.10255465178172037, "rouge2_recall_stderr": 0.0028918136354984434, "rougeL_fmeasure": 0.06664007036780817, "rougeL_fmeasure_stderr": 0.001225946991287631, "rougeL_precision": 0.04253166976477524, "rougeL_precision_stderr": 0.0009520378507813956, "rougeL_recall": 0.23570739197296856, "rougeL_recall_stderr": 0.004138563159116714, "rougeLsum_fmeasure": 0.06858259906736544, "rougeLsum_fmeasure_stderr": 0.001295048790148881, "rougeLsum_precision": 0.044154443741561675, "rougeLsum_precision_stderr": 0.0010805238879022037, "rougeLsum_recall": 0.24018067638000298, "rougeLsum_recall_stderr": 0.004203191799368478}}, "2": {"PALM_prompt": {"bleu": 0.28354223508209625, "bleu_stderr": 0.02196438336730763, "rouge1_fmeasure": 0.077057196050893, "rouge1_fmeasure_stderr": 0.001405552196581262, "rouge1_precision": 0.04907071394368477, "rouge1_precision_stderr": 0.0011814228439300668, "rouge1_recall": 0.2717922146470884, "rouge1_recall_stderr": 0.004679127784588707, "rouge2_fmeasure": 0.029195418367829688, "rouge2_fmeasure_stderr": 0.0008404145834110129, "rouge2_precision": 0.01847026867964925, "rouge2_precision_stderr": 0.0006271549211076853, "rouge2_recall": 0.10837787945721133, "rouge2_recall_stderr": 0.00292786577048572, "rougeL_fmeasure": 0.0683478121462329, "rougeL_fmeasure_stderr": 0.001219408641805486, "rougeL_precision": 0.043547243362200026, "rougeL_precision_stderr": 0.0010632111900203151, "rougeL_recall": 0.24217304687370458, "rougeL_recall_stderr": 0.0041386950242398935, "rougeLsum_fmeasure": 0.07117154749767428, "rougeLsum_fmeasure_stderr": 0.001278540354605195, "rougeLsum_precision": 0.04538438937137174, "rougeLsum_precision_stderr": 0.0010999742522302452, "rougeLsum_recall": 0.2509319606971942, "rougeLsum_recall_stderr": 0.004265166783401968}}, "3": {"PALM_prompt": {"bleu": 0.26231700778830913, "bleu_stderr": 0.01677381795342571, "rouge1_fmeasure": 0.07701202526647727, "rouge1_fmeasure_stderr": 0.0013264916492560968, "rouge1_precision": 0.048773010954172226, "rouge1_precision_stderr": 0.0010801404950998403, "rouge1_recall": 0.2723524322573744, "rouge1_recall_stderr": 0.004614644088480283, "rouge2_fmeasure": 0.028476680174906326, "rouge2_fmeasure_stderr": 0.0007831062590315599, "rouge2_precision": 0.01830389408829641, "rouge2_precision_stderr": 0.000765761231033471, "rouge2_recall": 0.10724024189536346, "rouge2_recall_stderr": 0.0029417688916459905, "rougeL_fmeasure": 0.06751038060872314, "rougeL_fmeasure_stderr": 0.0011522535843318086, "rougeL_precision": 0.04285349136828219, "rougeL_precision_stderr": 0.0009848666531091866, "rougeL_recall": 0.23931793383909106, "rougeL_recall_stderr": 0.0039933449779696656, "rougeLsum_fmeasure": 0.0713319417621141, "rougeLsum_fmeasure_stderr": 0.0012205962369251139, "rougeLsum_precision": 0.045262801919532414, "rougeLsum_precision_stderr": 0.001021866086917732, "rougeLsum_recall": 0.25210574997713586, "rougeLsum_recall_stderr": 0.004210876653447827}}, "4": {"PALM_prompt": {"bleu": 0.2909737297069181, "bleu_stderr": 0.016893325417647067, "rouge1_fmeasure": 0.07880513839875916, "rouge1_fmeasure_stderr": 0.0013577530127801877, "rouge1_precision": 0.05081036279065346, "rouge1_precision_stderr": 0.001340290442932179, "rouge1_recall": 0.27689930067824825, "rouge1_recall_stderr": 0.004592552339097503, "rouge2_fmeasure": 