OUTPUT of run 2 of at most 100: ``` Layer 0 singular spectra (first 60, then indices 80, 120, 200, 300, 383): W_q [4.368 3.131 3.099 3.042 2.994 2.951 2.926 2.881 2.859 2.857 2.804 2.766 2.762 2.692 2.683 2.661 2.637 2.588 2.539 2.509 2.458 2.433 2.403 2.299 2.272 2.173 2.098 2.048 1.954 1.887 1.877 1.865 1.738 1.725 1.665 1.605 1.552 1.535 1.476 1.424 1.378 1.341 1.306 1.264 1.24 1.213 1.154 1.122 1.109 1.068 1.053 1.031 1.017 1.007 0.985 0.957 0.944 0.913 0.888 0.879] tail [0.6556 0.4899 0.3045 0.1303 0.0006] W_k [4.163 3.952 3.467 3.427 3.364 3.335 3.279 3.256 3.214 3.163 3.115 3.029 2.977 2.901 2.881 2.832 2.798 2.76 2.658 2.584 2.528 2.449 2.353 2.301 2.281 2.206 2.121 2.074 1.917 1.896 1.813 1.753 1.713 1.676 1.573 1.53 1.47 1.431 1.415 1.355 1.3 1.26 1.187 1.169 1.161 1.116 1.108 1.104 1.072 1.032 1.003 0.988 0.965 0.953 0.931 0.911 0.907 0.885 0.873 0.866] tail [0.6608 0.5097 0.317 0.1399 0.0003] W_v [5.208 3.364 3.006 2.962 2.815 2.639 2.527 2.46 2.37 2.301 2.25 2.23 2.151 2.135 2.044 2.039 2.001 1.981 1.91 1.875 1.822 1.773 1.749 1.708 1.634 1.611 1.556 1.487 1.453 1.402 1.268 1.213 1.187 1.13 1.109 1.08 1.026 0.992 0.987 0.959 0.933 0.931 0.908 0.892 0.887 0.866 0.864 0.853 0.843 0.827 0.811 0.793 0.783 0.781 0.759 0.748 0.721 0.715 0.71 0.701] tail [0.5718 0.4517 0.2931 0.1256 0.0007] W_o [7.361 6.099 3.167 2.452 2.321 2.202 2.171 2.16 2.105 2.072 2.021 1.995 1.975 1.967 1.935 1.922 1.905 1.885 1.857 1.84 1.815 1.784 1.759 1.744 1.712 1.694 1.634 1.611 1.585 1.553 1.511 1.466 1.456 1.446 1.397 1.392 1.37 1.336 1.321 1.305 1.294 1.27 1.248 1.218 1.197 1.182 1.167 1.148 1.136 1.099 1.095 1.077 1.049 1.044 1.011 1.005 0.998 0.985 0.966 0.941] tail [0.6103 0.3674 0.1734 0.059 0.0002] W_fc [18.118 12.76 11.337 9.323 9.267 9.021 8.898 8.817 8.609 8.4 8.189 8.06 7.819 7.724 7.62 7.543 7.226 7.081 6.979 6.833 6.779 6.682 6.592 6.402 6.39 6.306 6.207 6.18 5.974 5.864 5.789 5.741 5.646 5.579 5.512 5.353 5.196 5.193 5.122 5.053 5.029 4.849 4.646 4.573 4.452 4.163 4.133 3.899 3.888 3.557 3.455 3.347 3.241 3.008 2.921 2.833 2.796 2.657 2.612 2.555] tail [1.7122 1.2795 0.9228 0.6587 0.4311] W_proj [10.899 7.702 7.173 6.638 6.541 6.288 6.201 6.133 5.941 5.917 5.758 5.641 5.552 5.476 5.393 5.363 5.259 5.206 5.171 5.153 5.057 5.018 4.96 4.88 4.821 4.788 4.77 4.73 4.643 4.579 4.546 4.529 4.505 4.493 4.44 4.415 4.369 4.315 4.293 4.261 4.243 4.185 4.134 4.104 4.064 4.049 4.026 3.988 3.967 3.95 3.933 3.901 3.857 3.804 3.778 3.754 3.74 3.713 