File size: 28,011 Bytes
8263782 |
1 |
{"GEM/web_nlg_en": {"0": {"PALM_prompt": {"bleu": 0.32917980943922837, "bleu_stderr": 0.028436979522399264, "rouge1_fmeasure": 0.11092359878172349, "rouge1_fmeasure_stderr": 0.0021083743244835647, "rouge1_precision": 0.07454038083418474, "rouge1_precision_stderr": 0.001766454284632946, "rouge1_recall": 0.30290353802392916, "rouge1_recall_stderr": 0.004835002819067963, "rouge2_fmeasure": 0.05212321093989505, "rouge2_fmeasure_stderr": 0.0013412640798549406, "rouge2_precision": 0.03405534335780326, "rouge2_precision_stderr": 0.0009795255375204467, "rouge2_recall": 0.14725829338052365, "rouge2_recall_stderr": 0.0033496237842817397, "rougeL_fmeasure": 0.1059469350929842, "rougeL_fmeasure_stderr": 0.0019397934888169463, "rougeL_precision": 0.07086594439969038, "rougeL_precision_stderr": 0.0016082892660627373, "rougeL_recall": 0.29226896530687846, "rougeL_recall_stderr": 0.00466575095416219, "rougeLsum_fmeasure": 0.1060475373688485, "rougeLsum_fmeasure_stderr": 0.0019809059744097615, "rougeLsum_precision": 0.0711343154862909, "rougeLsum_precision_stderr": 0.001649638611842115, "rougeLsum_recall": 0.2903231316909427, "rougeLsum_recall_stderr": 0.004582555178091478}}, "1": {"PALM_prompt": {"bleu": 0.5285748263104777, "bleu_stderr": 0.05154539253732287, "rouge1_fmeasure": 0.1218665522358165, "rouge1_fmeasure_stderr": 0.0019158434392141466, "rouge1_precision": 0.07826126347996014, "rouge1_precision_stderr": 0.001412323883175065, "rouge1_recall": 0.38231102690350377, "rouge1_recall_stderr": 0.0053881864820272286, "rouge2_fmeasure": 0.058822699744876736, "rouge2_fmeasure_stderr": 0.0012405607498159584, "rouge2_precision": 0.037391127717872924, "rouge2_precision_stderr": 0.0008664479676328054, "rouge2_recall": 0.19839458948033734, "rouge2_recall_stderr": 0.004010973785657583, "rougeL_fmeasure": 0.11559328292690177, "rougeL_fmeasure_stderr": 0.001764131335387481, "rougeL_precision": 0.07407975935897594, "rougeL_precision_stderr": 0.0012854765812861943, "rougeL_recall": 0.3629315958992141, "rougeL_recall_stderr": 0.005069344323262684, "rougeLsum_fmeasure": 0.11544948693652962, "rougeLsum_fmeasure_stderr": 0.0017842924340211324, "rougeLsum_precision": 0.07415121638849131, "rougeLsum_precision_stderr": 0.001318142684082723, "rougeLsum_recall": 0.3614370946625198, "rougeLsum_recall_stderr": 0.004964006794864235}}, "2": {"PALM_prompt": {"bleu": 0.6066453441760302, "bleu_stderr": 0.02113523644793951, "rouge1_fmeasure": 0.12529215862638043, "rouge1_fmeasure_stderr": 0.0018259072915703836, "rouge1_precision": 0.08017864305448325, "rouge1_precision_stderr": 0.0014244751923494794, "rouge1_recall": 0.40480204267227765, "rouge1_recall_stderr": 0.005138963278066339, "rouge2_fmeasure": 0.059482069465634994, "rouge2_fmeasure_stderr": 0.0011595782202414246, "rouge2_precision": 0.03746116833011539, "rouge2_precision_stderr": 0.0008054382500683709, "rouge2_recall": 0.2078327469176558, "rouge2_recall_stderr": 0.003905498610846629, "rougeL_fmeasure": 0.11765794068577944, "rougeL_fmeasure_stderr": 0.0016600536639441099, "rougeL_precision": 0.0752258255540516, "rougeL_precision_stderr": 0.0013017426600869019, "rougeL_recall": 0.37921672738638584, "rougeL_recall_stderr": 0.004724125031038082, "rougeLsum_fmeasure": 0.11886578443137756, "rougeLsum_fmeasure_stderr": 0.001710265056607144, "rougeLsum_precision": 0.07612402487838264, "rougeLsum_precision_stderr": 0.0013499817758363393, "rougeLsum_recall": 0.3833520840989382, "rougeLsum_recall_stderr": 