0.029911000513183756, "rouge2_fmeasure_stderr": 0.0008104665489277981, "rouge2_precision": 0.02000297688991883, "rouge2_precision_stderr": 0.0009665376160012339, "rouge2_recall": 0.1098959472411629, "rouge2_recall_stderr": 0.0029265187187869925, "rougeL_fmeasure": 0.06910981407367636, "rougeL_fmeasure_stderr": 0.001165837956941959, "rougeL_precision": 0.04469978216410197, "rougeL_precision_stderr": 0.001228415770141008, "rougeL_recall": 0.24325022171643784, "rougeL_recall_stderr": 0.003959401381097609, "rougeLsum_fmeasure": 0.07342750134261251, "rougeLsum_fmeasure_stderr": 0.0012548900975791317, "rougeLsum_precision": 0.047479319186822515, "rougeLsum_precision_stderr": 0.0012835127261851631, "rougeLsum_recall": 0.25785008091098, "rougeLsum_recall_stderr": 0.004217817751472416}}, "5": {"PALM_prompt": {"bleu": 0.275592913998383, "bleu_stderr": 0.010675634701516879, "rouge1_fmeasure": 0.07936270839708998, "rouge1_fmeasure_stderr": 0.0013461361721732127, "rouge1_precision": 0.050188036684923054, "rouge1_precision_stderr": 0.0010589097851446405, "rouge1_recall": 0.28261873177986385, "rouge1_recall_stderr": 0.004706167216581722, "rouge2_fmeasure": 0.029335540062254653, "rouge2_fmeasure_stderr": 0.00077573226541773, "rouge2_precision": 0.018414057787839, "rouge2_precision_stderr": 0.0005667544666951426, "rouge2_recall": 0.11149049346375932, "rouge2_recall_stderr": 0.002957507274568516, "rougeL_fmeasure": 0.06878175079936166, "rougeL_fmeasure_stderr": 0.0011305052098162837, "rougeL_precision": 0.04348160210158506, "rougeL_precision_stderr": 0.0009025307067376957, "rougeL_recall": 0.2468060892224703, "rougeL_recall_stderr": 0.004013696328255729, "rougeLsum_fmeasure": 0.07315850261422062, "rougeLsum_fmeasure_stderr": 0.0012299656434355821, "rougeLsum_precision": 0.046368846883703685, "rougeLsum_precision_stderr": 0.0009967098378676012, "rougeLsum_recall": 0.25991163700009623, "rougeLsum_recall_stderr": 0.0042359736213800495}}}, "GEM/wiki_lingua_en": {"0": {"tldr_en": {"bleu": 1.4498797804800554, "bleu_stderr": 0.030033155430787307, "rouge1_fmeasure": 0.17351939045996323, "rouge1_fmeasure_stderr": 0.0018292808999788056, "rouge1_precision": 0.1508638357497856, "rouge1_precision_stderr": 0.0019477340479700028, "rouge1_recall": 0.24849466528414227, "rouge1_recall_stderr": 0.0025701642988402284, "rouge2_fmeasure": 0.033112770233475294, "rouge2_fmeasure_stderr": 0.0007842583345225733, "rouge2_precision": 0.028565167813151382, "rouge2_precision_stderr": 0.0007213658899061881, "rouge2_recall": 0.04860070927724722, "rouge2_recall_stderr": 0.0012792940293872988, "rougeL_fmeasure": 0.1300215852768387, "rougeL_fmeasure_stderr": 0.0012668686976567704, "rougeL_precision": 0.11205417400994366, "rougeL_precision_stderr": 0.001364543594879335, "rougeL_recall": 0.19083146969499562, "rougeL_recall_stderr": 0.0020249121235725457, "rougeLsum_fmeasure": 0.16176041579500514, "rougeLsum_fmeasure_stderr": 0.0016905736140852296, "rougeLsum_precision": 0.14039987541721508, "rougeLsum_precision_stderr": 0.001794428495675629, "rougeLsum_recall": 0.23240706811113432, "rougeLsum_recall_stderr": 0.002406095416292502}}, "1": {"tldr_en": {"bleu": 1.5523030824089203, "bleu_stderr": 0.0404745781112937, "rouge1_fmeasure": 0.17598744290815144, "rouge1_fmeasure_stderr": 