{"data":[{"project":"default","id":"cap.cardinality-cases-kai-1-kai","benchmark":"cap.cardinality","split":"cases","system":"kai","version":"kai-1@0834a74f2d140642a453373da09e5a128e3d2c8d9e8bb1dcaf86d7210a4ecdfc","baseline":false,"study":"","embedder":"","reader":"","k":null,"temperature":null,"max_tokens":null,"prompt":"","prompt_digest":"","dataset_digest":"","store_digest":"","facts_digest":"","commit":"5138293ef73d16d2f6300d3cd3386a8ed11fe5bc","questions":null,"answered":null,"when":"","ended":"2026-10-03T00:20:27Z","by":"dgx","completion":"unknown","measures":[{"category":"all","metric":"k1000.accall","value":0.005,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.accuracy","value":0.005,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.accall","value":0.03,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.accuracy","value":0.03,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.brier","value":1.0824,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.ece","value":0.253,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.nll","value":11.9627,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.recall1","value":0.03,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.recall20","value":0.655,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.recall5","value":0.135,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.bi.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.brier","value":0.9982,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.ece","value":0.0261,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.nll","value":6.6967,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.recall1","value":0.005,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.recall20","value":0.1,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.recall5","value":0.035,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.brier","value":1.118,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.ece","value":0.2307,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.nll","value":26.6337,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.recall20","value":0.04,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.recall5","value":0.025,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.sl.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k1000.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.accall","value":0.02,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.accuracy","value":0.02,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.brier","value":1.0923,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.ece","value":0.2318,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.nll","value":17.9489,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.recall1","value":0.02,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.recall20","value":0.405,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.recall5","value":0.075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.bi.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.brier","value":1,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.ece","value":0.01,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.nll","value":8.7704,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.recall20","value":0.035,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.recall5","value":0.01,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.brier","value":1.1302,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.ece","value":0.2365,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.nll","value":27.2571,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.recall20","value":0.015,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.recall5","value":0.01,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.sl.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k10000.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.brier","value":1.0965,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.ece","value":0.1965,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.nll","value":21.4707,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.recall20","value":0.26,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.recall5","value":0.04,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.bi.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.brier","value":1.001,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.ece","value":0.0047,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.nll","value":11.2211,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.recall20","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.recall5","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.accall","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.accuracy","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.brier","value":1.1271,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.ece","value":0.2351,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.nll","value":27.1246,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.recall1","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.recall20","value":0.02,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.recall5","value":0.02,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.sl.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k100000.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.accall","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.accuracy","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.accall","value":0.8025,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.accuracy","value":0.8025,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.brier","value":0.3219,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.ece","value":0.1376,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.nll","value":4.2481,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.recall1","value":0.8025,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.recall20","value":0.8575,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.recall5","value":0.8525,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.bi.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.brier","value":0.15,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.ece","value":0.0409,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.nll","value":0.5465,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.recall1","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.recall20","value":0.9875,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.recall5","value":0.9675,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.accall","value":0.19,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.accuracy","value":0.19,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.brier","value":1.3418,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.ece","value":0.6058,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.nll","value":22.1634,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.recall1","value":0.19,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.recall20","value":0.2,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.recall5","value":0.1975,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.sl.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k150.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.accall","value":0.935,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.accuracy","value":0.935,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.brier","value":0.1135,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.ece","value":0.0516,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.nll","value":0.5631,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.recall1","value":0.935,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.recall20","value":1,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.recall5","value":0.99,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k16.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.accall","value":0.9675,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.accuracy","value":0.9675,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.brier","value":0.0564,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.ece","value":0.028,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.nll","value":0.2471,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.recall1","value":0.9675,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.recall20","value":1,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.recall5","value":1,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k4.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.accall","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.accuracy","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.accall","value":0.8875,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.accuracy","value":0.8875,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.brier","value":0.1987,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.ece","value":0.09,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.nll","value":1.7013,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.recall1","value":0.8875,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.recall20","value":0.965,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.recall5","value":0.9525,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.bi.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.brier","value":0.1638,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.ece","value":0.0747,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.nll","value":0.7448,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.recall1","value":0.9075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.recall20","value":0.9925,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.recall5","value":0.9725,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.accall","value":0.075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.accuracy","value":0.075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.brier","value":1.568,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.ece","value":0.7244,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.nll","value":25.4405,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.recall1","value":0.075,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.recall20","value":0.08,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.recall5","value":0.0775,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.rejected","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.sl.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null},{"category":"all","metric":"k77.unanswered","value":0,"lo":null,"hi":null,"n":null,"of":null}],"notes":"device cuda; dtype f32; host dgx; corrected: ended and commit as hanzo-inc/kai 5138293 results/cardinality/cardinality.json records them, runs 2026-10-03 00:11:42Z and 00:20:27Z on weights 0834a74f","revision":"corrected","canonical":true,"visibility":"public","trainable":false,"publishable":false}],"total":1,"completion":{"complete":0,"partial":0,"inconsistent":0,"unknown":1},"gaps":[{"field":"answered","runs":1},{"field":"dataset_digest","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":"questions","runs":1},{"field":"reader","runs":1},{"field":"store_digest","runs":1},{"field":"temperature","runs":1},{"field":"when","runs":1}]}