pythia-jlens-artifacts / results /d2_device_crossvalidation.json
AnonymousInterpScience's picture
Anonymous artifact snapshot
fe984e5
Raw History Blame Contribute Delete
10.3 kB
{
"experiment": "D2 \u2014 cross-validating CUDA-scored reads against CPU-scored reads",
"status": "DIAGNOSTIC. Joins two stored results files; measures nothing new.",
"why": "CONTEXT.md \u00a77: E65's CUDA scoring branch is wrong by 37% on an arm with no operator, and ladder410/ladder1b are CUDA-scored and feed the asymptotes and the interaction. No CUDA-vs-CPU cross-check had ever been run.",
"inputs": [
"results/ladder410/*.json (device=cuda)",
"results/e48_crossover_410m.json (device=cpu)"
],
"shared_conditions": {
"lenses": "results/e28_<corpus>_410m_n200_s<seed>.pt",
"band": [
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
],
"K": [
1,
2,
5,
10,
20,
50,
100
],
"admitted_sets": [
"multihop",
"multilingual",
"order-ops",
"poetry",
"typo"
],
"rung": 200
},
"ladder_devices_seen": [
"cuda"
],
"tolerance": 1e-06,
"tolerance_derivation": "one (item,intermediate) pair flipping at one k moves a set mean by >= 1/(893*7) = 1.6e-4, so 1e-6 separates fp accumulation from a single changed decision",
"n_cells_compared": 30,
"max_abs_diff": 0.0002773862864289822,
"worst_cell": {
"corpus": "USPTO_Backgrounds",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.1574542057301317,
"cpu_e48": 0.1571768194437027,
"abs_diff": 0.0002773862864289822,
"rel_diff": 0.0017648040430563356
},
"per_cell": [
{
"corpus": "Github",
"seed": 0,
"aggregation": "persist",
"cuda_ladder": 0.02704486567527056,
"cpu_e48": 0.02704486474394798,
"abs_diff": 9.313225780849255e-10,
"rel_diff": 3.443620764616077e-08
},
{
"corpus": "Github",
"seed": 0,
"aggregation": "min",
"cuda_ladder": 0.10961288312183959,
"cpu_e48": 0.10961288511753083,
"abs_diff": 1.995691237266506e-09,
"rel_diff": 1.8206721181790396e-08
},
{
"corpus": "Github",
"seed": 1,
"aggregation": "persist",
"cuda_ladder": 0.029035190998443535,
"cpu_e48": 0.02903519310057163,
"abs_diff": 2.1021280965927147e-09,
"rel_diff": 7.23993151797337e-08
},
{
"corpus": "Github",
"seed": 1,
"aggregation": "min",
"cuda_ladder": 0.11494814133537667,
"cpu_e48": 0.1149481400847435,
"abs_diff": 1.2506331803496806e-09,
"rel_diff": 1.0879977522278075e-08
},
{
"corpus": "Github",
"seed": 2,
"aggregation": "persist",
"cuda_ladder": 0.027610584854015283,
"cpu_e48": 0.027610587328672408,
"abs_diff": 2.4746571250511273e-09,
"rel_diff": 8.962710918073453e-08
},
{
"corpus": "Github",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.11328674435083355,
"cpu_e48": 0.11328674778342247,
"abs_diff": 3.432588921437052e-09,
"rel_diff": 3.0300004092264633e-08
},
{
"corpus": "Pile-CC",
"seed": 0,
"aggregation": "persist",
"cuda_ladder": 0.04509159454277583,
"cpu_e48": 0.04509159799199551,
"abs_diff": 3.449219680706772e-09,
"rel_diff": 7.649362263273668e-08
},
{
"corpus": "Pile-CC",
"seed": 0,
"aggregation": "min",
"cuda_ladder": 0.14523330333509615,
"cpu_e48": 0.1452333066612482,
"abs_diff": 3.3261520482330553e-09,
"rel_diff": 2.2902129853664994e-08
},
{
"corpus": "Pile-CC",
"seed": 1,
"aggregation": "persist",
