Spaces:
Sleeping
Sleeping
File size: 122,768 Bytes
5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d 9bfc4be 5f2263d | 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 | """
Data loader for swap experiment results (domain-agnostic).
Reads from output/<domain>_batch/ directory structure:
- _swaps/runs/{run_id}/by_source/*/to_*.json - Swap details (run-specific)
- _swaps/runs/{run_id}/config_resolved.json - Domain & model metadata
- _swaps/_analysis_v3/*.json - Analysis data
- */manifest.json - Neuronpedia URLs
Supports multiple experiment runs and auto-detects domain fields
(e.g. state/capital, character/author) from config_resolved.json.
Multi-dataset discovery:
DemoRegistry scans output/* for run_manifest.json files that carry
"display_demo": true and groups them by dataset (graphs_root). The
demo can then switch both dataset and run without restarting the server.
"""
from pathlib import Path
from typing import Any, Dict, List, Optional, Set
import os
import json
import csv
import re
from functools import lru_cache
from urllib.parse import parse_qsl, urlencode, urlsplit, urlunsplit
# ---------------------------------------------------------------------------
# Control variant helpers
# ---------------------------------------------------------------------------
_VARIANT_SUFFIX_RE = re.compile(r"__(?:r\d+|add_.+|ctrl_.+|m_tuned)$")
# ---------------------------------------------------------------------------
# Logit-shift regime classification (Section 4.4 of METHODOLOGY_REPORT.md)
# ---------------------------------------------------------------------------
_REGIME_FLAT_THRESHOLD = 1.0 # |delta| below this is considered FLAT
_REGIME_INFO = {
"A": {"label": "Clean Redirection", "short": "A"},
"B": {"label": "Both Boosted", "short": "B"},
"C": {"label": "Differential Disruption", "short": "C"},
"D": {"label": "Generic Disruption", "short": "D"},
"E": {"label": "Pure Suppression", "short": "E"},
}
def _classify_regime(pos0: Dict[str, Any]) -> Optional[str]:
"""Classify a swap into a logit-shift regime from position-0 comparison.
Returns one of A/B/C/D/E or None if classification is impossible.
"""
if not pos0:
return None
tgt = pos0.get("target_logit_delta")
src = pos0.get("source_logit_delta")
flip = pos0.get("flip_at_0", False)
if tgt is None or src is None:
return None
tgt_up = tgt > _REGIME_FLAT_THRESHOLD
tgt_down = tgt < -_REGIME_FLAT_THRESHOLD
src_down = src < -_REGIME_FLAT_THRESHOLD
src_up = src > _REGIME_FLAT_THRESHOLD
if tgt_up and src_down and flip:
return "A"
if tgt_up and src_up:
return "B"
if tgt_down and src_down and flip:
return "C"
if tgt_down and src_down and not flip:
return "D"
if not tgt_up and not tgt_down and src_down and flip:
return "E"
return None
def _parse_to_slug(filename_stem: str) -> tuple:
"""
Extract ``(entity_slug, variant_suffix)`` from a swap result filename stem.
Normal files: ``to_georgia_savannah`` -> ``("georgia_savannah", "")``
Variant files: ``to_georgia_savannah__r0`` -> ``("georgia_savannah", "r0")``
``to_georgia_savannah__add_source`` -> ``("georgia_savannah", "add_source")``
Double suffix: ``to_buzz__add_sound__m_tuned`` -> ``("buzz", "add_sound__m_tuned")``
"""
raw = filename_stem.replace("to_", "", 1)
# Handle double suffix: strip __m_tuned, parse the inner variant, re-append
m_tuned_tail = ""
if raw.endswith("__m_tuned") and "__add_" in raw:
m_tuned_tail = "__m_tuned"
raw = raw[: -len("__m_tuned")]
m = _VARIANT_SUFFIX_RE.search(raw)
if m:
return raw[: m.start()], raw[m.start() + 2 :] + m_tuned_tail
return raw, ""
def _is_control_variant(filename_stem: str) -> bool:
"""Return True if the filename contains a control variant suffix."""
return bool(_VARIANT_SUFFIX_RE.search(filename_stem.replace("to_", "", 1)))
# ---------------------------------------------------------------------------
# Run metadata helpers
# ---------------------------------------------------------------------------
_CONTROL_MODE_LABELS = {
"labeled": "Labeled",
"random_feature_matched": "Random x3",
"additivity": "Field Additivity",
"m_tuned": "Adaptive M",
}
# Run-id / experiment_name substrings that hide a run from the demo
# selector. These are calibration sweeps or "labeled" control runs that
# would otherwise dominate the dropdown without adding signal for the
# end user. The cross-run-best aggregator applies the exact same filter
# so "Best across runs" is restricted to runs the user can inspect.
_RUN_BLACKLIST_SUBSTRINGS = (
"sweep_", # m-sweep calibration runs
"entropy_study", # entropy calibration runs
"highm_", # high-M ablations
"_labeled", # *_labeled control replicas (kept on disk for analysis)
)
def _is_dropdown_run(run_id: str, manifest: Dict[str, Any]) -> bool:
"""Return True if a run should appear in the demo run-selector.
The aim is to keep only the three "headline" run families:
- Random x3 (control_mode == random_feature_matched)
- Field Additivity (control_mode == additivity)
- Classic / canonical swap runs (everything else not blacklisted)
"""
rid = run_id.lower()
exp_name = (
manifest.get("config", {}).get("experiment_name", "") or ""
).lower()
for blocked in _RUN_BLACKLIST_SUBSTRINGS:
if blocked in rid or blocked in exp_name:
return False
return True
def _extract_control_mode(manifest: Dict[str, Any], run_dir: Optional[Path] = None) -> str:
"""Return the control mode from run metadata.
Checks the manifest config first, then falls back to
``config_resolved.json`` on disk (which contains the full YAML config).
"""
mode = manifest.get("config", {}).get("control", {}).get("mode", "")
if mode:
return mode
if run_dir is not None:
resolved_path = run_dir / "config_resolved.json"
if resolved_path.exists():
try:
with open(resolved_path, "r", encoding="utf-8") as fh:
resolved = json.load(fh)
mode = resolved.get("control", {}).get("mode", "")
if mode:
return mode
except (json.JSONDecodeError, IOError):
pass
return "labeled"
def _semantic_run_label(
run_id: str, manifest: Dict[str, Any], run_dir: Optional[Path] = None,
) -> str:
"""Build a concise, human-readable label that distinguishes control modes."""
mode = _extract_control_mode(manifest, run_dir=run_dir)
mode_label = _CONTROL_MODE_LABELS.get(mode, mode.replace("_", " ").title())
exp_name = manifest.get("config", {}).get("experiment_name", "")
if exp_name.startswith("fullscale_"):
return mode_label
if exp_name:
pretty = exp_name.replace("_", " ").title()
pretty = pretty.replace("50states", "50 States").replace("6states", "6 States")
return pretty
pretty = run_id.replace("_", " ").title()
pretty = pretty.replace("50states", "50 States").replace("6states", "6 States")
pretty = pretty.replace("Gpu", "GPU")
return pretty
def _is_subpath(child: Path, parent: Path) -> bool:
"""Return True if *child* is equal to or nested under *parent*."""
try:
child.relative_to(parent)
return True
except ValueError:
return False
# ---------------------------------------------------------------------------
# DemoRegistry -- multi-dataset discovery
# ---------------------------------------------------------------------------
class DemoRegistry:
"""Scan an output root for demo-enabled runs and manage the active dataset.
Discovery rules:
- Walk output_root/*/_swaps/runs/*/run_manifest.json
- Keep only manifests where display_demo == True
- Derive dataset_dir from manifest["config"]["inputs"]["graphs_root"]
falling back to the grandparent of the run dir (_swaps/runs/<id> -> dataset)
- Group runs by resolved dataset_dir (normalised to an absolute Path)
The active dataset+run is exposed through ``active_loader``.
"""
DEFAULT_DATASET_ID = "usa_states_batch"
# Default run for the homepage: the field-additivity sweep on USA
# states (paper's main result, drives the 73% T5 / 83% T3+ headline
# numbers). Falls through to ``_pick_default_run``'s fallback chain
# for other datasets that don't ship this exact run id.
DEFAULT_RUN_ID = "fullscale_usa_field_add"
# Datasets present in the output root but hidden from the public demo.
# The data still exists on disk (kept for research / local use); it is
# just not surfaced in the dataset dropdown. Overrides:
# DEMO_SHOW_HIDDEN=1 -> show everything (local research)
# DEMO_HIDDEN_DATASETS=a,b,c -> replace this default list
HIDDEN_DATASET_IDS = frozenset({"sounds_colors_batch"})
@classmethod
def _resolve_hidden_ids(cls) -> Set[str]:
if (os.environ.get("DEMO_SHOW_HIDDEN") or "").strip().lower() in (
"1", "true", "yes",
):
return set()
env = (os.environ.get("DEMO_HIDDEN_DATASETS") or "").strip()
if env:
return {s.strip() for s in env.split(",") if s.strip()}
return set(cls.HIDDEN_DATASET_IDS)
def __init__(self, output_root: Path, initial_data_dir: Optional[Path] = None):
self.output_root = Path(output_root)
# {dataset_id: {"dir": Path, "label": str, "runs": [run_dict, ...]}}
self._datasets: Dict[str, Dict[str, Any]] = {}
self._active_dataset_id: Optional[str] = None
self.active_loader: Optional["DataLoader"] = None
self._best_aggregator = None # lazy init after scan
# Cached entity-slug intersection per dataset. A slug only stays
# if every dropdown-visible run has a ``by_source/<slug>/`` dir
# with at least one ``to_*.json`` swap. Used to propagate manual
# entity deletions (e.g. flawed targets) across all runs.
self._allowed_slugs_cache: Dict[str, Set[str]] = {}
self._scan()
if self._datasets:
# If a specific data_dir was given, prefer the dataset that matches it
preferred_id = self._resolve_preferred_dataset_id(initial_data_dir)
self._activate_dataset(preferred_id)
def _resolve_preferred_dataset_id(self, initial_data_dir: Optional[Path]) -> str:
"""Pick a deterministic default dataset for the demo."""
if initial_data_dir is not None:
initial_data_dir = Path(initial_data_dir).resolve()
for ds_id, ds in self._datasets.items():
if ds["dir"].resolve() == initial_data_dir:
return ds_id
env_dataset_id = (os.environ.get("DEMO_DEFAULT_DATASET") or "").strip()
if env_dataset_id and env_dataset_id in self._datasets:
return env_dataset_id
if self.DEFAULT_DATASET_ID in self._datasets:
return self.DEFAULT_DATASET_ID
# Fall back to the first dataset after label sort.
return next(iter(self._datasets))
# ------------------------------------------------------------------
# Discovery
# ------------------------------------------------------------------
def _scan(self) -> None:
"""Walk output_root looking for demo-enabled run_manifest.json files."""
if not self.output_root.exists():
return
datasets: Dict[Path, Dict[str, Any]] = {}
for manifest_path in sorted(
self.output_root.glob("*/_swaps/runs/*/run_manifest.json"),
reverse=True,
):
try:
with open(manifest_path, "r", encoding="utf-8") as fh:
manifest = json.load(fh)
except (json.JSONDecodeError, IOError):
continue
if not manifest.get("display_demo"):
continue
# Resolve dataset dir: prefer the explicit graphs_root in config
graphs_root_raw = (
manifest.get("config", {})
.get("inputs", {})
.get("graphs_root", "")
)
dataset_dir = None
if graphs_root_raw:
candidate = Path(graphs_root_raw)
if not candidate.is_absolute():
candidate = self.output_root.parent / candidate
resolved_root = self.output_root.resolve()
if candidate.exists() and _is_subpath(candidate.resolve(), resolved_root):
dataset_dir = candidate
if dataset_dir is None:
# Physical fallback: output/<ds>/_swaps/runs/<run_id>/run_manifest.json
dataset_dir = manifest_path.parent.parent.parent.parent
dataset_dir = dataset_dir.resolve()
if dataset_dir not in datasets:
label = self._derive_dataset_label(dataset_dir.name)
datasets[dataset_dir] = {
"dir": dataset_dir,
"label": label,
"runs": [],
}
run_id = manifest_path.parent.name
swap_count = sum(
1 for _ in (manifest_path.parent / "by_source").glob("*/to_*.json")
) if (manifest_path.parent / "by_source").exists() else 0
run_dir = manifest_path.parent
datasets[dataset_dir]["runs"].append({
"id": run_id,
"manifest": manifest,
"swap_count": swap_count,
"control_mode": _extract_control_mode(manifest, run_dir=run_dir),
"semantic_label": _semantic_run_label(run_id, manifest, run_dir=run_dir),
})
# Build ordered dict: datasets sorted by label, runs already reverse-sorted
hidden = self._resolve_hidden_ids()
for dataset_dir, ds in sorted(datasets.items(), key=lambda x: x[1]["label"]):
ds_id = dataset_dir.name
if ds_id in hidden:
continue
self._datasets[ds_id] = ds
@staticmethod
def _derive_dataset_label(folder_name: str) -> str:
"""Turn 'book_characters_authors_batch' -> 'Book Characters Authors'."""