0.004787520387833171}}, "3": {"PALM_prompt": {"bleu": 0.6647933502215259, "bleu_stderr": 0.04923433986433192, "rouge1_fmeasure": 0.12320395767637426, "rouge1_fmeasure_stderr": 0.0017848979409527468, "rouge1_precision": 0.0781723421483203, "rouge1_precision_stderr": 0.0013023299811420648, "rouge1_recall": 0.4046567231900049, "rouge1_recall_stderr": 0.005137890673994109, "rouge2_fmeasure": 0.05931767519712951, "rouge2_fmeasure_stderr": 0.0011649351056668448, "rouge2_precision": 0.037339890555931926, "rouge2_precision_stderr": 0.0008144074272058679, "rouge2_recall": 0.2106193724526676, "rouge2_recall_stderr": 0.003927640735394401, "rougeL_fmeasure": 0.1150473458151772, "rougeL_fmeasure_stderr": 0.001619592187307252, "rougeL_precision": 0.07298298160971306, "rougeL_precision_stderr": 0.0011729650830812643, "rougeL_recall": 0.37541756571956564, "rougeL_recall_stderr": 0.004598099062061018, "rougeLsum_fmeasure": 0.1167920444605632, "rougeLsum_fmeasure_stderr": 0.0016685972133379994, "rougeLsum_precision": 0.0741403065968848, "rougeLsum_precision_stderr": 0.0012209548996141538, "rougeLsum_recall": 0.38296199684709586, "rougeLsum_recall_stderr": 0.004753402389332228}}, "4": {"PALM_prompt": {"bleu": 0.6566656848041873, "bleu_stderr": 0.03433459283803238, "rouge1_fmeasure": 0.12360753221007545, "rouge1_fmeasure_stderr": 0.001737583890202225, "rouge1_precision": 0.07771155841369212, "rouge1_precision_stderr": 0.0012374008770267796, "rouge1_recall": 0.4160226119034251, "rouge1_recall_stderr": 0.005178203274221526, "rouge2_fmeasure": 0.05886517252410502, "rouge2_fmeasure_stderr": 0.0011242425541824303, "rouge2_precision": 0.036708952345678336, "rouge2_precision_stderr": 0.0007671406759682118, "rouge2_recall": 0.2142530872032143, "rouge2_recall_stderr": 0.003938344188890309, "rougeL_fmeasure": 0.11421873690833906, "rougeL_fmeasure_stderr": 0.0015701261033135242, "rougeL_precision": 0.07189004426671647, "rougeL_precision_stderr": 0.001119612418604389, "rougeL_recall": 0.381893221081078, "rougeL_recall_stderr": 0.004577229345577291, "rougeLsum_fmeasure": 0.11777877062717519, "rougeLsum_fmeasure_stderr": 0.0016474868402205588, "rougeLsum_precision": 0.0741079421040528, "rougeLsum_precision_stderr": 0.0011766632070640108, "rougeLsum_recall": 0.39536320623753024, "rougeLsum_recall_stderr": 0.004834163190323313}}, "5": {"PALM_prompt": {"bleu": 0.7608515369903156, "bleu_stderr": 0.05544338902785279, "rouge1_fmeasure": 0.12542945154484628, "rouge1_fmeasure_stderr": 0.0017246642532628591, "rouge1_precision": 0.07874251282939974, "rouge1_precision_stderr": 0.0012390781355337446, "rouge1_recall": 0.4274915903037139, "rouge1_recall_stderr": 0.0052178214889688, "rouge2_fmeasure": 0.059919950381812144, "rouge2_fmeasure_stderr": 0.0011224239943820847, "rouge2_precision": 0.03725892588478996, "rouge2_precision_stderr": 0.0007691554935322187, "rouge2_recall": 0.22193078536394883, "rouge2_recall_stderr": 0.003999018222823969, "rougeL_fmeasure": 0.11503193912763204, "rougeL_fmeasure_stderr": 0.0015332329154476773, "rougeL_precision": 0.07225670171843512, "rougeL_precision_stderr": 0.001101933354024317, "rougeL_recall": 0.390252551254915, "rougeL_recall_stderr": 0.004591599013709285, "rougeLsum_fmeasure": 0.11871364855558929, "rougeLsum_fmeasure_stderr": 0.0016202511632686142, "rougeLsum_precision": 0.07457349456330893, "rougeLsum_precision_stderr": 0.0011662890645907052, "rougeLsum_recall": 0.40390693081462914, "rougeLsum_recall_stderr": 