pythia-jlens-artifacts / results /cv8_positional_extrapolation_github.json
AnonymousInterpScience's picture
Anonymous artifact snapshot
fe984e5
Raw History Blame Contribute Delete
10.4 kB
{
"experiment": "CV8 \u2014 positional extrapolation: is the corpus effect an extrapolation artifact?",
"prereg": "docs/experiments/preregs/CV8_positional_extrapolation.md",
"status": "PRE-REGISTERED",
"model": "EleutherAI/pythia-410m-deduped",
"band": [
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
],
"device": "cpu",
"smoke": false,
"n_sequences_for_G": 128,
"prefix_corpus": "Github",
"probe_positions": [
16,
24,
32,
48,
64,
80,
96,
112,
126,
133,
142,
150,
160,
177
],
"p_read": 142,
"in_support_positions": [
16,
24,
32,
48,
64,
80,
96,
112,
126
],
"out_of_support_positions": [
133,
142,
150,
160,
177
],
"alignment_by_corpus_by_position": {
"Github": {
"16": 0.8199622410112316,
"24": 0.8649228662025228,
"32": 0.8823327600490843,
"48": 0.9208955206003056,
"64": 0.9392306171621081,
"80": 0.9484260866168507,
"96": 0.9504478884880948,
"112": 0.9423929234741516,
"126": 0.9310935599306583,
"133": 0.8484558265334022,
"142": 0.8928635661375124,
"150": 0.8991461850577471,
"160": 0.863023156475018,
"177": 0.8091136683178206
},
"Pile-CC": {
"16": 0.7388377966330032,
"24": 0.7623900261264125,
"32": 0.7401483099189051,
"48": 0.8078829412811408,
"64": 0.8026415395543312,
"80": 0.8074120082320871,
"96": 0.8188838080903893,
"112": 0.8275872410947619,
"126": 0.8261497692178901,
"133": 0.8825382212439484,
"142": 0.8410997638017006,
"150": 0.784536015586303,
"160": 0.737921293108437,
"177": 0.7043631819379887
},
"StackExchange": {
"16": 0.7868950844126045,
"24": 0.8172987029771465,
"32": 0.7958271283933123,
"48": 0.8631015421781737,
"64": 0.8521835524590238,
"80": 0.8567536478592028,
"96": 0.865747722698461,
"112": 0.869334769490467,
"126": 0.8632586365641963,
"133": 0.8908658940965233,
"142": 0.8680869951100778,
"150": 0.8141380848365287,
"160": 0.7637898622813123,
"177": 0.72162493963662
},
"USPTO_Backgrounds": {
"16": 0.7250339105496532,
"24": 0.7466318620527843,
"32": 0.7217538165312588,
"48": 0.7893456531556848,
"64": 0.7835063101091788,
"80": 0.7881745586910742,
"96": 0.8035122217701587,
"112": 0.8106891547649236,
"126": 0.8059748422771726,
"133": 0.8468851304825057,
"142": 0.8175291862732885,
"150": 0.765009421906371,
"160": 0.7211020727707582,
"177": 0.6890500060955719
},
"Wikipedia_en": {
"16": 0.714737290764607,
"24": 0.7414217829325931,
"32": 0.7282756412878298,
"48": 0.7893573136174032,
"64": 0.7913875814589846,
"80": 0.8014941703581413,
"96": 0.8124232223799928,
"112": 0.8217904107612356,
"126": 0.820110692737911,
"133": 0.8634235496340769,
"142": 0.8326451279763215,
"150": 0.7839842876667025,
"160": 0.7441040846658752,
"177": 0.7115479365349467
},
"OOD_arXiv_2023": {
"16": 0.7314797724797859,
"24": 0.7538180293815879,
"32": 0.7296685194098056,
"48": 0.7965723238808289,
"64": 0.7904637129490234,
"80": 0.792464811600057,
"96": 0.8055670247052423,
"112": 0.8121151935399167,
"126": 0.8085203811055055,
"133": 0.8561239410946141,
"142": 0.8231678802397187,
"150": 0.7658832922857227,