3.692 3.683] tail [3.1846 2.4346 1.3916 0.6756 0.2813] Layer 0 individual-head spectra (top 12): H 0 QK [0.8532 0.6126 0.3681 0.2867 0.2786 0.2441 0.2317 0.2099 0.2016 0.1885 0.1818 0.1728] OV [3.6115 3.3924 2.8811 2.3143 2.1276 1.8367 1.6785 1.5294 1.377 1.3026 1.1515 1.1013] H 1 QK [0.838 0.3105 0.2193 0.1245 0.0908 0.0863 0.0659 0.0389 0.0378 0.0336 0.0332 0.0297] OV [5.7206 3.3585 1.1458 0.5449 0.4435 0.3867 0.3839 0.3541 0.331 0.3111 0.2885 0.2674] H 2 QK [0.5432 0.1831 0.1574 0.1329 0.0717 0.0659 0.0464 0.0456 0.0362 0.0332 0.0321 0.0292] OV [3.9425 2.3569 1.0061 0.5213 0.5 0.4496 0.3863 0.3309 0.3101 0.3051 0.2774 0.2548] H 3 QK [1.114 1.0924 1.0705 1.064 1.0301 0.9924 0.9677 0.9537 0.9464 0.9002 0.8883 0.8685] OV [4.4261 3.911 3.6812 3.5972 3.5958 3.4449 3.343 3.2829 3.0733 3.0135 2.9723 2.9144] H 4 QK [0.5096 0.2267 0.1852 0.1237 0.1013 0.0658 0.0655 0.0489 0.0484 0.0454 0.0401 0.035 ] OV [2.4485 1.3166 0.666 0.4611 0.4377 0.3631 0.3568 0.3142 0.2959 0.2618 0.2484 0.2372] H 5 QK [0.1805 0.0481 0.0241 0.0205 0.0202 0.0191 0.0188 0.0181 0.0179 0.0172 0.0169 0.0164] OV [24.7988 12.8996 5.8492 4.2294 3.6654 2.3878 1.919 1.6636 1.3942 1.1336 0.8624 0.7865] Layer 0 QK Frobenius cosines [[ 1. 0.032 0.067 -0.132 0.114 0.017] [ 0.032 1. 0.58 -0.022 0.419 -0.008] [ 0.067 0.58 1. -0.028 0.451 -0.053] [-0.132 -0.022 -0.028 1. -0.032 -0.013] [ 0.114 0.419 0.451 -0.032 1. 0.029] [ 0.017 -0.008 -0.053 -0.013 0.029 1. ]] Layer 0 OV Frobenius cosines [[ 1. 0.026 0.027 -0.041 0.033 0.037] [ 0.026 1. 0.801 -0.004 0.7 0.415] [ 0.027 0.801 1. 0.005 0.668 0.52 ] [-0.041 -0.004 0.005 1. 0.009 -0.013] [ 0.033 0.7 0.668 0.009 1. 0.369] [ 0.037 0.415 0.52 -0.013 0.369 1. ]] LayerNorm largest absolute bias coordinates: ln1_b.0 [(312, -0.09539999812841415), (146, 0.09059999883174896), (12, 0.08560000360012054), (104, -0.08410000056028366), (289, 0.08259999752044678), (156, -0.07800000160932541), (303, -0.07760000228881836), (328, 0.07620000094175339)] ln2_b.0 [(99, -0.6236000061035156), (58, 0.4636000096797943), (153, 0.21320000290870667), (39, -0.20200000703334808), (18, 0.18160000443458557), (29, -0.1720000058412552), (383, -0.16519999504089355), (131, 0.16019999980926514)] ln1_b.1 [(39, -1.5676000118255615), (58, 1.3651000261306763), (99, -1.3355000019073486), (184, -0.5616999864578247), (383, -0.46219998598098755), (21, -0.42989999055862427), (153, 0.2969000041484833), (266, 0.27730000019073486)] ln2_b.1 [(58, 0.4169999957084656), (375, 