"cuda_ladder": 0.04611681202160461,
"cpu_e48": 0.046116814273409544,
"abs_diff": 2.2518049369590898e-09,
"rel_diff": 4.882828470346999e-08
},
{
"corpus": "Pile-CC",
"seed": 1,
"aggregation": "min",
"cuda_ladder": 0.1453225617696132,
"cpu_e48": 0.1453225638717413,
"abs_diff": 2.1021281104705025e-09,
"rel_diff": 1.4465256147873893e-08
},
{
"corpus": "Pile-CC",
"seed": 2,
"aggregation": "persist",
"cuda_ladder": 0.045189316916678636,
"cpu_e48": 0.04518931868951768,
"abs_diff": 1.7728390422355744e-09,
"rel_diff": 3.923137355568962e-08
},
{
"corpus": "Pile-CC",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.14300646082098994,
"cpu_e48": 0.14300646334886552,
"abs_diff": 2.5278755755309135e-09,
"rel_diff": 1.767665262348415e-08
},
{
"corpus": "StackExchange",
"seed": 0,
"aggregation": "persist",
"cuda_ladder": 0.04637858061386006,
"cpu_e48": 0.04637858076021075,
"abs_diff": 1.4635068851243105e-10,
"rel_diff": 3.1555663436339706e-09
},
{
"corpus": "StackExchange",
"seed": 0,
"aggregation": "min",
"cuda_ladder": 0.14699955502791065,
"cpu_e48": 0.1469995591789484,
"abs_diff": 4.151037763522325e-09,
"rel_diff": 2.8238436813739703e-08
},
{
"corpus": "StackExchange",
"seed": 1,
"aggregation": "persist",
"cuda_ladder": 0.04555068582828556,
"cpu_e48": 0.04555068640038371,
"abs_diff": 5.72098153572842e-10,
"rel_diff": 1.2559594569973873e-08
},
{
"corpus": "StackExchange",
"seed": 1,
"aggregation": "min",
"cuda_ladder": 0.14416987733649356,
"cpu_e48": 0.14416988678276538,
"abs_diff": 9.446271820312546e-09,
"rel_diff": 6.552180924263437e-08
},
{
"corpus": "StackExchange",
"seed": 2,
"aggregation": "persist",
"cuda_ladder": 0.045260619930922986,
"cpu_e48": 0.04526062114164233,
"abs_diff": 1.2107193442245645e-09,
"rel_diff": 2.6749949816986367e-08
},
{
"corpus": "StackExchange",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.14551294004278523,
"cpu_e48": 0.14551294930279254,
"abs_diff": 9.26000731649168e-09,
"rel_diff": 6.36369983624129e-08
},
{
"corpus": "USPTO_Backgrounds",
"seed": 0,
"aggregation": "persist",
"cuda_ladder": 0.04964359144547156,
"cpu_e48": 0.04964358932338655,
"abs_diff": 2.1220850146552728e-09,
"rel_diff": 4.274640580139482e-08
},
{
"corpus": "USPTO_Backgrounds",
"seed": 0,
"aggregation": "min",
"cuda_ladder": 0.15465880749481065,
"cpu_e48": 0.15465880744159222,
"abs_diff": 5.3218429663104416e-11,
"rel_diff": 3.4410215973767067e-10
},
{
"corpus": "USPTO_Backgrounds",
"seed": 1,
"aggregation": "persist",
"cuda_ladder": 0.04933665746024677,
"cpu_e48": 0.04933666000142693,
"abs_diff": 2.541180162130008e-09,
"rel_diff": 5.150693545238999e-08
},
{
"corpus": "USPTO_Backgrounds",
"seed": 1,
"aggregation": "min",
"cuda_ladder": 0.15324432841901267,
"cpu_e48": 0.153244324401021,
"abs_diff": 4.017991661608988e-09,
"rel_diff": 2.6219513690402082e-08
},
{
"corpus": "USPTO_Backgrounds",
"seed": 2,
"aggregation": "persist",
"cuda_ladder": 0.04986586142331362,
"cpu_e48": 0.04986586356535554,
"abs_diff": 2.142041918840043e-09,
"rel_diff": 4.295607787946208e-08
},
{
"corpus": "USPTO_Backgrounds",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.1574542057301317,
"cpu_e48": 0.1571768194437027,
