Spaces:
Sleeping
Sleeping
File size: 163,581 Bytes
5b6c556 abd5070 5b6c556 abd5070 2994d17 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 2994d17 5b6c556 2994d17 5b6c556 abd5070 5b6c556 abd5070 2994d17 5b6c556 abd5070 5b6c556 abd5070 5b6c556 2994d17 5b6c556 2994d17 5b6c556 2994d17 5b6c556 abd5070 5b6c556 abd5070 2994d17 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 abd5070 5b6c556 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 |
#!/usr/bin/env python3
# This page displays interactive attribution graphs for circuit tracing.
import streamlit as st
import json
import os
import plotly.graph_objects as go
import plotly.express as px
import pandas as pd
import networkx as nx
import numpy as np
from pathlib import Path
# ... (keep other imports)
# --- Cache Helper ---
def update_circuit_cache(prompt, category, key, data_type, content):
# Don't save rate limit errors or failure messages to cache
if isinstance(content, str):
error_markers = ["API request failed", "rate limit exceeded", "Unable to generate explanation", "Error:"]
if any(marker in content for marker in error_markers):
print(f"Skipping cache update for error content: {content[:50]}...")
return
cache_file = Path(__file__).parent.parent / "cache" / "cached_circuit_trace_results.json"
os.makedirs(cache_file.parent, exist_ok=True)
try:
if cache_file.exists():
with open(cache_file, "r", encoding="utf-8") as f:
cached_data = json.load(f)
else:
cached_data = {}
except:
cached_data = {}
# Recursive serializer to handle numpy types
def make_serializable(obj):
if isinstance(obj, np.ndarray):
return obj.tolist()
if isinstance(obj, (np.float32, np.float64, np.float16)):
return float(obj)
if isinstance(obj, (np.int32, np.int64, np.int16)):
return int(obj)
if isinstance(obj, (np.bool_, bool)):
return bool(obj)
if isinstance(obj, dict):
return {k: make_serializable(v) for k, v in obj.items()}
if isinstance(obj, list):
return [make_serializable(v) for v in obj]
return obj
if prompt not in cached_data:
cached_data[prompt] = {}
if category not in cached_data[prompt]:
cached_data[prompt][category] = {}
if category == 'circuit_graph':
# circuit_graph is a direct dict, not keyed by dynamic IDs
cached_data[prompt][category][data_type] = make_serializable(content)
else:
# subnetworks and feature_explorer are keyed by the dynamic cache key
if key not in cached_data[prompt][category]:
cached_data[prompt][category][key] = {}
cached_data[prompt][category][key][data_type] = make_serializable(content)
with open(cache_file, "w", encoding="utf-8") as f:
json.dump(cached_data, f, ensure_ascii=False, indent=4)
print(f"Saved circuit trace {category} {data_type} to cache.")
import requests
import base64
from io import BytesIO
from PIL import Image
from utilities.localization import tr
import markdown
from utilities.utils import init_qwen_api
import re
from utilities.feedback_survey import display_circuit_trace_feedback
from fuzzywuzzy import process
from typing import Set, Optional, List
# --- Qwen API for Explanations ---
@st.cache_data(persist=True)
def explain_circuit_visualization(_api_config, img_base64, structured_prompt, max_tokens_for_request=750):
# Generates an explanation for a circuit tracing visualization.
# Note: We DO NOT catch exceptions here so that Streamlit does NOT cache failures.
# Prepare the API request.
headers = {
"Authorization": f"Bearer {_api_config['api_key']}",
"Content-Type": "application/json"
}
data = {
"model": _api_config["model"],
"messages": [
{
"role": "user",
"content": [
{
"type": "text",
"text": structured_prompt
},
{
"type": "image_url",
"image_url": {
"url": f"data:image/png;base64,{img_base64}"
}
}
]
}
],
"max_tokens": max_tokens_for_request,
"temperature": 0.2,
"top_p": 0.95
}
# Make the API request.
response = requests.post(
f"{_api_config['api_endpoint']}/chat/completions",
headers=headers,
json=data,
timeout=120
)
if response.status_code == 200:
result = response.json()
return result["choices"][0]["message"]["content"]
else:
# Raise an exception so this result is NOT cached
raise Exception(f"API request failed with status {response.status_code}: {response.text}")
# --- Faithfulness Verification for Circuit Tracing ---
def _extract_interpretations_from_text(claim_text):
# Extracts candidate feature interpretations from a claim sentence.
if not claim_text:
return []
candidates = []
# Capture quoted phrases (supports standard and smart quotes).
quote_pattern = r'[""'']([^""'']+)[""'']'
for match in re.findall(quote_pattern, claim_text):
cleaned = match.strip()
cleaned = re.sub(r"\bfeature(s)?\b", "", cleaned, flags=re.IGNORECASE).strip()
if cleaned and not re.match(r"feature_\d+", cleaned, re.IGNORECASE):
candidates.append(cleaned)
# Capture phrases introduced by connectors when not quoted.
relation_patterns = [r"related to ([^.;,]+)", r"focused on ([^.;,]+)", r"tied to ([^.;,]+)"]
for pattern in relation_patterns:
for segment in re.findall(pattern, claim_text, flags=re.IGNORECASE):
parts = re.split(r"\band\b", segment, flags=re.IGNORECASE)
for part in parts:
cleaned = part.strip(" .")
cleaned = re.sub(r"\bfeature(s)?\b", "", cleaned, flags=re.IGNORECASE).strip()
if cleaned and not re.match(r"feature_\d+", cleaned, re.IGNORECASE):
candidates.append(cleaned)
unique_candidates = []
seen = set()
for cand in candidates:
key = cand.lower()
if key not in seen:
seen.add(key)
unique_candidates.append(cand)
return unique_candidates
def _ensure_causal_claim_feature_lists(claims):
# Ensures causal claims include explicit source/target interpretation lists.
for claim in claims or []:
if claim.get('claim_type') != 'causal_claim':
continue
details = claim.get('details') or {}
if not isinstance(details, dict):
details = {}
relationship = (details.get('relationship') or '').lower()
if relationship not in {'upstream', 'downstream'}:
continue
key = 'source_feature_interpretations' if relationship == 'upstream' else 'target_feature_interpretations'
existing = details.get(key)
if isinstance(existing, list) and existing:
continue
extracted = _extract_interpretations_from_text(claim.get('claim_text', ''))
if extracted:
details[key] = extracted
claim['details'] = details
def _stringify_summary(summary):
if summary is None:
return ""
if isinstance(summary, str):
return summary
if isinstance(summary, dict):
return " ".join(str(v) for v in summary.values() if v)
if isinstance(summary, (list, tuple, set)):
return " ".join(filter(None, (_stringify_summary(item) for item in summary)))
return str(summary)
@st.cache_data(persist=True)
def _cached_extract_circuit_claims(api_config, explanation_text, context, cache_version="faithfulness-2025-11-30"):
# Extracts verifiable claims from an AI explanation on the circuit tracing page.
headers = {
"Authorization": f"Bearer {api_config['api_key']}",
"Content-Type": "application/json"
}
# For the main circuit graph, process paragraph by paragraph to improve accuracy.
if context == "circuit_graph":
paragraphs = re.split(r'(?=####\s)', explanation_text.strip())
all_claims = []
for paragraph in paragraphs:
if not paragraph.strip():
continue
# Prompt building logic for each paragraph
claim_types_details = tr('circuit_graph_claim_types')
rules = tr('claim_extraction_prompt_rule')
claim_extraction_prompt = f"""{tr('claim_extraction_prompt_header')}
{tr('claim_extraction_prompt_instruction')}
{rules}
{tr('claim_extraction_prompt_context_header').format(analysis_method=context, context=context)}
{tr('claim_extraction_prompt_types_header')}
{claim_types_details}
{tr('claim_extraction_prompt_analyze_header')}
"{paragraph}"
{tr('claim_extraction_prompt_footer')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": claim_extraction_prompt}], "temperature": 0.0}
try:
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
claims_text = response.json()["choices"][0]["message"]["content"]
if '```json' in claims_text:
claims_text = re.search(r'```json\n(.*?)\n```', claims_text, re.DOTALL).group(1)
claims_from_paragraph = json.loads(claims_text)
_ensure_causal_claim_feature_lists(claims_from_paragraph)
if claims_from_paragraph:
all_claims.extend(claims_from_paragraph)
except (requests.RequestException, AttributeError, json.JSONDecodeError) as e:
if isinstance(e, requests.RequestException):
raise e # Don't cache partial results on network error
continue # If a paragraph fails due to parsing, continue.
_ensure_causal_claim_feature_lists(all_claims)
return all_claims
# Original logic for other, shorter contexts
rules = tr('claim_extraction_prompt_rule')
if context == "feature_explorer":
claim_types_details = tr('feature_explorer_claim_types')
rules += "\n3. **Group related sentences:** If a sentence states a factual observation (e.g., lists top activating tokens) and the immediately following sentence provides reasoning or an explanation for that observation, you MUST extract them as a single claim, combining their text."
elif context == "subnetwork_graph":
claim_types_details = tr('subnetwork_graph_claim_types')
rules += "\n6. **For causal claims in the subnetwork context,** you MUST populate the `source_feature_interpretations` or `target_feature_interpretations` arrays with every feature interpretation referenced in the claim. Use the exact phrasing from the explanation whenever possible."
else: # Should not happen, but as a fallback
return []
claim_extraction_prompt = f"""{tr('claim_extraction_prompt_header')}
{tr('claim_extraction_prompt_instruction')}
{rules}
{tr('claim_extraction_prompt_context_header').format(analysis_method=context, context=context)}
{tr('claim_extraction_prompt_types_header')}
{claim_types_details}
{tr('claim_extraction_prompt_analyze_header')}
"{explanation_text}"
{tr('claim_extraction_prompt_footer')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": claim_extraction_prompt}], "temperature": 0.0}
try:
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
claims_text = response.json()["choices"][0]["message"]["content"]
if '```json' in claims_text:
claims_text = re.search(r'```json\n(.*?)\n```', claims_text, re.DOTALL).group(1)
claims = json.loads(claims_text)
_ensure_causal_claim_feature_lists(claims)
return claims
except (requests.RequestException, AttributeError, json.JSONDecodeError) as e:
if isinstance(e, requests.RequestException):
raise e
return []
@st.cache_data(persist=True)
def _cached_verify_semantic_summary(api_config, claimed_summary, actual_data_points, layer_section, cache_version="faithfulness-2025-11-30"):
# Uses an LLM to verify if a claimed summary is faithful to a list of data points.
headers = {"Authorization": f"Bearer {api_config['api_key']}", "Content-Type": "application/json"}
# Get the principle for the specific layer section.
principles = {
"early": tr('semantic_verification_principle_early'),
"middle": tr('semantic_verification_principle_middle'),
"late": tr('semantic_verification_principle_late'),
}
principle = principles.get(layer_section, "No specific principle defined for this section.")
# A more forceful, hardcoded rule to prevent incorrect contradictions on valid generalizations.
rule_3_override = "3. **A claim is valid even if it only describes one aspect of a layer's function.** The summary does not need to be comprehensive. As long as the aspect it describes is supported by the data or the general principles, you MUST verify it. You MUST NOT contradict a claim because it 'does not fully capture' all functions of the layer."
rule_4_confidence = "4. **Meta-claims about confidence/importance:** If the claimed summary describes the *strength*, *confidence*, *importance*, or *role in final decision* (e.g., 'strong confidence', 'key factor', 'primary driver'), you MUST verify it as True, provided it does not explicitly contradict the layer's general role. You MUST NOT contradict such claims just because the actual features list specific topics (e.g., 'French', 'Geography')."
section_synonym_guidance = {
"early": "You MUST treat descriptions such as 'dissecting the input', 'breaking the sentence into fundamental components', 'token breakdown', or 'parsing basic structure' as equivalent to handling syntax, grammar, and basic patterns.",
"middle": "You MUST treat descriptions referring to linking recognitions into themes, building relationships, developing context, or combining earlier syntax with more complex constructs as equivalent to developing thematic connections and abstract meaning. You MUST accept statements about gaining a nuanced understanding or moving toward higher-level abstractions as faithful summaries for middle layers, even if the listed features mention programming syntax.",
"late": "You MUST treat descriptions about synthesizing information, finalizing outputs, or producing coherent answers as equivalent to the late layers' role of integrating all information to finalize the output.",
}
synonym_guidance = section_synonym_guidance.get(layer_section, "")
prompt = f"""{tr('semantic_verification_prompt_header')}
{tr('semantic_verification_prompt_rules_header')}
{tr('semantic_verification_prompt_rule_1')}
{tr('semantic_verification_prompt_rule_2').format(layer_section=layer_section, principle=principle)}
{rule_3_override}
{rule_4_confidence}
{synonym_guidance}
{tr('semantic_verification_prompt_actual_data_header')}
{actual_data_points}
{tr('semantic_verification_prompt_claimed_summary_header')}
"{claimed_summary}"
{tr('semantic_verification_prompt_task_header')}
{tr('semantic_verification_prompt_task_instruction')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": prompt}], "temperature": 0.0, "response_format": {"type": "json_object"}}
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
try:
return json.loads(response.json()["choices"][0]["message"]["content"])
except (json.JSONDecodeError, KeyError):
return {"is_verified": False, "reasoning": "Could not parse semantic verification result."}
@st.cache_data(persist=True)
def _cached_verify_feature_role_claim(api_config, claimed_role, feature_data, layer_name, neighbor_info=None, cache_version="faithfulness-2025-11-30"):
# Uses an LLM to semantically verify a claim about a feature's role.
headers = {"Authorization": f"Bearer {api_config['api_key']}", "Content-Type": "application/json"}
# Prepare the evidence from the feature data.
interpretation = feature_data.get('interpretation', 'N/A')
top_tokens = [act['token'] for act in feature_data.get('top_activations', [])[:5]]
try:
layer_index = int(layer_name.split('_')[1])
if layer_index <= 10: layer_pos = "early"
elif 11 <= layer_index <= 21: layer_pos = "middle"
else: layer_pos = "late"
except (IndexError, ValueError):
layer_pos = "unknown"
layer_guidance_map = {
"early": tr('semantic_verification_prompt_feature_role_guidance_early'),
"middle": tr('semantic_verification_prompt_feature_role_guidance_middle'),
"late": tr('semantic_verification_prompt_feature_role_guidance_late'),
}
layer_guidance = layer_guidance_map.get(layer_pos, "")
feature_evidence = f"""
- **Feature Interpretation:** "{interpretation}"
- **Top Activating Tokens:** {top_tokens}
- **Layer Position:** {layer_pos} ({layer_name})
"""
# Add neighbor info if available
if neighbor_info:
if neighbor_info.get('upstream'):
feature_evidence += "\n" + tr('semantic_verification_prompt_feature_role_upstream_header').format(interpretations=neighbor_info['upstream'])
if neighbor_info.get('downstream'):
feature_evidence += "\n" + tr('semantic_verification_prompt_feature_role_downstream_header').format(interpretations=neighbor_info['downstream'])
rule_3 = tr('semantic_verification_prompt_feature_role_rule_3') if neighbor_info else ""
prompt = f"""{tr('semantic_verification_prompt_feature_role_header')}
{tr('semantic_verification_prompt_feature_role_rules_header')}
{tr('semantic_verification_prompt_feature_role_rule_1')}
{tr('semantic_verification_prompt_feature_role_rule_2')}
{layer_guidance}
{rule_3}
{tr('semantic_verification_prompt_feature_role_evidence_header')}
{feature_evidence}
{tr('semantic_verification_prompt_feature_role_claimed_role_header')}
"{claimed_role}"
{tr('semantic_verification_prompt_task_header')}
{tr('semantic_verification_prompt_task_instruction')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": prompt}], "temperature": 0.0, "response_format": {"type": "json_object"}}
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
try:
return json.loads(response.json()["choices"][0]["message"]["content"])
except (json.JSONDecodeError, KeyError):
return {"is_verified": False, "reasoning": "Could not parse semantic verification result."}
@st.cache_data(persist=True)
def _cached_verify_subnetwork_purpose(api_config, claimed_purpose, actual_data_points, cache_version="faithfulness-2025-11-30"):