name = folder_name.replace("_batch", "").replace("_", " ").title()
return name
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
def list_datasets(self) -> List[Dict[str, Any]]:
"""Return summary list of all demo datasets."""
result = []
for ds_id, ds in self._datasets.items():
result.append({
"id": ds_id,
"label": ds["label"],
"run_count": len(ds["runs"]),
"is_active": ds_id == self._active_dataset_id,
})
return result
def list_runs_for_dataset(self, dataset_id: str) -> List[Dict[str, Any]]:
"""Return run summaries for a specific dataset (filtered for the UI)."""
ds = self._datasets.get(dataset_id)
if not ds:
return []
result = []
for r in ds["runs"]:
manifest = r["manifest"]
if not _is_dropdown_run(r["id"], manifest):
continue
result.append({
"id": r["id"],
"dataset_id": dataset_id,
"dataset_label": ds["label"],
"name": self._format_run_name(r["id"]),
"semantic_label": r.get("semantic_label", ""),
"control_mode": r.get("control_mode", "labeled"),
"swap_count": r["swap_count"],
"status": manifest.get("status", "unknown"),
"timestamp": manifest.get("timestamp_started", manifest.get("timestamp", "")),
"has_trajectory": (
(ds["dir"] / "_swaps" / "runs" / r["id"] / "_trajectory_analysis").exists()
),
})
return result
def list_all_runs(self) -> List[Dict[str, Any]]:
"""Return all demo-enabled runs across all datasets, grouped by dataset."""
all_runs = []
for ds_id in self._datasets:
all_runs.extend(self.list_runs_for_dataset(ds_id))
return all_runs
def set_dataset(self, dataset_id: str) -> bool:
"""Switch the active dataset (picks its best run automatically)."""
if dataset_id not in self._datasets:
return False
self._activate_dataset(dataset_id)
return True
def set_run(self, run_id: str) -> bool:
"""Switch run within the current dataset."""
if self.active_loader is None:
return False
return self.active_loader.set_run(run_id)
def set_run_global(self, run_id: str) -> bool:
"""Switch to any run across all datasets.
Finds which dataset owns the run, activates that dataset, then
selects the specific run.
"""
for ds_id, ds in self._datasets.items():
for r in ds["runs"]:
if r["id"] == run_id:
if ds_id != self._active_dataset_id:
self._activate_dataset(ds_id)
self.active_loader.set_run(run_id)
return True
return False
@property
def active_dataset_id(self) -> Optional[str]:
return self._active_dataset_id
@property
def best_aggregator(self):
"""Lazy-init cross-run best aggregator."""
if self._best_aggregator is None:
from .cross_run_best import CrossRunBestAggregator
self._best_aggregator = CrossRunBestAggregator(self)
return self._best_aggregator
def get_best_cross_run_matrix(self) -> Dict[str, Any]:
"""Return best-per-cell matrix + winners for the active dataset."""
if not self._active_dataset_id:
return {"matrix": {}, "winners": {}, "considered_runs": []}
return self.best_aggregator.get_best_matrix(self._active_dataset_id)
def get_allowed_slugs(self, dataset_id: str) -> Optional[Set[str]]:
"""Return the entity-slug intersection across visible runs.
A slug is "allowed" only if every dropdown-visible run for the
dataset has a non-empty ``by_source/<slug>/`` directory. This
propagates manual entity deletions in any one run to all the
others (the user removes flawed targets in Field Additivity and
the Classic Swap matrix automatically hides the same rows /
columns).
Returns ``None`` when the dataset is unknown or has no visible
runs (so the caller should fall back to no filtering).
"""
if dataset_id in self._allowed_slugs_cache:
return self._allowed_slugs_cache[dataset_id]
ds = self._datasets.get(dataset_id)
if not ds:
return None
visible_runs = [
r for r in ds["runs"]
if _is_dropdown_run(r["id"], r.get("manifest", {}))
]
if not visible_runs:
return None
per_run_slugs: List[Set[str]] = []
for r in visible_runs:
run_dir = ds["dir"] / "_swaps" / "runs" / r["id"]
slugs = self._collect_run_slugs(run_dir)
if slugs:
per_run_slugs.append(slugs)
if not per_run_slugs:
return None
allowed = set.intersection(*per_run_slugs)
self._allowed_slugs_cache[dataset_id] = allowed
return allowed
@staticmethod
def _collect_run_slugs(run_dir: Path) -> Set[str]:
"""Return the set of entity slugs that have at least one swap file
in ``run_dir/by_source/<slug>/to_*.json`` (canonical or variant).
An empty source folder (because every ``to_*.json`` was deleted)
is treated as "entity removed" and excluded.
"""
by_source = run_dir / "by_source"
if not by_source.exists():
return set()
slugs: Set[str] = set()
for source_dir in by_source.iterdir():
if not source_dir.is_dir():
continue
if any(source_dir.glob("to_*.json")):
slugs.add(source_dir.name)
return slugs
def get_swap_detail_for_run(
self, run_id: str, from_slug: str, to_slug: str,
variant: Optional[str] = None,
) -> Optional[Dict]:
"""Load a swap detail from a specific run (for cross-run best mode)."""
for ds in self._datasets.values():
for r in ds["runs"]:
if r["id"] == run_id:
loader = DataLoader(ds["dir"], run_id=run_id)
return loader.get_swap_detail(
from_slug, to_slug, variant=variant
)
return None
# ------------------------------------------------------------------
# Internal helpers
# ------------------------------------------------------------------
def _activate_dataset(self, dataset_id: str) -> None:
ds = self._datasets[dataset_id]
best_run_id = self._pick_default_run(ds["runs"])
allowed = self.get_allowed_slugs(dataset_id)
self.active_loader = DataLoader(
ds["dir"], run_id=best_run_id, allowed_slugs=allowed,
)
self._active_dataset_id = dataset_id
@classmethod
def _pick_default_run(cls, runs: List[Dict[str, Any]]) -> Optional[str]:
"""Choose the default run for a dataset.
Priority:
1. ``DEMO_DEFAULT_RUN`` env var (exact run id match)
2. Class constant ``DEFAULT_RUN_ID``
3. Largest "classic" swap run that's visible in the dropdown
4. Largest visible run regardless of mode
5. First run in list (newest)
"""
if not runs:
return None
env_run = (os.environ.get("DEMO_DEFAULT_RUN") or "").strip()
if env_run:
for r in runs:
if r["id"] == env_run:
return r["id"]
if cls.DEFAULT_RUN_ID:
for r in runs:
if r["id"] == cls.DEFAULT_RUN_ID:
return r["id"]
visible = [
r for r in runs
if _is_dropdown_run(r["id"], r.get("manifest", {}))
]
if visible:
classic = [
r for r in visible
if r.get("control_mode") not in ("random_feature_matched", "additivity")
]
pool = classic or visible
return max(pool, key=lambda r: r.get("swap_count", 0))["id"]
return runs[0]["id"]
@staticmethod
def _format_run_name(run_id: str) -> str:
name = run_id.replace("_", " ").title()
name = name.replace("50states", "50 States").replace("6states", "6 States")
return name
class DataLoader:
"""Load and cache swap experiment data with multi-run support."""
def __init__(
self,
data_dir: Path,
run_id: Optional[str] = None,
allowed_slugs: Optional[Set[str]] = None,
):
"""Initialise a loader for one dataset directory.
``allowed_slugs`` is an optional whitelist of entity slugs. When
provided, every public matrix / state getter prunes its output
to only those slugs. This is how cross-run entity deletions
propagate: ``DemoRegistry`` computes the intersection of slugs
across all visible runs and forwards it here.
"""
self.data_dir = Path(data_dir)
self.base_swaps_dir = self.data_dir / "_swaps"
self.run_id = run_id
self.allowed_slugs: Optional[Set[str]] = (
set(allowed_slugs) if allowed_slugs is not None else None
)
self._set_swaps_dir()
self._matrix_cache: Dict[Optional[str], Dict] = {}
self._flip_matrix_cache: Dict[Optional[str], Dict] = {}
self._regime_matrix_cache: Dict[Optional[str], Dict] = {}
self._vsmax_matrix_cache: Dict[Optional[str], Dict] = {}
self._variant_index_cache: Optional[Dict] = None
self._states_cache: Optional[List[Dict]] = None
self._analysis_cache: Optional[Dict] = None
self._stats_cache: Optional[Dict] = None
self._domain_config_cache: Optional[Dict] = None
self._gpu_batch_features_index: Optional[Dict[str, List]] = None
def _set_swaps_dir(self):
"""Set swaps_dir based on current run_id."""
if self.run_id:
run_dir = self.base_swaps_dir / "runs" / self.run_id
if run_dir.exists():
self.swaps_dir = run_dir
else:
# Fallback to base swaps dir
self.swaps_dir = self.base_swaps_dir
else:
# Check if runs directory exists and has runs
runs_dir = self.base_swaps_dir / "runs"
if runs_dir.exists():
# Prefer the full 50-state run on a fresh load.
default_run_id = self._get_default_run_id()
if default_run_id:
self.run_id = default_run_id
self.swaps_dir = self.base_swaps_dir / "runs" / default_run_id
else:
self.swaps_dir = self.base_swaps_dir
else:
self.swaps_dir = self.base_swaps_dir
def _get_default_run_id(self) -> Optional[str]:
"""Choose the default run for a fresh app load (most recent non-legacy).
``list_runs`` already filters out sweeps and ``*_labeled`` controls,
so we only need to drop legacy runs here.
"""
runs = self.list_runs()
if not runs:
return None
for run in runs:
if "legacy" not in run["id"].lower():
return run["id"]
return runs[0]["id"]
def list_runs(self) -> List[Dict[str, Any]]:
"""List available experiment runs."""
runs = []
runs_dir = self.base_swaps_dir / "runs"
if not runs_dir.exists():
return runs
for run_dir in sorted(runs_dir.iterdir(), reverse=True):
if not run_dir.is_dir():
continue
by_source = run_dir / "by_source"
if not by_source.exists():
continue
# Load run manifest for metadata
manifest_path = run_dir / "run_manifest.json"
manifest = {}
if manifest_path.exists():
try:
with open(manifest_path, 'r', encoding='utf-8') as f:
manifest = json.load(f)
except (json.JSONDecodeError, IOError):
pass
if not _is_dropdown_run(run_dir.name, manifest):
continue
# Count swaps
swap_count = sum(1 for _ in by_source.glob("*/to_*.json"))
# Check for trajectory data
has_trajectory = (run_dir / "_trajectory_analysis").exists()
# Get config info
config = manifest.get('config', {})
ct_config = config.get('ct_steering', {})
runs.append({
'id': run_dir.name,
'name': self._format_run_name(run_dir.name),
'semantic_label': _semantic_run_label(run_dir.name, manifest, run_dir=run_dir),
'control_mode': _extract_control_mode(manifest, run_dir=run_dir),
'swap_count': swap_count,
'has_trajectory': has_trajectory,
'timestamp': manifest.get('timestamp_started', ''),
'status': manifest.get('status', 'unknown'),
'M_ablate': ct_config.get('M_ablate', -2),
'M_amplify': ct_config.get('M_amplify', 20),
'is_current': run_dir.name == self.run_id,
})
return runs
def _format_run_name(self, run_id: str) -> str:
"""Format run ID into human-readable name."""
