{"data":[{"project":"default","id":"massive-harness-kai-1.0834a74f-kai","benchmark":"massive","split":"harness","system":"kai","version":"kai-1@0834a74f2d140642a453373da09e5a128e3d2c8d9e8bb1dcaf86d7210a4ecdfc","baseline":false,"study":"","embedder":"","reader":"","k":null,"temperature":null,"max_tokens":null,"prompt":"","prompt_digest":"","dataset_digest":"dde140a63fc527cbeb8f7ada29088bf17064d4bcf30c9680d6423696f0378673","store_digest":"","facts_digest":"","commit":"2c74e2d61addc885e847358975fb866b03cf17dc","questions":5100,"answered":5100,"when":"","ended":"2026-10-03T00:06:45Z","by":"","completion":"complete","measures":[{"category":"all","metric":"macro_accuracy","value":0.9096,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_brier","value":0.1358,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_ece","value":0.06,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_aurc","value":0.0112,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"all","metric":"macro_zero_prob","value":0,"lo":null,"hi":null,"n":5100,"of":5100},{"category":"af","metric":"accuracy","value":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"brier","value":0.1488,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"ece","value":0.0555,"lo":null,"hi":null,"n":100,"of":100},{"category":"af","metric":"aurc","value":0.0093,"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":0.85,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"brier","value":0.2193,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"ece","value":0.0846,"lo":null,"hi":null,"n":100,"of":100},{"category":"am","metric":"aurc","value":0.0217,"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":0.1228,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"ece","value":0.0636,"lo":null,"hi":null,"n":100,"of":100},{"category":"ar","metric":"aurc","value":0.0065,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"brier","value":0.075,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"ece","value":0.0451,"lo":null,"hi":null,"n":100,"of":100},{"category":"az","metric":"aurc","value":0.0042,"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":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"brier","value":0.1423,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"ece","value":0.0619,"lo":null,"hi":null,"n":100,"of":100},{"category":"bn","metric":"aurc","value":0.0126,"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":0.85,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"brier","value":0.1932,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"ece","value":0.0744,"lo":null,"hi":null,"n":100,"of":100},{"category":"cy","metric":"aurc","value":0.023,"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":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"brier","value":0.1216,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"ece","value":0.051,"lo":null,"hi":null,"n":100,"of":100},{"category":"da","metric":"aurc","value":0.0065,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"brier","value":0.1143,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"ece","value":0.0466,"lo":null,"hi":null,"n":100,"of":100},{"category":"de","metric":"aurc","value":0.0062,"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":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"brier","value":0.0984,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"ece","value":0.0412,"lo":null,"hi":null,"n":100,"of":100},{"category":"el","metric":"aurc","value":0.0048,"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":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"brier","value":0.1035,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"ece","value":0.044,"lo":null,"hi":null,"n":100,"of":100},{"category":"en","metric":"aurc","value":0.0046,"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":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"brier","value":0.1293,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"ece","value":0.049,"lo":null,"hi":null,"n":100,"of":100},{"category":"es","metric":"aurc","value":0.0124,"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":0.0889,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"ece","value":0.0318,"lo":null,"hi":null,"n":100,"of":100},{"category":"fa","metric":"aurc","value":0.0037,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"brier","value":0.1092,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"ece","value":0.0513,"lo":null,"hi":null,"n":100,"of":100},{"category":"fi","metric":"aurc","value":0.0192,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"brier","value":0.1374,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"ece","value":0.0738,"lo":null,"hi":null,"n":100,"of":100},{"category":"fr","metric":"aurc","value":0.011,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"brier","value":0.1455,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"ece","value":0.063,"lo":null,"hi":null,"n":100,"of":100},{"category":"he","metric":"aurc","value":0.0124,"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":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"brier","value":0.1248,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"ece","value":0.0588,"lo":null,"hi":null,"n":100,"of":100},{"category":"hi","metric":"aurc","value":0.0066,"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":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"brier","value":0.1636,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"ece","value":0.0788,"lo":null,"hi":null,"n":100,"of":100},{"category":"hu","metric":"aurc","value":0.015,"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":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"brier","value":0.1602,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"ece","value":0.0619,"lo":null,"hi":null,"n":100,"of":100},{"category":"hy","metric":"aurc","value":0.0116,"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":0.96,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"brier","value":0.0753,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"ece","value":0.0517,"lo":null,"hi":null,"n":100,"of":100},{"category":"id","metric":"aurc","value":0.0055,"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":0.1946,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"ece","value":0.0808,"lo":null,"hi":null,"n":100,"of":100},{"category":"is","metric":"aurc","value":0.017,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"brier","value":0.1255,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"ece","value":0.0541,"lo":null,"hi":null,"n":100,"of":100},{"category":"it","metric":"aurc","value":0.0107,"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":0.069,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"ece","value":0.0368,"lo":null,"hi":null,"n":100,"of":100},{"category":"ja","metric":"aurc","value":0.0022,"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":0.89,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"brier","value":0.1866,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"ece","value":0.0811,"lo":null,"hi":null,"n":100,"of":100},{"category":"jv","metric":"aurc","value":0.0149,"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":0.83,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"brier","value":0.2399,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"ece","value":0.0852,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"aurc","value":0.0362,"lo":null,"hi":null,"n":100,"of":100},{"category":"ka","