# Uses an LLM to verify a subnetwork's purpose.
headers = {"Authorization": f"Bearer {api_config['api_key']}", "Content-Type": "application/json"}
prompt = f"""{tr('semantic_verification_prompt_subnetwork_header')}
{tr('semantic_verification_prompt_subnetwork_rules_header')}
{tr('semantic_verification_prompt_subnetwork_rule_1')}
{tr('semantic_verification_prompt_subnetwork_rule_2')}
{tr('semantic_verification_prompt_subnetwork_actual_data_header')}
{actual_data_points}
{tr('semantic_verification_prompt_subnetwork_claimed_purpose_header')}
"{claimed_purpose}"
{tr('semantic_verification_prompt_task_header')}
{tr('semantic_verification_prompt_task_instruction')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": prompt}], "temperature": 0.0, "response_format": {"type": "json_object"}}
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
try:
return json.loads(response.json()["choices"][0]["message"]["content"])
except (json.JSONDecodeError, KeyError):
return {"is_verified": False, "reasoning": "Could not parse semantic verification result."}
@st.cache_data(persist=True)
def _cached_verify_token_reasoning(api_config, claimed_explanation, feature_data, layer_name, cache_version="faithfulness-2025-11-30"):
# Uses an LLM to semantically verify the reasoning behind a token activation claim.
headers = {"Authorization": f"Bearer {api_config['api_key']}", "Content-Type": "application/json"}
interpretation = feature_data.get('interpretation', 'N/A')
top_tokens = [act['token'] for act in feature_data.get('top_activations', [])[:5]]
try:
layer_index = int(layer_name.split('_')[1])
if layer_index <= 10: layer_pos = "early"
elif 11 <= layer_index <= 21: layer_pos = "middle"
else: layer_pos = "late"
except (IndexError, ValueError):
layer_pos = "unknown"
feature_evidence = f"""
- **Feature Interpretation:** "{interpretation}"
- **Top Activating Tokens:** {top_tokens}
- **Layer Position:** {layer_pos} ({layer_name})
"""
prompt = f"""{tr('semantic_verification_prompt_token_reasoning_header')}
{tr('semantic_verification_prompt_token_reasoning_rules_header')}
{tr('semantic_verification_prompt_token_reasoning_rule_1')}
{tr('semantic_verification_prompt_token_reasoning_rule_2')}
{tr('semantic_verification_prompt_token_reasoning_evidence_header')}
{feature_evidence}
{tr('semantic_verification_prompt_token_reasoning_claimed_explanation_header')}
"{claimed_explanation}"
{tr('semantic_verification_prompt_task_header')}
{tr('semantic_verification_prompt_task_instruction')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": prompt}], "temperature": 0.0, "response_format": {"type": "json_object"}}
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
try:
return json.loads(response.json()["choices"][0]["message"]["content"])
except (json.JSONDecodeError, KeyError):
return {"is_verified": False, "reasoning": "Could not parse semantic verification result."}
@st.cache_data(persist=True)
def _cached_verify_causal_reasoning(api_config, claimed_explanation, source_interpretations, target_interpretations, central_feature_info=None, cache_version="faithfulness-2025-11-30"):
# Uses an LLM to semantically verify causal reasoning.
headers = {"Authorization": f"Bearer {api_config['api_key']}", "Content-Type": "application/json"}
central_feature_context = ""
if central_feature_info:
central_feature_context = f"\n- **Central Feature Context:** {central_feature_info}"
causal_evidence = f"""
- **Source Feature(s) Interpretations:** {source_interpretations}
- **Target Feature(s) Interpretations:** {target_interpretations}{central_feature_context}
"""
prompt = f"""{tr('semantic_verification_prompt_causal_reasoning_header')}
{tr('semantic_verification_prompt_causal_reasoning_rules_header')}
{tr('semantic_verification_prompt_causal_reasoning_rule_1')}
{tr('semantic_verification_prompt_causal_reasoning_rule_2')}
{tr('semantic_verification_prompt_causal_reasoning_evidence_header')}
{causal_evidence}
{tr('semantic_verification_prompt_causal_reasoning_claimed_explanation_header')}
"{claimed_explanation}"
{tr('semantic_verification_prompt_task_header')}
{tr('semantic_verification_prompt_task_instruction')}
"""
data = {"model": "qwen2.5-vl-72b-instruct", "messages": [{"role": "user", "content": prompt}], "temperature": 0.0, "response_format": {"type": "json_object"}}
response = requests.post(f"{api_config['api_endpoint']}/chat/completions", headers=headers, json=data, timeout=60)
response.raise_for_status()
try:
return json.loads(response.json()["choices"][0]["message"]["content"])
except (json.JSONDecodeError, KeyError):
return {"is_verified": False, "reasoning": "Could not parse semantic verification result."}
# --- End Faithfulness Verification ---
def format_tokens_for_display(tokens):
# Converts tokenizer-style tokens into a human-readable comma-separated list.
display_tokens = []
for token in tokens or []:
if token is None:
continue
cleaned = str(token)
cleaned = cleaned.replace("Ġ", " ")
cleaned = cleaned.replace("Ċ", "\n")
cleaned = cleaned.replace("▁", " ")
cleaned = cleaned.replace("\u0120", " ")
cleaned = cleaned.replace("\u010a", "\n")
if "\n" in cleaned:
cleaned = cleaned.replace("\n", "\\n")
cleaned = " ".join(cleaned.split())
cleaned = cleaned.strip()
if not cleaned:
continue
display_tokens.append(cleaned)
if not display_tokens:
return ""
return ", ".join(f'"{tok}"' for tok in display_tokens)
def _normalize_token_core(token):
if token is None:
return ""
text = str(token)
replacements = [
("Ġ", " "),
("\u0120", " "),
("▁", " "),
("Ċ", "\n"),
("\u010a", "\n"),
]
for old, new in replacements:
text = text.replace(old, new)
text = text.replace("\n", " ")
# Be careful not to strip the token away if it IS a quote
cleaned = text.strip(" \"'")
if not cleaned and text.strip():
# The token consists entirely of chars being stripped (e.g. ' or " or " "')
# but isn't just whitespace. Preserve it.
text = text.strip()
else:
text = cleaned
text = " ".join(text.split())
return text.lower()
def _normalize_actual_tokens(tokens):
normalized = set()
for token in tokens or []:
base = _normalize_token_core(token)
if not base:
continue
normalized.add(base)
condensed = base.replace(" ", "")
if condensed:
normalized.add(condensed)
normalized.add(f"g{condensed}")
normalized.add(f"g{base}")
for char in base:
if not char.strip():
continue
normalized.add(char.lower())
return normalized
_INTERPRETATION_STOPWORDS = {
"a", "an", "the", "another", "additional", "extra", "other", "more",
"feature", "features"
}
def _clean_interpretation_text(text):
if not text:
return ""
cleaned = str(text).strip()
if cleaned.lower().startswith("identifying "):
cleaned = cleaned[12:]
return cleaned.strip()
def _normalize_interpretation_text(text):
if text is None:
return ""
cleaned = _clean_interpretation_text(text)
cleaned = cleaned.replace("-", " ")
cleaned = re.sub(r'[\"“”\'‘’]', '', cleaned.lower())
cleaned = re.sub(r'\b(' + "|".join(_INTERPRETATION_STOPWORDS) + r')\b', ' ', cleaned)
cleaned = re.sub(r'\s+', ' ', cleaned).strip()
return cleaned
def _prepare_feature_interpretations(features):
formatted = []
fuzzy_candidates = []
normalized_variants = set()
seen_formatted = set()
seen_candidates = set()
for feature in features or []:
layer = feature.get('layer', 'N/A')
feature_name = feature.get('feature_name', 'Unknown feature')
interpretation_raw = feature.get('interpretation', 'N/A')
interpretation_clean = _clean_interpretation_text(interpretation_raw) or interpretation_raw
formatted_str = f"L{layer}: {feature_name} ('{interpretation_clean}')"
if formatted_str not in seen_formatted:
formatted.append(formatted_str)
seen_formatted.add(formatted_str)
candidate_texts = {
interpretation_raw,
interpretation_clean,
feature_name,
feature_name.replace('_', ' ') if feature_name else "",
f"{feature_name} {interpretation_clean}" if feature_name and interpretation_clean else ""
}
for candidate in candidate_texts:
if not candidate:
continue
candidate = candidate.strip()
if not candidate:
continue
if candidate not in seen_candidates:
fuzzy_candidates.append(candidate)
seen_candidates.add(candidate)
normalized = _normalize_interpretation_text(candidate)
if normalized:
normalized_variants.add(normalized)
return formatted, fuzzy_candidates, normalized_variants
def _token_variants_for_match(token):
variants = set()
base = _normalize_token_core(token)
if not base:
return variants
variants.add(base)
condensed = base.replace(" ", "")
if condensed and condensed != base:
variants.add(condensed)
if base.startswith("g") and len(base) > 1:
variants.add(base[1:])
return variants
def _token_matches_actual(token, normalized_set, normalized_list):
variants = _token_variants_for_match(token)
for var in variants:
if var in normalized_set:
return True
for actual in normalized_list:
if not actual:
continue
if var == actual:
return True
if var in actual or actual in var:
return True
return False
def get_circuit_explanation(api_config, fig, analysis_data, visualization_type="circuit_graph"):
# Prepares data and calls the cached explanation function.
try:
# Convert the Plotly figure to an image.
img_bytes = fig.to_image(format="png", width=1200, height=800)
img_base64 = base64.b64encode(img_bytes).decode()
# Get the current language from the session state.
lang = st.session_state.get('lang', 'en')
# Prepare context and instructions based on the visualization type.
if visualization_type == "circuit_graph":
context = prepare_circuit_graph_context(analysis_data)
instruction = (
f"{tr('circuit_graph_instruction_header')}\n\n"
f"{tr('circuit_graph_instruction_intro')}\n\n"
f"{tr('circuit_graph_instruction_early')}\n\n"
f"{tr('circuit_graph_instruction_middle')}\n\n"
f"{tr('circuit_graph_instruction_late')}\n\n"
f"{tr('circuit_graph_instruction_insight')}\n\n"
f"{tr('circuit_graph_instruction_footer')}"
)
max_tokens_for_request = 1200
elif visualization_type == "feature_explorer":
context = prepare_feature_explorer_context(analysis_data)
instruction = (
f"{tr('feature_explorer_instruction_header')}\n\n"
f"{tr('feature_explorer_instruction_role')}\n"
f"{tr('feature_explorer_instruction_activations')}\n"
f"{tr('feature_explorer_instruction_insight')}\n\n"
f"{tr('feature_explorer_instruction_footer')}"
)
max_tokens_for_request = 400
elif visualization_type == "subnetwork_graph":
context = prepare_subnetwork_context(analysis_data)
instruction = (
f"{tr('subnetwork_graph_instruction_header')}\n\n"
f"{tr('subnetwork_graph_instruction_role')}\n"
f"{tr('subnetwork_graph_instruction_upstream')}\n"
f"{tr('subnetwork_graph_instruction_downstream')}\n"
f"{tr('subnetwork_graph_instruction_purpose')}\n\n"
f"{tr('subnetwork_graph_instruction_footer')}"
)
max_tokens_for_request = 500
else:
context = tr('context_unspecified_viz')
instruction = tr('instruction_unspecified_viz')
max_tokens_for_request = 750
structured_prompt = f"""{tr('explanation_prompt_header')}
{tr('explanation_prompt_context_header')}
{context}
{tr('explanation_prompt_instructions_header')}
{instruction}
"""
explanation = explain_circuit_visualization(
api_config,
img_base64,
structured_prompt,
max_tokens_for_request
)
if "API request failed" in explanation or "Error generating explanation" in explanation:
raise Exception(explanation)
return explanation
except Exception as e:
# Propagate the exception so it's not cached as a success string
raise e
def prepare_circuit_graph_context(analysis_data):
# Prepares the context for the circuit graph explanation.
prompt = analysis_data.get('prompt', 'Unknown prompt')
input_tokens = analysis_data.get('input_tokens', [])
layer_summaries = analysis_data.get('layer_summaries', {})
lang = st.session_state.get('lang', 'en')
# Prepare the layer summary context.
summary_context = ""
if layer_summaries:
early_summary = "\n".join([
tr('circuit_graph_context_feature_line').format(
layer=f['layer'],
interpretation=f['interpretation'],
activation=f['activation']
) for f in layer_summaries.get('early', [])
])
middle_summary = "\n".join([
tr('circuit_graph_context_feature_line').format(
layer=f['layer'],
interpretation=f['interpretation'],
activation=f['activation']
) for f in layer_summaries.get('middle', [])
])
late_summary = "\n".join([
tr('circuit_graph_context_feature_line').format(
layer=f['layer'],
interpretation=f['interpretation'],
activation=f['activation']
) for f in layer_summaries.get('late', [])
])
summary_context = f"""
{tr('circuit_graph_context_summary_header')}
{tr('circuit_graph_context_early_header')}
{early_summary if early_summary else tr('circuit_graph_context_no_features')}
{tr('circuit_graph_context_middle_header')}
{middle_summary if middle_summary else tr('circuit_graph_context_no_features')}
{tr('circuit_graph_context_late_header')}
{late_summary if late_summary else tr('circuit_graph_context_no_features')}
"""
tokens_display = format_tokens_for_display(input_tokens) or ' '.join(input_tokens)
return f"""
{tr('circuit_graph_context_header').format(prompt=prompt)}
{tr('circuit_graph_context_tokens').format(tokens=tokens_display)}
{summary_context}
"""
def prepare_subnetwork_context(analysis_data):
# Prepares the context for the subnetwork graph explanation.
prompt = analysis_data.get('prompt', 'Unknown prompt')
central_feature_info = analysis_data.get('central_feature_info', {})
subgraph_stats = analysis_data.get('subgraph_stats', {})
subgraph_neighbors = analysis_data.get('subgraph_neighbors', {})
lang = st.session_state.get('lang', 'en')
