fittingcorpus / results /cv8_positional_extrapolation.json
Tarush-AI's picture AnonymousInterpScience's picture
Duplicate from AnonymousInterpScience/pythia-jlens-artifacts
3b151ef
Raw History Blame Contribute Delete
10.5 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": "Pile-CC",
"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": 0.6996252853536956,
"24": 0.7195726800527922,
"32": 0.7419920931750147,
"48": 0.7890212811960222,
"64": 0.7941302109133536,
"80": 0.8161464913360663,
"96": 0.8183671340138088,
"112": 0.8229096214682946,
"126": 0.8238566421789222,
"133": 0.7944473162024065,
"142": 0.8388732396731006,
"150": 0.8087969634902892,
"160": 0.7728870509309129,
"177": 0.7256898382376776
},
"Pile-CC": {
"16": 0.8655362229126606,
"24": 0.8794407644581727,
"32": 0.9045990717899588,
"48": 0.9609246582531124,
"64": 0.9600123937570153,
"80": 0.9735561251166504,
"96": 0.9736115131621349,
"112": 0.9587698643649143,
"126": 0.9315715191158161,
"133": 0.8844150766354874,
"142": 0.8829077025264984,
"150": 0.812117650005082,
"160": 0.750308456655454,
"177": 0.6929892962431711
},
"StackExchange": {
"16": 0.8036795767735414,
"24": 0.823674148552886,
"32": 0.8445555689011814,
"48": 0.8826739930722733,
"64": 0.8859453230341394,
"80": 0.898833215588331,
"96": 0.904113715176498,
"112": 0.8884164602671668,
"126": 0.8710085702476517,
"133": 0.8486331253630209,
"142": 0.8570446480097353,
"150": 0.7892426785265432,
"160": 0.7309038381482675,
"177": 0.6775588355322955
},
"USPTO_Backgrounds": {
"16": 0.7907249900285207,
"24": 0.8090482826892232,
"32": 0.8344965816173697,
"48": 0.8756122774890299,
"64": 0.8785620695065707,
"80": 0.8933423001911186,
"96": 0.8913479278432859,
"112": 0.8787880476980111,
"126": 0.8573401928967778,
"133": 0.827182917778604,
"142": 0.8381376088564241,
"150": 0.7732960251599772,
"160": 0.7201225439786406,
"177": 0.6684383565635394
},
"Wikipedia_en": {
"16": 0.8116737040343527,
"24": 0.8243563752414974,
"32": 0.8515287521634032,
"48": 0.9105007011285444,
"64": 0.9187448191728964,
"80": 0.929348254154488,
"96": 0.925303206745383,
"112": 0.9200323122113367,
"126": 0.9014905392400295,
"133": 0.8630461395728837,
"142": 0.8731856807813917,
"150": 0.8077957514475436,
"160": 0.7556025082473946,
"177": 0.6978742364606902
},
"OOD_arXiv_2023": {
"16": 0.7916525423537559,
"24": 0.8102770192869214,
"32": 0.8335794211271619,
"48": 0.8782709376178942,
"64": 0.8771631566419636,
"80": 0.8915520081502442,
"96": 0.8949200856470252,
"112": 0.8814222544321175,
"126": 0.8556494233519906,
"133": 0.8358392088399716,
"142": 0.8390375985784466,
"150": 0.7702991962986523,
"160": 0.7144830722075272,
"177": 0.6603754919946767
},
"OOD_CommonPile": {
"16": 0.781236677486555,
"24": 0.7988611712590944,
"32": 0.8232537507439353,
"48": 0.8654877791238893,
"64": 0.8656894410263634,
"80": 0.8812748843918979,
"96": 0.8850409193797165,
"112": 0.8723165202545101,
"126": 0.8502470394124511,
"133": 0.833852988611821,
"142": 0.8369051828581651,
"150": 0.7682897033917105,
"160": 0.7138111346577654,
"177": 0.6595057776158877
},
"OOD_News_2024": {
"16": 0.827480165969923,
"24": 0.837726723624863,
"32": 0.8632292155316605,
"48": 0.9210869498301496,
"64": 0.9247255673704081,
"80": 0.9362241489898391,
"96": 0.9349798506919959,
"112": 0.9225137382747429,
"126": 0.8975991050504516,
"133": 0.8515537919652221,
"142": 0.8455001686641376,
"150": 0.7801679271847001,
"160": 0.7259258340690871,
"177": 0.6713449714937855
}
},
"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": 0.14285714285714285,
"exact_two_sided_p": 0.7520337301587302,
"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": [
0.8388732396731006,
0.8369051828581651,
0.8455001686641376,
0.8390375985784466,
0.8829077025264984,
0.8570446480097353,
0.8381376088564241,
0.8731856807813917
],
"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": 0.01784753488872414,
"between_corpus_sd_in_support_p64": 0.049280116927836765,
"ratio": 0.3621650272230308,
"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": 4.8365533800165394e-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": 0.7941302109133536,
"max_abs_vs_random": 0.00020370669755364056,
"per_corpus_in_support": {
"Github": 0.7941302109133536,
"Pile-CC": 0.9600123937570153,
"StackExchange": 0.8859453230341394,
"USPTO_Backgrounds": 0.8785620695065707,
"Wikipedia_en": 0.9187448191728964,
"OOD_arXiv_2023": 0.8771631566419636,
"OOD_CommonPile": 0.8656894410263634,
"OOD_News_2024": 0.9247255673704081
},
"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 = 0.7520 >= 0.20, rho = +0.143, 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": "2026-08-31T03:39:53+00:00",
"script": {
"path": "experiments/cv8_positional_extrapolation.py",
"sha256": "7bc5d7908a74bbd6713744b4af830c963396d9402c51b1fb289d7c92d23edb8d",
"n_lines": 408
},
"git": {
"commit": "99a6dd48ac17cd06f63d77e1e2fa51296cd189f7",
"commit_short": "99a6dd48ac17",
"git_repository_present": true,
"branch": "pythia",
"dirty": true,
"dirty_files": [
"PAPER_REWRITE_SCOPE.md",
"xperiments/e66b_determinism_floor.py"
],
"dirty_diff_sha256": "01b0d0f1f9e05dca6bc8dca13e923a4920f8bf53c42c0d1a3cca9d266754795c",
"describe": "99a6dd4-dirty"
},
"argv": [
"experiments/cv8_positional_extrapolation.py"
],
"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": "a950c3b88b5401cefca4116adf35303854131b5bd45dfd084c0efe00325e4501"
}
}