hslee1981 commited on
Commit
3ed6290
·
verified ·
1 Parent(s): 76932f1

T18 Phase 1 Tier 1: config snapshot

Browse files
Files changed (1) hide show
  1. config.json +77 -0
config.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_type": "sklearn-logistic-regression",
3
+ "embedding_model": "openai-text-embedding-3-small",
4
+ "embedding_dim": 1536,
5
+ "max_pages_per_doc": 20,
6
+ "skip_boilerplate": true,
7
+ "operating_threshold": 0.436,
8
+ "decision_band": {
9
+ "reject_below": 0.3,
10
+ "fast_pass_at_or_above": 0.85,
11
+ "escalate_between": "(0.30, 0.85)"
12
+ },
13
+ "training_data": {
14
+ "source": "setfit_experiments PostgreSQL DB + multi-signal Phase 0 relabeling",
15
+ "n_pages": 7129,
16
+ "n_documents": 465,
17
+ "binary_class_balance": {
18
+ "positive": 3014,
19
+ "negative": 4115
20
+ },
21
+ "split": {
22
+ "train": 298,
23
+ "val": 64,
24
+ "test": 65,
25
+ "train_pos": 201,
26
+ "val_pos": 39,
27
+ "test_pos": 47
28
+ }
29
+ },
30
+ "test_metrics": {
31
+ "name": "logreg_tuned (TEST set)",
32
+ "threshold": 0.4359872072086175,
33
+ "accuracy": 0.9076923076923077,
34
+ "f1": 0.94,
35
+ "roc_auc": 0.9550827423167849,
36
+ "brier_score": 0.13433461274707392,
37
+ "ece": 0.27835753511850964,
38
+ "confusion_matrix": [
39
+ [
40
+ 12,
41
+ 6
42
+ ],
43
+ [
44
+ 0,
45
+ 47
46
+ ]
47
+ ]
48
+ },
49
+ "validation_metrics": {
50
+ "name": "LogReg @ best-threshold",
51
+ "threshold": 0.4359872072086175,
52
+ "accuracy": 0.859375,
53
+ "f1": 0.8941176470588236,
54
+ "roc_auc": 0.8748717948717949,
55
+ "brier_score": 0.15576505514468417,
56
+ "ece": 0.19068488965598734,
57
+ "confusion_matrix": [
58
+ [
59
+ 17,
60
+ 8
61
+ ],
62
+ [
63
+ 1,
64
+ 38
65
+ ]
66
+ ]
67
+ },
68
+ "candidates_compared": [
69
+ "logreg_05",
70
+ "logreg_tuned",
71
+ "logreg_platt_05",
72
+ "logreg_platt_tuned",
73
+ "mlp_05",
74
+ "mlp_tuned"
75
+ ],
76
+ "winner_selection_rule": "max F1 across LogReg, LogReg+Platt, MLP at best threshold; simplicity tiebreak to LogReg"
77
+ }