0.004850186326089204}}}, "GEM/wiki_lingua_en": {"0": {"tldr_en": {"bleu": 1.7311283501769414, "bleu_stderr": 0.07292848467922458, "rouge1_fmeasure": 0.18112864366064585, "rouge1_fmeasure_stderr": 0.0019261005266957636, "rouge1_precision": 0.15412231045043578, "rouge1_precision_stderr": 0.0019133423603737852, "rouge1_recall": 0.26596610436897833, "rouge1_recall_stderr": 0.0029078617781084343, "rouge2_fmeasure": 0.03900064299554168, "rouge2_fmeasure_stderr": 0.0009076451372655918, "rouge2_precision": 0.032745788295643126, "rouge2_precision_stderr": 0.000790295587407252, "rouge2_recall": 0.06015567293207339, "rouge2_recall_stderr": 0.0015850260882465787, "rougeL_fmeasure": 0.13940156755327857, "rougeL_fmeasure_stderr": 0.0013668263941440067, "rougeL_precision": 0.1171936085278983, "rougeL_precision_stderr": 0.00132039683220814, "rougeL_recall": 0.21005905052835785, "rougeL_recall_stderr": 0.002357155817310683, "rougeLsum_fmeasure": 0.16698321861271978, "rougeLsum_fmeasure_stderr": 0.0017578260329471065, "rougeLsum_precision": 0.14187748986714835, "rougeLsum_precision_stderr": 0.0017420221740131032, "rougeLsum_recall": 0.2462694168340833, "rougeLsum_recall_stderr": 0.002709379098155295}}, "1": {"tldr_en": {"bleu": 3.5922679202027847, "bleu_stderr": 0.05984463327930867, "rouge1_fmeasure": 0.24327622952771533, "rouge1_fmeasure_stderr": 0.002037288100861591, "rouge1_precision": 0.2123430992404537, "rouge1_precision_stderr": 0.002316190617323844, "rouge1_recall": 0.349869999562048, "rouge1_recall_stderr": 0.0028776097446615145, "rouge2_fmeasure": 0.06561109570973458, "rouge2_fmeasure_stderr": 0.0011754883309790891, "rouge2_precision": 0.05726169522995292, "rouge2_precision_stderr": 0.0011305457211571272, "rouge2_recall": 0.09761319314398223, "rouge2_recall_stderr": 0.001931346687472591, "rougeL_fmeasure": 0.17044218669835878, "rougeL_fmeasure_stderr": 0.001393233918351284, "rougeL_precision": 0.147593790999516, "rougeL_precision_stderr": 0.0015843013179494662, "rougeL_recall": 0.25163325238017475, "rougeL_recall_stderr": 0.0023185700212327545, "rougeLsum_fmeasure": 0.22970291581408894, "rougeLsum_fmeasure_stderr": 0.0019241182984497916, "rougeLsum_precision": 0.20040929059968782, "rougeLsum_precision_stderr": 0.0021906873980096493, "rougeLsum_recall": 0.33105684407128727, "rougeLsum_recall_stderr": 0.0027539492668926546}}, "2": {"tldr_en": {"bleu": 3.9658166802535484, "bleu_stderr": 0.07925927931568665, "rouge1_fmeasure": 0.24678912656203086, "rouge1_fmeasure_stderr": 0.002037101161511086, "rouge1_precision": 0.22135115607796563, "rouge1_precision_stderr": 0.002466762662359004, "rouge1_recall": 0.3499365920126553, "rouge1_recall_stderr": 0.0029103670069339765, "rouge2_fmeasure": 0.06998330324965386, "rouge2_fmeasure_stderr": 0.0012429781942009475, "rouge2_precision": 0.06270201642870808, "rouge2_precision_stderr": 0.001276661389687198, "rouge2_recall": 0.10236690363201062, "rouge2_recall_stderr": 0.0019960722912629107, "rougeL_fmeasure": 0.17719154028942327, "rougeL_fmeasure_stderr": 0.0014768440178427862, "rougeL_precision": 0.15820160856971907, "rougeL_precision_stderr": 0.001803134257764531, "rougeL_recall": 0.25667905638991584, "rougeL_recall_stderr": 0.002404233442497606, "rougeLsum_fmeasure": 0.2338283238156954, "rougeLsum_fmeasure_stderr": 0.0019162970817694011, "rougeLsum_precision": 0.20961741686683294, "rougeLsum_precision_stderr": 0.0023343675964983485, "rougeLsum_recall": 0.332399007963623, "rougeLsum_recall_stderr": 0.0027852830969350697}}, "3": {"tldr_en": {"bleu": 