Spaces:
Sleeping
Sleeping
Update clinicbench_result.json
Browse files- clinicbench_result.json +23 -0
clinicbench_result.json
CHANGED
@@ -12,6 +12,7 @@
|
|
12 |
"General LLMs": "No",
|
13 |
"Medical LLMs": "No",
|
14 |
"SOTA": "Yes",
|
|
|
15 |
"Verified": "Yes"
|
16 |
},
|
17 |
"MedQA": {
|
@@ -56,6 +57,7 @@
|
|
56 |
"General LLMs": "Yes",
|
57 |
"Medical LLMs": "No",
|
58 |
"SOTA": "No",
|
|
|
59 |
"Verified": "Yes"
|
60 |
},
|
61 |
"MedQA": {
|
@@ -121,6 +123,7 @@
|
|
121 |
"General LLMs": "Yes",
|
122 |
"Medical LLMs": "No",
|
123 |
"SOTA": "No",
|
|
|
124 |
"Verified": "Yes"
|
125 |
},
|
126 |
"MedQA": {
|
@@ -186,6 +189,7 @@
|
|
186 |
"General LLMs": "Yes",
|
187 |
"Medical LLMs": "No",
|
188 |
"SOTA": "No",
|
|
|
189 |
"Verified": "Yes"
|
190 |
},
|
191 |
"MedQA": {
|
@@ -251,6 +255,7 @@
|
|
251 |
"General LLMs": "Yes",
|
252 |
"Medical LLMs": "No",
|
253 |
"SOTA": "No",
|
|
|
254 |
"Verified": "Yes"
|
255 |
},
|
256 |
"MedQA": {
|
@@ -316,6 +321,7 @@
|
|
316 |
"General LLMs": "Yes",
|
317 |
"Medical LLMs": "No",
|
318 |
"SOTA": "No",
|
|
|
319 |
"Verified": "Yes"
|
320 |
},
|
321 |
"MedQA": {
|
@@ -381,6 +387,7 @@
|
|
381 |
"General LLMs": "Yes",
|
382 |
"Medical LLMs": "No",
|
383 |
"SOTA": "No",
|
|
|
384 |
"Verified": "Yes"
|
385 |
},
|
386 |
"MedQA": {
|
@@ -446,6 +453,7 @@
|
|
446 |
"General LLMs": "Yes",
|
447 |
"Medical LLMs": "No",
|
448 |
"SOTA": "No",
|
|
|
449 |
"Verified": "Yes"
|
450 |
},
|
451 |
"MedQA": {
|
@@ -511,6 +519,7 @@
|
|
511 |
"General LLMs": "Yes",
|
512 |
"Medical LLMs": "No",
|
513 |
"SOTA": "No",
|
|
|
514 |
"Verified": "Yes"
|
515 |
},
|
516 |
"MedQA": {
|
@@ -576,6 +585,7 @@
|
|
576 |
"General LLMs": "Yes",
|
577 |
"Medical LLMs": "No",
|
578 |
"SOTA": "No",
|
|
|
579 |
"Verified": "Yes"
|
580 |
},
|
581 |
"MedQA": {
|
@@ -641,6 +651,7 @@
|
|
641 |
"General LLMs": "Yes",
|
642 |
"Medical LLMs": "No",
|
643 |
"SOTA": "No",
|
|
|
644 |
"Verified": "Yes"
|
645 |
},
|
646 |
"MedQA": {
|
@@ -706,6 +717,7 @@
|
|
706 |
"General LLMs": "Yes",
|
707 |
"Medical LLMs": "No",
|
708 |
"SOTA": "No",
|
|
|
709 |
"Verified": "Yes"
|
710 |
},
|
711 |
"MedQA": {
|
@@ -771,6 +783,7 @@
|
|
771 |
"General LLMs": "No",
|
772 |
"Medical LLMs": "Yes",
|
773 |
"SOTA": "No",
|
|
|
774 |
"Verified": "Yes"
|
775 |
},
|
776 |
"MedQA": {
|
@@ -836,6 +849,7 @@
|
|
836 |
"General LLMs": "No",
|
837 |
"Medical LLMs": "Yes",
|
838 |
"SOTA": "No",
|
|
|
839 |
"Verified": "Yes"
|
840 |
},
|
841 |
"MedQA": {
|
@@ -901,6 +915,7 @@
|
|
901 |
"General LLMs": "No",
|
902 |
"Medical LLMs": "Yes",
|
903 |
"SOTA": "No",
|
|
|
904 |
"Verified": "Yes"
|
905 |
},
|
906 |
"MedQA": {
|
@@ -966,6 +981,7 @@
|
|
966 |
"General LLMs": "No",
|
967 |