0.001808190292239866, "rouge1_precision": 0.1526735082069478, "rouge1_precision_stderr": 0.001954192368311485, "rouge1_recall": 0.2537630901501927, "rouge1_recall_stderr": 0.0025247263534353263, "rouge2_fmeasure": 0.032548106421086444, "rouge2_fmeasure_stderr": 0.0007854290741659407, "rouge2_precision": 0.02844271878006758, "rouge2_precision_stderr": 0.0007501808391968073, "rouge2_recall": 0.048050899811526195, "rouge2_recall_stderr": 0.0013077335984725508, "rougeL_fmeasure": 0.1291623611792601, "rougeL_fmeasure_stderr": 0.0012144030507005796, "rougeL_precision": 0.11085969871259992, "rougeL_precision_stderr": 0.0013092492625755249, "rougeL_recall": 0.1910548183701386, "rougeL_recall_stderr": 0.0019406911815812453, "rougeLsum_fmeasure": 0.1656292822984552, "rougeLsum_fmeasure_stderr": 0.0016884673231307584, "rougeLsum_precision": 0.14352989085571205, "rougeLsum_precision_stderr": 0.0018308845070724087, "rougeLsum_recall": 0.23982592472606548, "rougeLsum_recall_stderr": 0.0024037588777864357}}, "2": {"tldr_en": {"bleu": 1.5504469719150438, "bleu_stderr": 0.046981957687555205, "rouge1_fmeasure": 0.17851912473986262, "rouge1_fmeasure_stderr": 0.0017588268154772792, "rouge1_precision": 0.15614318865894197, "rouge1_precision_stderr": 0.001953522003890445, "rouge1_recall": 0.2564233598947695, "rouge1_recall_stderr": 0.002443975404517576, "rouge2_fmeasure": 0.032531974500306655, "rouge2_fmeasure_stderr": 0.0007707378614567543, "rouge2_precision": 0.028786254184172754, "rouge2_precision_stderr": 0.0007381734830468564, "rouge2_recall": 0.04721098878117776, "rouge2_recall_stderr": 0.0012484536682449941, "rougeL_fmeasure": 0.13175959208331914, "rougeL_fmeasure_stderr": 0.00119267469976916, "rougeL_precision": 0.11436547586059817, "rougeL_precision_stderr": 0.0013634627836511027, "rougeL_recall": 0.19403348190827804, "rougeL_recall_stderr": 0.001915506475404068, "rougeLsum_fmeasure": 0.16809587331436024, "rougeLsum_fmeasure_stderr": 0.001653116049291498, "rougeLsum_precision": 0.14706172179649213, "rougeLsum_precision_stderr": 0.0018475712783318878, "rougeLsum_recall": 0.24190155350034248, "rougeLsum_recall_stderr": 0.0023184025713278256}}, "3": {"tldr_en": {"bleu": 1.6506772563162972, "bleu_stderr": 0.07202666343435976, "rouge1_fmeasure": 0.14977645750144936, "rouge1_fmeasure_stderr": 0.00199726722683507, "rouge1_precision": 0.13699405852667912, "rouge1_precision_stderr": 0.0022271962965683227, "rouge1_recall": 0.21313164031151458, "rouge1_recall_stderr": 0.0028175993079765913, "rouge2_fmeasure": 0.027873427098971903, "rouge2_fmeasure_stderr": 0.0007427085100927844, "rouge2_precision": 0.025286085868956577, "rouge2_precision_stderr": 0.0007388515719204758, "rouge2_recall": 0.04046828319806029, "rouge2_recall_stderr": 0.001196737089891652, "rougeL_fmeasure": 0.11096435645199118, "rougeL_fmeasure_stderr": 0.0014009503741229052, "rougeL_precision": 0.10087113357194523, "rougeL_precision_stderr": 0.0016060316815470442, "rougeL_recall": 0.1621097484851492, "rougeL_recall_stderr": 0.0021946464603937125, "rougeLsum_fmeasure": 0.1408419182005074, "rougeLsum_fmeasure_stderr": 0.0018690086358082113, "rougeLsum_precision": 0.1288247035213342, "rougeLsum_precision_stderr": 0.0020910254870389346, "rougeLsum_recall": 0.2007771126062864, "rougeLsum_recall_stderr": 