3.971856308559798, "bleu_stderr": 0.09289145672499902, "rouge1_fmeasure": 0.2013934895923683, "rouge1_fmeasure_stderr": 0.0024101386411320806, "rouge1_precision": 0.1877914020690942, "rouge1_precision_stderr": 0.002780603639421943, "rouge1_recall": 0.28469073541532774, "rouge1_recall_stderr": 0.003552554685317243, "rouge2_fmeasure": 0.05612430703445463, "rouge2_fmeasure_stderr": 0.0011944949124968374, "rouge2_precision": 0.05231468260770844, "rouge2_precision_stderr": 0.00130393473012376, "rouge2_recall": 0.08172649028680513, "rouge2_recall_stderr": 0.0018880261158435462, "rougeL_fmeasure": 0.14593849406958256, "rougeL_fmeasure_stderr": 0.0017576382763014175, "rougeL_precision": 0.1363160410891958, "rougeL_precision_stderr": 0.0021057536768160825, "rougeL_recall": 0.21040942089474482, "rougeL_recall_stderr": 0.002825298497100027, "rougeLsum_fmeasure": 0.19069625305891352, "rougeLsum_fmeasure_stderr": 0.002278271153942546, "rougeLsum_precision": 0.17789698988425304, "rougeLsum_precision_stderr": 0.0026448838544979162, "rougeLsum_recall": 0.2699836129539428, "rougeLsum_recall_stderr": 0.003393162716447727}}, "4": {"tldr_en": {"bleu": 0.8216828826403728, "bleu_stderr": 0.06116206229347704, "rouge1_fmeasure": 0.06294177276794305, "rouge1_fmeasure_stderr": 0.0021234318631902995, "rouge1_precision": 0.05986373469352374, "rouge1_precision_stderr": 0.0022413247620520016, "rouge1_recall": 0.09299451528533728, "rouge1_recall_stderr": 0.003182662484214728, "rouge2_fmeasure": 0.01701967942257221, "rouge2_fmeasure_stderr": 0.0008072426334728445, "rouge2_precision": 0.01607251699061772, "rouge2_precision_stderr": 0.0008627575848833899, "rouge2_recall": 0.026217246289428207, "rouge2_recall_stderr": 0.0013573347317069598, "rougeL_fmeasure": 0.04627660806628133, "rougeL_fmeasure_stderr": 0.001562406165753214, "rougeL_precision": 0.04416852187112644, "rougeL_precision_stderr": 0.0016929957101032731, "rougeL_recall": 0.06977139563193656, "rougeL_recall_stderr": 0.002453474360518891, "rougeLsum_fmeasure": 0.05949644164272974, "rougeLsum_fmeasure_stderr": 0.002009657736950104, "rougeLsum_precision": 0.056539524010580496, "rougeLsum_precision_stderr": 0.002122275707889583, "rougeLsum_recall": 0.08802275906234883, "rougeLsum_recall_stderr": 0.00301921948656467}}, "5": {"tldr_en": {"bleu": 2.939501013520158e-06, "bleu_stderr": 6.1416042954126946e-06, "rouge1_fmeasure": 0.010348132353778797, "rouge1_fmeasure_stderr": 0.0009643798928416376, "rouge1_precision": 0.010026792511549231, "rouge1_precision_stderr": 0.0009722843833023145, "rouge1_recall": 0.014997054249515373, "rouge1_recall_stderr": 0.0014200291615747175, "rouge2_fmeasure": 0.002947850654970828, "rouge2_fmeasure_stderr": 0.00038101201967536574, "rouge2_precision": 0.002614111292262381, "rouge2_precision_stderr": 0.0003582186570242107, "rouge2_recall": 0.0043272321523977674, "rouge2_recall_stderr": 0.000556694180882867, "rougeL_fmeasure": 0.007846213305722558, "rougeL_fmeasure_stderr": 0.0007330941188437008, "rougeL_precision": 0.007696209279533862, "rougeL_precision_stderr": 0.0007541223407519663, "rougeL_recall": 0.01149712833393674, "rougeL_recall_stderr": 0.001106103576275103, "rougeLsum_fmeasure": 0.009790499222718256, "rougeLsum_fmeasure_stderr": 0.0009156513115428641, "rougeLsum_precision": 0.009502293292635411, "rougeLsum_precision_stderr": 0.0009264781441363666, "rougeLsum_recall": 0.01418906709799213, "rougeLsum_recall_stderr": 0.0013478881437725126}}}, "e2e_nlg_cleaned": {"0": {"generate_text_restaurant": {"bleu": 