"Medical LLMs": "Yes",
|
968 |
"SOTA": "No",
|
|
|
969 |
"Verified": "Yes"
|
970 |
},
|
971 |
"MedQA": {
|
@@ -1031,6 +1047,7 @@
|
|
1031 |
"General LLMs": "No",
|
1032 |
"Medical LLMs": "Yes",
|
1033 |
"SOTA": "No",
|
|
|
1034 |
"Verified": "Yes"
|
1035 |
},
|
1036 |
"MedQA": {
|
@@ -1096,6 +1113,7 @@
|
|
1096 |
"General LLMs": "No",
|
1097 |
"Medical LLMs": "Yes",
|
1098 |
"SOTA": "No",
|
|
|
1099 |
"Verified": "Yes"
|
1100 |
},
|
1101 |
"MedQA": {
|
@@ -1161,6 +1179,7 @@
|
|
1161 |
"General LLMs": "No",
|
1162 |
"Medical LLMs": "Yes",
|
1163 |
"SOTA": "No",
|
|
|
1164 |
"Verified": "Yes"
|
1165 |
},
|
1166 |
"MedQA": {
|
@@ -1226,6 +1245,7 @@
|
|
1226 |
"General LLMs": "No",
|
1227 |
"Medical LLMs": "Yes",
|
1228 |
"SOTA": "No",
|
|
|
1229 |
"Verified": "Yes"
|
1230 |
},
|
1231 |
"MedQA": {
|
@@ -1291,6 +1311,7 @@
|
|
1291 |
"General LLMs": "No",
|
1292 |
"Medical LLMs": "Yes",
|
1293 |
"SOTA": "No",
|
|
|
1294 |
"Verified": "Yes"
|
1295 |
},
|
1296 |
"MedQA": {
|
@@ -1356,6 +1377,7 @@
|
|
1356 |
"General LLMs": "No",
|
1357 |
"Medical LLMs": "Yes",
|
1358 |
"SOTA": "No",
|
|
|
1359 |
"Verified": "Yes"
|
1360 |
},
|
1361 |
"MedQA": {
|
@@ -1421,6 +1443,7 @@
|
|
1421 |
"General LLMs": "No",
|
1422 |
"Medical LLMs": "Yes",
|
1423 |
"SOTA": "No",
|
|
|
1424 |
"Verified": "Yes"
|
1425 |
},
|
1426 |
"MedQA": {
|
|
|
12 |
"General LLMs": "No",
|
13 |
"Medical LLMs": "No",
|
14 |
"SOTA": "Yes",
|
15 |
+
"OpenSource": "Yes",
|
16 |
"Verified": "Yes"
|
17 |
},
|
18 |
"MedQA": {
|
|
|
57 |
"General LLMs": "Yes",
|
58 |
"Medical LLMs": "No",
|
59 |
"SOTA": "No",
|
60 |
+
"OpenSource": "No",
|
61 |
"Verified": "Yes"
|
62 |
},
|
63 |
"MedQA": {
|
|
|
123 |
"General LLMs": "Yes",
|
124 |
"Medical LLMs": "No",
|
125 |
"SOTA": "No",
|
126 |
+
"OpenSource": "No",
|
127 |
"Verified": "Yes"
|
128 |
},
|
129 |
"MedQA": {
|
|
|
189 |
"General LLMs": "Yes",
|
190 |
"Medical LLMs": "No",
|
191 |
"SOTA": "No",
|
192 |
+
"OpenSource": "No",
|
193 |
"Verified": "Yes"
|
194 |
},
|
195 |
"MedQA": {
|
|
|
255 |
"General LLMs": "Yes",
|
256 |
"Medical LLMs": "No",
|
257 |
"SOTA": "No",
|
258 |
+
"OpenSource": "Yes",
|
259 |
"Verified": "Yes"
|
260 |
},
|
261 |
"MedQA": {
|
|
|
321 |
"General LLMs": "Yes",
|
322 |
"Medical LLMs": "No",
|
323 |
"SOTA": "No",
|
324 |
+
"OpenSource": "Yes",
|
325 |
"Verified": "Yes"
|
326 |
},
|
327 |
"MedQA": {
|
|
|
387 |
"General LLMs": "Yes",
|
388 |
"Medical LLMs": "No",
|
389 |
"SOTA": "No",
|
390 |
+
"OpenSource": "Yes",
|
391 |
"Verified": "Yes"
|
392 |
},
|
393 |
"MedQA": {
|
|
|
453 |
"General LLMs": "Yes",
|
454 |
"Medical LLMs": "No",
|
455 |
"SOTA": "No",
|
456 |
+
"OpenSource": "Yes",
|
457 |
"Verified": "Yes"
|
458 |
