{"data":[{"project":"default","id":"massive-harness-kai-1.4.c1856a77-kai","benchmark":"massive","split":"harness","system":"kai","version":"kai-1.4@c1856a77730f33781dd4ca662a629051d445b50a52d3ea012c9492395461f52a","baseline":false,"study":"","embedder":"","reader":"","k":null,"temperature":null,"max_tokens":null,"prompt":"","prompt_digest":"","dataset_digest":"dde140a63fc527cbeb8f7ada29088bf17064d4bcf30c9680d6423696f0378673","store_digest":"","facts_digest":"","commit":"5f616dd13167bc407bd1a56645c0982f297e9e4a","questions":5100,"answered":5100,"when":"","ended":"2026-10-16T13:53:18Z","by":"","completion":"complete","key":"a6d479f90159bba6ba290268cc70c064df937c8ed6040784de793a81e9c6ca31","measures":[{"category":"all","metric":"macro_accuracy","value":0.9073,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_brier","value":0.1464,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_ece","value":0.0635,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_aurc","value":0.0182,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_zero_prob","value":0.0004,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"af","metric":"accuracy","value":0.88,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"brier","value":0.1914,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"ece","value":0.0806,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"aurc","value":0.0216,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"accuracy","value":0.86,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"brier","value":0.1987,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"ece","value":0.0613,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"aurc","value":0.0255,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"brier","value":0.1237,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"ece","value":0.0626,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"aurc","value":0.0081,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"accuracy","value":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"brier","value":0.0929,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"ece","value":0.0521,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"aurc","value":0.004,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"accuracy","value":0.89,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"brier","value":0.1578,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"ece","value":0.0741,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"aurc","value":0.0173,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"accuracy","value":0.86,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"brier","value":0.2062,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"ece","value":0.0705,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"aurc","value":0.0471,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"brier","value":0.1425,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"ece","value":0.057,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"aurc","value":0.0087,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"brier","value":0.1032,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"ece","value":0.0388,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"aurc","value":0.0063,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"brier","value":0.1363,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"ece","value":0.0637,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"aurc","value":0.0112,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"accuracy","value":0.95,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"brier","value":0.0975,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"ece","value":0.0493,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"aurc","value":0.0089,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"brier","value":0.1342,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"ece","value":0.0724,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"aurc","value":0.0224,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"accuracy","value":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"brier","value":0.0837,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"ece","value":0.0392,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"aurc","value":0.0077,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"accuracy","value":0.96,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"brier","value":0.0871,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"ece","value":0.0482,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"aurc","value":0.016,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"brier","value":0.1586,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"ece","value":0.0905,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"aurc","value":0.0203,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"brier","value":0.1395,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"ece","value":0.0702,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"aurc","value":0.0146,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"brier","value":0.145,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"ece","value":0.0524,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"aurc","value":0.0099,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"accuracy","value":0.87,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"brier","value":0.2056,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"ece","value":0.0982,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"aurc","value":0.0352,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"accuracy","value":0.88,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"brier","value":0.1635,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"ece","value":0.0641,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"aurc","value":0.0125,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"brier","value":0.0962,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"ece","value":0.0353,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"aurc","value":0.0078,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"accuracy","value":0.85,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"brier","value":0.2212,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"ece","value":0.0955,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"aurc","value":0.0226,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"brier","value":0.1165,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"ece","value":0.0516,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"aurc","value":0.0065,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"accuracy","value":0.97,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"brier","value":0.0652,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"ece","value":0.0384,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"aurc","value":0.004,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"accuracy","value":0.84,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"brier","value":0.2467,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"ece","value":0.0848,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"aurc","value":0.0287,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"accuracy","value":0.8,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"brier","value":0.277,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"ece","value":0.0752,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"aurc","value":0.0769,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"zero_prob","value":0.01,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"accuracy","value":0.85,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"brier","value":0.2362,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"ece","value":0.0982,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"aurc","value":0.0384,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"accuracy","value":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"brier","value":0.1894