"abs_diff": 0.0002773862864289822,
"rel_diff": 0.0017648040430563356
},
{
"corpus": "Wikipedia_en",
"seed": 0,
"aggregation": "persist",
"cuda_ladder": 0.03809930235147477,
"cpu_e48": 0.038099304959177974,
"abs_diff": 2.6077032061477823e-09,
"rel_diff": 6.844490231362073e-08
},
{
"corpus": "Wikipedia_en",
"seed": 0,
"aggregation": "min",
"cuda_ladder": 0.12926935620073762,
"cpu_e48": 0.12926936261355876,
"abs_diff": 6.412821135226565e-09,
"rel_diff": 4.960820573082906e-08
},
{
"corpus": "Wikipedia_en",
"seed": 1,
"aggregation": "persist",
"cuda_ladder": 0.03906317305352006,
"cpu_e48": 0.039063173532485965,
"abs_diff": 4.789659016624093e-10,
"rel_diff": 1.2261315667660452e-08
},
{
"corpus": "Wikipedia_en",
"seed": 1,
"aggregation": "min",
"cuda_ladder": 0.13214441066873925,
"cpu_e48": 0.1321444157510996,
"abs_diff": 5.0823603381378035e-09,
"rel_diff": 3.846065162307483e-08
},
{
"corpus": "Wikipedia_en",
"seed": 2,
"aggregation": "persist",
"cuda_ladder": 0.037466185060994965,
"cpu_e48": 0.03746618703007698,
"abs_diff": 1.9690820154960598e-09,
"rel_diff": 5.2556242617251837e-08
},
{
"corpus": "Wikipedia_en",
"seed": 2,
"aggregation": "min",
"cuda_ladder": 0.1300129838021738,
"cpu_e48": 0.13001298978924752,
"abs_diff": 5.9870737256773054e-09,
"rel_diff": 4.6049811910197724e-08
}
],
"VERDICT": "DIVERGENT \u2014 max |CUDA - CPU| = 2.774e-04 over 30 cells (worst: USPTO_Backgrounds s2, min, 0.18% relative). STOP. Every ladder-derived number in the paper, meaning the per-corpus asymptotes, the flatness in N and the 1B replication, must be re-sourced from a CPU scoring pass before it is cited.",
"provenance": {
"schema_version": 1,
"experiment": "D2 \u2014 cross-validating CUDA-scored reads against CPU-scored reads",
"utc": "2026-08-20T00:11:35+00:00",
"script": {
"path": "experiments/d2_device_crossvalidation.py",
"sha256": "e1033efb2cbd9ea71d60d628b4f33ee8fcf866946445cec4a1d054a28046c916",
"n_lines": 125
},
"git": {
"commit": "2dfeda4b4d89650f970dc362f8183e2f6077c9fe",
"commit_short": "2dfeda4b4d89",
"branch": "pythia",
"dirty": true,
"dirty_files": [
"experiments/d2_device_crossvalidation.py",
"figures/fB_band_width.png",
"figures/make_figures.py",
"igures/f0_grid_single.png",
"paper/paper.tex",
"paper/paper_5pp.bbl",
"paper/paper_5pp.blg",
"paper/paper_5pp.tex",
"paper/paper_9pp.tex",
"paper/refs.bib",
"paper/tables_auto.tex",
"tools/make_paper_tables.py"
],
"dirty_diff_sha256": "0b3cf18514db7d8a624429659f62d811e911c001c158012b785b6f8927d12df6",
"describe": "2dfeda4-dirty"
},
"argv": [
"experiments/d2_device_crossvalidation.py"
],
"cwd": ".",
"inputs": [
{
"path": "results/e48_crossover_410m.json",
"sha256": "7335aab33c2ea0b6bc3e9deec90181823ab2fb68202626a66c8658ad2cae4051",
"bytes": 116036,
"exists": true
}
],
"env": {
"python": "3.14.2",
"torch": "2.13.0",
"transformers": "5.14.1",
"numpy": "2.5.1",
"platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
"machine": "arm64",
"hostname": "<redacted for double-blind review>",
"tf32": {
"allow_tf32_matmul": false,
"allow_tf32_cudnn": true,
"NVIDIA_TF32_OVERRIDE": null
}
},
"payload_sha256": "e3c1b0af60b548841f62ecd52aba7dcc0e8eb5db7f7528a2aee67a2044f53858"
}
}