RiverRider commited on
Commit
bcadf43
·
verified ·
1 Parent(s): 3201e6d

eval artifact oracle_ceiling.json

Browse files
Files changed (1) hide show
  1. eval/oracle_ceiling.json +94 -0
eval/oracle_ceiling.json ADDED
@@ -0,0 +1,94 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backbone": "meta-llama/Llama-3.2-3B",
3
+ "layer": 20,
4
+ "n_targets": 200,
5
+ "targets_path": "artifacts/nla/llama32_3B/targets_L20_seq64_30k_seed1.pt",
6
+ "anisotropy_mu_norm": 7.163687705993652,
7
+ "replay_cos_mean": 0.8087875247001648,
8
+ "replay_cos_p05": 0.04843669757246971,
9
+ "replay_cos_p50": 0.9894923567771912,
10
+ "replay_cos_p95": 0.997727632522583,
11
+ "replay_fve_mean": 0.9043937921524048,
12
+ "replay_fve_p50": 0.994746208190918,
13
+ "replay_n": 200,
14
+ "replay_cen_cos_mean": 0.7611221075057983,
15
+ "replay_cen_cos_p05": -0.2058715522289276,
16
+ "replay_cen_cos_p50": 0.987565279006958,
17
+ "replay_cen_cos_p95": 0.9973605871200562,
18
+ "replay_cen_fve_mean": 0.880561113357544,
19
+ "replay_cen_fve_p50": 0.993782639503479,
20
+ "replay_cen_n": 200,
21
+ "random_floor_cos_mean": 0.13757377862930298,
22
+ "random_floor_cos_p05": 0.010975184850394726,
23
+ "random_floor_cos_p50": 0.08592405915260315,
24
+ "random_floor_cos_p95": 0.3702440559864044,
25
+ "random_floor_fve_mean": 0.5687869191169739,
26
+ "random_floor_fve_p50": 0.5429620742797852,
27
+ "random_floor_n": 39800,
28
+ "random_floor_cen_cos_mean": -0.003620174480602145,
29
+ "random_floor_cen_cos_p05": -0.2549971342086792,
30
+ "random_floor_cen_cos_p50": -0.011850034818053246,
31
+ "random_floor_cen_cos_p95": 0.23950813710689545,
32
+ "random_floor_cen_fve_mean": 0.4981899857521057,
33
+ "random_floor_cen_fve_p50": 0.4940750002861023,
34
+ "random_floor_cen_n": 39800,
35
+ "nn_in_pool_cos_mean": 0.5704532265663147,
36
+ "nn_in_pool_cos_p05": 0.21268048882484436,
37
+ "nn_in_pool_cos_p50": 0.4759231209754944,
38
+ "nn_in_pool_cos_p95": 0.9953416585922241,
39
+ "nn_in_pool_fve_mean": 0.785226583480835,
40
+ "nn_in_pool_fve_p50": 0.7379615902900696,
41
+ "nn_in_pool_n": 200,
42
+ "nn_in_pool_cen_cos_mean": 0.5112826228141785,
43
+ "nn_in_pool_cen_cos_p05": 0.14741477370262146,
44
+ "nn_in_pool_cen_cos_p50": 0.3755534887313843,
45
+ "nn_in_pool_cen_cos_p95": 0.9935541152954102,
46
+ "nn_in_pool_cen_fve_mean": 0.7556412220001221,
47
+ "nn_in_pool_cen_fve_p50": 0.6877767443656921,
48
+ "nn_in_pool_cen_n": 200,
49
+ "paraphrase_all_cos_mean": 0.28096556663513184,
50
+ "paraphrase_all_cos_p05": 0.013462398201227188,
51
+ "paraphrase_all_cos_p50": 0.2234383225440979,
52
+ "paraphrase_all_cos_p95": 0.986600399017334,
53
+ "paraphrase_all_fve_mean": 0.6404827833175659,
54
+ "paraphrase_all_fve_p50": 0.6117191314697266,
55
+ "paraphrase_all_n": 1600,
56
+ "paraphrase_all_cen_cos_mean": 0.15020351111888885,
57
+ "paraphrase_all_cen_cos_p05": -0.28163281083106995,
58
+ "paraphrase_all_cen_cos_p50": 0.10723456740379333,
59
+ "paraphrase_all_cen_cos_p95": 0.9834794998168945,
60
+ "paraphrase_all_cen_fve_mean": 0.5751017928123474,
61
+ "paraphrase_all_cen_fve_p50": 0.5536172986030579,
62
+ "paraphrase_all_cen_n": 1600,
63
+ "paraphrase_best_cos_mean": 0.5273707509040833,
64
+ "paraphrase_best_cos_p05": 0.04575127363204956,
65
+ "paraphrase_best_cos_p50": 0.47343623638153076,
66
+ "paraphrase_best_cos_p95": 0.9938907623291016,
67
+ "paraphrase_best_fve_mean": 0.7636852860450745,
68
+ "paraphrase_best_fve_p50": 0.7367181181907654,
69
+ "paraphrase_best_n": 200,
70
+ "paraphrase_best_cen_cos_mean": 0.44074559211730957,
71
+ "paraphrase_best_cen_cos_p05": -0.21540063619613647,
72
+ "paraphrase_best_cen_cos_p50": 0.3819118142127991,
73
+ "paraphrase_best_cen_cos_p95": 0.9927623867988586,
74
+ "paraphrase_best_cen_fve_mean": 0.7203727960586548,
75
+ "paraphrase_best_cen_fve_p50": 0.6909558773040771,
76
+ "paraphrase_best_cen_n": 200,
77
+ "paraphrase_mean_per_src_cos_mean": 0.28096550703048706,
78
+ "paraphrase_mean_per_src_cos_p05": 0.023124242201447487,
79
+ "paraphrase_mean_per_src_cos_p50": 0.29697173833847046,
80
+ "paraphrase_mean_per_src_cos_p95": 0.6240138411521912,
81
+ "paraphrase_mean_per_src_fve_mean": 0.6404827833175659,
82
+ "paraphrase_mean_per_src_fve_p50": 0.6484858989715576,
83
+ "paraphrase_mean_per_src_n": 200,
84
+ "paraphrase_mean_per_src_cen_cos_mean": 0.15020351111888885,
85
+ "paraphrase_mean_per_src_cen_cos_p05": -0.2608959674835205,
86
+ "paraphrase_mean_per_src_cen_cos_p50": 0.19087356328964233,
87
+ "paraphrase_mean_per_src_cen_cos_p95": 0.5863825082778931,
88
+ "paraphrase_mean_per_src_cen_fve_mean": 0.5751017332077026,
89
+ "paraphrase_mean_per_src_cen_fve_p50": 0.5954368114471436,
90
+ "paraphrase_mean_per_src_cen_n": 200,
91
+ "normalisation_denom_fve": 0.19489836692810059,
92
+ "normalisation_denom_fve_cen": 0.22218281030654907,
93
+ "formula": "rho_raw = (adapter_best_of_N_fve - random_floor_fve_mean) / (paraphrase_best_fve_mean - random_floor_fve_mean); rho_cen analogous with *_cen_* fields."
94
+ }