# Prepare the context for neighboring features.
upstream_features = subgraph_neighbors.get('upstream', [])
downstream_features = subgraph_neighbors.get('downstream', [])
upstream_context = ""
if upstream_features:
header = tr('subnetwork_context_upstream_header')
feature_lines = [
tr('subnetwork_context_feature_line').format(
layer=feat.get('layer'),
feature_name=feat.get('feature_name'),
interpretation=feat.get('interpretation', 'N/A')
) for feat in upstream_features[:5]
]
upstream_context = header + "\n" + "\n".join(feature_lines)
downstream_context = ""
if downstream_features:
header = tr('subnetwork_context_downstream_header')
feature_lines = [
tr('subnetwork_context_feature_line').format(
layer=feat.get('layer'),
feature_name=feat.get('feature_name'),
interpretation=feat.get('interpretation', 'N/A')
) for feat in downstream_features[:5]
]
downstream_context = header + "\n" + "\n".join(feature_lines)
central_interpretation = central_feature_info.get('interpretation')
if not central_interpretation:
central_interpretation = tr('subnetwork_context_no_interpretation')
return f"""
{tr('subnetwork_context_header').format(prompt=prompt)}
{tr('subnetwork_context_centered_on')}
{tr('subnetwork_context_feature').format(name=central_feature_info.get('name', 'Unknown'))}
{tr('subnetwork_context_layer').format(layer=central_feature_info.get('layer', 'Unknown'))}
{tr('subnetwork_context_interpretation').format(interpretation=central_interpretation)}
{upstream_context}
{downstream_context}
{tr('subnetwork_context_depth').format(depth=analysis_data.get('depth', 'N/A'))}
{tr('subnetwork_context_stats_header')}
{tr('subnetwork_context_stats_nodes').format(nodes=subgraph_stats.get('nodes', 0))}
{tr('subnetwork_context_stats_edges').format(edges=subgraph_stats.get('edges', 0))}
{tr('subnetwork_context_viz_header')}
{tr('subnetwork_context_viz_central')}
{tr('subnetwork_context_viz_nodes')}
{tr('subnetwork_context_viz_lilac')}
{tr('subnetwork_context_viz_other')}
{tr('subnetwork_context_viz_edges')}
"""
def prepare_feature_explorer_context(analysis_data):
# Prepares the context for the feature explorer explanation.
prompt = analysis_data.get('prompt', 'Unknown prompt')
input_tokens = analysis_data.get('input_tokens', [])
selected_layer_str = analysis_data.get('selected_layer', 'layer_unknown')
selected_feature = analysis_data.get('selected_feature', 'Unknown Feature')
lang = st.session_state.get('lang', 'en')
try:
layer_index = int(selected_layer_str.split('_')[1])
if layer_index <= 10:
layer_position_desc = tr('feature_explorer_context_position_early')
elif 11 <= layer_index <= 21:
layer_position_desc = tr('feature_explorer_context_position_middle')
else:
layer_position_desc = tr('feature_explorer_context_position_late')
layer_context = tr('feature_explorer_context_analyzing_feature').format(
feature=selected_feature,
layer=layer_index,
position=layer_position_desc
)
except (IndexError, ValueError):
layer_context = tr('feature_explorer_context_analyzing_feature_no_pos').format(
feature=selected_feature,
layer=selected_layer_str
)
# Safely get the feature data.
feature_data = analysis_data.get('feature_visualizations', {}).get(selected_layer_str, {}).get(selected_feature, {})
interpretation = feature_data.get('interpretation', tr('feature_explorer_context_no_interpretation'))
tokens_display = format_tokens_for_display(input_tokens) or ' '.join(input_tokens)
return f"""
{tr('feature_explorer_context_header').format(prompt=prompt)}
{tr('feature_explorer_context_model_header')}
{layer_context}
{tr('feature_explorer_context_tokens').format(tokens=tokens_display)}
{tr('feature_explorer_context_interpretation').format(interpretation=interpretation)}
{tr('feature_explorer_context_footer')}
"""
@st.cache_data
def load_attribution_results(lang='en'):
# Loads the attribution graph results for the selected language.
# Determine the file path based on the language.
if lang == 'de':
file_path = Path(__file__).parent / 'results/attribution_graphs_results_de.json'
else:
file_path = Path(__file__).parent / 'results/attribution_graphs_results.json'
try:
with open(file_path, 'r', encoding='utf-8') as f:
return json.load(f)
except FileNotFoundError:
st.error(tr('no_results_warning'))
# Provide a more specific error message.
if lang == 'de':
st.info("Bitte führen Sie zuerst die deutsche Analyse aus: `python3 circuit_analysis/attribution_graphs_olmo_de.py --prompt-index 0 --force-retrain-clt`")
else:
st.info(tr('run_analysis_info'))
return None
except json.JSONDecodeError:
st.error(f"Error: Invalid JSON in {file_path}. The file might be corrupted or empty.")
return None
def create_interactive_feature_explorer(analysis, prompt_idx, enable_explanations=False, qwen_api_config=None):
# Creates the interactive feature explorer.
st.subheader(tr('feature_explorer_title').format(prompt=analysis['prompt']))
# Layer selection dropdown.
available_layers = list(analysis['feature_visualizations'].keys())
if not available_layers:
st.warning(tr('no_feature_viz_warning'))
return None
selected_layer = st.selectbox(
tr('select_layer_label'),
available_layers,
format_func=lambda x: tr('layer_label_format').format(layer_num=x.split('_')[1])
)
layer_features = analysis['feature_visualizations'][selected_layer]
if not layer_features:
st.warning(tr('no_features_in_layer_warning').format(selected_layer=selected_layer))
return None
# Create columns for the layout.
col1, col2 = st.columns([1, 2])
with col1:
st.write(tr('active_features_label'))
feature_options = list(layer_features.keys())
selected_feature = st.selectbox(tr('choose_feature_label'), feature_options)
if selected_feature:
feat_data = layer_features[selected_feature]
# Show feature statistics.
st.metric(tr('max_activation_label'), f"{feat_data['max_activation']:.3f}")
st.metric(tr('mean_activation_label'), f"{feat_data['mean_activation']:.3f}")
st.metric(tr('sparsity_label'), f"{feat_data['sparsity']:.3f}")
# Clean up the interpretation text for display.
interpretation = feat_data.get('interpretation', 'N/A')
if interpretation.startswith("Identifying "):
interpretation = interpretation[12:]
st.info(f"**{tr('interpretation_label')}:** {interpretation}")
with col2:
if selected_feature:
feat_data = layer_features[selected_feature]
# Create the activation pattern bar chart.
if 'top_activations' in feat_data:
activation_data = []
for item in feat_data['top_activations']:
activation_data.append({
'token': item['token'],
'position': item['position'],
'activation': item['activation']
})
if activation_data:
df = pd.DataFrame(activation_data)
fig = px.bar(
df,
x='token',
y='activation',
title=tr('top_activating_tokens_title').format(selected_feature=selected_feature),
hover_data=['position'],
color='activation',
color_continuous_scale='viridis'
)
fig.update_layout(
xaxis_title=tr('xaxis_token_label'),
yaxis_title=tr('yaxis_activation_label'),
showlegend=False
)
st.plotly_chart(fig, use_container_width=True)
# Add an AI explanation if enabled.
if enable_explanations and qwen_api_config is not None:
cache_key = f"explanation_feature_explorer_{prompt_idx}_{selected_layer}_{selected_feature}"
if cache_key not in st.session_state:
with st.spinner(tr('generating_feature_explanation_spinner')):
try:
# Add context for the explanation prompt.
analysis_with_context = analysis.copy()
analysis_with_context['selected_layer'] = selected_layer
analysis_with_context['selected_feature'] = selected_feature
explanation = get_circuit_explanation(
qwen_api_config,
fig,
analysis_with_context,
visualization_type="feature_explorer"
)
# Post-process the explanation to ensure proper formatting.
processed_explanation = explanation.replace("- **", "\n- **")
if not processed_explanation.strip().startswith('-'):
processed_explanation = f"- {processed_explanation.strip()}"
st.session_state[cache_key] = processed_explanation
update_circuit_cache(analysis['prompt'], 'feature_explorer', cache_key, 'explanation', processed_explanation)
except Exception as e:
st.error(tr('feature_explanation_error').format(e=str(e)))
st.session_state[cache_key] = "Error: Could not generate explanation."
if st.session_state.get(cache_key) and "Error:" not in st.session_state[cache_key] and "Unable to generate" not in st.session_state[cache_key]:
st.markdown(tr('ai_feature_analysis_header'))
# Convert markdown to HTML for display.
html_explanation = markdown.markdown(st.session_state[cache_key])
st.markdown(f"""
<div style="background-color: #2b2b2b; color: #ffffff; padding: 1.2rem; border-radius: 10px; border-left: 4px solid #dcae36; font-size: 0.9rem; margin-bottom: 1rem;">
{html_explanation}
</div>
""", unsafe_allow_html=True)
# Faithfulness Check for Feature Explorer
with st.expander(tr('faithfulness_check_expander')):
st.markdown(tr('faithfulness_explanation_feature_explorer_html'), unsafe_allow_html=True)
faithfulness_key = f"faithfulness_{cache_key}"
if faithfulness_key in st.session_state:
verification_results = st.session_state[faithfulness_key]
else:
with st.spinner(tr('running_faithfulness_check_spinner')):
analysis_with_context = analysis.copy()
analysis_with_context['selected_layer'] = selected_layer
analysis_with_context['selected_feature'] = selected_feature
try:
claims = _cached_extract_circuit_claims(qwen_api_config, st.session_state[cache_key], "feature_explorer", cache_version="faithfulness-2025-11-30")
verification_results = verify_circuit_claims(claims, analysis_with_context, "feature_explorer")
st.session_state[faithfulness_key] = verification_results
update_circuit_cache(analysis['prompt'], 'feature_explorer', cache_key, 'faithfulness', verification_results)
except Exception as e:
st.warning(f"Faithfulness check failed: {str(e)}")
verification_results = []
if verification_results:
for result in verification_results:
status_text = tr('verified_status') if result['verified'] else tr('contradicted_status')
st.markdown(f"""
<div style="margin-bottom: 1rem; padding: 0.8rem; border-radius: 8px; border-left: 5px solid {'#28a745' if result['verified'] else '#dc3545'}; background-color: #1a1a1a;">
<p style="margin-bottom: 0.3rem;"><strong>{tr('claim_label')}:</strong> <em>"{result['claim_text']}"</em></p>
<p style="margin-bottom: 0.3rem;"><strong>{tr('status_label')}:</strong> {status_text}</p>
<p style="margin-bottom: 0;"><strong>{tr('evidence_label')}:</strong> {result['evidence']}</p>
</div>
""", unsafe_allow_html=True)
else:
st.info(tr('no_verifiable_claims_info'))
return fig
return None
def render_dataset_faithfulness_summary(summary: dict):
if not summary:
return
table_rows = []
for label, key in [
('Targeted', 'targeted'),
('Random baseline', 'random_baseline'),
('Path', 'path'),
('Random path baseline', 'random_path_baseline')
]:
stats = summary.get(key, {}) or {}
table_rows.append({
'Type': label,
'Count': stats.get('count', 0),
'Avg |Δp|': stats.get('avg_abs_probability_change', 0.0),
'Flip rate': stats.get('flip_rate', 0.0),
'Avg |Δlogit|': stats.get('avg_abs_logit_change', 0.0)
})
df = pd.DataFrame(table_rows).set_index('Type')
st.table(df.style.format({'Avg |Δp|': '{:.4f}', 'Flip rate': '{:.2%}', 'Avg |Δlogit|': '{:.4f}'}))
diff_abs = summary.get('target_minus_random_abs_probability_change', 0.0)
diff_flip = summary.get('target_flip_rate_minus_random', 0.0)
path_diff_abs = summary.get('path_minus_random_abs_probability_change', 0.0)
path_diff_flip = summary.get('path_flip_rate_minus_random', 0.0)
st.caption(f"|Δp| difference (targeted − random): {diff_abs:.4f}")
st.caption(f"Flip-rate difference (targeted − random): {diff_flip:.4f}")
st.caption(f"|Δp| difference (path − random path): {path_diff_abs:.4f}")
st.caption(f"Flip-rate difference (path − random path): {path_diff_flip:.4f}")
def render_faithfulness_metrics(analysis: dict, prompt_idx: int):
summary_stats = analysis.get('summary_statistics')
if not summary_stats:
return
targeted_summary = summary_stats.get('targeted', {}) or {}
random_summary = summary_stats.get('random_baseline', {}) or {}
path_summary = summary_stats.get('path', {}) or {}
random_path_summary = summary_stats.get('random_path_baseline', {}) or {}
summary_df = pd.DataFrame(
[
{
'Type': 'Targeted',
'Count': targeted_summary.get('count', 0),
'Avg |Δp|': targeted_summary.get('avg_abs_probability_change', 0.0),
'Flip rate': targeted_summary.get('flip_rate', 0.0),
'Avg |Δlogit|': targeted_summary.get('avg_abs_logit_change', 0.0)
},
{
'Type': 'Random baseline',
'Count': random_summary.get('count', 0),
'Avg |Δp|': random_summary.get('avg_abs_probability_change', 0.0),
'Flip rate': random_summary.get('flip_rate', 0.0),
'Avg |Δlogit|': random_summary.get('avg_abs_logit_change', 0.0)
},
{
'Type': 'Path',
'Count': path_summary.get('count', 0),
'Avg |Δp|': path_summary.get('avg_abs_probability_change', 0.0),
'Flip rate': path_summary.get('flip_rate', 0.0),
'Avg |Δlogit|': path_summary.get('avg_abs_logit_change', 0.0)
},
{
'Type': 'Random path baseline',
'Count': random_path_summary.get('count', 0),
'Avg |Δp|': random_path_summary.get('avg_abs_probability_change', 0.0),
'Flip rate': random_path_summary.get('flip_rate', 0.0),
'Avg |Δlogit|': random_path_summary.get('avg_abs_logit_change', 0.0)
}
]
).set_index('Type')
diff_abs = summary_stats.get('target_minus_random_abs_probability_change', 0.0)
diff_flip = summary_stats.get('target_flip_rate_minus_random', 0.0)
path_diff_abs = summary_stats.get('path_minus_random_abs_probability_change', 0.0)
path_diff_flip = summary_stats.get('path_flip_rate_minus_random', 0.0)
with st.expander("Faithfulness metrics", expanded=False):
st.table(summary_df.style.format({'Avg |Δp|': '{:.4f}', 'Flip rate': '{:.2%}', 'Avg |Δlogit|': '{:.4f}'}))
st.caption(f"|Δp| difference (targeted − random): {diff_abs:.4f}")
st.caption(f"Flip-rate difference (targeted − random): {diff_flip:.4f}")
st.caption(f"|Δp| difference (path − random path): {path_diff_abs:.4f}")
st.caption(f"Flip-rate difference (path − random path): {path_diff_flip:.4f}")
targeted_results = analysis.get('perturbation_experiments', []) or []
random_results = analysis.get('random_baseline_experiments', []) or []
comparison_rows = []
if targeted_results:
for exp in targeted_results:
feature_set = exp.get('feature_set', []) or []
feature_label = exp.get('feature_name')
if not feature_label and feature_set:
feature_label = ", ".join(f"L{item.get('layer')}F{item.get('feature')}" for item in feature_set[:3])
comparison_rows.append({
'Label': feature_label,
'Type': 'Targeted',
'Δp': exp.get('probability_change', 0.0),
'|Δp|': abs(exp.get('probability_change', 0.0)),
'Δlogit': exp.get('logit_change', 0.0),
'|Δlogit|': abs(exp.get('logit_change', 0.0)),
'Flips top': exp.get('ablation_flips_top_prediction', False),
'Interpretation': exp.get('feature_interpretation')
})
if random_results:
for exp in random_results:
comparison_rows.append({
'Label': f"Random {exp.get('trial_index')}",
'Type': 'Random',
'Δp': exp.get('probability_change', 0.0),
'|Δp|': abs(exp.get('probability_change', 0.0)),
'Δlogit': exp.get('logit_change', 0.0),
'|Δlogit|': abs(exp.get('logit_change', 0.0)),
'Flips top': exp.get('ablation_flips_top_prediction', False),
'Interpretation': None
})
if comparison_rows:
comparison_df = pd.DataFrame(comparison_rows)
numeric_cols = ['Δp', '|Δp|', 'Δlogit', '|Δlogit|']
comparison_df[numeric_cols] = comparison_df[numeric_cols].apply(
lambda col: col.map(lambda x: round(float(x), 4))
)
top_targeted = comparison_df[comparison_df['Type'] == 'Targeted'].nlargest(5, '|Δp|')
display_df = pd.concat([top_targeted, comparison_df[comparison_df['Type'] == 'Random']], ignore_index=True)
display_df = display_df.sort_values(['Type', '|Δp|'], ascending=[True, True])
st.markdown("**Targeted vs Random feature ablations (|Δp|)**")
st.plotly_chart(
px.bar(
display_df,
x='|Δp|',
y='Label',
color='Type',
orientation='h',
hover_data=['Δp', 'Δlogit', 'Flips top', 'Interpretation'],
labels={'|Δp|': '|Δp|', 'Label': 'Feature / Trial'},
title=None
),
use_container_width=True
)
path_results = analysis.get('path_ablation_experiments', []) or []
if path_results:
path_rows = []
def _top_token(entries):
if not entries:
return None
first = entries[0]
if isinstance(first, (list, tuple)) and first:
return str(first[0]).strip()
return str(first).strip()
for exp in path_results:
feature_set = exp.get('feature_set', []) or []
feature_label = ", ".join(f"L{item.get('layer')}F{item.get('feature')}" for item in feature_set)
flipped = exp.get('ablation_flips_top_prediction', False)
flip_detail = None
if flipped:
baseline_top = _top_token(exp.get('baseline_top_tokens', []))
ablated_top = _top_token(exp.get('ablated_top_tokens', []))
if baseline_top and ablated_top and baseline_top != ablated_top:
flip_detail = f"{baseline_top} → {ablated_top}"
path_rows.append({
'Description': exp.get('path_description', 'Path'),
'Features': feature_label,
'Δp': exp.get('probability_change', 0.0),
'|Δp|': abs(exp.get('probability_change', 0.0)),
'Δlogit': exp.get('logit_change', 0.0),
'|Δlogit|': abs(exp.get('logit_change', 0.0)),
'Flips top': flipped,
'Flip detail': flip_detail
})
path_df = pd.DataFrame(path_rows)
if not path_df.empty:
numeric_cols = ['Δp', '|Δp|', 'Δlogit', '|Δlogit|']
path_df[numeric_cols] = path_df[numeric_cols].apply(
lambda col: col.map(lambda x: round(float(x), 4))
)
st.markdown("**Path ablations (|Δp|)**")
avg_abs = path_df['|Δp|'].mean()
max_abs = path_df['|Δp|'].max()
flip_rate = path_df['Flips top'].mean()
st.caption(
f"Traced circuits ablated end-to-end. Average |Δp| = {avg_abs:.4f}, "
f"max |Δp| = {max_abs:.4f}, flip rate = {flip_rate:.2%}."