0.17854969577688523, "bleu_stderr": 0.02350515901837014, "rouge1_fmeasure": 0.14982661873128114, "rouge1_fmeasure_stderr": 0.0010486948402501617, "rouge1_precision": 0.34481003507654684, "rouge1_precision_stderr": 0.0024960642876163215, "rouge1_recall": 0.11140479511426793, "rouge1_recall_stderr": 0.0012326825192403048, "rouge2_fmeasure": 0.010042498177274786, "rouge2_fmeasure_stderr": 0.0004526496809829142, "rouge2_precision": 0.02652033223020393, "rouge2_precision_stderr": 0.0012104924394046484, "rouge2_recall": 0.007738642998342415, "rouge2_recall_stderr": 0.00040331722748865453, "rougeL_fmeasure": 0.13556698518148783, "rougeL_fmeasure_stderr": 0.0010029610186637107, "rougeL_precision": 0.31148977629184244, "rougeL_precision_stderr": 0.0024243327495898086, "rougeL_recall": 0.10191142585459571, "rougeL_recall_stderr": 0.0012109507603402815, "rougeLsum_fmeasure": 0.1364054799839495, "rougeLsum_fmeasure_stderr": 0.0010086900095262318, "rougeLsum_precision": 0.3173006007013496, "rougeLsum_precision_stderr": 0.002513826319580817, "rougeLsum_recall": 0.10030998557950792, "rougeLsum_recall_stderr": 0.0010758913304541808}}, "1": {"generate_text_restaurant": {"bleu": 8.550714951198435, "bleu_stderr": 0.06794955954281734, "rouge1_fmeasure": 0.3940891272584714, "rouge1_fmeasure_stderr": 0.00233253448293986, "rouge1_precision": 0.37441894592411723, "rouge1_precision_stderr": 0.00285043399066679, "rouge1_recall": 0.4681809355112701, "rouge1_recall_stderr": 0.0029387631347472464, "rouge2_fmeasure": 0.17310173953900088, "rouge2_fmeasure_stderr": 0.0017673218319637287, "rouge2_precision": 0.16514863563264978, "rouge2_precision_stderr": 0.0019360690941877992, "rouge2_recall": 0.20649495583013436, "rouge2_recall_stderr": 0.0021624061718912763, "rougeL_fmeasure": 0.2901588169021053, "rougeL_fmeasure_stderr": 0.0017017464729554115, "rougeL_precision": 0.2743117208263101, "rougeL_precision_stderr": 0.0021002662608094564, "rougeL_recall": 0.34924701312209255, "rougeL_recall_stderr": 0.0023705538319379695, "rougeLsum_fmeasure": 0.327671306182913, "rougeLsum_fmeasure_stderr": 0.0022171207606254796, "rougeLsum_precision": 0.31201556645452433, "rougeLsum_precision_stderr": 0.002622342685847327, "rougeLsum_recall": 0.3885350242195471, "rougeLsum_recall_stderr": 0.0027432106106179788}}, "2": {"generate_text_restaurant": {"bleu": 11.867888027230435, "bleu_stderr": 0.15989930331829488, "rouge1_fmeasure": 0.4442817836593718, "rouge1_fmeasure_stderr": 0.0019988950330102873, "rouge1_precision": 0.44246406218448636, "rouge1_precision_stderr": 0.002320846131050864, "rouge1_recall": 0.4822237492014395, "rouge1_recall_stderr": 0.0028480697292934715, "rouge2_fmeasure": 0.20571413885055867, "rouge2_fmeasure_stderr": 0.001815833869403785, "rouge2_precision": 0.20447987966953268, "rouge2_precision_stderr": 0.0019087131351150518, "rouge2_recall": 0.22575098541036898, "rouge2_recall_stderr": 0.0022679440038537117, "rougeL_fmeasure": 0.3232871465388061, "rougeL_fmeasure_stderr": 0.0017540023019821938, "rougeL_precision": 0.3220661789895318, "rougeL_precision_stderr": 0.0019807481977887516, "rougeL_recall": 0.35167106294020406, "rougeL_recall_stderr": 0.002412719701186727, "rougeLsum_fmeasure": 0.37187691948404467, "rougeLsum_fmeasure_stderr": 0.0020334507090512886, "rougeLsum_precision": 0.37040704761920407, "rougeLsum_precision_stderr": 0.002271535999762247, "rougeLsum_recall": 0.4036635693660681, "rougeLsum_recall_stderr": 0.0027100283713887775}}, "3": {"generate_text_restaurant": {"bleu": 