# full_50states_v1 -> "Full 50 States v1"
# pilot_6states_v1 -> "Pilot 6 States v1"
# gpu_test_8 -> "GPU Test 8"
name = run_id.replace('_', ' ').title()
name = name.replace('50states', '50 States')
name = name.replace('6states', '6 States')
name = name.replace('Gpu', 'GPU')
return name
def set_run(self, run_id: str) -> bool:
"""Switch to a different run."""
run_dir = self.base_swaps_dir / "runs" / run_id
if not run_dir.exists():
return False
self.run_id = run_id
self._set_swaps_dir()
self._clear_caches()
return True
def get_current_run(self) -> Optional[str]:
"""Get the current run ID."""
return self.run_id
def _clear_caches(self):
"""Clear all cached data."""
self._matrix_cache = {}
self._flip_matrix_cache = {}
self._regime_matrix_cache = {}
self._vsmax_matrix_cache = {}
self._variant_index_cache = None
self._states_cache = None
self._analysis_cache = None
self._stats_cache = None
self._domain_config_cache = None
self._gpu_batch_features_index = None
self._methodology_stats = None
# ------------------------------------------------------------------
# Cross-run entity filtering
# ------------------------------------------------------------------
def _apply_slug_filter(self, matrix: Dict[str, Dict]) -> Dict[str, Dict]:
"""Return a copy of *matrix* containing only allowed_slugs.
Filters both the row dimension (sources) and column dimension
(targets). When ``allowed_slugs`` is None this is a no-op (the
original dict is returned untouched, since callers don't mutate
the result).
"""
if self.allowed_slugs is None:
return matrix
allowed = self.allowed_slugs
filtered: Dict[str, Dict] = {}
for from_slug, row in matrix.items():
if from_slug not in allowed:
continue
filtered[from_slug] = {
to_slug: val
for to_slug, val in row.items()
if to_slug in allowed
}
return filtered
# Per-domain blacklists of common words that should be ignored by the
# word-boundary tier heuristic (e.g. "city" in "Salt Lake City" would
# otherwise match generic geography text like "is a city").
TIER_WORD_BLACKLIST: Dict[str, List[str]] = {
'usa_states': ['city'],
}
# ------------------------------------------------------------------
# Domain & model configuration (loaded from config_resolved.json)
# ------------------------------------------------------------------
def get_domain_config(self) -> Dict[str, Any]:
"""Return domain and model metadata for the active run.
Reads config_resolved.json once and caches the result. Falls back to
USA-states defaults when the file is missing (legacy runs).
"""
if self._domain_config_cache is not None:
return self._domain_config_cache
config = self._load_config_resolved()
if not config:
self._domain_config_cache = self._build_fallback_domain_config()
return self._domain_config_cache
entities = config.get('_entities', [])
entity_index = {e['slug']: e for e in entities}
concept_fields = config.get('swap', {}).get('concept_fields', [])
entity_fields: List[str] = []
if entities:
entity_fields = [k for k in entities[0] if k != 'slug']
primary_field = entity_fields[0] if entity_fields else 'slug'
for f in entity_fields:
if f not in concept_fields:
primary_field = f
break
answer_field = concept_fields[-1] if concept_fields else ''
ct = config.get('ct_steering', {})
model_id = ct.get('model_id', '')
transcoder_set = ct.get('transcoder_set', '')
np_model = model_id.split('/')[-1] if model_id else 'gemma-2-2b'
experiment_name = config.get('experiment_name', '')
is_usa_states = 'usa_states' in experiment_name or 'state' in concept_fields
display_name = experiment_name.replace('_swap', '').replace('_', ' ').title()
blacklist: List[str] = []
for key, words in self.TIER_WORD_BLACKLIST.items():
if key in experiment_name:
blacklist = [w.lower() for w in words]
break
self._domain_config_cache = {
'experiment_name': experiment_name,
'display_name': display_name,
'concept_fields': concept_fields,
'answer_field': answer_field,
'primary_field': primary_field,
'entity_fields': entity_fields,
'entity_index': entity_index,
'entity_count': len(entities),
'model_id': model_id,
'transcoder_set': transcoder_set,
'np_model': np_model,
'is_usa_states': is_usa_states,
'tier_word_blacklist': blacklist,
'm_ablate': ct.get('M_ablate', -2),
'm_amplify': ct.get('M_amplify', 20),
'temperature': ct.get('temperature', 0.3),
'n_tokens': ct.get('n_tokens', 10),
}
return self._domain_config_cache
def _load_config_resolved(self) -> Optional[Dict]:
"""Load config_resolved.json from the active run directory."""
path = self.swaps_dir / "config_resolved.json"
if path.exists():
try:
with open(path, 'r', encoding='utf-8') as f:
return json.load(f)
except (json.JSONDecodeError, IOError):
pass
return None
def _build_fallback_domain_config(self) -> Dict[str, Any]:
"""Fallback domain config for legacy runs without config_resolved.json."""
return {
'experiment_name': 'usa_states_swap',
'display_name': 'USA States',
'concept_fields': ['state', 'capital'],
'answer_field': 'capital',
'primary_field': 'city',
'entity_fields': ['city', 'state', 'capital'],
'entity_index': {},
'entity_count': 0,
'model_id': 'google/gemma-2-2b',
'transcoder_set': 'mntss/clt-gemma-2-2b-2.5M',
'np_model': 'gemma-2-2b',
'is_usa_states': True,
'tier_word_blacklist': ['city'],
'm_ablate': -2,
'm_amplify': 20,
'temperature': 0.3,
'n_tokens': 10,
}
@staticmethod
def _slug_to_label(slug: str) -> str:
"""Convert a slug to a human-readable label."""
return slug.replace('_', ' ').title()
def _make_abbr(self, slug: str, label: str) -> str:
"""Generate a short abbreviation for matrix headers."""
dc = self.get_domain_config()
if dc['is_usa_states']:
return self._state_to_abbr(label)
parts = slug.split('_')
if len(parts) >= 2:
return ''.join(p[0].upper() for p in parts[:3])
return slug[:3].upper()
# ------------------------------------------------------------------
# Variant index
# ------------------------------------------------------------------
def _build_variant_index(self) -> Dict:
"""Group swap files by canonical (from_slug, to_slug) pair.
Returns ``{from_slug: {to_slug: [record, ...]}}``.
Each record is::
{
"path": Path,
"variant_suffix": str, # "" for canonical, "r0" / "add_book" / etc.
"is_canonical": bool,
}
"""
if self._variant_index_cache is not None:
return self._variant_index_cache
index: Dict[str, Dict[str, list]] = {}
by_source_dir = self.swaps_dir / "by_source"
if not by_source_dir.exists():
self._variant_index_cache = index
return index
for source_dir in by_source_dir.iterdir():
if not source_dir.is_dir():
continue
from_slug = source_dir.name
if from_slug not in index:
index[from_slug] = {}
for swap_file in source_dir.glob("to_*.json"):
to_slug, variant_suffix = _parse_to_slug(swap_file.stem)
if not to_slug:
continue
index[from_slug].setdefault(to_slug, []).append({
"path": swap_file,
"variant_suffix": variant_suffix,
"is_canonical": variant_suffix == "",
})
self._variant_index_cache = index
return index
def get_variants_for_pair(
self, from_slug: str, to_slug: str
) -> List[Dict[str, Any]]:
"""Return lightweight variant descriptors for a single pair.
Each item includes ``variant_suffix``, ``tier``, ``flip_position``,
and control metadata extracted from the swap JSON.
"""
idx = self._build_variant_index()
records = idx.get(from_slug, {}).get(to_slug, [])
variants: List[Dict[str, Any]] = []
for rec in records:
if rec["is_canonical"]:
continue
try:
with open(rec["path"], "r", encoding="utf-8") as fh:
data = json.load(fh)
except (json.JSONDecodeError, IOError):
continue
ctrl = data.get("metadata", {}).get("control", {})
variants.append({
"variant_suffix": rec["variant_suffix"],
"tier": self._get_tier_from_swap(data),
"flip_position": self._get_flip_position_from_swap(data),
"control_mode": ctrl.get("control_mode", ""),
"concept_subsets_used": ctrl.get("concept_subsets_used", []),
"replicate_id": ctrl.get("replicate_id"),
})
return variants
def get_run_variant_suffixes(self) -> List[str]:
"""Return sorted list of distinct variant suffixes across the entire run."""
idx = self._build_variant_index()
suffixes: set = set()
for targets in idx.values():
for records in targets.values():
for r in records:
if r["variant_suffix"]:
suffixes.add(r["variant_suffix"])
return sorted(suffixes)
def _best_variant_path(
self, from_slug: str, to_slug: str
) -> Optional[Path]:
"""Return the path of the best variant file for a pair (highest tier)."""
idx = self._build_variant_index()
records = idx.get(from_slug, {}).get(to_slug, [])
variants = [r for r in records if not r["is_canonical"]]
if not variants:
return None
best_path = None
best_tier = -1
for rec in variants:
try:
with open(rec["path"], "r", encoding="utf-8") as fh:
data = json.load(fh)
tier = self._get_tier_from_swap(data) or 0
if tier > best_tier:
best_tier = tier
best_path = rec["path"]
except (json.JSONDecodeError, IOError):
continue
return best_path
# ------------------------------------------------------------------
# Matrix
# ------------------------------------------------------------------
def get_matrix(self, variant: Optional[str] = None) -> Dict[str, Dict[str, Optional[int]]]:
"""Build tier matrix dynamically from individual swap JSON files.
When *variant* is None (default), canonical files take priority and
variant-only pairs aggregate using the best (highest) tier.
When *variant* is set, only files matching that suffix are used.
"""
if variant in self._matrix_cache:
return self._apply_slug_filter(self._matrix_cache[variant])
idx = self._build_variant_index()
matrix: Dict[str, Dict[str, Optional[int]]] = {}
for from_slug, targets in idx.items():
matrix[from_slug] = {}
for to_slug, records in targets.items():
if variant is not None:
matched = [r for r in records if r["variant_suffix"] == variant]
if matched:
try:
with open(matched[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = self._get_tier_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
pass
continue
canonical = [r for r in records if r["is_canonical"]]
variants = [r for r in records if not r["is_canonical"]]
if canonical:
try:
with open(canonical[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = self._get_tier_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
pass
elif variants:
best_tier: Optional[float] = None
for rec in variants:
try:
with open(rec["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
tier = self._get_tier_from_swap(swap_data)
if tier is not None and (best_tier is None or tier > best_tier):
best_tier = tier
except (json.JSONDecodeError, IOError):
continue
if best_tier is not None:
matrix[from_slug][to_slug] = best_tier
if variant is None:
work_dir = self.swaps_dir / "work"
if work_dir.exists():
for swap_file in work_dir.glob("*.json"):
try:
with open(swap_file, "r", encoding="utf-8") as f:
swap_data = json.load(f)
swap_id = swap_file.stem
if "__to__" in swap_id:
from_slug, to_slug = swap_id.split("__to__")
if from_slug not in matrix:
matrix[from_slug] = {}
if to_slug not in matrix[from_slug]:
matrix[from_slug][to_slug] = self._get_tier_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
continue
self._matrix_cache[variant] = matrix
return self._apply_slug_filter(matrix)
def _get_tier_from_swap(self, swap_data: Dict) -> Optional[float]:
"""Extract tier from swap data, computing if necessary.
Tier semantics (domain-agnostic):
5 = PERFECT -- target answer detected in steered output via any of:
a) full answer match (strict or fuzzy punctuation)
b) steered first token is a substring of the target answer
c) any word (len>=3) of the target answer appears in the
steered output (e.g. "Harper" in "Nelle Harper...")