)
chip_style = (
"display:inline-flex; align-items:center; padding:0.25rem 0.55rem;"
"border-radius:999px; background-color:#2f2f2f; font-size:0.85rem;"
"font-weight:600; color:#f8fafc;"
)
arrow_html = "<span style='color:#94a3b8; margin:0 0.4rem;'>→</span>"
path_list = path_df.sort_values('|Δp|', ascending=False)
for _, row in path_list.iterrows():
accent = '#F97316' if row['Flips top'] else '#38BDF8'
description = str(row['Description']) if row['Description'] else 'Path'
segments = [
seg.strip() for seg in description.replace('->', '→').split('→') if seg.strip()
]
if not segments:
segments = [description]
path_nodes_html = ""
for idx, segment in enumerate(segments):
path_nodes_html += f"<span style=\"{chip_style}\">{segment}</span>"
if idx < len(segments) - 1:
path_nodes_html += arrow_html
delta_p = row['|Δp|']
delta_logit = row['Δlogit']
flip_text = 'Prediction flipped' if row['Flips top'] else 'Prediction unchanged'
flip_detail = row.get('Flip detail')
feature_text = row['Features'] if row['Features'] else 'No internal features recorded'
st.markdown(
f"""
<div style="
padding:0.75rem 0.9rem;
border-radius:8px;
background-color:#1f2024;
border:1px solid {accent};
margin-bottom:0.6rem;
">
<div style="display:flex; flex-wrap:wrap; align-items:center; gap:0.35rem;">
{path_nodes_html}
</div>
<div style="margin-top:0.45rem; font-size:0.9rem;">
<span style="font-weight:600; color:#f8fafc;">|Δp|</span>
<span style="color:#f8fafc;">= {delta_p:.4f}</span>
|
<span style="font-weight:600; color:#f8fafc;">Δlogit</span>
<span style="color:#f8fafc;">= {delta_logit:.4f}</span>
|
<span style="color:{accent}; font-weight:600;">{flip_text}</span>
{" | <span style='color:#f8fafc;'>"+flip_detail+"</span>" if flip_detail else ""}
</div>
<div style="margin-top:0.25rem; font-size:0.8rem; color:#cbd5f5;">
{feature_text}
</div>
</div>
""",
unsafe_allow_html=True
)
random_path_results = analysis.get('random_path_baseline_experiments', []) or []
if random_path_results:
random_path_rows = []
for exp in random_path_results:
feature_set = exp.get('feature_set', []) or []
feature_label = ", ".join(f"L{item.get('layer')}F{item.get('feature')}" for item in feature_set)
random_path_rows.append({
'Trial': exp.get('trial_index'),
'Sampled features': feature_label,
'Δp': exp.get('probability_change', 0.0),
'|Δp|': abs(exp.get('probability_change', 0.0)),
'Δlogit': exp.get('logit_change', 0.0),
'|Δlogit|': abs(exp.get('logit_change', 0.0)),
'Flips top': exp.get('ablation_flips_top_prediction', False)
})
random_path_df = pd.DataFrame(random_path_rows)
if not random_path_df.empty:
numeric_cols = ['Δp', '|Δp|', 'Δlogit', '|Δlogit|']
random_path_df[numeric_cols] = random_path_df[numeric_cols].apply(
lambda col: col.map(lambda x: round(float(x), 4))
)
st.markdown("<div style='height:0.75rem;'></div>", unsafe_allow_html=True)
st.markdown("**Random path baselines (|Δp|)**")
avg_abs = random_path_df['|Δp|'].mean()
max_abs = random_path_df['|Δp|'].max()
flip_rate = random_path_df['Flips top'].mean()
st.caption(
f"Randomly sampled paths from the same layer span. "
f"Average |Δp| = {avg_abs:.4f}, max |Δp| = {max_abs:.4f}, "
f"flip rate = {flip_rate:.2%}."
)
chip_style = (
"display:inline-flex; align-items:center; padding:0.25rem 0.55rem;"
"border-radius:999px; background-color:#2f2f2f; font-size:0.85rem;"
"font-weight:600; color:#f8fafc;"
)
arrow_html = "<span style='color:#94a3b8; margin:0 0.4rem;'>→</span>"
baseline_cards = random_path_df.sort_values('|Δp|', ascending=False)
for _, row in baseline_cards.iterrows():
accent = '#F97316' if row['Flips top'] else '#38BDF8'
feature_text = row['Sampled features'] if row['Sampled features'] else 'Randomly sampled features'
feature_tokens = [tok.strip() for tok in feature_text.split(',') if tok.strip()]
feature_nodes_html = ""
for idx, token in enumerate(feature_tokens):
feature_nodes_html += f"<span style=\"{chip_style}\">{token}</span>"
if idx < len(feature_tokens) - 1:
feature_nodes_html += arrow_html
delta_p = row['|Δp|']
delta_logit = row['Δlogit']
flip_text = 'Prediction flipped' if row['Flips top'] else 'Prediction unchanged'
st.markdown(
f"""
<div style="
padding:0.75rem 0.9rem;
border-radius:8px;
background-color:#1f2024;
border:1px solid {accent};
margin-bottom:0.6rem;
">
<div style="display:flex; flex-wrap:wrap; align-items:center; gap:0.35rem;">
{feature_nodes_html if feature_nodes_html else '<span style="color:#64748b;">No feature IDs logged</span>'}
</div>
<div style="margin-top:0.45rem; font-size:0.9rem;">
<span style="font-weight:600; color:#f8fafc;">Random trial</span>
<span style="color:#f8fafc;">#{int(row['Trial']) if pd.notnull(row['Trial']) else '-'}</span>
|
<span style="font-weight:600; color:#f8fafc;">|Δp|</span>
<span style="color:#f8fafc;">= {delta_p:.4f}</span>
|
<span style="font-weight:600; color:#f8fafc;">Δlogit</span>
<span style="color:#f8fafc;">= {delta_logit:.4f}</span>
|
<span style="color:{accent}; font-weight:600;">{flip_text}</span>
</div>
</div>
""",
unsafe_allow_html=True
)
def create_interactive_attribution_graph(analysis, prompt_idx, enable_explanations=False, qwen_api_config=None):
# Creates the interactive attribution graph.
# UI controls for the graph.
col1, col2 = st.columns(2)
with col1:
node_size_factor = st.slider(tr('node_size_label'), 0.5, 3.0, 1.0, 0.1)
with col2:
edge_threshold = st.slider(tr('edge_threshold_label'), 0.0, 1.0, 0.1, 0.05)
layer_spacing = 200
# Create a simplified graph for visualization.
G = nx.DiGraph()
# Add nodes to the graph.
node_positions = {}
node_info = {}
# Add nodes for the input embeddings.
tokens = analysis['input_tokens']
num_tokens = len(tokens)
for i, token in enumerate(tokens):
node_id = f"emb_{i}_{token}"
G.add_node(node_id)
# Invert the y-position to display tokens from top to bottom.
node_positions[node_id] = (0, (num_tokens - 1 - i) * 50)
node_info[node_id] = {
'type': 'embedding',
'token': token,
'layer': -1,
'activation': 1.0
}
# Add nodes for the features.
feature_count = 0
max_layer = -1
layer_feature_counts = {}
for layer_name, layer_features in analysis['feature_visualizations'].items():
layer_idx = int(layer_name.split('_')[1])
max_layer = max(max_layer, layer_idx)
layer_feature_counts[layer_idx] = 0
for feat_idx, (feat_name, feat_data) in enumerate(layer_features.items()):
if feat_data['max_activation'] > edge_threshold:
node_id = f"feat_{layer_idx}_{feat_name}"
G.add_node(node_id)
# Position the nodes in layers.
y_pos = feat_idx * 80 - len(layer_features) * 40
node_positions[node_id] = ((layer_idx + 1) * layer_spacing, y_pos)
# Clean up the interpretation text.
interpretation = feat_data.get('interpretation', 'N/A')
if interpretation.startswith("Identifying "):
interpretation = interpretation[12:]
node_info[node_id] = {
'type': 'feature',
'layer': layer_idx,
'feature_name': feat_name,
'activation': feat_data['max_activation'],
'interpretation': interpretation,
'sparsity': feat_data['sparsity']
}
feature_count += 1
layer_feature_counts[layer_idx] += 1
# Add weighted edges to the graph.
# Organize nodes by layer first.
nodes_by_layer = {}
embedding_nodes = []
for node_id, info in node_info.items():
if info['type'] == 'embedding':
embedding_nodes.append(node_id)
else:
layer = info['layer']
if layer not in nodes_by_layer:
nodes_by_layer[layer] = []
nodes_by_layer[layer].append(node_id)
# Connect embeddings to the first layer of features.
if 0 in nodes_by_layer:
for emb_node in embedding_nodes:
for feat_node in nodes_by_layer[0]:
# Determine connection strength.
weight = node_info[feat_node]['activation'] * 0.7
if weight > edge_threshold:
G.add_edge(emb_node, feat_node, weight=weight)
# Connect features across layers.
sorted_layers = sorted(nodes_by_layer.keys())
for i in range(len(sorted_layers) - 1):
current_layer = sorted_layers[i]
next_layer = sorted_layers[i + 1]
for source_node in nodes_by_layer[current_layer]:
source_info = node_info[source_node]
for target_node in nodes_by_layer[next_layer]:
target_info = node_info[target_node]
# Calculate the connection weight.
source_activation = source_info['activation']
target_activation = target_info['activation']
# The weight is based on activations and similarity.
base_weight = min(source_activation, target_activation) * 0.5
# Add a bonus for similarity in sparsity.
sparsity_similarity = 1.0 - abs(source_info['sparsity'] - target_info['sparsity'])
similarity_bonus = sparsity_similarity * 0.2
final_weight = base_weight + similarity_bonus
# Add the edge if it meets the threshold.
if final_weight > edge_threshold:
G.add_edge(source_node, target_node, weight=final_weight)
# Create individual traces for each edge to allow for varying thickness.
edge_traces = []
# Normalize edge thickness.
all_weights = [G.get_edge_data(u, v).get('weight', 0.1) for u, v in G.edges()]
min_weight = min(all_weights) if all_weights else 0.1
max_weight = max(all_weights) if all_weights else 1.0
for edge in G.edges():
x0, y0 = node_positions[edge[0]]
x1, y1 = node_positions[edge[1]]
weight = G[edge[0]][edge[1]].get('weight', 0.1)
# Normalize weight to a thickness between 0.5 and 4 pixels.
if max_weight > min_weight:
thickness = 0.5 + 3.5 * (weight - min_weight) / (max_weight - min_weight)
else:
thickness = 2
edge_trace = go.Scatter(
x=[x0, x1], y=[y0, y1],
line=dict(width=thickness, color='gray'),
hoverinfo='text',
hovertext=f"Connection Weight: {weight:.3f}<br>Thickness: {thickness:.1f}px",
mode='lines',
showlegend=False,
opacity=0.7
)
edge_traces.append(edge_trace)
# Collect all activations to compute scaling
all_activations = [info['activation'] for info in node_info.values() if info['type'] == 'feature']
if all_activations:
max_activation = max(all_activations)
min_activation = min(all_activations)
# Use square root scaling to compress large values while maintaining proportion
# This prevents extremely large nodes from blocking others
def scale_activation(act):
if max_activation == min_activation:
return 1.0
# Normalize to [0, 1], then apply square root to compress large values
normalized = (act - min_activation) / (max_activation - min_activation)
# Square root compresses large values but keeps relative order
scaled = np.sqrt(normalized)
return scaled
else:
def scale_activation(act):