0.3919999897480011), (76, -0.38920000195503235), (39, -0.36880001425743103), (319, 0.35519999265670776), (23, -0.3400999903678894), (358, 0.3149000108242035), (123, 0.31470000743865967)] ln1_b.2 [(99, -1.6589000225067139), (58, 1.3317999839782715), (39, -1.1991000175476074), (383, -0.5070000290870667), (184, -0.5045999884605408), (21, -0.45339998602867126), (63, -0.3244999945163727), (266, 0.29910001158714294)] ln2_b.2 [(236, -0.5170000195503235), (73, -0.48030000925064087), (322, 0.4489000141620636), (214, -0.42320001125335693), (97, 0.40139999985694885), (382, -0.39169999957084656), (184, -0.3831999897956848), (364, 0.3749000132083893)] ln1_b.3 [(58, 0.8241000175476074), (99, -0.7638999819755554), (5, 0.4359999895095825), (383, -0.39739999175071716), (266, 0.3425000011920929), (21, -0.31520000100135803), (184, -0.2888999879360199), (45, 0.2888000011444092)] ln2_b.3 [(21, -0.5401999950408936), (275, 0.39320001006126404), (215, -0.35199999809265137), (382, -0.3296999931335449), (364, 0.31869998574256897), (168, 0.31369999051094055), (86, -0.3041999936103821), (22, -0.3012000024318695)] Affine QK algebra using learned LN scale and bias: head | ||A||, top SV, stable rank, tr(A), ||key-linear coefficient||, coefficient fraction in top key singular direction L0H0 [ 0.2934 0.2481 1.3984 -0.1047 0.0429 0.6945] L0H1 [ 0.159 0.1502 1.1208 -0.0985 0.0125 0.6069] L0H2 [ 0.1074 0.1024 1.1007 -0.0873 0.0071 0.8765] L0H3 [0.6976 0.2435 8.2055 1.3415 0.0428 0.1444] L0H4 [ 0.0885 0.0796 1.2366 -0.0459 0.0167 0.9422] L0H5 [ 0.0339 0.0322 1.1088 -0.0086 0.0053 0.9713] L1H0 [ 1.9781 1.1849 2.7872 -1.7482 3.9745 0.9862] L1H1 [ 1.8428 1.1538 2.5506 -1.5574 3.6885 0.9761] L1H2 [ 1.7812 1.1758 2.2947 -1.8874 3.0513 0.9744] L1H3 [ 1.7874 1.1874 2.2662 -1.5116 3.3904 0.9776] L1H4 [ 1.6294 1.059 2.3672 -2.0109 2.3285 0.9148] L1H5 [ 1.8633 1.1462 2.6427 -1.6298 3.895 0.9789] L2H0 [ 1.6311 0.6887 5.6082 -1.5548 2.02 0.9774] L2H1 [ 2.1541 0.5775 13.9127 -0.2158 1.8804 0.88 ] L2H2 [ 1.7201 0.518 11.0249 -1.2972 1.6082 0.9745] L2H3 [ 1.623 0.502 10.4524 -0.5951 1.0595 0.7388] L2H4 [ 2.2151 0.8129 7.4259 -0.979 2.3054 0.9497] L2H5 [ 2.2641 0.6152 13.5436 -0.2664 2.0234 0.926 ] L3H0 [ 2.0288 0.4354 21.7128 2.1778 0.5004 0.7025] L3H1 [ 1.2158 0.3451 12.4148 0.4312 0.2796 0.4056] L3H2 [ 1.4366 0.383 14.0691 0.3494 0.4953 0.8805] L3H3 [ 1.8151 0.3215 31.869 5.8301 0.2202 0.205 ] L3H4 [0.6225 0.2358 6.9675 0.0021 0.1441 0.2275] L3H5 [ 1.6651 0.4542 13.4423 1.4368 0.4934 0.861 ] ```