"160": 0.7208183538688316,
"177": 0.6866824113541039
},
"OOD_CommonPile": {
"16": 0.7274252644273965,
"24": 0.7487318124617317,
"32": 0.7244637094978772,
"48": 0.7910418690218445,
"64": 0.786705942925807,
"80": 0.7893214492024896,
"96": 0.8042403273110403,
"112": 0.8103238690236166,
"126": 0.8063952401312486,
"133": 0.8562419725480844,
"142": 0.8221690899817258,
"150": 0.7646649434641972,
"160": 0.7195846531182958,
"177": 0.6856726030758462
},
"OOD_News_2024": {
"16": 0.6930364953152686,
"24": 0.714579362305328,
"32": 0.6958332436342336,
"48": 0.759202188142498,
"64": 0.7562811167705011,
"80": 0.7640614480227068,
"96": 0.7759056468667923,
"112": 0.7856216585197024,
"126": 0.7848264476594219,
"133": 0.8403997315239466,
"142": 0.7984675936770528,
"150": 0.7452955111728733,
"160": 0.7056315182522144,
"177": 0.675980543031208
}
},
"read_scores_min_from_r9": {
"Github": 0.24049734957516195,
"Pile-CC": 0.26239902824163436,
"StackExchange": 0.26430597975850106,
"USPTO_Backgrounds": 0.26112590581178663,
"Wikipedia_en": 0.24160844944417476,
"OOD_arXiv_2023": 0.2692730328068137,
"OOD_CommonPile": 0.2528811445459723,
"OOD_News_2024": 0.25006329715251924
},
"PRIMARY": {
"statistic": "Spearman rho between alignment at p_read and the stored min read score",
"rho": -0.047619047619047616,
"exact_two_sided_p": 0.9348710317460317,
"n_permutations": 40320,
"critical_abs_rho_at_alpha_0.05": 0.7142857142857143,
"minimum_detectable": "n=8 detects only |rho| >= 0.714 at alpha=0.05 two-sided; a moderate relationship would NOT be detected by this design",
"corpora_in_order": [
"Github",
"OOD_CommonPile",
"OOD_News_2024",
"OOD_arXiv_2023",
"Pile-CC",
"StackExchange",
"USPTO_Backgrounds",
"Wikipedia_en"
],
"alignment_at_p_read": [
0.8928635661375124,
0.8221690899817258,
0.7984675936770528,
0.8231678802397187,
0.8410997638017006,
0.8680869951100778,
0.8175291862732885,
0.8326451279763215
],
"read_score": [
0.24049734957516195,
0.2528811445459723,
0.25006329715251924,
0.2692730328068137,
0.26239902824163436,
0.26430597975850106,
0.26112590581178663,
0.24160844944417476
]
},
"SECONDARY_alignment_spread": {
"between_corpus_sd_at_p_read": 0.030232561695707077,
"between_corpus_sd_in_support_p64": 0.05773061221794703,
"ratio": 0.5236833723774111,
"note": "descriptive; no threshold is attached, per the registration"
},
"controls": {
"C1_scale_invariance": {
"required": "cos_F unchanged to < 1e-9 when G is multiplied by alpha in {0.5, 2, 100}; the control for the target-count artifact that motivated a scale-free metric. AMENDED from '== 0.0', which is unreachable in floating point.",
"max_abs_difference": 7.949996216893851e-11,
"tolerance": 1e-09,
"fires": true
},
"C2_metric_is_not_vacuous": {
"required": "in-support alignment >= 0.30 for all 8 corpora AND |alignment| vs a Gaussian matrix <= 0.05. IF THIS FAILS THE INSTRUMENT IS BROKEN AND NO VERDICT ISSUES.",
"min_in_support_p64": 0.7562811167705011,
"max_abs_vs_random": 0.00020370669755364056,
"per_corpus_in_support": {
"Github": 0.9392306171621081,
"Pile-CC": 0.8026415395543312,
"StackExchange": 0.8521835524590238,
