SPIRA / python /artifacts /metrics_v3_ab.json
thefinalboss's picture
SPIRA v3: cortex representation repair — direct input pathway, multi-timescale trace banks, structured trigram inputs; controlled A/B (SNN alone 0.40% -> 2.07% in-vocab held-out vs bigram 4.94% at matched vocab 4096); paper, README and docs updated to third iteration
6923c51 verified
Raw History Blame Contribute Delete
4.12 kB
{
"experiment": "SPIRA v3 cortex representation repair \u2014 controlled A/B vs v2 baseline",
"date": "2026-10-22",
"protocol": {
"corpus": "same rebuilt corpus for both arms (train 13.27 MB / test 640 KB)",
"vocab_size": 4096,
"snn_tokens": 200000,
"epochs": 2,
"am_contexts": 50000,
"test_positions": 20000,
"vocab_coverage_test": 0.4524,
"headline_metric": "SNN alone, in-vocab accuracy on held-out test positions",
"bigram_reference": "patch bigram trained on train cols [0:600k], evaluated on exactly the same test positions at the same vocab 4096"
},
"headline": {
"snn_invocab_test_v2_baseline": 0.003978779840848806,
"snn_invocab_test_v3": 0.020667550839964632,
"ratio_v3_over_v2": "x5.2",
"bigram_4096_invocab_test": 0.04940318302387268,
"gap_to_bigram": "x12.4 (v2) -> x2.4 (v3); beat-the-bigram criterion NOT met",
"snn_invocab_trained_region_v2_baseline": 0.0376034093757834,
"snn_invocab_trained_region_v3": 0.1328653797944347,
"training_speed_tps_v2_baseline": "~530 (512.6 / 561.8 per epoch)",
"training_speed_tps_v3": 188.97756640356553,
"checkpoint_mb_v2_baseline": 14.6,
"checkpoint_mb_v3": 45.9,
"notes": [
"Free generation still incoherent in both arms.",
"At this small budget, fusion hurts on held-out: v3 fused in-vocab 0.001326259946949602 < SNN alone 0.020667550839964632."
]
},
"bigram_reference_detail": {
"bigram_4096_invocab_test20k": 0.04940318302387268,
"n_invocab": 9048
},
"baseline_arm_full": {
"label": "baseline",
"hist": [
{
"acc": 0.092515,
"acc_fast": 0.112785,
"tps": 512.6229927407619
},
{
"acc": 0.065475,
"acc_fast": 0.101045,
"tps": 561.7816027951075
}
],
"test": {
"n_positions": 20000,
"vocab_coverage": 0.4524,
"snn_acc": 0.0018,
"hier_acc": 0.00025,
"am_acc": 0.0001,
"fused_acc": 0.00035,
"snn_acc_invocab": 0.003978779840848806,
"hier_acc_invocab": 0.0005526083112290009,
"am_acc_invocab": 0.00022104332449160034,
"fused_acc_invocab": 0.0007736516357206013,
"corrupt_lanes": 0
},
"trained": {
"n_positions": 8000,
"vocab_coverage": 0.498625,
"snn_acc": 0.01875,
"hier_acc": 0.00675,
"am_acc": 0.011125,
"fused_acc": 0.017375,
"snn_acc_invocab": 0.0376034093757834,
"hier_acc_invocab": 0.013537227375282025,
"am_acc_invocab": 0.01779894710453748,
"fused_acc_invocab": 0.030834795688142392,
"corrupt_lanes": 0
},
"gen": "The old man looked at the retuiltIN S .\n IN S add\n r ifotIthing beithingtory rethingeek eIGHTND RIGHT it athingthingthing if t:\n #irlother:\n #\n f\n \" andeturotIe)\n a s\n fone)Imone)e the .\n .\n thingotICAPef g if##########IGHTtry: lf._lf._otI iflear a singl"
},
"v3_arm_full": {
"label": "v3",
"hist": [
{
"acc": 0.120205,
"acc_fast": 0.142835,
"tps": 188.97756640356553
},
{
"acc": 0.11196,
"acc_fast": 0.15781,
"tps": 188.5136375504029
}
],
"test": {
"n_positions": 20000,
"vocab_coverage": 0.4524,
"snn_acc": 0.00935,
"hier_acc": 0.0005,
"am_acc": 0.0001,
"fused_acc": 0.0006,
"snn_acc_invocab": 0.020667550839964632,
"hier_acc_invocab": 0.0011052166224580018,
"am_acc_invocab": 0.00022104332449160034,
"fused_acc_invocab": 0.001326259946949602,
"corrupt_lanes": 0
},
"trained": {
"n_positions": 8000,
"vocab_coverage": 0.498625,
"snn_acc": 0.06625,
"hier_acc": 0.005375,
"am_acc": 0.011125,
"fused_acc": 0.015875,
"snn_acc_invocab": 0.1328653797944347,
"hier_acc_invocab": 0.010779644021057909,
"am_acc_invocab": 0.01779894710453748,
"fused_acc_invocab": 0.02782652293807972,
"corrupt_lanes": 0
},
"gen": "The old man looked at the rrorone\nr hi \n r hi r hiand a a l on ar hi\n on alooloor hiloolectueEloo looynon at see a slectlectlector\non a##########yn a s##########ramelooErlooow, golectlectErlooow,loo.\n .\n lect lectlect rameloo a s"
}
}