0.0026617775847568633}}, "4": {"tldr_en": {"bleu": 0.35442363598057186, "bleu_stderr": 0.02957890609663373, "rouge1_fmeasure": 0.04990479829313592, "rouge1_fmeasure_stderr": 0.0017225289026089794, "rouge1_precision": 0.04727049455418174, "rouge1_precision_stderr": 0.0018100262381206956, "rouge1_recall": 0.07308789496556233, "rouge1_recall_stderr": 0.002548253609808092, "rouge2_fmeasure": 0.009704784607524384, "rouge2_fmeasure_stderr": 0.0005246010118344956, "rouge2_precision": 0.009152769715333476, "rouge2_precision_stderr": 0.0005809882470945591, "rouge2_recall": 0.014793686682183805, "rouge2_recall_stderr": 0.0008871836496692909, "rougeL_fmeasure": 0.038102997687537214, "rougeL_fmeasure_stderr": 0.0012904657107170763, "rougeL_precision": 0.03634445955852837, "rougeL_precision_stderr": 0.0014097418513319498, "rougeL_recall": 0.05712304517098294, "rougeL_recall_stderr": 0.002020929141181275, "rougeLsum_fmeasure": 0.04692854753900321, "rougeLsum_fmeasure_stderr": 0.001615455808763454, "rougeLsum_precision": 0.04450904725651575, "rougeLsum_precision_stderr": 0.0017072089167818546, "rougeLsum_recall": 0.06907832245228929, "rougeLsum_recall_stderr": 0.002416204133617309}}, "5": {"tldr_en": {"bleu": 2.3773909381181067e-07, "bleu_stderr": 6.777702484631335e-07, "rouge1_fmeasure": 0.007934022472178462, "rouge1_fmeasure_stderr": 0.0007646898711388704, "rouge1_precision": 0.007806943563100253, "rouge1_precision_stderr": 0.0008029596223294583, "rouge1_recall": 0.011916878778381854, "rouge1_recall_stderr": 0.0011710685254217557, "rouge2_fmeasure": 0.0014852201283601309, "rouge2_fmeasure_stderr": 0.00020361223200698468, "rouge2_precision": 0.0015499616053509758, "rouge2_precision_stderr": 0.00024432717225911417, "rouge2_recall": 0.0022995865319653413, "rouge2_recall_stderr": 0.00033582149369530104, "rougeL_fmeasure": 0.006092063289579814, "rougeL_fmeasure_stderr": 0.0005734478033484189, "rougeL_precision": 0.006059612662075617, "rougeL_precision_stderr": 0.0006187644415825575, "rougeL_recall": 0.009361683565350928, "rougeL_recall_stderr": 0.0009310373409583372, "rougeLsum_fmeasure": 0.007474906777946338, "rougeLsum_fmeasure_stderr": 0.0007195818405089297, "rougeLsum_precision": 0.0073509618326022316, "rougeLsum_precision_stderr": 0.0007548623257902688, "rougeLsum_recall": 0.011247516251388902, "rougeLsum_recall_stderr": 0.001102260125459754}}}, "e2e_nlg_cleaned": {"0": {"generate_text_restaurant": {"bleu": 0.8482338536273092, "bleu_stderr": 0.06579534941253831, "rouge1_fmeasure": 0.029524410502349083, "rouge1_fmeasure_stderr": 0.0012194408459950685, "rouge1_precision": 0.03689992434617125, "rouge1_precision_stderr": 0.00143080390296388, "rouge1_recall": 0.03176166527004807, "rouge1_recall_stderr": 0.0015278801549138494, "rouge2_fmeasure": 0.006282252522560226, "rouge2_fmeasure_stderr": 0.0005083233271996686, "rouge2_precision": 0.005295864917355121, "rouge2_precision_stderr": 0.000436477995087522, "rouge2_recall": 0.008570227243564495, "rouge2_recall_stderr": 0.0006945659775254231, "rougeL_fmeasure": 0.02704266874243934, "rougeL_fmeasure_stderr": 0.001092799039388727, "rougeL_precision": 0.034035125364177875, "rougeL_precision_stderr": 0.0013033143404442467, "rougeL_recall": 0.02894660702654024, "rougeL_recall_stderr": 0.0013685312959876327, "rougeLsum_fmeasure": 