},
|
459 |
"MedQA": {
|
|
|
519 |
"General LLMs": "Yes",
|
520 |
"Medical LLMs": "No",
|
521 |
"SOTA": "No",
|
522 |
+
"OpenSource": "Yes",
|
523 |
"Verified": "Yes"
|
524 |
},
|
525 |
"MedQA": {
|
|
|
585 |
"General LLMs": "Yes",
|
586 |
"Medical LLMs": "No",
|
587 |
"SOTA": "No",
|
588 |
+
"OpenSource": "Yes",
|
589 |
"Verified": "Yes"
|
590 |
},
|
591 |
"MedQA": {
|
|
|
651 |
"General LLMs": "Yes",
|
652 |
"Medical LLMs": "No",
|
653 |
"SOTA": "No",
|
654 |
+
"OpenSource": "Yes",
|
655 |
"Verified": "Yes"
|
656 |
},
|
657 |
"MedQA": {
|
|
|
717 |
"General LLMs": "Yes",
|
718 |
"Medical LLMs": "No",
|
719 |
"SOTA": "No",
|
720 |
+
"OpenSource": "Yes",
|
721 |
"Verified": "Yes"
|
722 |
},
|
723 |
"MedQA": {
|
|
|
783 |
"General LLMs": "No",
|
784 |
"Medical LLMs": "Yes",
|
785 |
"SOTA": "No",
|
786 |
+
"OpenSource": "Yes",
|
787 |
"Verified": "Yes"
|
788 |
},
|
789 |
"MedQA": {
|
|
|
849 |
"General LLMs": "No",
|
850 |
"Medical LLMs": "Yes",
|
851 |
"SOTA": "No",
|
852 |
+
"OpenSource": "Yes",
|
853 |
"Verified": "Yes"
|
854 |
},
|
855 |
"MedQA": {
|
|
|
915 |
"General LLMs": "No",
|
916 |
"Medical LLMs": "Yes",
|
917 |
"SOTA": "No",
|
918 |
+
"OpenSource": "Yes",
|
919 |
"Verified": "Yes"
|
920 |
},
|
921 |
"MedQA": {
|
|
|
981 |
"General LLMs": "No",
|
982 |
"Medical LLMs": "Yes",
|
983 |
"SOTA": "No",
|
984 |
+
"OpenSource": "Yes",
|
985 |
"Verified": "Yes"
|
986 |
},
|
987 |
"MedQA": {
|
|
|
1047 |
"General LLMs": "No",
|
1048 |
"Medical LLMs": "Yes",
|
1049 |
"SOTA": "No",
|
1050 |
+
"OpenSource": "Yes",
|
1051 |
"Verified": "Yes"
|
1052 |
},
|
1053 |
"MedQA": {
|
|
|
1113 |
"General LLMs": "No",
|
1114 |
"Medical LLMs": "Yes",
|
1115 |
"SOTA": "No",
|
1116 |
+
"OpenSource": "Yes",
|
1117 |
"Verified": "Yes"
|
1118 |
},
|
1119 |
"MedQA": {
|
|
|
1179 |
"General LLMs": "No",
|
1180 |
"Medical LLMs": "Yes",
|
1181 |
"SOTA": "No",
|
1182 |
+
"OpenSource": "Yes",
|
1183 |
"Verified": "Yes"
|
1184 |
},
|
1185 |
"MedQA": {
|
|
|
1245 |
"General LLMs": "No",
|
1246 |
"Medical LLMs": "Yes",
|
1247 |
"SOTA": "No",
|
1248 |
+
"OpenSource": "Yes",
|
1249 |
"Verified": "Yes"
|
1250 |
},
|
1251 |
"MedQA": {
|
|
|
1311 |
"General LLMs": "No",
|
1312 |
"Medical LLMs": "Yes",
|
1313 |
"SOTA": "No",
|
1314 |
+
"OpenSource": "Yes",
|
1315 |
"Verified": "Yes"
|
1316 |
},
|
1317 |
"MedQA": {
|
|
|
1377 |
"General LLMs": "No",
|
1378 |
"Medical LLMs": "Yes",
|
1379 |
"SOTA": "No",
|
1380 |
+
"OpenSource": "Yes",
|
1381 |
"Verified": "Yes"
|
1382 |
},
|
1383 |
"MedQA": {
|
|
|
1443 |
"General LLMs": "No",
|
1444 |
"Medical LLMs": "Yes",
|
1445 |
"SOTA": "No",
|
1446 |
+
"OpenSource": "Yes",
|
1447 |
"Verified": "Yes"
|
1448 |
},
|
1449 |
"MedQA": {
|