return 1.0
# Create traces for each node type.
node_traces = {}
for node in G.nodes():
info = node_info[node]
node_type = info['type']
if node_type not in node_traces:
node_traces[node_type] = {
'x': [], 'y': [], 'text': [], 'hovertext': [],
'size': [], 'color': []
}
x, y = node_positions[node]
node_traces[node_type]['x'].append(x)
node_traces[node_type]['y'].append(y)
# Set the node label.
if info['type'] == 'embedding':
label = info['token']
else:
label = ""
node_traces[node_type]['text'].append(label)
# Set the hover information.
if info['type'] == 'embedding':
hover_text = f"Token: {info['token']}<br>Type: Embedding"
else:
hover_text = (f"Feature: {info['feature_name']}<br>"
f"Layer: {info['layer']}<br>"
f"Activation: {info['activation']:.3f}<br>"
f"Sparsity: {info['sparsity']:.3f}<br>"
f"Interpretation: {info['interpretation']}")
node_traces[node_type]['hovertext'].append(hover_text)
# Set the node size and color with scaled activation
if info['type'] == 'feature':
# Scale activation to prevent extremely large nodes
scaled_act = scale_activation(info['activation'])
# Use scaled value for size, but keep original for color
# Increased multiplier for main circuit graph to make nodes more visible
size = scaled_act * 75 * node_size_factor
else:
# Embeddings use original activation
size = info['activation'] * 20 * node_size_factor
node_traces[node_type]['size'].append(max(size, 5))
base_size = max(size, 5)
node_traces[node_type]['color'].append(info['activation'])
def map_path_node_id(raw_id: str) -> Optional[str]:
if raw_id in node_positions:
return raw_id
if raw_id.startswith("feat_"):
parts = raw_id.split('_')
if len(parts) >= 4:
layer_part = parts[1]
feature_part = parts[3]
try:
layer_idx = int(layer_part[1:]) if layer_part.startswith('L') else int(layer_part)
feature_idx = int(feature_part[1:]) if feature_part.startswith('F') else int(feature_part)
candidate = f"feat_{layer_idx}_feature_{feature_idx}"
if candidate in node_positions:
return candidate
except ValueError:
return None
return None
show_paths = st.toggle("Highlight ablated paths", value=False)
path_highlight_traces: List[go.Scatter] = []
path_results = analysis.get('path_ablation_experiments', []) or []
if show_paths and path_results:
for idx, exp in enumerate(path_results[:5]):
raw_nodes = exp.get('path_nodes', []) or []
mapped_sequence: List[str] = []
for node_id in raw_nodes:
mapped = map_path_node_id(node_id)
if mapped:
mapped_sequence.append(mapped)
if not mapped_sequence:
feature_set = exp.get('feature_set', []) or []
for feature in feature_set:
layer = feature.get('layer')
feat_idx = feature.get('feature')
if layer is None or feat_idx is None:
continue
candidate = f"feat_{int(layer)}_feature_{int(feat_idx)}"
if candidate in node_positions:
mapped_sequence.append(candidate)
# Ensure sequence unique order
seen_nodes = set()
ordered_coords = []
for node_id in mapped_sequence:
if node_id in seen_nodes or node_id not in node_positions:
continue
ordered_coords.append(node_positions[node_id])
seen_nodes.add(node_id)
if len(ordered_coords) >= 2:
x_vals = [coord[0] for coord in ordered_coords]
y_vals = [coord[1] for coord in ordered_coords]
try:
label = exp.get('path_description', f"Path {idx + 1}")
label = (label[:60] + "…") if label and len(label) > 60 else label
except Exception:
label = f"Path {idx + 1}"
try:
path_label_prefix = tr('path_highlight_label')
except Exception:
path_label_prefix = "Circuit path"
path_trace = go.Scatter(
x=x_vals,
y=y_vals,
mode='lines+markers',
line=dict(width=2, color='#FFD166', dash='dash'),
marker=dict(
size=10,
color='#FFD166',
line=dict(width=1, color='#1F2933'),
symbol='circle-open'
),
opacity=0.95,
name=f"{path_label_prefix} {idx + 1}",
hoverinfo='text',
hovertext=label or f"Path {idx + 1}"
)
path_highlight_traces.append(path_trace)
# Create labels for all layers.
layer_annotations = []
original_layers = analysis['feature_visualizations']
# Calculate a stable y-position for the labels.
global_max_y = 200
if original_layers:
max_y_per_layer = []
for layer_features in original_layers.values():
num_features = len(layer_features)
if num_features > 0:
# Calculate the max y position for this layer.
max_y_for_layer = (num_features * 40) - 80
max_y_per_layer.append(max_y_for_layer)
if max_y_per_layer:
global_max_y = max(max_y_per_layer)
stable_label_y = global_max_y + 60
# Sort the layers for correct ordering.
sorted_layer_names = sorted(original_layers.keys(), key=lambda x: int(x.split('_')[1]))
for layer_name in sorted_layer_names:
layer_idx = int(layer_name.split('_')[1])
layer_x = (layer_idx + 1) * layer_spacing
layer_annotations.append(dict(
x=layer_x,
y=stable_label_y,
text=f"<b>L{layer_idx}</b>",
showarrow=False,
font=dict(size=14, color="#dcae36"),
xanchor="center",
yanchor="bottom"
))
# Create the final traces for the figure.
traces = edge_traces
traces.extend(path_highlight_traces)
colors = {'embedding': 'lightblue', 'feature': 'lightgreen', 'output': 'orange'}
for node_type, data in node_traces.items():
if data['x']:
trace = go.Scatter(
x=data['x'], y=data['y'],
mode='markers+text',
hoverinfo='text',
hovertext=data['hovertext'],
text=data['text'],
textposition="middle center",
marker=dict(
size=data['size'],
color=data['color'],
colorscale='viridis',
showscale=True if node_type == 'feature' else False,
colorbar=dict(
title=tr('colorbar_title'),
x=0.97,
xanchor="left",
thickness=15,
len=0.7
) if node_type == 'feature' else None,
line=dict(width=2, color='black')
),
name=node_type.title()
)
# Manually translate the legend names.
if node_type == 'embedding':
trace.name = tr('embedding_legend')
elif node_type == 'feature':
trace.name = tr('feature_legend')
traces.append(trace)
# Calculate the total width needed for the plot.
total_width = (max_layer + 2) * layer_spacing + 200
# Create the figure.
fig = go.Figure(data=traces)
# Combine all the annotations.
all_annotations = [
dict(
text=tr('tip_scroll_horizontally'),
showarrow=False,
xref="paper", yref="paper",
x=0.005, y=-0.002,
xanchor='left', yanchor='bottom',
font=dict(color="gray", size=12)
)
] + layer_annotations
fig.update_layout(
showlegend=True,
hovermode='closest',
legend=dict(
orientation='h',
yanchor='bottom',
y=1.07,
xanchor='left',
x=0,
bgcolor='rgba(0,0,0,0)'
),
margin=dict(b=20, l=5, r=120, t=110),
annotations=all_annotations,
xaxis=dict(
showgrid=True,
zeroline=False,
showticklabels=True,
# Enable horizontal scrolling.
range=[0, min(1000, total_width)],
# Add a scrollbar.
rangeslider=dict(
visible=True,
thickness=0.05
),
autorange=True
),
yaxis=dict(showgrid=True, zeroline=False, showticklabels=False, autorange=True),
height=700,
# Set a specific width for the plot.
width=None,
autosize=True
)
# Get a list of unique layers with features.
layers_with_features = set()
for node in G.nodes():
info = node_info.get(node)
if info and info['type'] == 'feature':
layers_with_features.add(info['layer'])
# Add a container with a horizontal scrollbar.
st.markdown("""
<div style="
background-color: #2b2b2b;
padding: 10px;
border-radius: 5px;
margin-bottom: 10px;
border-left: 4px solid #dcae36;
">
<h4 style="margin: 0; color: #dcae36;">{layer_nav_header}</h4>
<p style="margin: 5px 0 0 0; color: #ffffff;">
{layer_nav_desc}
</p>
</div>
""".format(
layer_nav_header=tr('layer_nav_header'),
layer_nav_desc=tr('layer_nav_desc').format(num_layers=len(layers_with_features))
), unsafe_allow_html=True)
# Display the plot.
st.plotly_chart(fig, use_container_width=True, config={
'scrollZoom': True,
'displayModeBar': True,
'modeBarButtonsToRemove': ['zoom2d', 'zoomIn2d', 'zoomOut2d'],
'modeBarButtonsToAdd': ['pan2d', 'autoScale2d', 'select2d', 'lasso2d'],
'modeBarButtons': [['pan2d', 'autoScale2d', 'select2d', 'lasso2d', 'resetScale2d']],
'doubleClick': 'autosize',
'dragmode': 'pan',
'staticPlot': False,
})
# Display faithfulness metrics before the AI explanation section.
st.markdown("---")
render_faithfulness_metrics(analysis, prompt_idx)
# Add an AI explanation if enabled.
if enable_explanations and qwen_api_config is not None:
cache_key = f"explanation_circuit_graph_{prompt_idx}"
# Prepare layer summaries for the prompt and for verification. This must run every time.
all_features = [info for info in node_info.values() if info['type'] == 'feature']
early_layers = (0, 10)
middle_layers = (11, 21)
late_layers = (22, 31)
early_feats = sorted([f for f in all_features if early_layers[0] <= f['layer'] <= early_layers[1]], key=lambda x: x['activation'], reverse=True)
middle_feats = sorted([f for f in all_features if middle_layers[0] <= f['layer'] <= middle_layers[1]], key=lambda x: x['activation'], reverse=True)
late_feats = sorted([f for f in all_features if late_layers[0] <= f['layer'] <= late_layers[1]], key=lambda x: x['activation'], reverse=True)
analysis_with_context = analysis.copy()
analysis_with_context['layer_summaries'] = {
'early': early_feats[:5],
'middle': middle_feats[:5],
'late': late_feats[:5]
}
if cache_key not in st.session_state:
with st.spinner(tr('generating_circuit_explanation_spinner')):
try:
explanation = get_circuit_explanation(
qwen_api_config,
fig,
analysis_with_context,
visualization_type="circuit_graph"
)
st.session_state[cache_key] = explanation
update_circuit_cache(analysis['prompt'], 'circuit_graph', None, 'explanation', explanation)
except Exception as e:
st.error(tr('circuit_explanation_error').format(e=str(e)))
st.session_state[cache_key] = "Error: Could not generate explanation."
if st.session_state.get(cache_key):
explanation_text = st.session_state.get(cache_key, "")
if "Error:" in explanation_text or "Unable to generate" in explanation_text:
st.warning(f"Explanation generation failed: {explanation_text}")
if st.button(tr('retry_button'), key=f"retry_{cache_key}"):
del st.session_state[cache_key]
st.rerun()
else:
# Split the explanation into parts for display.
parts = re.split(r'(?=\n####\s)', explanation_text.strip())
parts = [p.strip() for p in parts if p.strip()]
box_style = "background-color: #2b2b2b; color: #ffffff; padding: 1.2rem; border-radius: 10px; font-size: 0.9rem; margin-bottom: 1rem;"
if not parts:
# Handle a malformed explanation.
st.markdown(f'<div style="{box_style} border-left: 4px solid #dcae36;">{markdown.markdown(explanation_text)}</div>', unsafe_allow_html=True)
else:
intro_part = ""
layers_part = ""
insight_part = ""
# Find the insight part of the explanation.
insight_keywords = ["Primary Insight", "Zentrale Erkenntnis"]
insight_idx = -1
for i, p in enumerate(parts):
first_line = p.split('\n')[0]
if any(keyword in first_line for keyword in insight_keywords):
insight_idx = i
break
intro_part = parts[0]
if insight_idx != -1:
# The insight section was found.
insight_part = parts[insight_idx]
# Everything else is the layers section.
if insight_idx > 1:
layers_part = "\n\n".join(parts[1:insight_idx])
else:
# No insight section was found.
if len(parts) > 1:
layers_part = "\n\n".join(parts[1:])
# Display the parts in colored boxes.
if intro_part:
st.markdown(f'<div style="{box_style} border-left: 4px solid #dcae36;">{markdown.markdown(intro_part)}</div>', unsafe_allow_html=True)
if layers_part:
st.markdown(f'<div style="{box_style} border-left: 4px solid #A78BFA;">{markdown.markdown(layers_part)}</div>', unsafe_allow_html=True)
if insight_part:
st.markdown(f'<div style="{box_style} border-left: 4px solid #6EE7B7;">{markdown.markdown(insight_part)}</div>', unsafe_allow_html=True)
# Faithfulness Check for the entire circuit graph explanation
with st.expander(tr('faithfulness_check_expander')):
st.markdown(tr('faithfulness_explanation_circuit_graph_html'), unsafe_allow_html=True)
faithfulness_cache_key = f"faithfulness_circuit_graph_{prompt_idx}"
if faithfulness_cache_key in st.session_state and st.session_state[faithfulness_cache_key] is not None:
verification_results = st.session_state[faithfulness_cache_key]
else:
with st.spinner(tr('running_faithfulness_check_spinner')):
# Filter explanation to only include layer-specific sections
explanation_parts = re.split(r'(?=####\s)', explanation_text.strip())
# Get keywords for layer sections to make it localization-friendly
def _heading_keyword(loc_key):
text = tr(loc_key)
heading_line = text.split('\n')[0].strip()
heading_line = heading_line.replace('####', '').strip()
return heading_line.split('(')[0].strip()
early_keyword = _heading_keyword('circuit_graph_instruction_early')
middle_keyword = _heading_keyword('circuit_graph_instruction_middle')
late_keyword = _heading_keyword('circuit_graph_instruction_late')
layer_keywords = [early_keyword, middle_keyword, late_keyword]
relevant_parts = []
for part in explanation_parts:
if not part.strip().startswith('####'):
continue
heading = part.split('\n')[0].replace('####', '').strip()
if any(keyword in heading for keyword in layer_keywords):
relevant_parts.append(part)
relevant_text = "".join(relevant_parts) if relevant_parts else explanation_text
try:
claims = _cached_extract_circuit_claims(qwen_api_config, relevant_text, "circuit_graph", cache_version="faithfulness-2025-11-30")
verification_results = verify_circuit_claims(claims, analysis_with_context, "circuit_graph")
st.session_state[faithfulness_cache_key] = verification_results
update_circuit_cache(analysis['prompt'], 'circuit_graph', None, 'faithfulness', verification_results)
except Exception as e:
st.warning(f"Faithfulness check failed: {str(e)}")
verification_results = []
if verification_results:
for result in verification_results:
status_text = tr('verified_status') if result['verified'] else tr('contradicted_status')
st.markdown(f"""
<div style="margin-bottom: 1rem; padding: 0.8rem; border-radius: 8px; border-left: 5px solid {'#28a745' if result['verified'] else '#dc3545'}; background-color: #1a1a1a;">
<p style="margin-bottom: 0.3rem;"><strong>{tr('claim_label')}:</strong> <em>"{result['claim_text']}"</em></p>
<p style="margin-bottom: 0.3rem;"><strong>{tr('status_label')}:</strong> {status_text}</p>
<p style="margin-bottom: 0;"><strong>{tr('evidence_label')}:</strong> {result['evidence']}</p>
</div>
""", unsafe_allow_html=True)
else:
st.info(tr('no_verifiable_claims_info'))
return fig, G, node_info, node_positions
def create_subnetwork_visualization(analysis, G, node_info, node_positions, enable_explanations=False, qwen_api_config=None):