"USPTO_Backgrounds": 0.7835063101091788,
"Wikipedia_en": 0.7913875814589846,
"OOD_arXiv_2023": 0.7904637129490234,
"OOD_CommonPile": 0.786705942925807,
"OOD_News_2024": 0.7562811167705011
},
"per_corpus_vs_random": {
"Github": 8.715056257069297e-05,
"Pile-CC": 8.34612596484493e-05,
"StackExchange": 0.00020370669755364056,
"USPTO_Backgrounds": 7.459009373812329e-05,
"Wikipedia_en": 5.197126198105234e-05,
"OOD_arXiv_2023": 3.00259731882602e-05,
"OOD_CommonPile": -2.5453768332120378e-05,
"OOD_News_2024": -0.00011232876610156356
},
"fires": true
},
"C3_no_leakage": {
"required": "the items building G and the items scored are disjoint",
"n_overlap": 0,
"n_build": 128,
"n_score": 270,
"fires": true
},
"C4_probe_matches_the_real_read": {
"required": "p_read inside CV2's measured prefixed read range, and CV2's n_in_support == 0",
"p_read": 142,
"cv2_min": 133,
"cv2_max": 177,
"cv2_median": 142,
"cv2_n_in_support": 0,
"fires": true
},
"C5_band_rule": {
"required": "band == floor(0.38L)..floor(0.92L) intersected with layers strictly below the penultimate target layer, at L=24",
"band": [
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
],
"rule_band": [
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19,
20,
21
],
"target_layer": 22,
"fires": true
}
},
"controls_fired": {
"C1_scale_invariance": true,
"C2_metric_is_not_vacuous": true,
"C3_no_leakage": true,
"C4_probe_matches_the_real_read": true,
"C5_band_rule": true
},
"VERDICT": "EXTRAPOLATION DOES NOT EXPLAIN THE ORDERING \u2014 exact two-sided p = 0.9349 >= 0.20, rho = -0.048, below the alpha=0.05 critical |rho| of 0.714 for n=8. No STRONG monotone relationship between alignment at position 142 and the read. The positional hedge is retired; this design could not have detected a moderate relationship and the claim is bounded accordingly.",
"provenance": {
"schema_version": 1,
"experiment": "CV8",
"utc": "2026-08-31T05:58:45+00:00",
"script": {
"path": "experiments/cv8_positional_extrapolation.py",
"sha256": "7bc5d7908a74bbd6713744b4af830c963396d9402c51b1fb289d7c92d23edb8d",
"n_lines": 408
},
"git": {
"commit": "9770d208af381e9579ae259d49d95c0c0a5b9a33",
"commit_short": "9770d208af38",
"git_repository_present": true,
"branch": "pythia",
"dirty": true,
"dirty_files": [
"research-paper-writing/"
],
"dirty_diff_sha256": null,
"describe": "9770d20"
},
"argv": [
"experiments/cv8_positional_extrapolation.py",
"--prefix-corpus",
"Github",
"--out",
"results/cv8_positional_extrapolation_github.json"
],
"cwd": ".",
"inputs": [
{
"path": "results/r9_permutation_calibrated_min.json",
"sha256": "eea61cccda689ed7903ca0f24af1961beb18f104fc7226a21f61a2563e3a9ba2",
"bytes": 20471,
"exists": true
},
{
"path": "results/cv2_position_support.json",
"sha256": "e61248b05d2e8fa4db0262b325e2748579f873326414b1992d5bd4476c136b9c",
"bytes": 9152,
"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": false,
"NVIDIA_TF32_OVERRIDE": null
},
"threads": {
"torch_num_threads": 8,
"torch_num_interop_threads": 14,
"OMP_NUM_THREADS": null,
"os_cpu_count": 14
}
},
"payload_sha256": "db838d5ea323fb6ef2b73cec85b2d427d891d37340a801e138f2bf20f0b01d15"
}
}