2 = SUPPRESSED_ONLY -- source gone, no target signal
1 = SOURCE_PERSISTS -- source answer still in output
"""
if 'classification' in swap_data:
tier = swap_data['classification'].get('tier')
if tier is not None:
return float(tier)
evaluation = swap_data.get('evaluation', {})
exact = evaluation.get('exact_match', {})
hit = exact.get('steered_has_to_capital') or exact.get('steered_has_to_answer')
to_answer = evaluation.get('to_answer', '')
steered_out = evaluation.get('raw', {}).get('steered_output', '')
if not hit and to_answer and steered_out:
to_norm = to_answer.replace('.', '').replace('-', ' ').lower()
out_norm = steered_out.replace('.', '').replace('-', ' ').lower()
if to_norm and to_norm in out_norm:
hit = True
if not hit:
first_token = evaluation.get('first_token', {})
steered_first_tok = (first_token.get('steered', '') or '').strip()
if len(steered_first_tok) >= 2 and to_answer:
answer_norm = to_answer.replace('.', '').lower()
tok_norm = steered_first_tok.lower()
if tok_norm in answer_norm:
hit = True
if not hit and to_answer and steered_out:
blacklist = set(self.get_domain_config().get('tier_word_blacklist', []))
out_lower = steered_out.lower()
for word in to_answer.replace('.', '').split():
if len(word) >= 3 and word.lower() not in blacklist:
pattern = r'\b' + re.escape(word.lower()) + r'\b'
if re.search(pattern, out_lower):
hit = True
break
if hit:
return 5
if not exact.get('from_suppressed'):
return 1
return 2
@staticmethod
def _get_flip_position_from_swap(swap_data: Dict):
"""Extract logit flip position from swap data.
Returns:
int >= 0 -- generation position where the flip first occurs
None -- trajectory was tracked but flip never occurred
-1 -- no trajectory data available for this swap
"""
traj = swap_data.get('evaluation', {}).get('logit_trajectory')
if traj is None or not isinstance(traj, dict) or 'summary' not in traj:
return -1
return traj['summary'].get('flip_position')
def get_flip_matrix(self, variant: Optional[str] = None) -> Dict[str, Dict[str, Optional[int]]]:
"""Build flip-position matrix from swap JSON files.
Same structure as ``get_matrix()`` but values are the generation
position where the logit flip first occurs (0, 1, 2 ...) or
``None`` when no flip is observed.
When *variant* is None, canonical files take priority and variant-only
pairs pick the earliest non-null flip. When *variant* is set, only
files matching that suffix are used.
"""
if variant in self._flip_matrix_cache:
return self._apply_slug_filter(self._flip_matrix_cache[variant])
idx = self._build_variant_index()
matrix: Dict[str, Dict[str, Optional[int]]] = {}
for from_slug, targets in idx.items():
matrix[from_slug] = {}
for to_slug, records in targets.items():
if variant is not None:
matched = [r for r in records if r["variant_suffix"] == variant]
if matched:
try:
with open(matched[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = self._get_flip_position_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
pass
continue
canonical = [r for r in records if r["is_canonical"]]
variants = [r for r in records if not r["is_canonical"]]
if canonical:
try:
with open(canonical[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = self._get_flip_position_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
pass
elif variants:
best_flip = None
for rec in variants:
try:
with open(rec["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
fp = self._get_flip_position_from_swap(swap_data)
if fp is not None and fp >= 0:
if best_flip is None or fp < best_flip:
best_flip = fp
except (json.JSONDecodeError, IOError):
continue
matrix[from_slug][to_slug] = best_flip
if variant is None:
work_dir = self.swaps_dir / "work"
if work_dir.exists():
for swap_file in work_dir.glob("*.json"):
try:
with open(swap_file, "r", encoding="utf-8") as f:
swap_data = json.load(f)
swap_id = swap_file.stem
if "__to__" in swap_id:
from_slug, to_slug = swap_id.split("__to__")
if from_slug not in matrix:
matrix[from_slug] = {}
if to_slug not in matrix[from_slug]:
matrix[from_slug][to_slug] = self._get_flip_position_from_swap(swap_data)
except (json.JSONDecodeError, IOError):
continue
self._flip_matrix_cache[variant] = matrix
return self._apply_slug_filter(matrix)
# ------------------------------------------------------------------
# Regime & VsMax matrices
# ------------------------------------------------------------------
@staticmethod
def _get_regime_from_swap(swap_data: Dict) -> Optional[str]:
"""Extract logit-shift regime (A-E) from swap data."""
pos0 = swap_data.get('evaluation', {}).get('position_0_comparison', {})
return _classify_regime(pos0)
@staticmethod
def _get_vsmax_from_swap(swap_data: Dict) -> Optional[float]:
"""Extract best-trajectory vsMax from swap data.
Returns the ``best_target_minus_max`` float or None.
"""
contrast = (swap_data.get('evaluation', {})
.get('logit_trajectory', {})
.get('contrast_groups', {}))
if not isinstance(contrast, dict):
return None
grp = contrast.get('same_dataset', {})
if not isinstance(grp, dict):
return None
agg = grp.get('aggregate', {})
if not isinstance(agg, dict):
return None
return agg.get('best_target_minus_max')
def _build_overlay_matrix(
self,
extractor,
cache: Dict,
variant: Optional[str],
best_cmp,
) -> Dict[str, Dict]:
"""Generic matrix builder reused by regime and vsmax.
``extractor`` pulls a value from swap_data.
``best_cmp(current_best, candidate)`` returns True when candidate is
preferred over current_best (used for variant-only fallback).
"""
if variant in cache:
return self._apply_slug_filter(cache[variant])
idx = self._build_variant_index()
matrix: Dict[str, Dict] = {}
for from_slug, targets in idx.items():
matrix[from_slug] = {}
for to_slug, records in targets.items():
if variant is not None:
matched = [r for r in records if r["variant_suffix"] == variant]
if matched:
try:
with open(matched[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = extractor(swap_data)
except (json.JSONDecodeError, IOError):
pass
continue
canonical = [r for r in records if r["is_canonical"]]
variants_list = [r for r in records if not r["is_canonical"]]
if canonical:
try:
with open(canonical[0]["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
matrix[from_slug][to_slug] = extractor(swap_data)
except (json.JSONDecodeError, IOError):
pass
elif variants_list:
best_val = None
for rec in variants_list:
try:
with open(rec["path"], "r", encoding="utf-8") as f:
swap_data = json.load(f)
val = extractor(swap_data)
if val is not None and (best_val is None or best_cmp(best_val, val)):
best_val = val
except (json.JSONDecodeError, IOError):
continue
if best_val is not None:
matrix[from_slug][to_slug] = best_val
if variant is None:
work_dir = self.swaps_dir / "work"
if work_dir.exists():
for swap_file in work_dir.glob("*.json"):
try:
with open(swap_file, "r", encoding="utf-8") as f:
swap_data = json.load(f)
swap_id = swap_file.stem
if "__to__" in swap_id:
from_slug, to_slug = swap_id.split("__to__")
if from_slug not in matrix:
matrix[from_slug] = {}
if to_slug not in matrix[from_slug]:
matrix[from_slug][to_slug] = extractor(swap_data)
except (json.JSONDecodeError, IOError):
continue
cache[variant] = matrix
return self._apply_slug_filter(matrix)
def get_regime_matrix(self, variant: Optional[str] = None) -> Dict[str, Dict[str, Optional[str]]]:
"""Build regime matrix (A-E letters) from swap JSON files."""
_REGIME_RANK = {"A": 5, "B": 4, "C": 3, "D": 2, "E": 1}
return self._build_overlay_matrix(
self._get_regime_from_swap,
self._regime_matrix_cache,
variant,
best_cmp=lambda cur, cand: _REGIME_RANK.get(cand, 0) > _REGIME_RANK.get(cur, 0),
)
def get_vsmax_matrix(self, variant: Optional[str] = None) -> Dict[str, Dict[str, Optional[float]]]:
"""Build vsMax matrix from swap JSON files. Higher is better."""
return self._build_overlay_matrix(
self._get_vsmax_from_swap,
self._vsmax_matrix_cache,
variant,
best_cmp=lambda cur, cand: cand > cur,
)
def get_states(self) -> List[Dict[str, Any]]:
"""Get entity list with metadata from by_source directories.
Returns a list of dicts with backward-compatible keys (``state``,
``capital``, ``city``, ``abbr``) plus generic keys (``label``,
``answer``, ``fields``) so that frontends can work with any domain.
"""
if self._states_cache is not None:
return self._states_cache
dc = self.get_domain_config()
entity_index = dc['entity_index']
primary_field = dc['primary_field']
answer_field = dc['answer_field']
analysis = self.get_analysis()
archetypes = analysis.get('archetypes', {})
archetype_lookup: Dict[str, Dict] = {}
for archetype, entity_list in archetypes.items():
for ed in entity_list:
key = ed.get('state', ed.get(primary_field, ''))
archetype_lookup[key] = {
'archetype': archetype,
'native_prob': ed.get('native_prob', 0),
'supernodes': ed.get('supernodes', 0),
'src_tier': ed.get('src_tier', 0),
'tgt_tier': ed.get('tgt_tier', 0),
}
entities: List[Dict[str, Any]] = []
by_source_dir = self.swaps_dir / "by_source"
if by_source_dir.exists():
for source_dir in sorted(by_source_dir.iterdir()):
if not source_dir.is_dir():
continue
slug = source_dir.name
if self.allowed_slugs is not None and slug not in self.allowed_slugs:
continue
edata = entity_index.get(slug, {})
if not edata:
edata = self._get_entity_info_from_swap(source_dir)
label = (edata.get(primary_field, '')
or self._slug_to_label(slug))
answer = edata.get(answer_field, '')
abbr = self._make_abbr(slug, edata.get('state', label))
# Backward-compatible keys
state_val = edata.get('state', label)
capital_val = edata.get('capital', answer)
city_val = edata.get('city',
edata.get(primary_field, label))
arch_key = state_val if dc['is_usa_states'] else label
arch_data = archetype_lookup.get(arch_key, {})
neuronpedia_url = self._get_neuronpedia_url(slug)
logit_flags = (self._get_entity_logit_flags(slug, answer)
if self._find_state_dir(slug) else {})
entities.append({
'slug': slug,
'label': label,
'answer': answer,
'state': state_val,
'capital': capital_val,
'city': city_val,
'abbr': abbr,
'archetype': arch_data.get('archetype', 'Unknown'),
'native_prob': arch_data.get('native_prob', 0),
'supernodes': arch_data.get('supernodes', 0),
'src_tier': arch_data.get('src_tier', 0),
'tgt_tier': arch_data.get('tgt_tier', 0),
'neuronpedia_url': neuronpedia_url,
'fields': {k: v for k, v in edata.items()
if k != 'slug'},
**logit_flags,
})
self._states_cache = entities
return entities
def _get_entity_info_from_swap(self, source_dir: Path) -> Dict[str, str]:
"""Extract entity fields from a sample swap file's source section.
Tries canonical files first, falls back to variant files for
variant-only runs.
"""
for swap_file in sorted(source_dir.glob("to_*.json")):
try:
with open(swap_file, 'r', encoding='utf-8') as f:
swap_data = json.load(f)
source = swap_data.get('source', {})
if source:
return {k: v for k, v in source.items()
if k not in ('prompt', 'concept',
'neuronpedia_url')}
except (json.JSONDecodeError, IOError):
continue
return {}
def _load_entity_logit_metadata(self, entity_dir: Path,
answer: str) -> Dict[str, Any]:
"""Load top-logit metadata from graph.json.
``answer`` is the expected output token (capital, author, etc.)
used only for highlighting; the rest is model-agnostic.
