Ganesh01kumar02reddy commited on
Commit
7beea9b
·
verified ·
1 Parent(s): 0e50069

uploaded: Full reproducibility metadata for multi-model experiment

Browse files
Files changed (1) hide show
  1. experiment_config_multimodel.json +104 -0
experiment_config_multimodel.json ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "experiment": "multi-model intersectional bias study",
3
+ "n_observations": 384,
4
+ "methods": [
5
+ "SEAT",
6
+ "CEAT",
7
+ "MLM"
8
+ ],
9
+ "datasets": [
10
+ "CareerFamily",
11
+ "CompetenceWarmth",
12
+ "PleasantUnpleasant"
13
+ ],
14
+ "pairs": [
15
+ "Hindu_Female",
16
+ "Hindu_Male",
17
+ "LowerCaste_Female",
18
+ "LowerCaste_Male",
19
+ "Muslim_Female",
20
+ "Muslim_Male",
21
+ "UpperCaste_Female",
22
+ "UpperCaste_Male"
23
+ ],
24
+ "n_bootstrap_ceat": 3000,
25
+ "ias_threshold": 0.05,
26
+ "within_method_z_normalisation": true,
27
+ "seed": 42,
28
+ "encoder_models": {
29
+ "SEAT": [
30
+ "mpnet-multi",
31
+ "BGE-M3",
32
+ "E5-large",
33
+ "Qwen3-Emb-4B",
34
+ "IndicSBERT",
35
+ "MuRIL"
36
+ ],
37
+ "CEAT": [
38
+ "mBERT",
39
+ "MuRIL",
40
+ "ModernBERT",
41
+ "XLM-R-large",
42
+ "IndicBERT-v2"
43
+ ],
44
+ "MLM": [
45
+ "mBERT",
46
+ "MuRIL",
47
+ "ModernBERT",
48
+ "XLM-R-large",
49
+ "IndicBERT-v2"
50
+ ]
51
+ },
52
+ "ml_models": [
53
+ "Ridge",
54
+ "RandomForest",
55
+ "GradBoosting",
56
+ "ElasticNet",
57
+ "GaussianProcess",
58
+ "MLP"
59
+ ],
60
+ "best_ml_model": "Ridge",
61
+ "best_lodo_r2": 0.9405090356483208,
62
+ "best_lomo_r2": 0.9386120562963216,
63
+ "best_lomoo_r2": 0.9377524162718508,
64
+ "cv_axes": {
65
+ "LODO": "3 groups (datasets)",
66
+ "LOMO": "3 groups (methods)",
67
+ "LOMOO": "10 groups (encoder models)"
68
+ },
69
+ "feature_cols_real": [
70
+ "score_A",
71
+ "score_B",
72
+ "abs_score_A",
73
+ "abs_score_B",
74
+ "max_single",
75
+ "score_A_times_B",
76
+ "score_A_sq",
77
+ "score_B_sq"
78
+ ],
79
+ "feature_cols_dummy": [
80
+ "dummy_gaussian",
81
+ "dummy_uniform",
82
+ "dummy_permuted_target"
83
+ ],
84
+ "hardware": {
85
+ "gpu": "NVIDIA T4 (15 GB VRAM)",
86
+ "ram": "12.7 GB",
87
+ "platform": "Google Colab free tier"
88
+ },
89
+ "key_findings": {
90
+ "linearity": "Ridge (linear) R\u00b2=0.941 beats all non-linear models",
91
+ "career_family_scale_artifact": "CareerFamily LODO R\u00b2 was negative at 72 rows without normalisation; positive (0.72) at 384 rows with z-norm",
92
+ "snr": 53.6,
93
+ "ablation_p": "< 10^-100 for all 6 conditions",
94
+ "muslim_male_lir_only": {
95
+ "IAS": -0.0470513346414766,
96
+ "LIR": 0.0781596722336495
97
+ },
98
+ "encoder_heterogeneity": {
99
+ "easiest": "mBERT (LOMOO R\u00b2=0.983), XLM-R-large (0.985)",
100
+ "hardest": "Qwen3-Emb-4B (0.429), mpnet-multi (0.539)",
101
+ "interpretation": "SEAT-only models encode bias differently from CEAT/MLM models"
102
+ }
103
+ }
104
+ }