12.176808245577714, "bleu_stderr": 0.1522141950985559, "rouge1_fmeasure": 0.4492578440402226, "rouge1_fmeasure_stderr": 0.0019667066974463896, "rouge1_precision": 0.44578562677217626, "rouge1_precision_stderr": 0.0022755220565382653, "rouge1_recall": 0.4889266631673112, "rouge1_recall_stderr": 0.0028806835048615135, "rouge2_fmeasure": 0.21198726153120898, "rouge2_fmeasure_stderr": 0.0018195540133290194, "rouge2_precision": 0.20934965847705986, "rouge2_precision_stderr": 0.0018748523719357, "rouge2_recall": 0.23374791215018093, "rouge2_recall_stderr": 0.0023263783868214196, "rougeL_fmeasure": 0.3276087887468011, "rougeL_fmeasure_stderr": 0.0017546007227396715, "rougeL_precision": 0.32511799875091574, "rougeL_precision_stderr": 0.001967046506806585, "rougeL_recall": 0.35739711489377646, "rougeL_recall_stderr": 0.002465697341506471, "rougeLsum_fmeasure": 0.3771411500357944, "rougeLsum_fmeasure_stderr": 0.0020137589520510426, "rougeLsum_precision": 0.3741382823179302, "rougeLsum_precision_stderr": 0.002234766274837387, "rougeLsum_recall": 0.4108171966535684, "rougeLsum_recall_stderr": 0.0027621282248548955}}, "4": {"generate_text_restaurant": {"bleu": 12.402384292539924, "bleu_stderr": 0.135024423037405, "rouge1_fmeasure": 0.4552455893734124, "rouge1_fmeasure_stderr": 0.0019583919126399917, "rouge1_precision": 0.4500297698600483, "rouge1_precision_stderr": 0.002279184782036985, "rouge1_recall": 0.49411520793728253, "rouge1_recall_stderr": 0.0028065754284846934, "rouge2_fmeasure": 0.2155975402534293, "rouge2_fmeasure_stderr": 0.001873826925854458, "rouge2_precision": 0.2124489885242478, "rouge2_precision_stderr": 0.001924872387744921, "rouge2_recall": 0.23653217550233827, "rouge2_recall_stderr": 0.0023355573505469554, "rougeL_fmeasure": 0.33116918396735845, "rougeL_fmeasure_stderr": 0.0017792810189211944, "rougeL_precision": 0.32709295569865143, "rougeL_precision_stderr": 0.001956608790763274, "rougeL_recall": 0.3602236836443227, "rougeL_recall_stderr": 0.002427496751180958, "rougeLsum_fmeasure": 0.3808847920405725, "rougeLsum_fmeasure_stderr": 0.002065156905180787, "rougeLsum_precision": 0.37596847671524797, "rougeLsum_precision_stderr": 0.002251960910942368, "rougeLsum_recall": 0.41399734380099557, "rougeLsum_recall_stderr": 0.0027661212857413163}}, "5": {"generate_text_restaurant": {"bleu": 12.262420161138401, "bleu_stderr": 0.18903914838075944, "rouge1_fmeasure": 0.45590768378496416, "rouge1_fmeasure_stderr": 0.001962173998693794, "rouge1_precision": 0.4494732628845236, "rouge1_precision_stderr": 0.0023023795088449608, "rouge1_recall": 0.49537684561519607, "rouge1_recall_stderr": 0.002767197249479356, "rouge2_fmeasure": 0.21627498227149344, "rouge2_fmeasure_stderr": 0.0018499821709783027, "rouge2_precision": 0.21297442064638186, "rouge2_precision_stderr": 0.001926070075686657, "rouge2_recall": 0.23703970881860875, "rouge2_recall_stderr": 0.0022735756287929677, "rougeL_fmeasure": 0.3331238951307013, "rougeL_fmeasure_stderr": 0.0017799650141595914, "rougeL_precision": 0.3281425210661801, "rougeL_precision_stderr": 0.001972901049188069, "rougeL_recall": 0.3628730729029585, "rougeL_recall_stderr": 0.0024176721948621425, "rougeLsum_fmeasure": 0.38246870938641964, "rougeLsum_fmeasure_stderr": 0.0020417804934292076, "rougeLsum_precision": 0.3768630011135112, "rougeLsum_precision_stderr": 0.0022664490914563948, "rougeLsum_recall": 0.41605530731114854, "rougeLsum_recall_stderr": 0.0027097026120309876}}}, "gem_xsum": {"0": {"article_DOC_summary": {"bleu": 