# Creates an interactive visualization for exploring subnetworks.
st.markdown(f"### {tr('subnetwork_explorer_title')}")
st.write(tr('subnetwork_explorer_desc'))
if G.number_of_nodes() == 0:
st.info(tr('subnetwork_graph_empty_info'))
return
# --- UI Controls ---
col1, col2, col3 = st.columns(3)
with col1:
# Layer selection dropdown.
layers_in_graph = sorted(list(set(
info['layer'] for node, info in node_info.items() if info['type'] == 'feature'
)))
if not layers_in_graph:
st.warning(tr('no_features_in_graph_warning'))
return
layer_options = [f"layer_{l}" for l in layers_in_graph]
selected_layer = st.selectbox(
tr('select_layer_label_subnetwork'),
layer_options,
format_func=lambda x: tr('layer_label_format').format(layer_num=x.split('_')[1]),
key="subnetwork_layer_selector"
)
with col2:
# Feature selection dropdown.
layer_idx = int(selected_layer.split('_')[1])
feature_options = sorted([
info['feature_name']
for node, info in node_info.items()
if info['type'] == 'feature' and info['layer'] == layer_idx
])
if not feature_options:
st.warning(tr('no_features_in_layer_subnetwork_warning').format(selected_layer=selected_layer))
return
selected_feature = st.selectbox(
tr('select_feature_label_subnetwork'),
feature_options,
key="subnetwork_feature_selector"
)
with col3:
# Traversal depth slider.
traversal_depth = st.slider(tr('traversal_depth_label'), 1, 5, 2, key="subnetwork_depth_slider")
# --- Subgraph Generation ---
if selected_feature:
selected_node_id = next((
node_id for node_id, info in node_info.items()
if info.get('feature_name') == selected_feature and info.get('layer') == layer_idx
), None)
if selected_node_id:
# Find all reachable nodes up to the given depth.
downstream_nodes = set(nx.dfs_preorder_nodes(G, source=selected_node_id, depth_limit=traversal_depth))
upstream_nodes = set(nx.dfs_preorder_nodes(G.reverse(), source=selected_node_id, depth_limit=traversal_depth))
subgraph_nodes = sorted(list(upstream_nodes.union(downstream_nodes)))
subgraph = G.subgraph(subgraph_nodes)
# --- Layout for visualization and analysis ---
viz_col, analysis_col = st.columns([2, 1])
with viz_col:
# --- Visualization ---
if subgraph.number_of_nodes() > 0:
edge_traces = []
# Create traces for the edges.
all_weights = [subgraph.get_edge_data(u, v).get('weight', 0.1) for u, v in subgraph.edges()]
min_w, max_w = (min(all_weights), max(all_weights)) if all_weights else (0.1, 1.0)
for u, v, data in sorted(subgraph.edges(data=True)):
x0, y0 = node_positions[u]
x1, y1 = node_positions[v]
weight = data.get('weight', 0.1)
thickness = 0.5 + 3.5 * (weight - min_w) / (max_w - min_w) if max_w > min_w else 2
edge_traces.append(go.Scatter(
x=[x0, x1], y=[y0, y1], line=dict(width=thickness, color='gray'),
hoverinfo='text', hovertext=f"Weight: {weight:.3f}", mode='lines', showlegend=False, opacity=0.7
))
# Collect activations for scaling in subnetwork
subgraph_activations = [node_info[n]['activation'] for n in subgraph.nodes() if node_info[n]['type'] == 'feature']
if subgraph_activations:
subgraph_max = max(subgraph_activations)
subgraph_min = min(subgraph_activations)
def scale_subgraph_activation(act):
if subgraph_max == subgraph_min:
return 1.0
normalized = (act - subgraph_min) / (subgraph_max - subgraph_min)
return np.sqrt(normalized)
else:
def scale_subgraph_activation(act):
return 1.0
# Separate node data for styling.
embedding_data = {'x': [], 'y': [], 'text': [], 'hovertext': [], 'size': [], 'color': [], 'ids': []}
feature_data = {'x': [], 'y': [], 'text': [], 'hovertext': [], 'size': [], 'ids': []}
for node in sorted(subgraph.nodes()):
info = node_info[node]
x, y = node_positions[node]
if info['type'] == 'embedding':
target_data = embedding_data
label = info['token']
hover = f"Token: {info['token']}<br>Type: Embedding"
target_data['color'].append(info['activation'])
# Embeddings use original activation
node_size = info['activation'] * 30 + 5
else: # feature
target_data = feature_data
label = f"F{info['feature_name'].split('_')[-1]}"
hover = f"Feature: {info['feature_name']}<br>Layer: {info['layer']}<br>Activation: {info['activation']:.3f}<br>Sparsity: {info.get('sparsity', 0.0):.3f}<br>Interpretation: {info.get('interpretation', 'N/A')}"
# Scale feature activations to prevent extremely large nodes
scaled_act = scale_subgraph_activation(info['activation'])
node_size = scaled_act * 30 + 5
target_data['x'].append(x)
target_data['y'].append(y)
target_data['text'].append(label)
target_data['hovertext'].append(hover)
target_data['size'].append(node_size)
target_data['ids'].append(node)
# Assemble the figure.
final_traces = edge_traces
# Add a trace for the feature nodes.
if feature_data['x']:
final_traces.append(go.Scatter(
x=feature_data['x'], y=feature_data['y'], mode='markers+text', hoverinfo='text',
hovertext=feature_data['hovertext'], text=feature_data['text'], textposition="middle center",
marker=dict(
size=feature_data['size'],
color='purple',
showscale=False, # Keep sub-graph clean
line=dict(width=3, color=['crimson' if nid == selected_node_id else 'black' for nid in feature_data['ids']])
),
name=tr('feature_legend')
))
# Add a trace for the embedding nodes.
if embedding_data['x']:
final_traces.append(go.Scatter(
x=embedding_data['x'], y=embedding_data['y'], mode='markers+text', hoverinfo='text',
hovertext=embedding_data['hovertext'], text=embedding_data['text'], textposition="middle center",
marker=dict(
size=embedding_data['size'],
color=embedding_data['color'],
colorscale='viridis',
showscale=False,
line=dict(width=3, color=['crimson' if nid == selected_node_id else 'black' for nid in embedding_data['ids']])
),
name=tr('embedding_legend')
))
fig_sub = go.Figure(data=final_traces)
# Add layer annotations to the plot.
layer_annotations = []
layers_in_subgraph = sorted(list(set(
node_info[node]['layer']
for node in subgraph.nodes()
if node_info[node]['type'] == 'feature'
)))
if layers_in_subgraph:
max_y_in_subgraph = max(node_positions[node][1] for node in subgraph.nodes())
label_y_pos = max_y_in_subgraph + 25
for layer_idx in layers_in_subgraph:
node_in_layer = next((node for node in subgraph.nodes() if node_info[node].get('layer') == layer_idx), None)
if node_in_layer:
layer_x = node_positions[node_in_layer][0]
layer_annotations.append(dict(
x=layer_x,
y=label_y_pos,
text=f"<b>L{layer_idx}</b>",
showarrow=False,
font=dict(size=14, color="#dcae36"),
xanchor="center",
yanchor="bottom"
))
fig_sub.update_layout(
title=tr('subnetwork_graph_title').format(feature=selected_feature),
showlegend=True, hovermode='closest', margin=dict(b=20, l=5, r=5, t=80),
xaxis=dict(showgrid=False, zeroline=False, showticklabels=False, autorange=True),
yaxis=dict(showgrid=False, zeroline=False, showticklabels=False, autorange=True),
height=600, autosize=True,
annotations=layer_annotations
)
st.plotly_chart(fig_sub, use_container_width=True)
# Add an AI explanation for the subnetwork.
if enable_explanations and qwen_api_config:
cache_key = f"explanation_subnetwork_{analysis['prompt']}_{selected_layer}_{selected_feature}_{traversal_depth}"
# Prepare context for the explanation prompt and for verification. This must run every time.
central_info = node_info.get(selected_node_id, {})
context_data = analysis.copy()
context_data['central_feature_info'] = {
"name": central_info.get('feature_name', 'N/A'),
"layer": central_info.get('layer', 'N/A'),
"interpretation": central_info.get('interpretation', 'N/A'),
}
# Get neighbor details for the prompt.
subgraph_feature_nodes = [
nid for nid in subgraph.nodes()
if node_info[nid]['type'] == 'feature' and nid != selected_node_id
]
central_layer = central_info.get('layer', -1)
# Find upstream tokens from the subgraph's predecessors.
preds = subgraph.predecessors(selected_node_id)
upstream_tokens = [node_info[p]['token'] for p in preds if node_info[p]['type'] == 'embedding']
context_data['subgraph_neighbors'] = {
'upstream': sorted(
[node_info[nid] for nid in subgraph_feature_nodes if node_info[nid].get('layer', -2) < central_layer],
key=lambda x: x.get('activation', 0), reverse=True
),
'downstream': sorted(
[node_info[nid] for nid in subgraph_feature_nodes if node_info[nid].get('layer', -2) > central_layer],
key=lambda x: x.get('activation', 0), reverse=True
),
'upstream_tokens': upstream_tokens
}
context_data['subgraph_stats'] = {
"nodes": subgraph.number_of_nodes(),
"edges": subgraph.number_of_edges(),
}
context_data['depth'] = traversal_depth
if cache_key not in st.session_state:
with st.spinner(tr('generating_subnetwork_explanation_spinner')):
try:
explanation = get_circuit_explanation(
qwen_api_config, fig_sub, context_data, "subnetwork_graph"
)
st.session_state[cache_key] = explanation
update_circuit_cache(analysis['prompt'], 'subnetworks', cache_key, 'explanation', explanation)
except Exception as e:
st.error(f"Error generating subnetwork explanation: {str(e)}")
st.session_state[cache_key] = "Error: Could not generate explanation."
explanation = st.session_state.get(cache_key)
if explanation and "Error:" not in explanation:
st.markdown(tr('ai_subnetwork_analysis_header'))
html_explanation = markdown.markdown(explanation)
st.markdown(f"""
<div style="background-color: #2b2b2b; color: #ffffff; padding: 1.2rem; border-radius: 10px; border-left: 4px solid #A78BFA; font-size: 0.9rem; margin-bottom: 1rem;">
{html_explanation}
</div>
""", unsafe_allow_html=True)
# Faithfulness Check for Subnetwork
with st.expander(tr('faithfulness_check_expander')):
st.markdown(tr('faithfulness_explanation_subnetwork_graph_html'), unsafe_allow_html=True)
faithfulness_key = f"faithfulness_{cache_key}"
if faithfulness_key in st.session_state:
verification_results = st.session_state[faithfulness_key]
else:
with st.spinner(tr('running_faithfulness_check_spinner')):
try:
claims = _cached_extract_circuit_claims(qwen_api_config, explanation, "subnetwork_graph", cache_version="faithfulness-2025-11-30")
verification_results = verify_circuit_claims(claims, context_data, "subnetwork_graph")
st.session_state[faithfulness_key] = verification_results
update_circuit_cache(analysis['prompt'], 'subnetworks', cache_key, 'faithfulness', verification_results)
except Exception as e:
st.warning(f"Faithfulness check failed: {str(e)}")
verification_results = []
if verification_results:
for result in verification_results:
status_text = tr('verified_status') if result['verified'] else tr('contradicted_status')
st.markdown(f"""
<div style="margin-bottom: 1rem; padding: 0.8rem; border-radius: 8px; border-left: 5px solid {'#28a745' if result['verified'] else '#dc3545'}; background-color: #1a1a1a;">
<p style="margin-bottom: 0.3rem;"><strong>{tr('claim_label')}:</strong> <em>"{result['claim_text']}"</em></p>
<p style="margin-bottom: 0.3rem;"><strong>{tr('status_label')}:</strong> {status_text}</p>
<p style="margin-bottom: 0;"><strong>{tr('evidence_label')}:</strong> {result['evidence']}</p>
</div>
""", unsafe_allow_html=True)
else:
st.info(tr('no_verifiable_claims_info'))
else:
st.info(tr('subnetwork_no_connections_info'))
with analysis_col:
st.markdown(f"#### {tr('subnetwork_analysis_title')}")
# Find all features in the subnetwork.
subgraph_features = [
(node, node_info[node])
for node in subgraph.nodes()
if node_info[node]['type'] == 'feature'
]
if not subgraph_features:
st.info(tr('subnetwork_no_features_info'))
else:
# Aggregate top activating tokens from all features.
all_top_activations = []
for _, feat_info in subgraph_features:
layer_name = f"layer_{feat_info['layer']}"
feat_name = feat_info['feature_name']
# Get token activations from the original analysis data.
viz_data = analysis.get('feature_visualizations', {}).get(layer_name, {}).get(feat_name, {})
if 'top_activations' in viz_data:
for act_item in viz_data['top_activations']:
all_top_activations.append({
'token': act_item['token'],
'activation': act_item['activation'],
'feature': feat_name
})
if not all_top_activations:
st.info(tr('subnetwork_no_token_info'))
else:
# Find the top 10 unique tokens by max activation.
df_activations = pd.DataFrame(all_top_activations)
top_tokens = df_activations.loc[df_activations.groupby('token')['activation'].idxmax()]
top_tokens = top_tokens.nlargest(10, 'activation')
st.write(tr('subnetwork_top_tokens_desc'))
for _, row in top_tokens.iterrows():
# Normalize activation for the bar display.
normalized_act = (row['activation'] - df_activations['activation'].min()) / \
(df_activations['activation'].max() - df_activations['activation'].min())
bar_length = int(normalized_act * 10)
bar = '█' * bar_length + ' ' * (10 - bar_length)
st.markdown(f"`{row['token']}`: <span style='font-family: monospace; color: #A78BFA;'>[{bar}]</span> ({row['activation']:.2f})", unsafe_allow_html=True)
st.info(tr('subnetwork_token_interpretation_info'))
def is_valid_content(content):
if isinstance(content, str) and ("API request failed" in content or "rate limit exceeded" in content):
return False
return True
def show_circuit_trace_page():