0.027093010874833397, "rougeLsum_fmeasure_stderr": 0.0011035568012963972, "rougeLsum_precision": 0.034329573061074765, "rougeLsum_precision_stderr": 0.0013151068156701008, "rougeLsum_recall": 0.028756511365956524, "rougeLsum_recall_stderr": 0.0013753740219635173}}, "1": {"generate_text_restaurant": {"bleu": 5.942260098167613, "bleu_stderr": 0.09441092756090216, "rouge1_fmeasure": 0.3304930105126096, "rouge1_fmeasure_stderr": 0.0020464079472463886, "rouge1_precision": 0.3146822903840917, "rouge1_precision_stderr": 0.0026228566349168303, "rouge1_recall": 0.4073067647472486, "rouge1_recall_stderr": 0.002694916559116025, "rouge2_fmeasure": 0.12621157789802925, "rouge2_fmeasure_stderr": 0.00144122341707661, "rouge2_precision": 0.11973672411480318, "rouge2_precision_stderr": 0.0017049415225723842, "rouge2_recall": 0.15858199335292084, "rouge2_recall_stderr": 0.001880021556117094, "rougeL_fmeasure": 0.24390496755077615, "rougeL_fmeasure_stderr": 0.0014758735639314516, "rougeL_precision": 0.23121444753390857, "rougeL_precision_stderr": 0.0019617352937841724, "rougeL_recall": 0.30603012289790815, "rougeL_recall_stderr": 0.002249710012043417, "rougeLsum_fmeasure": 0.27519248986257294, "rougeLsum_fmeasure_stderr": 0.0018786355388105225, "rougeLsum_precision": 0.26253682646143983, "rougeLsum_precision_stderr": 0.0023660383137224746, "rougeLsum_recall": 0.33971109353202855, "rougeLsum_recall_stderr": 0.002493554468070354}}, "2": {"generate_text_restaurant": {"bleu": 6.725052864934353, "bleu_stderr": 0.14629714654304157, "rouge1_fmeasure": 0.3455500221300751, "rouge1_fmeasure_stderr": 0.0020567847742166303, "rouge1_precision": 0.3370432420794776, "rouge1_precision_stderr": 0.0029582437787025427, "rouge1_recall": 0.41475409018206294, "rouge1_recall_stderr": 0.002581973755669344, "rouge2_fmeasure": 0.14215332459928473, "rouge2_fmeasure_stderr": 0.0015526430769182552, "rouge2_precision": 0.13924246863995587, "rouge2_precision_stderr": 0.0018935171496830272, "rouge2_recall": 0.17241918013486687, "rouge2_recall_stderr": 0.0019222974911274624, "rougeL_fmeasure": 0.26044817619281624, "rougeL_fmeasure_stderr": 0.0015920982446906947, "rougeL_precision": 0.25285247263136995, "rougeL_precision_stderr": 0.002281200773310861, "rougeL_recall": 0.31635151085258584, "rougeL_recall_stderr": 0.002181518712740274, "rougeLsum_fmeasure": 0.29316637284801744, "rougeLsum_fmeasure_stderr": 0.0019474180340165936, "rougeLsum_precision": 0.28533453436961675, "rougeLsum_precision_stderr": 0.0026174427606121924, "rougeLsum_recall": 0.35223597019269687, "rougeLsum_recall_stderr": 0.002458693422243018}}, "3": {"generate_text_restaurant": {"bleu": 6.973175823947467, "bleu_stderr": 0.08432471729881086, "rouge1_fmeasure": 0.3491442563303595, "rouge1_fmeasure_stderr": 0.002084406661790611, "rouge1_precision": 0.3372869158678504, "rouge1_precision_stderr": 0.002968659112264888, "rouge1_recall": 0.41917585622339343, "rouge1_recall_stderr": 0.0024657865682947604, "rouge2_fmeasure": 0.14925241670212142, "rouge2_fmeasure_stderr": 0.0015639781562393712, "rouge2_precision": 0.1451623356626606, "rouge2_precision_stderr": 0.001878882685675287, "rouge2_recall": 0.1805945511839357, "rouge2_recall_stderr": 0.001899662372461791, "rougeL_fmeasure": 0.2637903997625913, "rougeL_fmeasure_stderr": 