2.24608084382355, "bleu_stderr": 0.11799890497604383, "rouge1_fmeasure": 0.22178670977262957, "rouge1_fmeasure_stderr": 0.0026603762671440385, "rouge1_precision": 0.17400888678487844, "rouge1_precision_stderr": 0.002464330237468343, "rouge1_recall": 0.3502023417511915, "rouge1_recall_stderr": 0.004626152702942569, "rouge2_fmeasure": 0.05332080380122619, "rouge2_fmeasure_stderr": 0.0017188362970072153, "rouge2_precision": 0.040704743736010573, "rouge2_precision_stderr": 0.001393829442555013, "rouge2_recall": 0.08799878519146338, "rouge2_recall_stderr": 0.0029196986234636744, "rougeL_fmeasure": 0.16609949129334142, "rougeL_fmeasure_stderr": 0.0020543947028805716, "rougeL_precision": 0.12992137120722907, "rougeL_precision_stderr": 0.0018781656373678345, "rougeL_recall": 0.2641999836224796, "rougeL_recall_stderr": 0.0036812097901456225, "rougeLsum_fmeasure": 0.17331928576093303, "rougeLsum_fmeasure_stderr": 0.0022434557390865763, "rougeLsum_precision": 0.13526811734802355, "rougeLsum_precision_stderr": 0.001975414967835506, "rougeLsum_recall": 0.27591607890974656, "rougeLsum_recall_stderr": 0.004020504841332198}}, "1": {"article_DOC_summary": {"bleu": 1.8189118554354626, "bleu_stderr": 0.10029587869243096, "rouge1_fmeasure": 0.19671334605636295, "rouge1_fmeasure_stderr": 0.0027188446769093686, "rouge1_precision": 0.14019002697714425, "rouge1_precision_stderr": 0.002033439876939429, "rouge1_recall": 0.3437328928644267, "rouge1_recall_stderr": 0.004622184360550014, "rouge2_fmeasure": 0.04451173090845151, "rouge2_fmeasure_stderr": 0.0016237255917685457, "rouge2_precision": 0.031342583459769556, "rouge2_precision_stderr": 0.0011487670391079187, "rouge2_recall": 0.0803191311481453, "rouge2_recall_stderr": 0.003010334414958687, "rougeL_fmeasure": 0.14825510655652663, "rougeL_fmeasure_stderr": 0.0020028300424767575, "rougeL_precision": 0.1053445488770993, "rougeL_precision_stderr": 0.001476955053927121, "rougeL_recall": 0.2613986979531857, "rougeL_recall_stderr": 0.0035961912243557873, "rougeLsum_fmeasure": 0.1592072439127538, "rougeLsum_fmeasure_stderr": 0.002270863529627419, "rougeLsum_precision": 0.11322117707364136, "rougeLsum_precision_stderr": 0.001678631461460253, "rougeLsum_recall": 0.27993778037490374, "rougeLsum_recall_stderr": 0.003988822898958396}}, "2": {"article_DOC_summary": {"bleu": 2.0454443359589503, "bleu_stderr": 0.08368970692051826, "rouge1_fmeasure": 0.20889368966335237, "rouge1_fmeasure_stderr": 0.0026887281040736378, "rouge1_precision": 0.1488455450923004, "rouge1_precision_stderr": 0.002009481565468108, "rouge1_recall": 0.3642751817540943, "rouge1_recall_stderr": 0.004565498697439836, "rouge2_fmeasure": 0.05083267238349886, "rouge2_fmeasure_stderr": 0.001670794310427006, "rouge2_precision": 0.03580310577134537, "rouge2_precision_stderr": 0.0011818056791973873, "rouge2_recall": 0.09140711571916593, "rouge2_recall_stderr": 0.003096432205185466, "rougeL_fmeasure": 0.15708491403271296, "rougeL_fmeasure_stderr": 0.0019982552541995204, "rougeL_precision": 0.11164126156649677, "rougeL_precision_stderr": 0.001473760236619727, "rougeL_recall": 0.27611637516436727, "rougeL_recall_stderr": 0.003575935893484929, "rougeLsum_fmeasure": 0.1672065849237029, "rougeLsum_fmeasure_stderr": 0.002247088064282535, "rougeLsum_precision": 0.1188385887857819, "rougeLsum_precision_stderr": 0.0016510150185409913, "rougeLsum_recall": 0.2937542835590118, "rougeLsum_recall_stderr": 0.003999046140359821}}, "3": {"article_DOC_summary": {"bleu": 