"""
graph_path = entity_dir / "00 Graph Generation" / "graph.json"
if not graph_path.exists():
return {}
try:
with open(graph_path, 'r', encoding='utf-8') as f:
graph = json.load(f)
except (json.JSONDecodeError, IOError):
return {}
import re
logits: List[Dict] = []
native_prob = 0.0
target_token = ''
for node in graph.get('nodes', []):
clerp = node.get('clerp', '')
prob = node.get('token_prob')
if clerp and prob is not None and clerp.startswith('Output '):
match = re.search(r'Output\s+"([^"]*)"', clerp)
if match:
token = match.group(1).strip()
if token:
logits.append({
'token': token,
'prob': prob,
'is_target': node.get('is_target_logit', False),
})
if node.get('is_target_logit') and not target_token:
native_prob = node.get('token_prob', 0)
match = re.search(r'Output\s+"([^"]*)"', clerp)
target_token = (match.group(1).strip() if match
else clerp.strip().strip("'"))
logits.sort(key=lambda x: x['prob'], reverse=True)
top_logits = logits[:5]
answer_lower = answer.lower() if answer else ''
answer_in_logits = any(
t and len(t) > 1 and t.lower() in answer_lower
for t in (l.get('token', '') for l in top_logits)
)
return {
'logits': top_logits,
'native_prob': native_prob,
'target_token': target_token,
'capital_is_top_logit': (
target_token.lower() in answer_lower
if answer and target_token else False),
'capital_in_logits': answer_in_logits,
}
def _get_entity_logit_flags(self, slug: str,
answer: str) -> Dict[str, Any]:
"""Return compact warning flags for the matrix entity list."""
entity_dir = self._find_state_dir(slug)
if not entity_dir:
return {}
metadata = self._load_entity_logit_metadata(entity_dir, answer)
return {
'capital_is_top_logit': metadata.get('capital_is_top_logit'),
'capital_in_logits': metadata.get('capital_in_logits'),
}
def get_analysis(self) -> Dict[str, Any]:
"""Load analysis summary data."""
if self._analysis_cache is not None:
return self._analysis_cache
analysis = {}
analysis_dir = self.swaps_dir / "_analysis_v3"
# Load tier summary
tier_summary_path = analysis_dir / "tier_summary.json"
if tier_summary_path.exists():
with open(tier_summary_path, 'r', encoding='utf-8') as f:
analysis['tier_summary'] = json.load(f)
# Load swap factor summary
factor_path = analysis_dir / "swap_factor_summary.json"
if factor_path.exists():
with open(factor_path, 'r', encoding='utf-8') as f:
factor_data = json.load(f)
analysis['correlations'] = factor_data.get('correlations', {})
analysis['archetypes'] = factor_data.get('archetypes', {})
analysis['insights'] = factor_data.get('insights', [])
analysis['anomalies'] = factor_data.get('anomalies', {})
self._analysis_cache = analysis
return analysis
def get_swap_detail(
self, from_slug: str, to_slug: str, variant: Optional[str] = None,
) -> Optional[Dict[str, Any]]:
"""Load detailed swap result with variant resolution.
Resolution order:
1. Explicit *variant* requested -> load that variant file.
2. Canonical file exists -> load it.
3. Otherwise -> load the best variant for the pair.
"""
data = None
swap_path = None
if variant:
swap_path = (
self.swaps_dir / "by_source" / from_slug
/ f"to_{to_slug}__{variant}.json"
)
else:
canonical = self.swaps_dir / "by_source" / from_slug / f"to_{to_slug}.json"
if canonical.exists():
swap_path = canonical
else:
swap_path = self._best_variant_path(from_slug, to_slug)
if swap_path is None or not swap_path.exists():
work_path = self.swaps_dir / "work" / f"{from_slug}__to__{to_slug}.json"
if work_path.exists():
swap_path = work_path
if swap_path is None or not swap_path.exists():
return None
with open(swap_path, "r", encoding="utf-8") as f:
data = json.load(f)
# Enrich source/target with per-entity metadata
data['source']['neuronpedia_url'] = self._get_neuronpedia_url(from_slug)
data['target']['neuronpedia_url'] = self._get_neuronpedia_url(to_slug)
for key, slug in (('source', from_slug), ('target', to_slug)):
pct = self._get_error_node_influence(slug)
if pct is not None:
data[key]['error_node_influence_pct'] = pct
# Inject derived metrics block (additive, null-safe)
ctrl = data.get("metadata", {}).get("control", {})
traj_summary = data.get("evaluation", {}).get("logit_trajectory", {}).get("summary", {})
contrast = data.get("evaluation", {}).get("logit_trajectory", {}).get("contrast_groups", {})
pos0 = data.get("evaluation", {}).get("position_0_comparison", {})
regime = _classify_regime(pos0)
regime_info = _REGIME_INFO.get(regime) if regime else None
data["_derived"] = {
"control_mode": ctrl.get("control_mode"),
"fields_used": ctrl.get("concept_subsets_used"),
"replicate_id": ctrl.get("replicate_id"),
"flip_position": traj_summary.get("flip_position"),
"gap_closure": traj_summary.get("gap_closure"),
"initial_gap": traj_summary.get("initial_gap"),
"best_gap": traj_summary.get("best_gap"),
"regime": regime,
"regime_label": regime_info["label"] if regime_info else None,
"target_logit_delta": pos0.get("target_logit_delta"),
"source_logit_delta": pos0.get("source_logit_delta"),
"vs_max": None,
"vs_topk": None,
"rank_in_group": None,
"vs_max_pos0": None,
"vs_topk_pos0": None,
"rank_in_group_pos0": None,
}
if isinstance(contrast, dict):
grp = contrast.get("same_dataset", {})
if isinstance(grp, dict):
agg = grp.get("aggregate", {}) if isinstance(grp, dict) else {}
data["_derived"]["vs_max"] = agg.get("best_target_minus_max")
data["_derived"]["vs_topk"] = agg.get("best_target_minus_topk")
data["_derived"]["rank_in_group"] = agg.get("best_rank_within")
data["_derived"]["vs_max_pos0"] = agg.get("initial_target_minus_max")
data["_derived"]["vs_topk_pos0"] = agg.get("initial_target_minus_topk")
data["_derived"]["rank_in_group_pos0"] = agg.get("initial_rank_within")
data["_derived"]["contrast_members"] = grp.get("members", [])
data["_derived"]["contrast_n"] = grp.get("n_members", 0)
data["_derived"]["contrast_topk_k"] = grp.get("topk_k", 3)
# Variant availability summary
variants = self.get_variants_for_pair(from_slug, to_slug)
data["_variants"] = variants
_, loaded_suffix = _parse_to_slug(swap_path.stem)
data["_loaded_variant"] = loaded_suffix or None
return data
def save_annotation(self, from_slug: str, to_slug: str,
tier: Optional[int] = None,
notes: Optional[str] = None) -> Dict[str, Any]:
"""
Save annotation (tier/notes) to the swap JSON file.
Args:
from_slug: Source state slug
to_slug: Target state slug
tier: New tier value (1-5), or None to keep existing
notes: Annotation notes, or None to keep existing
Returns:
Updated swap data with new stats
"""
# Find the swap file
swap_path = self.swaps_dir / "by_source" / from_slug / f"to_{to_slug}.json"
if not swap_path.exists():
swap_path = self.swaps_dir / "work" / f"{from_slug}__to__{to_slug}.json"
if not swap_path.exists():
raise FileNotFoundError(f"Swap file not found: {from_slug} -> {to_slug}")
# Load existing data
with open(swap_path, 'r', encoding='utf-8') as f:
data = json.load(f)
# Ensure classification section exists
if 'classification' not in data:
data['classification'] = {}
# Update tier if provided
if tier is not None:
old_tier = data['classification'].get('tier')
data['classification']['tier'] = tier
data['classification']['manually_edited'] = True
# Update notes if provided
if notes is not None:
data['classification']['notes'] = notes
data['classification']['manually_edited'] = True
# Save back to file
with open(swap_path, 'w', encoding='utf-8') as f:
json.dump(data, f, indent=2, ensure_ascii=False)
# Clear caches so stats are recalculated
self._matrix_cache = {}
self._flip_matrix_cache = {}
self._regime_matrix_cache = {}
self._vsmax_matrix_cache = {}
self._stats_cache = None
# Return updated data with fresh stats
return {
'swap': data,
'stats': self.get_stats(),
'matrix_update': {
'from': from_slug,
'to': to_slug,
'tier': data['classification'].get('tier'),
'manually_edited': True,
}
}
def get_annotated_swaps(self) -> List[Dict[str, str]]:
"""Get list of all manually annotated swaps."""
annotated = []
by_source_dir = self.swaps_dir / "by_source"
if by_source_dir.exists():
for source_dir in by_source_dir.iterdir():
if not source_dir.is_dir():
continue
from_slug = source_dir.name
for swap_file in source_dir.glob("to_*.json"):
if _is_control_variant(swap_file.stem):
continue
try:
with open(swap_file, 'r', encoding='utf-8') as f:
data = json.load(f)
if data.get('classification', {}).get('manually_edited'):
to_slug, _ = _parse_to_slug(swap_file.stem)
annotated.append({
'from': from_slug,
'to': to_slug,
'tier': data['classification'].get('tier'),
'notes': data['classification'].get('notes', ''),
})
except (json.JSONDecodeError, IOError):
continue
return annotated
def get_stats(self) -> Dict[str, Any]:
"""Get aggregate statistics - computed dynamically from matrix."""
if self._stats_cache is not None:
return self._stats_cache
matrix = self.get_matrix()
analysis = self.get_analysis()
# Count tiers from actual matrix data
tier_counts = {
'PERFECT': 0, # 5
'TARGET_STATE_CITY': 0, # 4
'TARGET_STATE_ONLY': 0, # 3
'SUPPRESSED_ONLY': 0, # 2
'SOURCE_PERSISTS': 0, # 1
}
tier_names = {5: 'PERFECT', 4: 'TARGET_STATE_CITY', 3: 'TARGET_STATE_ONLY',
2: 'SUPPRESSED_ONLY', 1: 'SOURCE_PERSISTS'}
total_swaps = 0
tier_sum = 0
perfect_count = 0
state_correct_count = 0 # T3+
suppressed_count = 0 # T2+
for from_slug, targets in matrix.items():
for to_slug, tier in targets.items():
if tier is not None and from_slug != to_slug:
total_swaps += 1
tier_sum += tier
tier_name = tier_names.get(tier, 'UNKNOWN')
if tier_name in tier_counts:
tier_counts[tier_name] += 1
if tier == 5:
perfect_count += 1
if tier >= 3:
state_correct_count += 1
if tier >= 2:
suppressed_count += 1
# Calculate rates
tier_rates = {}
for name, count in tier_counts.items():
tier_rates[name] = count / total_swaps if total_swaps > 0 else 0
# Flip stats from trajectory data
flip_matrix = self.get_flip_matrix()
flip_tracked = 0
flip_at_0 = 0
flip_at_01 = 0
for from_slug, targets in flip_matrix.items():
for to_slug, fp in targets.items():
if fp == -1:
continue
flip_tracked += 1
if fp is not None and fp <= 1:
flip_at_01 += 1
if fp is not None and fp == 0:
flip_at_0 += 1
aggregate = {
'perfect_rate': perfect_count / total_swaps if total_swaps > 0 else 0,
'state_correct_rate': state_correct_count / total_swaps if total_swaps > 0 else 0,
'suppression_rate': suppressed_count / total_swaps if total_swaps > 0 else 0,
'avg_tier': tier_sum / total_swaps if total_swaps > 0 else 0,
'flip_at_01_rate': flip_at_01 / flip_tracked if flip_tracked > 0 else 0,
'flip_at_0_rate': flip_at_0 / flip_tracked if flip_tracked > 0 else 0,
'flip_tracked': flip_tracked,
}
stats = {
'total_swaps': total_swaps,
'tier_counts': tier_counts,
'tier_rates': tier_rates,
'aggregate': aggregate,
'correlations': analysis.get('correlations', {}),
'insights': analysis.get('insights', []),
}
self._stats_cache = stats
return stats
def get_methodology_stats(self) -> Dict[str, Any]:
"""Compute dynamic stats for the About / methodology panel.
Scans the current run's swap files for regime distribution,
aggregate vsMax, and target recovery rates. Results are cached.