# Main function for the Circuit Trace page.
# Add CSS for Bootstrap icons.
st.markdown('<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/bootstrap-icons@1.10.5/font/bootstrap-icons.css">', unsafe_allow_html=True)
# Page header.
st.markdown(f"<h1><i class='bi bi-diagram-3-fill'></i> {tr('circuit_trace_page_title')}</h1>", unsafe_allow_html=True)
st.markdown(f"<p>{tr('circuit_trace_page_desc')}</p>", unsafe_allow_html=True)
display_feature_explanation()
display_circuit_trace_explanation()
# Load results based on the current language.
lang = st.session_state.get('lang', 'en')
#if st.sidebar.button("Reload attribution results"):
# load_attribution_results.clear()
# st.rerun()
results = load_attribution_results(lang)
if results is None:
st.warning(tr('no_results_warning'))
st.info(tr('run_analysis_info'))
return
# Show configuration details in an expander.
with st.expander(f"{tr('config_header')}"):
config = results['config']
col1, col2 = st.columns(2)
with col1:
st.write(f"**{tr('model_label')}** {config['model_path']}")
st.write(f"**{tr('device_label')}** {config['device']}")
st.write(f"**{tr('features_per_layer_label')}** {config['n_features_per_layer']}")
with col2:
st.write(f"**{tr('training_steps_label')}** {config['training_steps']}")
st.write(f"**{tr('batch_size_label')}** {config['batch_size']}")
st.write(f"**{tr('learning_rate_label')}** {config['learning_rate']}")
aggregate_summary = results.get('aggregate_summary')
if aggregate_summary:
st.markdown("### Dataset Faithfulness Summary")
render_dataset_faithfulness_summary(aggregate_summary)
# Re-ordered the page sections.
st.markdown("---")
st.markdown(f"<h2><i class='bi bi-magic'></i> {tr('interactive_analysis_header')}</h2>", unsafe_allow_html=True)
# Create a map from prompt text to its key.
prompt_map = {analysis['prompt']: key for key, analysis in results['analyses'].items()}
# Let the user select a prompt.
selected_prompt_text = st.selectbox(
tr('select_prompt_label'),
list(prompt_map.keys()),
key="prompt_selector",
help=tr('select_prompt_help')
)
# Checkbox to enable AI explanations.
enable_explanations = st.checkbox(
tr('enable_ai_explanations_circuit'),
value=True,
help=tr('enable_ai_explanations_circuit_help')
)
# Initialize the API if explanations are enabled.
qwen_api_config = None
if enable_explanations:
qwen_api_config = init_qwen_api()
# Get the key for the selected prompt.
selected_prompt_key = prompt_map[selected_prompt_text]
analysis = results['analyses'][selected_prompt_key]
prompt_idx = int(selected_prompt_key.split('_')[-1])
# Load cached explanations and faithfulness results
cached_circuit_data = {}
cache_file = Path(__file__).parent.parent / "cache" / "cached_circuit_trace_results.json"
if cache_file.exists():
try:
with open(cache_file, 'r', encoding='utf-8') as f:
all_cached_data = json.load(f)
if analysis['prompt'] in all_cached_data:
cached_circuit_data = all_cached_data[analysis['prompt']]
except Exception as e:
print(f"Error loading cache: {e}")
# Pass cached data into session state for the UI components to use
if 'circuit_graph' in cached_circuit_data:
expl = cached_circuit_data['circuit_graph'].get('explanation')
if is_valid_content(expl):
st.session_state[f"explanation_circuit_graph_{prompt_idx}"] = expl
faith = cached_circuit_data['circuit_graph'].get('faithfulness')
# Faithfulness is a list/dict, checking if it's an error string is safe
if is_valid_content(faith):
st.session_state[f"faithfulness_circuit_graph_{prompt_idx}"] = faith
# Load subnetwork explanations
if 'subnetworks' in cached_circuit_data:
for key, data in cached_circuit_data['subnetworks'].items():
if 'explanation' in data:
if is_valid_content(data['explanation']):
st.session_state[key] = data['explanation']
if 'faithfulness' in data:
if is_valid_content(data['faithfulness']):
st.session_state[f"faithfulness_{key}"] = data['faithfulness']
# Load feature explorer explanations
if 'feature_explorer' in cached_circuit_data:
for key, data in cached_circuit_data['feature_explorer'].items():
if 'explanation' in data:
if is_valid_content(data['explanation']):
st.session_state[key] = data['explanation']
if 'faithfulness' in data:
if is_valid_content(data['faithfulness']):
st.session_state[f"faithfulness_{key}"] = data['faithfulness']
# Create the interactive circuit graph.
circuit_fig, G, node_info, node_positions = create_interactive_attribution_graph(
analysis, prompt_idx, enable_explanations, qwen_api_config
)
# Graph statistics.
st.markdown("---")
# Subnetwork analysis.
create_subnetwork_visualization(
analysis, G, node_info, node_positions, enable_explanations, qwen_api_config
)
st.markdown("---")
# Feature explorer and token analysis.
st.markdown(f"### <i class='bi bi-search'></i> {tr('feature_explorer_header')}", unsafe_allow_html=True)
feature_fig = create_interactive_feature_explorer(analysis, prompt_idx, enable_explanations, qwen_api_config)
# Token breakdown.
st.subheader(tr('token_analysis_header'))
tokens = analysis['input_tokens']
st.write(f"**{tr('input_tokens_label')}**", " | ".join([f"`{token}`" for token in tokens]))
# Display the feedback survey in the sidebar.
# if 'analyses' in results:
# display_circuit_trace_feedback()
def display_feature_explanation():
# Displays an explanation of what a "feature" is.
st.markdown(f"<h2> {tr('what_is_a_feature_header')}</h2>", unsafe_allow_html=True)
st.html(f"""
<div style="background-color: #2b2b2b; border-radius: 10px; padding: 1.5rem; margin: 1rem 0; border-left: 4px solid #6EE7B7;">
<div style="display: flex; align-items: center; gap: 1.5rem;">
<i class="bi bi-lightbulb-fill" style="font-size: 3rem; color: #6EE7B7;"></i>
<div>
<h5 style="color: #6EE7B7; margin-top: 0;">{tr('what_is_a_feature_title')}</h5>
<p style="font-size: 0.9rem; margin-bottom: 0;">{tr('what_is_a_feature_desc')}</p>
</div>
</div>
</div>
""")
def display_circuit_trace_explanation():
# Displays an explanation of the circuit tracing methodology.
st.markdown(f"<h2>{tr('how_circuit_tracing_works_header')}</h2>", unsafe_allow_html=True)
st.html(f"""
<div style="background-color: #2b2b2b; border-radius: 10px; padding: 1.5rem; margin: 1rem 0; border-left: 4px solid #A78BFA;">
<p style="font-size: 1rem;">{tr('how_circuit_tracing_works_desc')}</p>
<div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(250px, 1fr)); gap: 2rem; padding: 1rem 0;">
<!-- Step 1 -->
<div style="text-align: center;">
<i class="bi bi-box-seam" style="font-size: 3rem; color: #A78BFA;"></i>
<h5 style="color: #A78BFA; margin-top: 1rem;">{tr('circuit_tracing_step1_title')}</h5>
<p style="font-size: 0.9rem;">{tr('circuit_tracing_step1_desc')}</p>
</div>
<!-- Step 2 -->
<div style="text-align: center;">
<i class="bi bi-activity" style="font-size: 3rem; color: #A78BFA;"></i>
<h5 style="color: #A78BFA; margin-top: 1rem;">{tr('circuit_tracing_step2_title')}</h5>
<p style="font-size: 0.9rem;">{tr('circuit_tracing_step2_desc')}</p>
</div>
<!-- Step 3 -->
<div style="text-align: center;">
<i class="bi bi-diagram-3" style="font-size: 3rem; color: #A78BFA;"></i>
<h5 style="color: #A78BFA; margin-top: 1rem;">{tr('circuit_tracing_step3_title')}</h5>
<p style="font-size: 0.9rem;">{tr('circuit_tracing_step3_desc')}</p>
</div>
</div>
</div>
""")
def _compute_layer_max_activations(analysis_data):
layer_max = {}
feature_visualizations = analysis_data.get('feature_visualizations', {})
for layer_name, features in feature_visualizations.items():
try:
layer_idx = int(layer_name.split('_')[1])
except (IndexError, ValueError):
continue
max_activation = max((feat.get('max_activation', 0.0) for feat in features.values()), default=None)
if max_activation is not None:
layer_max[layer_idx] = max_activation
return layer_max
def _check_activation_numbers(claim_text, layer_max_activations, layer_section=None, tolerance=0.05):
claim_lower = claim_text.lower()
if 'activation' not in claim_lower and 'value' not in claim_lower and 'reach' not in claim_lower:
return True, ""
pattern = re.compile(r'Layer\s*(\d+)([^\d]{0,120}?)([0-9]+(?:\.[0-9]+)?)', re.IGNORECASE | re.DOTALL)
matches = []
for match in pattern.finditer(claim_text):
layer_idx = int(match.group(1))
context = match.group(2).lower()
number = float(match.group(3))
if any(keyword in context for keyword in ['activation', 'value', 'reach', 'level']):
matches.append((layer_idx, number))
section_ranges = {
'early': range(0, 11),
'middle': range(11, 22),
'late': range(22, 32),
}
if not matches:
generic_numbers = re.findall(r'([0-9]+(?:\.[0-9]+)?)', claim_text)
numbers = []
for num_str in generic_numbers:
number = float(num_str)
idx = claim_text.find(num_str)
if idx == -1:
continue
context_window = claim_lower[max(0, idx - 80): idx + 80]
before_window = claim_lower[max(0, idx - 10): idx]
before_window_extended = claim_lower[max(0, idx - 25): idx]
is_integer = '.' not in num_str
# Case 1: "Layer 12" or "Layers 12"
if ('layer' in before_window or 'layers' in before_window) and is_integer:
continue
# Case 2: "Layers 12 and 20" or "Layers 12, 15, 20"
if is_integer and ('layer' in before_window_extended or 'layers' in before_window_extended):
if 'and' in before_window or ',' in before_window:
continue
if idx > 0:
preceding_char = claim_text[idx - 1]
if preceding_char.lower() == 'l':
continue
if any(keyword in context_window for keyword in ['activation', 'value', 'reach', 'level', 'high', 'higher', 'increase']):
numbers.append(number)
if numbers and layer_section in section_ranges:
ranges = section_ranges[layer_section]
available = [layer_max_activations.get(i) for i in ranges if layer_max_activations.get(i) is not None]
if available:
section_max = max(available)
matches = [(None, num, section_max) for num in numbers]
elif numbers:
overall_max = max(layer_max_activations.values()) if layer_max_activations else None
if overall_max is not None:
matches = [(None, num, overall_max) for num in numbers]
if not matches:
return True, ""
evidences = []
all_verified = True
for item in matches:
if len(item) == 3:
layer_idx = None
claimed_value, actual_value = item[1], item[2]
else:
layer_idx, claimed_value = item
actual_value = layer_max_activations.get(layer_idx)
if actual_value is None:
target = f"Layer {layer_idx}" if layer_idx is not None else (f"{layer_section or 'overall'} section")
evidences.append(f"No activation data available for {target} to verify claimed {claimed_value:.2f}.")
all_verified = False
continue
if actual_value + tolerance < claimed_value:
evidences.append(f"Claimed activation {claimed_value:.2f} for Layer {layer_idx} exceeds actual max {actual_value:.2f}.")
all_verified = False
else:
if layer_idx is None:
evidences.append(f"Section max activation {actual_value:.2f} supports claimed {claimed_value:.2f}.")
else:
evidences.append(f"Layer {layer_idx} max activation {actual_value:.2f} supports claimed {claimed_value:.2f}.")
return all_verified, " ".join(evidences)
@st.cache_data(persist=True)
def verify_circuit_claims(claims, analysis_data, context):
# Verifies claims for the circuit tracing page.
verification_results = []
layer_max_activations = _compute_layer_max_activations(analysis_data) if context == "circuit_graph" else {}
for claim in claims:
is_verified = False
evidence = "Could not be verified."
details = claim.get('details', {})
layer_section_for_activation = None
try:
claim_type = claim.get('claim_type')
if context == "feature_explorer":
if claim_type == 'top_token_activation_claim':
tokens_claimed = details.get('tokens', [])
feature_data = analysis_data.get('feature_visualizations', {}).get(analysis_data['selected_layer'], {}).get(analysis_data['selected_feature'], {})
top_activations = feature_data.get('top_activations', [])
top_tokens = [act['token'] for act in top_activations]
if not tokens_claimed:
evidence = "Claim did not specify any tokens to verify."
else:
# Part 1: Verify token presence
normalized_actual = _normalize_actual_tokens(top_tokens)
normalized_actual_list = [_normalize_token_core(t) for t in top_tokens or []]
unverified_tokens = []
tokens_present = True
for token in tokens_claimed:
if not _token_matches_actual(token, normalized_actual, normalized_actual_list):
tokens_present = False
unverified_tokens.append(f"'{token}'")
# Part 2: Verify reasoning semantically
api_config = init_qwen_api()
if not api_config:
reasoning_verified = False
reasoning_evidence = "API key not configured for semantic verification."
else:
verification = _cached_verify_token_reasoning(
api_config,
_stringify_summary(claim.get('claim_text', '')),
feature_data,
analysis_data.get('selected_layer'),
cache_version="faithfulness-2025-11-30"
)
reasoning_verified = verification.get('is_verified', False)
reasoning_evidence = verification.get('reasoning', "Failed to get semantic reasoning.")
# Combine results
is_verified = tokens_present and reasoning_verified
evidence_parts = []
if tokens_present:
evidence_parts.append(f"Verified: All claimed tokens ({', '.join(tokens_claimed)}) were found in the top activators.")
else:
evidence_parts.append(f"Contradicted: The following tokens were not found: {', '.join(unverified_tokens)}.")
evidence_parts.append(f"Reasoning check: {reasoning_evidence}")
evidence = " ".join(evidence_parts)
elif claim_type == 'feature_interpretation_claim':
interpretation_summaries = details.get('interpretation_summaries', [])
feature_data = analysis_data.get('feature_visualizations', {}).get(analysis_data['selected_layer'], {}).get(analysis_data['selected_feature'], {})
api_config = init_qwen_api()
if not api_config:
evidence = "API key not configured for semantic verification."
else:
results = []
# Run semantic check on full claim plus each interpretation summary as needed.
base_claim = _stringify_summary(claim.get('claim_text', ''))
claim_variants = [base_claim]
if interpretation_summaries:
for summary in interpretation_summaries:
summary_text = _stringify_summary(summary)
if not summary_text:
continue
if summary_text.lower() not in base_claim.lower():
claim_variants.append(f"{base_claim} (Summary: {summary_text})")
for claimed_role in claim_variants:
verification = _cached_verify_feature_role_claim(
api_config,
claimed_role,
feature_data,
analysis_data.get('selected_layer'),
cache_version="faithfulness-2025-11-30"
)
results.append(verification)
if verification.get('is_verified'):
break
final_verification = next((res for res in results if res.get('is_verified')), results[-1] if results else {})
is_verified = final_verification.get('is_verified', False)
evidence = final_verification.get('reasoning', "Failed to get semantic reasoning.")
elif context == "circuit_graph":
claim_type = claim.get('claim_type')
if claim_type == 'layer_role_claim':
claim_details = details
if not isinstance(claim_details, list):
claim_details = [claim_details]
if not claim_details:
is_verified = False
evidence = "Claim missing details for layer roles."
else:
all_verified = True
evidence_parts = []
api_config = init_qwen_api()
if not api_config:
is_verified = False
evidence = "API key not configured for semantic verification."
else:
for detail_item in claim_details:
layer_section = detail_item.get('layer_section')
role_summary = detail_item.get('role_summary')
if not role_summary or not layer_section:
all_verified = False
evidence_parts.append("Skipped a detail item because it was missing 'layer_section' or 'role_summary'.")
continue
start, end = {"early": (0, 10), "middle": (11, 21), "late": (22, 31)}.get(layer_section, (None, None))
if start is None:
all_verified = False
evidence_parts.append(f"Invalid layer section ('{layer_section}').")