2.0101663995130106, "bleu_stderr": 0.13035938633140529, "rouge1_fmeasure": 0.2005996992165322, "rouge1_fmeasure_stderr": 0.002922307185273832, "rouge1_precision": 0.1451633530537778, "rouge1_precision_stderr": 0.0022577436785926417, "rouge1_recall": 0.3454037481783916, "rouge1_recall_stderr": 0.005024002677948197, "rouge2_fmeasure": 0.047805324803931785, "rouge2_fmeasure_stderr": 0.001665552432783648, "rouge2_precision": 0.03399383589660795, "rouge2_precision_stderr": 0.0011970297160356436, "rouge2_recall": 0.08515533208064306, "rouge2_recall_stderr": 0.00305750154560904, "rougeL_fmeasure": 0.15011348459355398, "rougeL_fmeasure_stderr": 0.0021656916632942095, "rougeL_precision": 0.1082275772002201, "rougeL_precision_stderr": 0.0016303512277000601, "rougeL_recall": 0.26018618860799503, "rougeL_recall_stderr": 0.0038502171494847245, "rougeLsum_fmeasure": 0.1614133142157355, "rougeLsum_fmeasure_stderr": 0.00243279237750366, "rougeLsum_precision": 0.1163681459054013, "rougeLsum_precision_stderr": 0.0018244541208628394, "rougeLsum_recall": 0.27969335970329556, "rougeLsum_recall_stderr": 0.004315552235909818}}, "4": {"article_DOC_summary": {"bleu": 1.1210416443624807, "bleu_stderr": 0.18886560236043304, "rouge1_fmeasure": 0.05612514432374489, "rouge1_fmeasure_stderr": 0.0030731828239370837, "rouge1_precision": 0.04604504358511682, "rouge1_precision_stderr": 0.0026703224008911334, "rouge1_recall": 0.08906761715367195, "rouge1_recall_stderr": 0.0050274498862542885, "rouge2_fmeasure": 0.013937278041326827, "rouge2_fmeasure_stderr": 0.0011465942789645456, "rouge2_precision": 0.011276377923656494, "rouge2_precision_stderr": 0.0012571358140648242, "rouge2_recall": 0.023708158418892494, "rouge2_recall_stderr": 0.002002425619950796, "rougeL_fmeasure": 0.04184105680154411, "rougeL_fmeasure_stderr": 0.002285547204915818, "rougeL_precision": 0.03500836518053345, "rougeL_precision_stderr": 0.0021377284451202998, "rougeL_recall": 0.06645887444884206, "rougeL_recall_stderr": 0.0037898342177074525, "rougeLsum_fmeasure": 0.04557476167159472, "rougeLsum_fmeasure_stderr": 0.002503478149457618, "rougeLsum_precision": 0.0378191473406664, "rougeLsum_precision_stderr": 0.002272186759930814, "rougeLsum_recall": 0.07247078079866072, "rougeLsum_recall_stderr": 0.004137833255697601}}, "5": {"article_DOC_summary": {"bleu": 1.7628232823336373e-17, "bleu_stderr": 2.8521803760603906e-14, "rouge1_fmeasure": 0.0026984759176508304, "rouge1_fmeasure_stderr": 0.000738763092128848, "rouge1_precision": 0.0022472741378328047, "rouge1_precision_stderr": 0.0006307356242527256, "rouge1_recall": 0.004133903246439151, "rouge1_recall_stderr": 0.0011668769662793363, "rouge2_fmeasure": 0.000551357929598578, "rouge2_fmeasure_stderr": 0.00022851833366345053, "rouge2_precision": 0.00044268370843308837, "rouge2_precision_stderr": 0.00017799352097082119, "rouge2_recall": 0.0008499613285961788, "rouge2_recall_stderr": 0.0003602171813535447, "rougeL_fmeasure": 0.001991063584250406, "rougeL_fmeasure_stderr": 0.0005515138571257219, "rougeL_precision": 0.0016336554354190846, "rougeL_precision_stderr": 0.0004503176195113186, "rougeL_recall": 0.003027477992470196, "rougeL_recall_stderr": 0.0008730946685903953, "rougeLsum_fmeasure": 0.002097477776387118, "rougeLsum_fmeasure_stderr": 0.0005739873846094995, "rougeLsum_precision": 0.0017629252737840803, "rougeLsum_precision_stderr": 0.0005014954502254297, "rougeLsum_recall": 0.0031744555131817086, "rougeLsum_recall_stderr": 0.0008937483641872015}}}} |