"""
cache_key = "_methodology_stats"
if hasattr(self, cache_key) and getattr(self, cache_key) is not None:
return getattr(self, cache_key)
stats = self.get_stats()
dc = self.get_domain_config()
total_swaps = stats.get("total_swaps", 0)
aggregate = stats.get("aggregate", {})
regime_counts: Dict[str, int] = {}
vs_max_values: list = []
recovery_count = 0
recovery_total = 0
by_source = self.swaps_dir / "by_source"
if by_source.exists():
for swap_path in by_source.glob("*/to_*.json"):
if _is_control_variant(swap_path.stem):
continue
try:
with open(swap_path, "r", encoding="utf-8") as fh:
d = json.load(fh)
except (json.JSONDecodeError, IOError):
continue
pos0 = d.get("evaluation", {}).get("position_0_comparison", {})
regime = _classify_regime(pos0)
if regime:
regime_counts[regime] = regime_counts.get(regime, 0) + 1
contrast = (
d.get("evaluation", {})
.get("logit_trajectory", {})
.get("contrast_groups", {})
.get("same_dataset", {})
.get("aggregate", {})
)
vm = contrast.get("best_target_minus_max")
if vm is not None:
vs_max_values.append(vm)
traj_sum = d.get("evaluation", {}).get("logit_trajectory", {}).get("summary", {})
baseline = d.get("evaluation", {}).get("baseline_logits", {})
if traj_sum and baseline:
recovery_total += 1
tgt_baseline_logit = baseline.get("target", {}).get("logit")
best_gap = traj_sum.get("best_gap")
initial_gap = traj_sum.get("initial_gap")
if tgt_baseline_logit is not None and best_gap is not None:
recovery_count += 1
regime_total = sum(regime_counts.values()) or 1
result = {
"display_name": dc.get("display_name", ""),
"total_swaps": total_swaps,
"entity_count": dc.get("entity_count", 0),
"model_id": dc.get("model_id", ""),
"concept_fields": dc.get("concept_fields", []),
"answer_field": dc.get("answer_field", ""),
"m_ablate": dc.get("m_ablate", -2),
"m_amplify": dc.get("m_amplify", 20),
"n_tokens": dc.get("n_tokens", 10),
"temperature": dc.get("temperature", 0.3),
"hit_rate": aggregate.get("perfect_rate", 0),
"suppression_rate": aggregate.get("suppression_rate", 0),
"flip_at_01_rate": aggregate.get("flip_at_01_rate", 0),
"avg_tier": aggregate.get("avg_tier", 0),
"vs_max_mean": (
sum(vs_max_values) / len(vs_max_values)
if vs_max_values else None
),
"vs_max_positive_pct": (
sum(1 for v in vs_max_values if v > 0) / len(vs_max_values)
if vs_max_values else None
),
"regime_counts": regime_counts,
"regime_pcts": {
k: round(v / regime_total * 100, 1)
for k, v in regime_counts.items()
},
"regime_labels": {k: v["label"] for k, v in _REGIME_INFO.items()},
}
setattr(self, cache_key, result)
return result
def _get_gpu_batch_features_index(self) -> Dict[str, List]:
"""
Build and cache a swap_id -> feature_list index from GPU batch files.
Scans _swaps/_work/_gpu_batch_*/features.json which use the format:
{"global": [...], "per_prompt": {"swap_id": [...features], ...}}
"""
if self._gpu_batch_features_index is not None:
return self._gpu_batch_features_index
index: Dict[str, List] = {}
work_dir = self.base_swaps_dir / "_work"
if not work_dir.exists():
self._gpu_batch_features_index = index
return index
for batch_file in sorted(work_dir.glob("_gpu_batch_*/features.json")):
try:
with open(batch_file, 'r', encoding='utf-8') as f:
data = json.load(f)
for swap_id, feats in data.get('per_prompt', {}).items():
if isinstance(feats, list):
index[swap_id] = feats
except (json.JSONDecodeError, IOError):
pass
self._gpu_batch_features_index = index
return index
def _resolve_features_path(
self, from_slug: str, to_slug: str, variant: Optional[str] = None,
) -> Optional[Path]:
"""Find the features.json path for a swap, handling variant suffixes.
Resolution order:
1. Explicit variant -> ``work/{swap_id}__{variant}/features.json``
2. Base path (labeled runs) -> ``work/{swap_id}/features.json``
3. Best variant (mirroring ``get_swap_detail``) work dir.
"""
base_swap_id = f"{from_slug}__to__{to_slug}"
work_dir = self.swaps_dir / "work"
if variant:
p = work_dir / f"{base_swap_id}__{variant}" / "features.json"
return p if p.exists() else None
# Try the base path first (works for labeled runs with no suffix)
p = work_dir / base_swap_id / "features.json"
if p.exists():
return p
# Resolve via best variant (same logic as get_swap_detail)
best_path = self._best_variant_path(from_slug, to_slug)
if best_path is not None:
_, variant_suffix = _parse_to_slug(best_path.stem)
if variant_suffix:
p = work_dir / f"{base_swap_id}__{variant_suffix}" / "features.json"
if p.exists():
return p
return None
def get_swap_features(
self, from_slug: str, to_slug: str, variant: Optional[str] = None,
) -> Optional[Dict]:
"""
Get intervention features for a swap.
Tries locations in order:
1. Per-swap file via variant resolution (see ``_resolve_features_path``)
2. GPU batch index: _swaps/_work/_gpu_batch_N/features.json (per_prompt keyed)
Returns structured data with ablated/amplified features grouped by layer,
enriched with supernode_name from each entity's node_grouping.csv.
Also returns source/target grouping lists with ablated/amplified flags.
"""
swap_id = f"{from_slug}__to__{to_slug}"
raw_features: Optional[List] = None
features_path = self._resolve_features_path(from_slug, to_slug, variant)
if features_path is not None:
try:
with open(features_path, 'r', encoding='utf-8') as f:
raw_features = json.load(f)
except (json.JSONDecodeError, IOError):
pass
# Fall back to GPU batch index
if raw_features is None:
index = self._get_gpu_batch_features_index()
raw_features = index.get(swap_id)
if raw_features is None:
return None
np_model = self.get_domain_config().get('np_model', 'gemma-2-2b')
def _load_grouping_map(slug: str) -> Dict:
"""Return {(layer, index): supernode_name} for an entity."""
entity_dir = self._find_state_dir(slug)
if not entity_dir:
return {}
grouping_path = entity_dir / "02 Node Grouping" / "node_grouping.csv"
if not grouping_path.exists():
return {}
result = {}
try:
with open(grouping_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
try:
key = (int(row.get('layer', 0)), int(row.get('feature', 0)))
result[key] = row.get('supernode_name', '')
except (ValueError, TypeError):
pass
except IOError:
pass
return result
def _load_all_groupings(slug: str) -> List[Dict[str, Any]]:
"""Return sorted groupings with total feature counts for an entity."""
entity_dir = self._find_state_dir(slug)
if not entity_dir:
return []
grouping_path = entity_dir / "02 Node Grouping" / "node_grouping.csv"
if not grouping_path.exists():
return []
grouping_counts: Dict[str, int] = {}
try:
with open(grouping_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
sn = row.get('supernode_name', '').strip()
if sn:
grouping_counts[sn] = grouping_counts.get(sn, 0) + 1
except IOError:
pass
return [
{'name': name, 'feature_count': grouping_counts[name]}
for name in sorted(grouping_counts)
]
source_map = _load_grouping_map(from_slug)
target_map = _load_grouping_map(to_slug)
all_source_groupings = _load_all_groupings(from_slug)
all_target_groupings = _load_all_groupings(to_slug)
# Group features by type and layer
ablated = []
amplified = []
layer_counts: Dict = {'ablated': {}, 'amplified': {}}
ablated_grouping_names: set = set()
amplified_grouping_names: set = set()
for feat in raw_features:
layer = feat.get('layer', 0)
index = feat.get('index', 0)
M = feat.get('M', 0)
stored_activation = feat.get('stored_activation')
np_url = f"https://www.neuronpedia.org/{np_model}/{layer}-clt-hp/{index}"
if M < 0:
sn = source_map.get((layer, index), '')
feature_info = {
'layer': layer,
'index': index,
'M': M,
'stored_activation': stored_activation,
'neuronpedia_url': np_url,
'supernode_name': sn,
}
ablated.append(feature_info)
layer_counts['ablated'][layer] = layer_counts['ablated'].get(layer, 0) + 1
if sn:
ablated_grouping_names.add(sn)
else:
sn = target_map.get((layer, index), '')
feature_info = {
'layer': layer,
'index': index,
'M': M,
'stored_activation': stored_activation,
'neuronpedia_url': np_url,
'supernode_name': sn,
}
amplified.append(feature_info)
layer_counts['amplified'][layer] = layer_counts['amplified'].get(layer, 0) + 1
if sn:
amplified_grouping_names.add(sn)
ablated.sort(key=lambda x: (x['layer'], x['index']))
amplified.sort(key=lambda x: (x['layer'], x['index']))
source_groupings = [
{
'name': grouping['name'],
'feature_count': grouping['feature_count'],
'ablated': grouping['name'] in ablated_grouping_names,
}
for grouping in all_source_groupings
]
target_groupings = [
{
'name': grouping['name'],
'feature_count': grouping['feature_count'],
'amplified': grouping['name'] in amplified_grouping_names,
}
for grouping in all_target_groupings
]
return {
'swap_id': swap_id,
'ablated': ablated,
'amplified': amplified,
'layer_counts': layer_counts,
'summary': {
'ablate_count': len(ablated),
'amplify_count': len(amplified),
'total_count': len(ablated) + len(amplified),
},
'source_groupings': source_groupings,
'target_groupings': target_groupings,
}
def get_state_profile(self, slug: str) -> Optional[Dict]:
"""Get comprehensive entity profile with stats (domain-agnostic)."""
dc = self.get_domain_config()
entity_index = dc['entity_index']
edata = entity_index.get(slug, {})
primary_field = dc['primary_field']
answer_field = dc['answer_field']
label = edata.get(primary_field, '') or self._slug_to_label(slug)
answer = edata.get(answer_field, '')
entity_dir = self._find_state_dir(slug)
profile: Dict[str, Any] = {
'slug': slug,
'label': label,
'answer': answer,
'state': edata.get('state', label),
'city': edata.get('city', edata.get(primary_field, label)),
'capital': edata.get('capital', answer),
'fields': {k: v for k, v in edata.items() if k != 'slug'},
}
if entity_dir:
manifest = self._load_manifest(entity_dir)
if manifest:
np_data = manifest.get('neuronpedia', {})
profile['supernodes'] = np_data.get('supernodes', 0)
profile['pinned_nodes'] = np_data.get('pinned_nodes', 0)
profile['neuronpedia_url'] = np_data.get('url', '')
gq = manifest.get('graph_quality', {})
pct = gq.get('error_node_influence_pct')
if pct is not None:
profile['error_node_influence_pct'] = pct
if 'error_node_influence_pct' not in profile:
pct = self._get_error_node_influence(slug)
if pct is not None:
profile['error_node_influence_pct'] = pct
profile.update(
self._load_entity_logit_metadata(entity_dir, answer))
features_path = (entity_dir / "00 Graph Generation"
/ "selected_features_with_nodes.json")
if features_path.exists():
try:
with open(features_path, 'r', encoding='utf-8') as f:
features_data = json.load(f)
features = features_data.get('features', [])
profile['total_features'] = len(features)
layer_counts: Dict[int, int] = {}
for feat in features:
layer = feat.get('layer', 0)
layer_counts[layer] = layer_counts.get(layer, 0) + 1
profile['feature_layers'] = layer_counts
except (json.JSONDecodeError, IOError):
pass
grouping_path = (entity_dir / "02 Node Grouping"
/ "node_grouping.csv")
if grouping_path.exists():
try:
match_terms: set = set()
for cf in dc['concept_fields']:
val = edata.get(cf, '')
if val:
match_terms.add(val.lower())
if label:
match_terms.add(label.lower())
supernode_features = []
supernode_layer_counts: Dict[int, int] = {}
with open(grouping_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
sn = row.get('supernode_name', '').lower()
if any(t in sn or sn in t
for t in match_terms):
layer = int(row.get('layer', 0))
supernode_features.append({
'layer': layer,
'feature': row.get('feature', ''),
'supernode_name': row.get(
'supernode_name', ''),
})
supernode_layer_counts[layer] = (
supernode_layer_counts.get(layer, 0) + 1)
profile['supernode_features'] = supernode_features
profile['supernode_feature_count'] = len(
supernode_features)
profile['supernode_layer_counts'] = supernode_layer_counts
except (IOError, ValueError):
pass
# Attack/defense scores from matrix
matrix = self.get_matrix()
attack_tiers = []
if slug in matrix:
for to_slug, tier in matrix[slug].items():
if tier is not None and to_slug != slug:
attack_tiers.append(tier)
profile['attack_avg'] = (sum(attack_tiers) / len(attack_tiers)
if attack_tiers else 0)
profile['attack_count'] = len(attack_tiers)
attack_ok = len([t for t in attack_tiers if t >= 3])
profile['attack_success_rate'] = (attack_ok / len(attack_tiers)
if attack_tiers else 0)
defense_tiers = []
wrong_count = 0
for from_slug, targets in matrix.items():
if (from_slug != slug and slug in targets
and targets[slug] is not None):
tier = targets[slug]
defense_tiers.append(tier)
if tier == 2.5:
wrong_count += 1
profile['defense_avg'] = (sum(defense_tiers) / len(defense_tiers)
if defense_tiers else 0)
profile['defense_count'] = len(defense_tiers)
defense_ok = len([t for t in defense_tiers if t >= 3])
profile['defense_success_rate'] = (defense_ok / len(defense_tiers)
if defense_tiers else 0)
profile['wrong_state_rate'] = (wrong_count / len(defense_tiers)
if defense_tiers else 0)
profile['wrong_state_count'] = wrong_count
profile['has_token_overlap'] = False
if dc['is_usa_states']:
overlap_slugs = [
'colorado_colorado_springs', 'new_york_new_york_city',
'virginia_virginia_beach', 'idaho_idaho_falls',
'missouri_kansas_city', 'indiana_fort_wayne',
]
profile['has_token_overlap'] = slug in overlap_slugs
profile['swaps_as_target'] = self._get_swap_summaries_as_target(slug)
profile['swaps_as_source'] = self._get_swap_summaries_as_source(slug)
return profile
def _get_swap_summaries_as_target(self, slug: str) -> List[Dict]:
"""Swaps where this entity is the target."""