0.0016349236936735077, "rougeL_precision": 0.25377095839457386, "rougeL_precision_stderr": 0.0023013165756289485, "rougeL_recall": 0.31994498910057845, "rougeL_recall_stderr": 0.0020988836643955253, "rougeLsum_fmeasure": 0.29641103594998963, "rougeLsum_fmeasure_stderr": 0.0019700089086597617, "rougeLsum_precision": 0.2861590985165946, "rougeLsum_precision_stderr": 0.0026614220152062383, "rougeLsum_recall": 0.35633299447456107, "rougeLsum_recall_stderr": 0.00237175789815616}}, "4": {"generate_text_restaurant": {"bleu": 6.781202241557229, "bleu_stderr": 0.11394876149110739, "rouge1_fmeasure": 0.34660270706267965, "rouge1_fmeasure_stderr": 0.002015167448860447, "rouge1_precision": 0.3287498306441592, "rouge1_precision_stderr": 0.002864037922185099, "rouge1_recall": 0.4242438979853323, "rouge1_recall_stderr": 0.002418778880597351, "rouge2_fmeasure": 0.1484646710112249, "rouge2_fmeasure_stderr": 0.0015009486601977229, "rouge2_precision": 0.14188074782669924, "rouge2_precision_stderr": 0.0018111547132663394, "rouge2_recall": 0.18330145095889133, "rouge2_recall_stderr": 0.001867227120259638, "rougeL_fmeasure": 0.2623953647532703, "rougeL_fmeasure_stderr": 0.0015873390211834876, "rougeL_precision": 0.24760447580397063, "rougeL_precision_stderr": 0.002205771870408524, "rougeL_recall": 0.3245629843061881, "rougeL_recall_stderr": 0.002091486936788564, "rougeLsum_fmeasure": 0.2952178835224044, "rougeLsum_fmeasure_stderr": 0.0019315028392222534, "rougeLsum_precision": 0.2799336817988539, "rougeLsum_precision_stderr": 0.0025898070057799047, "rougeLsum_recall": 0.36159803652552097, "rougeLsum_recall_stderr": 0.002354029382709626}}, "5": {"generate_text_restaurant": {"bleu": 6.685901964010736, "bleu_stderr": 0.1354495963808724, "rouge1_fmeasure": 0.3437737218331649, "rouge1_fmeasure_stderr": 0.0019546229382639466, "rouge1_precision": 0.32176357670299083, "rouge1_precision_stderr": 0.0027967903182063224, "rouge1_recall": 0.4272486932644093, "rouge1_recall_stderr": 0.0023809505635255927, "rouge2_fmeasure": 0.14683833434343727, "rouge2_fmeasure_stderr": 0.0014540710021065215, "rouge2_precision": 0.13815978485481578, "rouge2_precision_stderr": 0.0017518294091340995, "rouge2_recall": 0.18494316023841778, "rouge2_recall_stderr": 0.0018444532176795417, "rougeL_fmeasure": 0.2596864980022702, "rougeL_fmeasure_stderr": 0.0015675414220731314, "rougeL_precision": 0.2417191397949724, "rougeL_precision_stderr": 0.00214960916866614, "rougeL_recall": 0.32601659789964693, "rougeL_recall_stderr": 0.0020925727851936926, "rougeLsum_fmeasure": 0.29423382905586287, "rougeLsum_fmeasure_stderr": 0.0018901601807881775, "rougeLsum_precision": 0.27513062879919625, "rougeLsum_precision_stderr": 0.0025215646737994962, "rougeLsum_recall": 0.3659104174188377, "rougeLsum_recall_stderr": 0.0023445078388895997}}}, "gem_xsum": {"0": {"article_DOC_summary": {"bleu": 1.445573424695658, "bleu_stderr": 0.15425934770761948, "rouge1_fmeasure": 0.19381060522505453, "rouge1_fmeasure_stderr": 0.0022326458513688033, "rouge1_precision": 0.14337493474921045, "rouge1_precision_stderr": 0.0018970632275809294, "rouge1_recall": 0.32850970396744716, "rouge1_recall_stderr": 0.0039197450040575135, "rouge2_fmeasure": 0.03756311678020454, "rouge2_fmeasure_stderr": 0.001356763612763011, "rouge2_precision": 