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"accuracy","value":0.84,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"brier","value":0.2351,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"ece","value":0.0837,"lo":null,"hi":null,"n":100,"of":100},{"category":"km","metric":"aurc","value":0.0316,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"brier","value":0.1591,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"ece","value":0.0757,"lo":null,"hi":null,"n":100,"of":100},{"category":"kn","metric":"aurc","value":0.0133,"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":0.1229,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"ece","value":0.0512,"lo":null,"hi":null,"n":100,"of":100},{"category":"ko","metric":"aurc","value":0.0075,"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":0.1669,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"ece","value":0.0783,"lo":null,"hi":null,"n":100,"of":100},{"category":"lv","metric":"aurc","value":0.0119,"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":0.87,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"brier","value":0.2184,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"ece","value":0.1029,"lo":null,"hi":null,"n":100,"of":100},{"category":"ml","metric":"aurc","value":0.0254,"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":0.88,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"brier","value":0.1975,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"ece","value":0.0968,"lo":null,"hi":null,"n":100,"of":100},{"category":"mn","metric":"aurc","value":0.0193,"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":0.1111,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"ece","value":0.0452,"lo":null,"hi":null,"n":100,"of":100},{"category":"ms","metric":"aurc","value":0.0117,"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":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"brier","value":0.1123,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"ece","value":0.0696,"lo":null,"hi":null,"n":100,"of":100},{"category":"my","metric":"aurc","value":0.0099,"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":0.94,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"brier","value":0.0879,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"ece","value":0.0369,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"aurc","value":0.0044,"lo":null,"hi":null,"n":100,"of":100},{"category":"nb","metric":"zero_prob","value":0,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"accuracy","value":0.92,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"brier","value":0.1198,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"ece","value":0.0492,"lo":null,"hi":null,"n":100,"of":100},{"category":"nl","metric":"aurc","value":0.006,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"brier","value":0.1045,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"ece","value":0.0411,"lo":null,"hi":null,"n":100,"of":100},{"category":"pl","metric":"aurc","value":0.0054,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"brier","value":0.1049,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"ece","value":0.0365,"lo":null,"hi":null,"n":100,"of":100},{"category":"pt","metric":"aurc","value":0.0069,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"brier","value":0.0938,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"ece","value":0.0505,"lo":null,"hi":null,"n":100,"of":100},{"category":"ro","metric":"aurc","value":0.005,"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":0.97,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"brier","value":0.0417,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"ece","value":0.0302,"lo":null,"hi":null,"n":100,"of":100},{"category":"ru","metric":"aurc","value":0.0008,"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":0.1254,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"ece","value":0.0483,"lo":null,"hi":null,"n":100,"of":100},{"category":"sl","metric":"aurc","value":0.0075,"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":0.1093,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"ece","value":0.0584,"lo":null,"hi":null,"n":100,"of":100},{"category":"sq","metric":"aurc","value":0.0047,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"brier","value":0.0977,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"ece","value":0.0366,"lo":null,"hi":null,"n":100,"of":100},{"category":"sv","metric":"aurc","value":0.0042,"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":0.84,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"brier","value":0.221,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"ece","value":0.0871,"lo":null,"hi":null,"n":100,"of":100},{"category":"sw","metric":"aurc","value":0.0191,"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":0.89,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"brier","value":0.1493,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"ece","value":0.0731,"lo":null,"hi":null,"n":100,"of":100},{"category":"ta","metric":"aurc","value":0.0095,"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":0.87,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"brier","value":0.194,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"ece","value":0.0825,"lo":null,"hi":null,"n":100,"of":100},{"category":"te","metric":"aurc","value":0.0222,"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":0.93,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"brier","value":0.0999,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"ece","value":0.039,"lo":null,"hi":null,"n":100,"of":100},{"category":"th","metric":"aurc","value":0.0056,"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":0.1482,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"ece","value":0.0666,"lo":null,"hi":null,"n":100,"of":100},{"category":"tl","metric":"aurc","value":0.0197,"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":0.9,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"brier","value":0.1743,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"ece","value":0.0748,"lo":null,"hi":null,"n":100,"of":100},{"category":"tr","metric":"aurc","value":0.0191,"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":0.1465,"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":0.0095,"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":0.91,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"brier","value":0.1386,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"ece","value":0.0697,"lo":null,"hi":null,"n":100,"of":100},{"category":"vi","metric":"aurc","value":0.0088,"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":0.0895,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"ece","value":0.0587,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-CN","metric":"aurc","value":0.0032,"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":0.95,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"brier","value":0.0689,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"ece","value":0.0325,"lo":null,"hi":null,"n":100,"of":100},{"category":"zh-TW","metric":"aurc","value":0.0029,"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 cuda; dtype bf16; preds sha256:368a5f76afa5171bd80e36a827debfd93803318a4e78d92ba6c4af493b8bf23d; scores sha256:1ec811cfb085ac209c7b094b408b00ef95d3d3c73b7ae7bdf6c52fc314d112e0","revision":"original","canonical":true,"visibility":"public","trainable":false,"publishable":false}],"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}]}