matrix = self.get_matrix()
entity_map = {s['slug']: s for s in self.get_states()}
summaries = []
for from_slug, targets in matrix.items():
if from_slug != slug and slug in targets and targets[slug] is not None:
tier = targets[slug]
e = entity_map.get(from_slug, {})
summaries.append({
'from_slug': from_slug,
'from_state': e.get('label', e.get('state', from_slug)),
'from_city': e.get('city', ''),
'tier': tier,
})
summaries.sort(key=lambda x: -x['tier'])
return summaries
def _get_swap_summaries_as_source(self, slug: str) -> List[Dict]:
"""Swaps where this entity is the source."""
matrix = self.get_matrix()
entity_map = {s['slug']: s for s in self.get_states()}
summaries = []
if slug in matrix:
for to_slug, tier in matrix[slug].items():
if to_slug != slug and tier is not None:
e = entity_map.get(to_slug, {})
summaries.append({
'to_slug': to_slug,
'to_state': e.get('label', e.get('state', to_slug)),
'to_city': e.get('city', ''),
'tier': tier,
})
summaries.sort(key=lambda x: -x['tier'])
return summaries
def _get_error_node_influence(self, slug: str) -> Optional[float]:
"""Return error_node_influence_pct for an entity, with CSV fallback."""
entity_dir = self._find_state_dir(slug)
if not entity_dir:
return None
manifest = self._load_manifest(entity_dir)
if manifest:
gq = manifest.get('graph_quality', {})
pct = gq.get('error_node_influence_pct')
if pct is not None:
return pct
# Fallback: compute from graph_feature_static_metrics.csv
csv_path = (entity_dir / "00 Graph Generation"
/ "graph_feature_static_metrics.csv")
if not csv_path.exists():
return None
try:
total = 0.0
error = 0.0
with open(csv_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
ni = float(row.get('node_influence', 0) or 0)
total += ni
feat = row.get('feature', '')
if feat == '-1':
error += ni
if total > 0:
return round(error / total * 100, 2)
except (IOError, ValueError):
pass
return None
def _find_state_dir(self, slug: str) -> Optional[Path]:
"""Find the state directory, handling case variations."""
state_dir = self.data_dir / slug
if state_dir.exists():
return state_dir
# Try case variations
for candidate in self.data_dir.iterdir():
if candidate.is_dir() and candidate.name.lower().replace(' ', '_') == slug.lower():
return candidate
return None
def _load_manifest(self, state_dir: Path) -> Optional[Dict]:
"""Load manifest.json from state directory."""
manifest_path = state_dir / "manifest.json"
if manifest_path.exists():
try:
with open(manifest_path, 'r', encoding='utf-8') as f:
return json.load(f)
except (json.JSONDecodeError, IOError):
pass
return None
@staticmethod
def _add_query_params(url: str, params: Dict[str, str], overwrite: bool = False) -> str:
"""
Add query params to a URL, preserving existing params.
Args:
url: Base URL
params: Params to add
overwrite: If True, overwrite existing keys. If False, keep existing keys.
"""
try:
parts = urlsplit(url)
except Exception:
return url
existing = parse_qsl(parts.query, keep_blank_values=True)
query_map = {}
for k, v in existing:
# Keep the first value for each key (Neuronpedia URLs typically do not repeat keys)
if k not in query_map:
query_map[k] = v
for k, v in (params or {}).items():
if k in query_map and not overwrite:
continue
query_map[k] = v
new_query = urlencode(query_map, doseq=False)
return urlunsplit((parts.scheme, parts.netloc, parts.path, new_query, parts.fragment))
def get_complete_subgraph_url(self, slug: str) -> Optional[Dict[str, Any]]:
"""
Return the Neuronpedia "complete subgraph" URL using the uploaded subgraph ID.
This uses neuronpedia.subgraph_id from manifest.json and appends it as the
`subgraph` query parameter to the manifest neuronpedia.url.
"""
state_dir = self._find_state_dir(slug)
if not state_dir:
return None
manifest = self._load_manifest(state_dir)
if not manifest:
return None
np_data = manifest.get('neuronpedia', {}) or {}
base_url = (np_data.get('url') or '').strip()
subgraph_id = (np_data.get('subgraph_id') or '').strip()
if not base_url or not subgraph_id:
return None
full_url = self._add_query_params(base_url, {"subgraph": subgraph_id}, overwrite=False)
return {
"url": full_url,
"base_url": base_url,
"subgraph_id": subgraph_id,
"mode": "complete",
}
def _get_neuronpedia_url(self, slug: str) -> Optional[str]:
"""Get Neuronpedia URL for a state slug."""
# Try exact match first
state_dir = self.data_dir / slug
if state_dir.exists():
manifest = self._load_manifest(state_dir)
if manifest:
return manifest.get('neuronpedia', {}).get('url')
# Try case variations
for candidate in self.data_dir.iterdir():
if candidate.is_dir() and candidate.name.lower().replace(' ', '_') == slug.lower():
manifest = self._load_manifest(candidate)
if manifest:
return manifest.get('neuronpedia', {}).get('url')
return None
def _slug_to_state_name(self, slug: str) -> str:
"""Convert slug to state name."""
# arizona_tucson -> Arizona
parts = slug.split('_')
if len(parts) >= 1:
state = parts[0].replace('_', ' ').title()
# Handle two-word states
if len(parts) >= 2 and parts[0].lower() in ['new', 'north', 'south', 'west', 'rhode']:
state = f"{parts[0].title()} {parts[1].title()}"
return state
return slug.title()
def _slug_to_city(self, slug: str) -> str:
"""Extract city from slug."""
# arizona_tucson -> Tucson
parts = slug.split('_')
state_parts = 1
if parts[0].lower() in ['new', 'north', 'south', 'west', 'rhode']:
state_parts = 2
city_parts = parts[state_parts:]
return ' '.join(p.title() for p in city_parts)
def _state_to_abbr(self, state_name: str) -> str:
"""Convert state name to abbreviation."""
abbrs = {
'Alabama': 'AL', 'Alaska': 'AK', 'Arizona': 'AZ', 'Arkansas': 'AR',
'California': 'CA', 'Colorado': 'CO', 'Connecticut': 'CT', 'Delaware': 'DE',
'Florida': 'FL', 'Georgia': 'GA', 'Hawaii': 'HI', 'Idaho': 'ID',
'Illinois': 'IL', 'Indiana': 'IN', 'Iowa': 'IA', 'Kansas': 'KS',
'Kentucky': 'KY', 'Louisiana': 'LA', 'Maine': 'ME', 'Maryland': 'MD',
'Massachusetts': 'MA', 'Michigan': 'MI', 'Minnesota': 'MN', 'Mississippi': 'MS',
'Missouri': 'MO', 'Montana': 'MT', 'Nebraska': 'NE', 'Nevada': 'NV',
'New Hampshire': 'NH', 'New Jersey': 'NJ', 'New Mexico': 'NM', 'New York': 'NY',
'North Carolina': 'NC', 'North Dakota': 'ND', 'Ohio': 'OH', 'Oklahoma': 'OK',
'Oregon': 'OR', 'Pennsylvania': 'PA', 'Rhode Island': 'RI', 'South Carolina': 'SC',
'South Dakota': 'SD', 'Tennessee': 'TN', 'Texas': 'TX', 'Utah': 'UT',
'Vermont': 'VT', 'Virginia': 'VA', 'Washington': 'WA', 'West Virginia': 'WV',
'Wisconsin': 'WI', 'Wyoming': 'WY'
}
return abbrs.get(state_name, state_name[:2].upper())
def _get_state_capital(self, state_name: str) -> str:
"""Get the capital city for a state."""
capitals = {
'Alabama': 'Montgomery', 'Alaska': 'Juneau', 'Arizona': 'Phoenix', 'Arkansas': 'Little Rock',
'California': 'Sacramento', 'Colorado': 'Denver', 'Connecticut': 'Hartford', 'Delaware': 'Dover',
'Florida': 'Tallahassee', 'Georgia': 'Atlanta', 'Hawaii': 'Honolulu', 'Idaho': 'Boise',
'Illinois': 'Springfield', 'Indiana': 'Indianapolis', 'Iowa': 'Des Moines', 'Kansas': 'Topeka',
'Kentucky': 'Frankfort', 'Louisiana': 'Baton Rouge', 'Maine': 'Augusta', 'Maryland': 'Annapolis',
'Massachusetts': 'Boston', 'Michigan': 'Lansing', 'Minnesota': 'Saint Paul', 'Mississippi': 'Jackson',
'Missouri': 'Jefferson City', 'Montana': 'Helena', 'Nebraska': 'Lincoln', 'Nevada': 'Carson City',
'New Hampshire': 'Concord', 'New Jersey': 'Trenton', 'New Mexico': 'Santa Fe', 'New York': 'Albany',
'North Carolina': 'Raleigh', 'North Dakota': 'Bismarck', 'Ohio': 'Columbus', 'Oklahoma': 'Oklahoma City',
'Oregon': 'Salem', 'Pennsylvania': 'Harrisburg', 'Rhode Island': 'Providence', 'South Carolina': 'Columbia',
'South Dakota': 'Pierre', 'Tennessee': 'Nashville', 'Texas': 'Austin', 'Utah': 'Salt Lake City',
'Vermont': 'Montpelier', 'Virginia': 'Richmond', 'Washington': 'Olympia', 'West Virginia': 'Charleston',
'Wisconsin': 'Madison', 'Wyoming': 'Cheyenne'
}
return capitals.get(state_name, '')
def get_simplified_subgraph_url(self, slug: str, max_features: int = 80,
max_url_length: int = 4000) -> Optional[Dict]:
"""
Generate a simplified subgraph URL with important features pinned.