,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"ece","value":0.0746,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"aurc","value":0.0276,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"brier","value":0.1293,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"ece","value":0.0707,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"aurc","value":0.0131,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"accuracy","value":0.89,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"brier","value":0.1751,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"ece","value":0.0752,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"aurc","value":0.0216,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"accuracy","value":0.85,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"brier","value":0.2259,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"ece","value":0.0857,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"aurc","value":0.0366,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"accuracy","value":0.86,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"brier","value":0.2211,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"ece","value":0.098,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"aurc","value":0.0274,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"brier","value":0.1072,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"ece","value":0.0559,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"aurc","value":0.0174,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"accuracy","value":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"brier","value":0.1344,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"ece","value":0.0741,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"aurc","value":0.0308,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"brier","value":0.1257,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"ece","value":0.0632,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"aurc","value":0.0103,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"zero_prob","value":0.01,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"accuracy","value":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"brier","value":0.1339,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"ece","value":0.0785,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"aurc","value":0.0082,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"brier","value":0.1337,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"ece","value":0.0583,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"aurc","value":0.0078,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"brier","value":0.1123,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"ece","value":0.0572,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"aurc","value":0.0134,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"accuracy","value":0.96,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"brier","value":0.069,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"ece","value":0.0367,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"aurc","value":0.0036,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"accuracy","value":0.98,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"brier","value":0.0309,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"ece","value":0.0208,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"aurc","value":0.0006,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"brier","value":0.1234,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"ece","value":0.052,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"aurc","value":0.0199,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"accuracy","value":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"brier","value":0.0927,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"ece","value":0.0536,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"aurc","value":0.0045,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"accuracy","value":0.95,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"brier","value":0.0954,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"ece","value":0.0539,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"aurc","value":0.0046,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"accuracy","value":0.83,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"brier","value":0.2559,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"ece","value":0.0824,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"aurc","value":0.0324,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"accuracy","value":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"brier","value":0.1647,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"ece","value":0.0574,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"aurc","value":0.0157,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"accuracy","value":0.88,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"brier","value":0.1954,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"ece","value":0.0669,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"aurc","value":0.024,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"brier","value":0.1315,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"ece","value":0.0552,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"aurc","value":0.0257,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"accuracy","value":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"brier","value":0.1388,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"ece","value":0.0668,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"aurc","value":0.0269,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"accuracy","value":0.88,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"brier","value":0.1745,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"ece","value":0.0719,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"aurc","value":0.0325,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"ur","metric":"accuracy","value":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"ur","metric":"brier","value":0.1569,"lo":null,"hi":null,"n":100,"of":100},{"category":"ur","metric":"ece","value":0.0592,"lo":null,"hi":null,"n":100,"of":100},{"category":"ur","metric":"aurc","value":0.0169,"lo":null,"hi":null,"n":100,"of":100},{"category":"ur","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"accuracy","value":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"brier","value":0.158,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"ece","value":0.0537,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"aurc","value":0.0124,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"accuracy","value":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"brier","value":0.103,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"ece","value":0.0564,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"aurc","value":0.0063,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"accuracy","value":0.96,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"brier","value":0.0637,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"ece","value":0.0367,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"aurc","value":0.0059,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100}],"notes":"device cpu; dtype as served; preds sha256:86282bbddaa627f8bb2465115a31c49043822d16c50ef284913c3032928d1809; scores sha256:520c84c6b7667ccb4d5df676739202a28db3f6b799139f4ae45e2fbefe3ff767","revision":"original","canonical":true,"visibility":"public","trainable":false,"publishable":true}],"total":1,"completion":{"complete":1,"partial":0,"inconsistent":0,"unknown":0},"gaps":[{"field":"by","runs":1},{"field":"embedder","runs":1},{"field":"facts_digest","runs":1},{"field":"k","runs":1},{"field":"max_tokens","runs":1},{"field":"prompt","runs":1},{"field":"prompt_digest","runs":1},{"field":"reader","runs":1},{"field":"store_digest","runs":1},{"field":"temperature","runs":1},{"field":"when","runs":1}]}