0.027496214834546444, "rouge2_precision_stderr": 0.001054159884756719, "rouge2_recall": 0.06556453127883762, "rouge2_recall_stderr": 0.0023771544908032247, "rougeL_fmeasure": 0.14442948981860032, "rougeL_fmeasure_stderr": 0.0016582345165837742, "rougeL_precision": 0.10679400111516424, "rougeL_precision_stderr": 0.0014304369917064012, "rougeL_recall": 0.2459644623091805, "rougeL_recall_stderr": 0.002986298895539128, "rougeLsum_fmeasure": 0.15408798170242552, "rougeLsum_fmeasure_stderr": 0.0018796051363025486, "rougeLsum_precision": 0.11360184621025315, "rougeLsum_precision_stderr": 0.0015482924171113445, "rougeLsum_recall": 0.26296431121251945, "rougeLsum_recall_stderr": 0.003409100925773213}}, "1": {"article_DOC_summary": {"bleu": 0.6954014401170389, "bleu_stderr": 0.0939232317796097, "rouge1_fmeasure": 0.1513724503049335, "rouge1_fmeasure_stderr": 0.0020050978617654432, "rouge1_precision": 0.10765659328853563, "rouge1_precision_stderr": 0.0015212603420465503, "rouge1_recall": 0.2657263624255442, "rouge1_recall_stderr": 0.0033066001915549596, "rouge2_fmeasure": 0.01966067423050901, "rouge2_fmeasure_stderr": 0.0009461105744381743, "rouge2_precision": 0.013942450754043968, "rouge2_precision_stderr": 0.0006842372730185846, "rouge2_recall": 0.03481401056485645, "rouge2_recall_stderr": 0.0016493879352824244, "rougeL_fmeasure": 0.11639992980160706, "rougeL_fmeasure_stderr": 0.0014853662011968952, "rougeL_precision": 0.08261246482339596, "rougeL_precision_stderr": 0.0011168888732104218, "rougeL_recall": 0.20552675444508914, "rougeL_recall_stderr": 0.0025401789847852084, "rougeLsum_fmeasure": 0.12287565931519653, "rougeLsum_fmeasure_stderr": 0.0016075317103237032, "rougeLsum_precision": 0.08719479126054427, "rougeLsum_precision_stderr": 0.0012099004109665159, "rougeLsum_recall": 0.21714465186994203, "rougeLsum_recall_stderr": 0.0027534050196141966}}, "2": {"article_DOC_summary": {"bleu": 0.7307303414586408, "bleu_stderr": 0.06155547874417727, "rouge1_fmeasure": 0.15344534516159183, "rouge1_fmeasure_stderr": 0.0020053949533959292, "rouge1_precision": 0.10867457736153135, "rouge1_precision_stderr": 0.001495555344909094, "rouge1_recall": 0.27143929169611797, "rouge1_recall_stderr": 0.003394856674159156, "rouge2_fmeasure": 0.02144461765365429, "rouge2_fmeasure_stderr": 0.0009703360723919554, "rouge2_precision": 0.01507662196122561, "rouge2_precision_stderr": 0.0006861991351095634, "rouge2_recall": 0.03867163671633521, "rouge2_recall_stderr": 0.0017747285897281744, "rougeL_fmeasure": 0.11854204244373989, "rougeL_fmeasure_stderr": 0.0014924813777114118, "rougeL_precision": 0.08378473061167403, "rougeL_precision_stderr": 0.0011056093599254957, "rougeL_recall": 0.21115579128906609, "rougeL_recall_stderr": 0.002637225983534165, "rougeLsum_fmeasure": 0.12557130842107206, "rougeLsum_fmeasure_stderr": 0.0016648364713290837, "rougeLsum_precision": 0.08877022608536633, "rougeLsum_precision_stderr": 0.0012265040039533249, "rougeLsum_recall": 0.22324911827513313, "rougeLsum_recall_stderr": 0.0029138573801630453}}, "3": {"article_DOC_summary": {"bleu": 0.7723989217062374, "bleu_stderr": 0.07506332455732345, "rouge1_fmeasure": 0.1459946014022398, "rouge1_fmeasure_stderr": 0.002181990692411414, "rouge1_precision": 0.10592564096937784, "rouge1_precision_stderr": 