Priority order for pinning:
1. Capital logit (layer 27) - always included
2. City embedding (layer -1, city token) - always included
3. State-related features - always included if space
4. Remaining features sorted by node_influence
Args:
slug: State slug (e.g., 'alabama_Birmingham')
max_features: Maximum features to include (default 80)
max_url_length: Maximum URL length (default 4000 chars)
Returns:
Dict with 'url', 'feature_count', 'supernode_count', 'url_length'
or None if data not found
"""
import csv
from urllib.parse import quote
state_dir = self._find_state_dir(slug)
if not state_dir:
return None
# Load manifest for base URL
manifest = self._load_manifest(state_dir)
if not manifest or 'neuronpedia' not in manifest:
return None
base_url = manifest['neuronpedia'].get('url', '')
if not base_url:
return None
dc = self.get_domain_config()
edata = dc['entity_index'].get(slug, {})
state_name = edata.get('state', self._slug_to_state_name(slug))
city_name = edata.get('city', self._slug_to_city(slug))
capital_name = edata.get('capital',
edata.get(dc['answer_field'], '')
or self._get_state_capital(state_name))
# Load graph.json to get the TOP logit by probability
# The metrics CSV has incomplete/empty data for layer 27
graph_path = state_dir / "00 Graph Generation" / "graph.json"
top_logit_node = None # Will store the full node dict for the top logit
top_logit_prob = 0
if graph_path.exists():
try:
with open(graph_path, 'r', encoding='utf-8') as f:
graph_data = json.load(f)
for node in graph_data.get('nodes', []):
clerp = node.get('clerp', '')
prob = node.get('token_prob', 0)
node_id = node.get('node_id', '')
# Output logits have format: "Output \" Token\" (p=0.123)"
if clerp.startswith('Output ') and prob and prob > top_logit_prob:
top_logit_prob = prob
# Build a node dict matching our format
top_logit_node = {
'node_id': node_id, # Use node_id directly from graph.json
'layer': '27',
'feature_id': node_id.split('_')[1] if '_' in node_id else '',
'ctx_idx': '8',
'token': clerp,
'influence': prob, # Use prob as influence
'supernode': 'output_logit',
}
except (json.JSONDecodeError, IOError):
pass
# Load graph metrics (has node_influence)
metrics_path = state_dir / "00 Graph Generation" / "graph_feature_static_metrics.csv"
if not metrics_path.exists():
return None
# Load node grouping for supernode assignments
# Key on {layer}_{feature} only, since a feature can appear at multiple positions
# in the metrics file but node_grouping only has the peak position
grouping_path = state_dir / "02 Node Grouping" / "node_grouping.csv"
supernode_map = {} # {layer}_{feature} -> supernode_name
if grouping_path.exists():
try:
with open(grouping_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
layer = row.get('layer', '0')
feature = row.get('feature', '0')
key = f"{layer}_{feature}"
supernode_map[key] = row.get('supernode_name', '')
except (IOError, csv.Error):
pass
# Parse metrics file
embeddings = [] # Layer -1 (priority)
output_logits = [] # Layer 27 (priority)
priority_nodes = [] # State/city related
other_nodes = [] # Everything else
all_supernode_nodes = {} # supernode_name -> list[node] (features only; excludes embeddings/logits)
try:
with open(metrics_path, 'r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
layer = row.get('layer', '0')
feature = row.get('feature', '0')
ctx_idx = row.get('ctx_idx', '0')
token = row.get('token', '').strip()
feature_id = row.get('id', feature) # For embeddings
# Skip residual connections (feature=-1 in layers > -1)
if feature == '-1' and layer != '-1':
continue
# Parse node_influence for sorting
try:
influence = float(row.get('node_influence', 0) or 0)
except (ValueError, TypeError):
influence = 0
# Build node_id based on layer type
# IMPORTANT: Use feature_id (the 'id' column) for ALL node IDs
# The 'feature' column contains internal IDs that Neuronpedia doesn't recognize
if layer == '-1':
# Embedding: E_{id}_{ctx_idx}
node_id = f"E_{feature_id}_{ctx_idx}"
else:
# Regular feature: {layer}_{id}_{ctx_idx}
node_id = f"{layer}_{feature_id}_{ctx_idx}"
# Get supernode from grouping or infer from token
# Key on {layer}_{feature_id} without position to match all occurrences
grouping_key = f"{layer}_{feature_id}"
supernode = supernode_map.get(grouping_key, '')
# For embeddings (layer -1), infer supernode from token
if layer == '-1' and not supernode:
clean_token = token.strip()
if clean_token:
# Use cleaned token as supernode name
supernode = clean_token.replace('<', '').replace('>', '')
# For output logits (layer 27), assign to output group
if layer == '27' and not supernode:
supernode = 'output_logit'
node = {
'node_id': node_id,
'layer': layer,
'feature': feature,
'feature_id': feature_id,
'ctx_idx': ctx_idx,
'token': token,
'influence': influence,
'supernode': supernode,
}
# Track full membership for each supernode (features only)
if supernode and layer not in ('-1', '27'):
if supernode not in all_supernode_nodes:
all_supernode_nodes[supernode] = []
all_supernode_nodes[supernode].append(node)
# Categorize by priority
if layer == '-1':
# Exclude functional embeddings (bos, The, of, the, etc.)
excluded_embeddings = {'bos', 'The', 'the', 'of', 'a', 'an', 'is'}
clean_token = token.strip()
if clean_token in excluded_embeddings or clean_token.replace('<', '').replace('>', '') in excluded_embeddings:
continue # Skip this embedding
embeddings.append(node)
# Check if this is city or state embedding
if city_name.lower() in token.lower():
node['priority'] = 'city'
elif state_name.lower() in token.lower():
node['priority'] = 'state'
elif layer == '27':
output_logits.append(node)
elif (city_name.lower() in token.lower() or
state_name.lower() in token.lower() or
(capital_name and capital_name.lower() in token.lower()) or
'capital' in supernode.lower() or
state_name.lower() in supernode.lower() or
city_name.lower() in supernode.lower() or
(capital_name and capital_name.lower() in supernode.lower())):
priority_nodes.append(node)
else:
other_nodes.append(node)
except (IOError, csv.Error) as e:
return None
# Sort each category by node_influence (descending)
embeddings.sort(key=lambda x: -x['influence'])
priority_nodes.sort(key=lambda x: -x['influence'])
other_nodes.sort(key=lambda x: -x['influence'])
# Use the top logit from graph.json (highest probability)
# This is more reliable than the metrics CSV which has incomplete layer 27 data
top_logit = top_logit_node
# Identify supernodes whose NAME contains the state or capital name.
# For these, we try to include the ENTIRE supernode group (all member nodes),
# as long as there is capacity (max_features).
import re
def _norm_name(s: str) -> str:
s = (s or '').lower().replace('_', ' ')
s = re.sub(r'[^a-z0-9 ]+', ' ', s)
return ' '.join(s.split())
norm_state = _norm_name(state_name)
norm_capital = _norm_name(capital_name)
forced_supernodes = [] # [(priority, -max_influence, name)]
if norm_state or norm_capital:
for name, nodes in (all_supernode_nodes or {}).items():
if not name or len(nodes) < 2:
continue
norm_group = _norm_name(name)
match_capital = bool(norm_capital) and (norm_capital in norm_group)
match_state = bool(norm_state) and (norm_state in norm_group)
if not (match_capital or match_state):
continue
max_inf = 0
try:
max_inf = max((n.get('influence') or 0) for n in nodes)
except ValueError:
max_inf = 0
forced_priority = 0 if match_capital else 1
forced_supernodes.append((forced_priority, -max_inf, name))
forced_supernodes.sort()
# Build forced nodes (include whole matching supernode groups if they fit)
forced_nodes = []
forced_included = 0
forced_omitted = 0
forced_node_ids = set()
reserved = len(embeddings) + (1 if top_logit else 0)
remaining_forced = max_features - reserved
if remaining_forced > 0 and forced_supernodes:
for _, __, name in forced_supernodes:
group_nodes = all_supernode_nodes.get(name, [])
group_nodes_sorted = sorted(group_nodes, key=lambda x: -(x.get('influence') or 0))
unique_group = [n for n in group_nodes_sorted if n.get('node_id') and n['node_id'] not in forced_node_ids]
if len(unique_group) <= remaining_forced:
forced_nodes.extend(unique_group)
forced_node_ids.update(n['node_id'] for n in unique_group)
forced_included += 1
remaining_forced -= len(unique_group)
else:
forced_omitted += 1
# Build candidate nodes list (after reserving space for forced groups)
# 1. Semantic embeddings (layer -1) - always included
# 2. Top output logit (layer 27) - always included
# 3. Forced groups (state/capital-named supernodes) - include whole group if it fits
# 4. Priority nodes (state/city/capital related)
# 5. Fill remaining with other high-influence nodes
candidates = []
candidates.extend(priority_nodes)
remaining_slots = max_features - len(embeddings) - (1 if top_logit else 0) - len(forced_nodes) - len(priority_nodes)
if remaining_slots > 0:
candidates.extend(other_nodes[:remaining_slots])
# First pass: only keep candidates that belong to supernodes with 2+ members
temp_supernode_groups = {}
for node in candidates:
supernode = node.get('supernode', '')
if not supernode:
continue
if supernode not in temp_supernode_groups:
temp_supernode_groups[supernode] = []
temp_supernode_groups[supernode].append(node)
valid_supernodes = {k for k, v in temp_supernode_groups.items() if len(v) >= 2}
# Build final selected list (deduplicated by node_id, preserving priority order)
selected = []
seen_node_ids = set()
def _add_node(node: Dict[str, Any]) -> None:
node_id = node.get('node_id', '')
if not node_id or node_id in seen_node_ids:
return
seen_node_ids.add(node_id)
selected.append(node)
for node in embeddings:
_add_node(node)
if top_logit:
_add_node(top_logit)
for node in forced_nodes:
_add_node(node)
for node in candidates:
if node.get('supernode', '') in valid_supernodes:
_add_node(node)
# Enforce max_features as a hard cap (forced groups are added before candidates,
# so truncation only drops lower-priority tail nodes).
if len(selected) > max_features:
selected = selected[:max_features]
# Build pinnedIds and supernodes
pinned_ids = []
supernode_groups = {}
for node in selected:
pinned_ids.append(node['node_id'])
# Embeddings (layer -1) and output logits (layer 27) are pinned but NOT grouped
if node['layer'] in ('-1', '27'):
continue
supernode = node['supernode']
if supernode:
if supernode not in supernode_groups:
supernode_groups[supernode] = []
supernode_groups[supernode].append(node['node_id'])
# Filter supernodes to only groups with 2+ members
supernode_groups = {k: v for k, v in supernode_groups.items() if len(v) >= 2}
# Build supernodes JSON array
supernodes_array = []
for name, ids in sorted(supernode_groups.items()):
supernodes_array.append([name] + ids)
# Encode and build URL
pinned_param = ','.join(pinned_ids)
supernodes_json = json.dumps(supernodes_array, separators=(',', ':'))
separator = '&' if '?' in base_url else '?'
full_url = f"{base_url}{separator}pinnedIds={quote(pinned_param)}&supernodes={quote(supernodes_json)}"
# If URL too long, reduce features
while len(full_url) > max_url_length and len(selected) > 20:
# Remove lowest influence non-priority nodes
selected = selected[:-5]
pinned_ids = [n['node_id'] for n in selected]
supernode_groups = {}
for node in selected:
# Skip embeddings and logits for grouping
if node['layer'] in ('-1', '27'):
continue
if node['supernode']:
if node['supernode'] not in supernode_groups:
supernode_groups[node['supernode']] = []
supernode_groups[node['supernode']].append(node['node_id'])
# Filter to 2+ node groups only
supernode_groups = {k: v for k, v in supernode_groups.items() if len(v) >= 2}
supernodes_array = [[name] + ids for name, ids in sorted(supernode_groups.items())]
pinned_param = ','.join(pinned_ids)
supernodes_json = json.dumps(supernodes_array, separators=(',', ':'))
full_url = f"{base_url}{separator}pinnedIds={quote(pinned_param)}&supernodes={quote(supernodes_json)}"
return {
'url': full_url,
'base_url': base_url,
'feature_count': len(selected),
'supernode_count': len(supernode_groups),
'url_length': len(full_url),
'embeddings_count': len([n for n in selected if n['layer'] == '-1']),
'logits_count': len([n for n in selected if n['layer'] == '27']),
'priority_count': len([n for n in selected if n in priority_nodes]),
'forced_supernodes_included': forced_included,
'forced_supernodes_omitted': forced_omitted,
}
|