OUTPUT of run 4 of at most 100: ``` L0H2 stacked Q/K spectrum: [5.0883 4.7015 4.4796 4.4228 4.3022 4.145 4.0379 3.9839 3.8495 3.7937 3.6476 3.5694 3.4852 3.2257 2.9036 2.8416 2.6236 2.4409 2.3484 2.1949 1.9175 1.8722 1.6528 1.5108 1.4407 1.4151 1.2959 1.2137 1.1136 1.089 0.9981 0.9436 0.8861 0.8692 0.8258 0.7734 0.7053 0.6979 0.6678 0.6553 0.6213 0.5938 0.5758 0.5637 0.5248 0.5048 0.4956 0.476 ] L0 OV PRODUCTS H0: top_s=[8.074 3.3277 2.097 0.6725 0.5606 0.3153 0.2578 0.1926], top2_energy=0.9326, top2_input_overlap_QK32=0.8858 read [(162, -0.3592), (223, 0.3113), (311, 0.3029), (300, -0.294), (59, -0.2334), (351, -0.2276), (152, 0.2016), (178, -0.1997)] write [(222, -0.9217), (360, 0.2573), (254, 0.143), (25, -0.0994), (197, 0.0816), (103, -0.0705), (273, -0.067), (136, 0.0491)] H1: top_s=[5.6611 3.8272 1.0646 0.5023 0.4285 0.2701 0.2169 0.1664], top2_energy=0.9623, top2_input_overlap_QK32=0.9510 read [(162, -0.3438), (223, 0.3314), (300, -0.2948), (311, 0.248), (178, -0.2459), (59, -0.2304), (114, -0.1999), (152, 0.1835)] write [(222, -0.8721), (360, 0.3589), (103, -0.1802), (254, 0.1698), (25, -0.1215), (197, 0.0829), (136, 0.0666), (72, 0.0544)] H2: top_s=[0.7873 0.718 0.6827 0.6713 0.6378 0.6013 0.5692 0.5583], top2_energy=0.1760, top2_input_overlap_QK32=0.1426 read [(301, -0.4618), (59, 0.2464), (240, 0.2027), (40, -0.1828), (311, 0.1796), (103, -0.1778), (278, -0.1744), (4, 0.1699)] write [(210, 0.2205), (302, 0.2109), (347, -0.1602), (217, -0.1502), (39, 0.1303), (311, -0.1297), (299, 0.1286), (119, 0.1233)] H3: top_s=[1.0319 0.9651 0.6173 0.5046 0.4477 0.227 0.2017 0.1835], top2_energy=0.6400, top2_input_overlap_QK32=0.4414 read [(239, 0.3383), (360, 0.3223), (35, 0.3201), (180, -0.2328), (273, -0.1996), (320, 0.1901), (76, -0.1603), (279, 0.1587)] write [(360, 0.6568), (78, 0.4234), (103, -0.3748), (222, -0.2994), (254, 0.1448), (197, 0.1127), (307, -0.0935), (25, -0.0861)] H4: top_s=[3.3958 1.6424 1.0818 0.6073 0.379 0.3274 0.2314 0.2045], top2_energy=0.8745, top2_input_overlap_QK32=0.8987 read [(162, 0.3448), (223, -0.3362), (300, 0.3015), (311, -0.271), (59, 0.2592), (178, 0.2212), (319, 0.1843), (114, 0.1833)] write [(222, 0.8891), (360, -0.3487), (103, 0.1699), (254, -0.1412), (25, 0.0857), (197, -0.0684), (273, 0.0593), (136, -0.0432)] H5: top_s=[8.3316 3.605 1.6787 0.7676 0.5106 0.3909 0.3445 0.3052], top2_energy=0.9507, top2_input_overlap_QK32=0.9317 read [(223, -0.3183), (60, -0.2921), (311, -0.2893), (300, 0.2715), (162, 0.243), (246, 0.2315), (59, 0.208), (319, 0.2076)] write [(222, 0.6501), (78, 0.6419), (72, -0.3032), (360, 0.1722), (103, -0.088), (273, -0.0784), (25, -0.0756), (254, 0.0672)] Input top-2 subspace overlaps: [[1. 0.7903 0.014 0.2851 0.7882 0.7202] [0.7903 1. 0.0109 0.1098 0.8786 0.8835] [0.014 0.0109 1. 0.0671 0.0084 0.011 ] [0.2851 0.1098 0.0671 1. 0.1033 0.0597] [0.7882 0.8786 0.0084 0.1033 1. 0.8349] [0.7202 0.8835 0.011 0.0597 0.8349 1. ]] Output top-2 subspace overlaps: [[0. 0. 0. 0. 0. 0. ] [0. 0. 0. 0. 0. 0. ] [0. 0. 0.0001 0.0001 0. 0. ] [0. 0. 0.0001 0.0002 0. 0. ] [0. 0. 0. 0. 0. 0. ] [0. 0. 0. 0. 0. 0. ]] L0 MLP ROW GEOMETRY F nearest cosine percentiles [0.3033 0.3919 0.4239 0.4632 0.5537 0.8533 0.9778 0.9873] k=16: inertia=1108.373, sizes=[361 342 270 106 103 77 64 41 31 29 27 25 21 17 13 9] k=24: inertia=1077.770, sizes=[272 162 153 136 130 113 111 99 57 34 31 28 28 27 21 21 20 16 15 14 13 13 13 9] k=28: inertia=1063.817, sizes=[222 160 155 151 141 109 106 99 89 38 27 27 25 21 19 18 15 14 14 13 13 12 10 9 9 8 7 5] k=40: inertia=1027.594, sizes=[162 135 123 104 99 95 69 64 63 60 55 52 51 46 34 33 30 20 20 19 18 14 13 13 12 12 12 11 10 10 9 9 9 8 8 8 8 7 6 5] F_times_OV_H2 nearest cosine percentiles [0.4643 0.642 0.6856 0.737 0.7833 0.8506 0.9891 0.9926] k=16: inertia=982.551, sizes=[143 132 112 112 111 104 102 90 89 88 84 79 78 77 75 60] k=24: inertia=929.819, sizes=[100 90 86 78 78 78 77 72 70 64 64 63 59 59 58 57 55 55 52 51 47 47 44 32] k=28: inertia=906.507, sizes=[85 71 67 66 66 64 64 64 64 62 58 58 57 56 56 56 54 53 52 51 48 45 42 39 39 39 38 22] k=40: inertia=857.700, sizes=[55 53 53 51 50 50 50 50 48 48 47 46 46 45 43 42 42 42 41 40 40 39 37 36 36 35 34 34 34 33 29 29 29 26 25 24 23 19 18 14] MLP READ/WRITE PRODUCTS: spectra and trace (no ReLU) L0: norm=182.768, trcos=0.0304, top_s=[126.4401 39.8521 32.4939 30.8028 28.0393 27.0859 26.1037 25.6267 23.9582 23.8516] L1: norm=173.862, trcos=-0.1114, top_s=[76.2158 32.9514 27.9667 26.8703 25.256 24.4786 24.094 23.1971 23.0743 22.4406] L2: norm=207.921, trcos=-0.2183, top_s=[68.0228 29.9279 27.2496 26.6717 26.1322 25.8145 25.3264 25.0969 24.4946 24.238 ] L3: norm=232.570, trcos=-0.2033, top_s=[133.2948 55.4455 48.8196 45.1388 38.6838 38.2303 37.4412 36.2021 35.7122 35.1299] ```