Karthkal commited on
Commit
215b300
·
verified ·
1 Parent(s): 59f06d4

Round thresholds to 3 decimal, recompute recall/FPR

Browse files
Files changed (1) hide show
  1. README.md +8 -8
README.md CHANGED
@@ -29,7 +29,7 @@ flag = (pi >= pi_thresh) OR (pi >= pi_lower_bound AND toxic >= toxic_thresh)
29
 
30
  | Parameter | Value |
31
  |:----------|------:|
32
- | `pi_thresh` | 0.996056 |
33
  | `pi_lower_bound` | 0.50 |
34
  | `toxic_thresh` | 1.00 |
35
 
@@ -37,14 +37,14 @@ flag = (pi >= pi_thresh) OR (pi >= pi_lower_bound AND toxic >= toxic_thresh)
37
 
38
  | Dataset | Recall | FPR |
39
  |:--------|-------:|----:|
40
- | test (262K) | 43.28% | 0.122% |
41
- | customer_test (1.4M) | 42.24% | 0.376% |
42
 
43
  ### Thresholds at 0.5% FPR
44
 
45
  | Parameter | Value |
46
  |:----------|------:|
47
- | `pi_thresh` | 0.990000 |
48
  | `pi_lower_bound` | 0.50 |
49
  | `toxic_thresh` | 0.95 |
50
 
@@ -57,7 +57,7 @@ flag = (pi >= pi_thresh) OR (pi >= pi_lower_bound AND toxic >= toxic_thresh)
57
 
58
  | Parameter | Value |
59
  |:----------|------:|
60
- | `pi_thresh` | 0.981736 |
61
  | `pi_lower_bound` | 0.50 |
62
  | `toxic_thresh` | 0.90 |
63
 
@@ -72,7 +72,7 @@ flag = (pi >= pi_thresh) OR (pi >= pi_lower_bound AND toxic >= toxic_thresh)
72
 
73
  | Test FPR | pi-mmbert-v2 Recall | pi-mmbert-v3.5 Recall | Δ |
74
  |:---------|--------------------:|----------------------:|--:|
75
- | 0.122% | 35.31% | **43.28%** | +7.97pp |
76
  | 0.644% | 60.46% | **69.83%** | +9.37pp |
77
  | 1.008% | 67.13% | **75.26%** | +8.13pp |
78
 
@@ -118,7 +118,7 @@ chunk_overlap = 200
118
  stride = max_length - chunk_overlap # 312
119
 
120
  # --- Tiered thresholds (0.1% FPR — PI-only, no tier rescue) ---
121
- # pi_thresh = 0.9961
122
  # pi_lower_bound = 0.5
123
  # toxic_thresh = 1.00 # effectively disabled
124
  # --- Tiered thresholds (0.5% FPR) ---
@@ -126,7 +126,7 @@ pi_thresh = 0.990
126
  pi_lower_bound = 0.5
127
  toxic_thresh = 0.95
128
  # --- Tiered thresholds (1% FPR) ---
129
- # pi_thresh = 0.9817
130
  # pi_lower_bound = 0.5
131
  # toxic_thresh = 0.90
132
  # --- Thresholds for POV (test: recall=97.34%, FPR=9.336%) ---
 
29
 
30
  | Parameter | Value |
31
  |:----------|------:|
32
+ | `pi_thresh` | 0.996 |
33
  | `pi_lower_bound` | 0.50 |
34
  | `toxic_thresh` | 1.00 |
35
 
 
37
 
38
  | Dataset | Recall | FPR |
39
  |:--------|-------:|----:|
40
+ | test (262K) | 43.59% | 0.127% |
41
+ | customer_test (1.4M) | 42.42% | 0.388% |
42
 
43
  ### Thresholds at 0.5% FPR
44
 
45
  | Parameter | Value |
46
  |:----------|------:|
47
+ | `pi_thresh` | 0.990 |
48
  | `pi_lower_bound` | 0.50 |
49
  | `toxic_thresh` | 0.95 |
50
 
 
57
 
58
  | Parameter | Value |
59
  |:----------|------:|
60
+ | `pi_thresh` | 0.982 |
61
  | `pi_lower_bound` | 0.50 |
62
  | `toxic_thresh` | 0.90 |
63
 
 
72
 
73
  | Test FPR | pi-mmbert-v2 Recall | pi-mmbert-v3.5 Recall | Δ |
74
  |:---------|--------------------:|----------------------:|--:|
75
+ | 0.127% | 35.31% | **43.59%** | +8.28pp |
76
  | 0.644% | 60.46% | **69.83%** | +9.37pp |
77
  | 1.008% | 67.13% | **75.26%** | +8.13pp |
78
 
 
118
  stride = max_length - chunk_overlap # 312
119
 
120
  # --- Tiered thresholds (0.1% FPR — PI-only, no tier rescue) ---
121
+ # pi_thresh = 0.996
122
  # pi_lower_bound = 0.5
123
  # toxic_thresh = 1.00 # effectively disabled
124
  # --- Tiered thresholds (0.5% FPR) ---
 
126
  pi_lower_bound = 0.5
127
  toxic_thresh = 0.95
128
  # --- Tiered thresholds (1% FPR) ---
129
+ # pi_thresh = 0.982
130
  # pi_lower_bound = 0.5
131
  # toxic_thresh = 0.90
132
  # --- Thresholds for POV (test: recall=97.34%, FPR=9.336%) ---