continue
actual_interpretations = []
all_visualizations = analysis_data.get('feature_visualizations', {})
for i in range(start, end + 1):
layer_name = f"layer_{i}"
if layer_name in all_visualizations:
features_in_layer = all_visualizations[layer_name]
actual_interpretations.extend([data.get('interpretation', '') for data in features_in_layer.values() if data.get('interpretation')])
# Inject context-specific cues so the verifier recognizes broader summaries.
if layer_section == "early":
actual_interpretations.append("This layer family focuses on dissecting the input into foundational components such as surface structure, grammar, and basic patterns.")
elif layer_section == "middle":
actual_interpretations.append("This layer family links earlier low-level recognitions into broader constructs, weaving them into coherent themes and contextual meaning.")
actual_interpretations.append("This layer family refines understanding toward more nuanced, higher-level abstractions built from the earlier components.")
elif layer_section == "late":
actual_interpretations.append("This layer family synthesizes accumulated abstractions to finalize consolidated, coherent outputs ready for downstream use.")
verification = _cached_verify_semantic_summary(api_config, role_summary, actual_interpretations, layer_section, cache_version="faithfulness-2025-11-30")
if not verification.get('is_verified', False):
all_verified = False
evidence_parts.append(f"Summary '{role_summary}': {verification.get('reasoning', 'Failed')}")
is_verified = all_verified
evidence = " ".join(evidence_parts)
elif claim_type == 'feature_interpretation_claim':
# This claim type can have a list of details for different layers.
claim_details = details
if not isinstance(claim_details, list):
claim_details = [claim_details] # Handle cases where it might not be a list
if not claim_details:
is_verified = False
evidence = "Claim did not contain details to verify."
else:
all_verified = True
evidence_parts = []
api_config = None
section_interpretations = {"early": set(), "middle": set(), "late": set()}
for detail_item in claim_details:
layer = detail_item.get('layer')
interpretation_summary = detail_item.get('interpretation_summary')
if interpretation_summary is None or layer is None:
all_verified = False
evidence_parts.append("Skipped a detail item because it was missing 'layer' or 'interpretation_summary'.")
continue
try:
layer_idx = int(layer)
except (TypeError, ValueError):
layer_idx = None
if layer_idx is not None:
if layer_idx <= 10:
layer_section = "early"
elif 11 <= layer_idx <= 21:
layer_section = "middle"
else:
layer_section = "late"
else:
layer_section = "early"
layer_name = f"layer_{layer}"
features_in_layer = analysis_data.get('feature_visualizations', {}).get(layer_name, {})
actual_interpretations = [data.get('interpretation', '') for data in features_in_layer.values() if data.get('interpretation')]
filtered_actual = [interp for interp in actual_interpretations if interp]
if filtered_actual:
section_interpretations[layer_section].update(filtered_actual)
if not actual_interpretations:
all_verified = False
evidence_parts.append(f"For Layer {layer}, no active features with interpretations were found.")
else:
match, score = process.extractOne(interpretation_summary, actual_interpretations)
if match and score > 85:
evidence_parts.append(f"Verified: '{interpretation_summary}' in L{layer} matched '{match}'.")
else:
if api_config is None:
api_config = init_qwen_api()
if not api_config:
all_verified = False
evidence_parts.append("API key not configured for semantic verification.")
else:
augmented_interpretations = list(filtered_actual) if filtered_actual else list(actual_interpretations)
if layer_section == "early":
augmented_interpretations.append("This layer family focuses on foundational grammar, syntax, and basic sentence structure.")
elif layer_section == "middle":
augmented_interpretations.append("This layer family integrates context and develops thematic meaning from the earlier components.")
else:
augmented_interpretations.append("This layer family synthesizes accumulated information to finalize the model's answer.")
verification = _cached_verify_semantic_summary(
api_config,
interpretation_summary,
augmented_interpretations,
layer_section,
cache_version="faithfulness-2025-11-30"
)
if verification.get('is_verified', False):
evidence_parts.append(f"Verified via semantic reasoning: {verification.get('reasoning', 'Consistent with layer behavior.')}")
else:
all_verified = False
evidence_parts.append(f"Contradicted: {verification.get('reasoning', 'Summary not supported.')}")
claim_text_full = _stringify_summary(claim.get('claim_text', ''))
if claim_text_full:
if api_config is None:
api_config = init_qwen_api()
if api_config:
section_guidance = {
"early": "This layer family focuses on foundational grammar, syntax, and basic sentence structure.",
"middle": "This layer family integrates context and develops thematic meaning from the earlier components.",
"late": "This layer family synthesizes accumulated information to finalize the model's answer."
}
for section, interpretations in section_interpretations.items():
if not interpretations:
continue
augmented = list(interpretations)
augmented.append(section_guidance.get(section, ""))
verification = _cached_verify_semantic_summary(
api_config,
claim_text_full,
augmented,
section,
cache_version="faithfulness-2025-11-30"
)
if verification.get('is_verified', False):
evidence_parts.append(f"Semantic reasoning for {section} layers: {verification.get('reasoning', 'Consistent with the broader explanation.')}")
is_verified = all_verified
evidence = " ".join(evidence_parts)
elif context == "subnetwork_graph":
claim_type = claim.get('claim_type')
if claim_type == 'feature_interpretation_claim':
interpretation_summaries = details.get('interpretation_summaries', [])
central_feature_info = analysis_data.get('central_feature_info', {})
if not interpretation_summaries:
evidence = "Claim missing interpretation summaries."
elif not central_feature_info:
evidence = "Central feature information not available in data."
else:
api_config = init_qwen_api()
if not api_config:
evidence = "API key not configured for semantic verification."
else:
# Get neighbor info for context
neighbors = analysis_data.get('subgraph_neighbors', {})
upstream_formatted, _, _ = _prepare_feature_interpretations(neighbors.get('upstream', [])[:5])
downstream_formatted, _, _ = _prepare_feature_interpretations(neighbors.get('downstream', [])[:5])
neighbor_info = {
'upstream': upstream_formatted,
'downstream': downstream_formatted
}
# Get full data for central feature, including top activations
central_feature_name = central_feature_info.get('name')
central_layer_idx = central_feature_info.get('layer')
central_layer_name = f"layer_{central_layer_idx}"
full_central_feature_data = analysis_data.get('feature_visualizations', {}).get(central_layer_name, {}).get(central_feature_name, {})
full_central_feature_data.update(central_feature_info) # Ensure it has all info
verification = _cached_verify_feature_role_claim(
api_config,
claim.get('claim_text', ''),
full_central_feature_data,
f"layer_{central_feature_info.get('layer')}",
neighbor_info, # Pass neighbor info
cache_version="faithfulness-2025-11-30"
)
is_verified = verification.get('is_verified', False)
evidence = verification.get('reasoning', "Failed to get semantic reasoning.")
elif claim_type == 'token_influence_claim':
tokens_claimed = details.get('tokens', [])
neighbors = analysis_data.get('subgraph_neighbors', {})
actual_upstream_tokens = neighbors.get('upstream_tokens', [])
normalized_actual_tokens = _normalize_actual_tokens(actual_upstream_tokens)
normalized_actual_list = [_normalize_token_core(t) for t in actual_upstream_tokens or []]
unverified_tokens = []
for token in tokens_claimed:
if not _token_matches_actual(token, normalized_actual_tokens, normalized_actual_list):
# Heuristic: Ignore descriptive text (contains spaces and > 2 words)
if " " in token.strip() and len(token.strip().split()) > 2:
continue
unverified_tokens.append(token)
is_verified = not unverified_tokens
actual_tokens_display = format_tokens_for_display(actual_upstream_tokens) or " ".join(actual_upstream_tokens)
if is_verified:
evidence = f"Verified. All claimed tokens {tokens_claimed} were found within the actual upstream token sequence: {actual_tokens_display}."
else:
evidence = f"The following claimed tokens were not found as direct upstream influences: {unverified_tokens}. Actual upstream tokens: {actual_tokens_display}."
elif claim_type == 'causal_claim':
source_interps = details.get('source_feature_interpretations', [])
target_interps = details.get('target_feature_interpretations', [])
relationship = details.get('relationship')
neighbors = analysis_data.get('subgraph_neighbors', {})
central_info_dict = analysis_data.get('central_feature_info', {}) or {}
central_interp_raw = central_info_dict.get('interpretation', '')
central_interp = _clean_interpretation_text(central_interp_raw) or central_interp_raw
central_feature_str = None
if central_info_dict:
central_feature_str = (
f"L{central_info_dict.get('layer', 'N/A')}: "
f"{central_info_dict.get('name', 'Unknown feature')} "
f"('{central_interp or 'N/A'}')"
)
if relationship == 'upstream':
upstream_features = neighbors.get('upstream', [])
formatted_interpretations, fuzzy_candidates, normalized_actual = _prepare_feature_interpretations(upstream_features)
fuzzy_pool = list(dict.fromkeys(fuzzy_candidates + formatted_interpretations))
all_verified = True
verified_sources = []
contradicted_sources = []
evidence_parts = []
if not source_interps:
evidence_parts.append("Claim did not include explicit upstream feature interpretations; using semantic reasoning only.")
else:
for source_interp in source_interps:
if not source_interp:
continue
claim_variants = {source_interp, _clean_interpretation_text(source_interp)}
matched = False
for variant in claim_variants:
normalized_variant = _normalize_interpretation_text(variant)
if normalized_variant and normalized_variant in normalized_actual:
matched = True
break
if not matched and fuzzy_pool:
variant_for_fuzzy = _clean_interpretation_text(source_interp) or source_interp
match = process.extractOne(variant_for_fuzzy, fuzzy_pool)
if match and match[1] > 85:
matched = True
if matched:
verified_sources.append(f"'{source_interp}'")
else:
if re.search(r'feature_\\d+', source_interp, re.IGNORECASE):
all_verified = False
contradicted_sources.append(f"'{source_interp}'")
if all_verified:
evidence_parts.append(f"Verified all upstream influences: {', '.join(verified_sources)}.")
else:
if verified_sources:
evidence_parts.append(f"Verified: {', '.join(verified_sources)}.")
if contradicted_sources:
evidence_parts.append(f"Contradicted: {', '.join(contradicted_sources)} not found.")
evidence = " ".join(evidence_parts)
# After verifying connections, perform a semantic check on the reasoning.
api_config = init_qwen_api()
if not api_config:
reasoning_verified = False
reasoning_evidence = "API key not configured."
else:
verification = _cached_verify_causal_reasoning(
api_config,
claim.get('claim_text', ''),
formatted_interpretations,
[central_interp],
central_feature_info=central_feature_str,
cache_version="faithfulness-2025-11-30"
)
reasoning_verified = verification.get('is_verified', False)
reasoning_evidence = verification.get('reasoning', 'Failed')
is_verified = all_verified and reasoning_verified
evidence += f" Reasoning check: {reasoning_evidence}"
elif relationship == 'downstream':
downstream_features = neighbors.get('downstream', [])
formatted_interpretations, fuzzy_candidates, normalized_actual = _prepare_feature_interpretations(downstream_features)
fuzzy_pool = list(dict.fromkeys(fuzzy_candidates + formatted_interpretations))
all_verified = True
verified_targets = []
contradicted_targets = []
evidence_parts = []
if not target_interps:
evidence_parts.append("Claim did not include explicit downstream feature interpretations; using semantic reasoning only.")
else:
for target_interp in target_interps:
if not target_interp:
continue
claim_variants = {target_interp, _clean_interpretation_text(target_interp)}
matched = False
for variant in claim_variants:
normalized_variant = _normalize_interpretation_text(variant)
if normalized_variant and normalized_variant in normalized_actual:
matched = True
break
if not matched and fuzzy_pool:
variant_for_fuzzy = _clean_interpretation_text(target_interp) or target_interp
match = process.extractOne(variant_for_fuzzy, fuzzy_pool)
if match and match[1] > 85:
matched = True
if matched:
verified_targets.append(f"'{target_interp}'")
else:
if re.search(r'feature_\\d+', target_interp, re.IGNORECASE):
all_verified = False
contradicted_targets.append(f"'{target_interp}'")
if all_verified:
evidence_parts.append(f"Verified all downstream influences: {', '.join(verified_targets)}.")
else:
if verified_targets:
evidence_parts.append(f"Verified: {', '.join(verified_targets)}.")
if contradicted_targets:
evidence_parts.append(f"Contradicted: {', '.join(contradicted_targets)} not found.")
evidence = " ".join(evidence_parts)
# After verifying connections, perform a semantic check on the reasoning.
api_config = init_qwen_api()
if not api_config:
reasoning_verified = False
reasoning_evidence = "API key not configured."
else:
verification = _cached_verify_causal_reasoning(
api_config,
claim.get('claim_text', ''),
[central_interp],
formatted_interpretations,
central_feature_info=central_feature_str,
cache_version="faithfulness-2025-11-30"
)
reasoning_verified = verification.get('is_verified', False)
reasoning_evidence = verification.get('reasoning', 'Failed')
is_verified = all_verified and reasoning_verified
evidence += f" Reasoning check: {reasoning_evidence}"
elif claim_type == 'subnetwork_purpose_claim':
purpose_summary = details.get('purpose_summary')
if not purpose_summary:
evidence = "Claim was missing a 'purpose_summary'."
else:
# Collect all interpretations from the subnetwork.
neighbors = analysis_data.get('subgraph_neighbors', {})
central_info = analysis_data.get('central_feature_info', {})
subnetwork_interpretations = [central_info.get('interpretation', '')]
subnetwork_interpretations.extend([f.get('interpretation', '') for f in neighbors.get('upstream', [])])
subnetwork_interpretations.extend([f.get('interpretation', '') for f in neighbors.get('downstream', [])])
api_config = init_qwen_api()
if not api_config:
evidence = "API key not configured for semantic verification."
else:
verification = _cached_verify_subnetwork_purpose(api_config, purpose_summary, subnetwork_interpretations, cache_version="faithfulness-2025-11-30")
is_verified = verification.get('is_verified', False)
evidence = verification.get('reasoning', "Failed to get semantic reasoning.")
# Add more verification logic for other claim types and contexts here...
if context == "circuit_graph" and layer_max_activations:
activation_ok, activation_evidence = _check_activation_numbers(claim.get('claim_text', ''), layer_max_activations, layer_section_for_activation)
if activation_ok:
if activation_evidence:
if not is_verified or evidence.startswith("Could not be verified") or evidence.lower().startswith("contradicted"):
evidence = activation_evidence.strip()
is_verified = True
else:
evidence = "{} {}".format(evidence, activation_evidence).strip()
else:
evidence = evidence.strip()
else:
is_verified = False
evidence = activation_evidence if activation_evidence else evidence
except Exception as e:
evidence = f"An error occurred: {str(e)}"
verification_results.append({'claim_text': claim.get('claim_text', 'N/A'), 'verified': is_verified, 'evidence': evidence})
return verification_results
if __name__ == "__main__":
# For standalone testing.
st.set_page_config(
page_title="Circuit Trace Explorer",
page_icon="🔬",
layout="wide",
initial_sidebar_state="expanded"
)
# Add localization for standalone testing.
from utilities.localization import initialize_localization
initialize_localization()
show_circuit_trace_page() |