0.0017525879317443233, "rouge1_recall": 0.2539427012445784, "rouge1_recall_stderr": 0.003731447303357977, "rouge2_fmeasure": 0.020061440587479485, "rouge2_fmeasure_stderr": 0.0009777698919290968, "rouge2_precision": 0.01420534449614851, "rouge2_precision_stderr": 0.0007006470449943444, "rouge2_recall": 0.03590630051582267, "rouge2_recall_stderr": 0.0017733541653792099, "rougeL_fmeasure": 0.11265406503599139, "rougeL_fmeasure_stderr": 0.001630908636011174, "rougeL_precision": 0.08153731863940099, "rougeL_precision_stderr": 0.0012887659869021155, "rougeL_recall": 0.19693769942475764, "rougeL_recall_stderr": 0.002869678698254578, "rougeLsum_fmeasure": 0.11915978365542443, "rougeLsum_fmeasure_stderr": 0.0017414042260366922, "rougeLsum_precision": 0.08628916736533274, "rougeLsum_precision_stderr": 0.0013712543798638922, "rougeLsum_recall": 0.2081960052350543, "rougeLsum_recall_stderr": 0.0030602079181311892}}, "4": {"article_DOC_summary": {"bleu": 0.4202554532269484, "bleu_stderr": 0.12262307961277104, "rouge1_fmeasure": 0.040215681107462056, "rouge1_fmeasure_stderr": 0.002237517556474153, "rouge1_precision": 0.034471583104804415, "rouge1_precision_stderr": 0.002238315408267895, "rouge1_recall": 0.06381728541655798, "rouge1_recall_stderr": 0.0036282870714125083, "rouge2_fmeasure": 0.004935566104695442, "rouge2_fmeasure_stderr": 0.0005753566907154545, "rouge2_precision": 0.003871537312889549, "rouge2_precision_stderr": 0.0005719133532575095, "rouge2_recall": 0.008269838331766569, "rouge2_recall_stderr": 0.0009564670794954605, "rougeL_fmeasure": 0.030441951036000715, "rougeL_fmeasure_stderr": 0.0016864228566178527, "rougeL_precision": 0.026758973645767164, "rougeL_precision_stderr": 0.0018801400712648681, "rougeL_recall": 0.04838251241637687, "rougeL_recall_stderr": 0.0027529301388328257, "rougeLsum_fmeasure": 0.03296335745988539, "rougeLsum_fmeasure_stderr": 0.001842268360788242, "rougeLsum_precision": 0.028758380610237905, "rougeLsum_precision_stderr": 0.0019838383896546154, "rougeLsum_recall": 0.052394609604491466, "rougeLsum_recall_stderr": 0.00300800960671939}}, "5": {"article_DOC_summary": {"bleu": 1.2660127317243813e-38, "bleu_stderr": 3.8267567150622576e-33, "rouge1_fmeasure": 0.002135746093866511, "rouge1_fmeasure_stderr": 0.0005838656686162581, "rouge1_precision": 0.002427203615941276, "rouge1_precision_stderr": 0.0006831484429133202, "rouge1_recall": 0.0019840191933539033, "rouge1_recall_stderr": 0.0005303913662852268, "rouge2_fmeasure": 0.00030743447999977873, "rouge2_fmeasure_stderr": 0.00013023642706646356, "rouge2_precision": 0.0003678072143665929, "rouge2_precision_stderr": 0.00015846001855024936, "rouge2_recall": 0.0002682301974754805, "rouge2_recall_stderr": 0.0001131215031105469, "rougeL_fmeasure": 0.001686259301474979, "rougeL_fmeasure_stderr": 0.0004921843202278661, "rougeL_precision": 0.0019066685986482705, "rougeL_precision_stderr": 0.0005793006239621258, "rougeL_recall": 0.001578844696073063, "rougeL_recall_stderr": 0.00044617009211618964, "rougeLsum_fmeasure": 0.001735086296720847, "rougeLsum_fmeasure_stderr": 0.0004903196095708464, "rougeLsum_precision": 0.0019704357925048386, "rougeLsum_precision_stderr": 0.0005792873025022089, "rougeLsum_recall": 0.00161831401010321, "rougeLsum_recall_stderr": 0.0004437097896608957}}}} |