Alasil commited on
Commit
5c06010
·
verified ·
1 Parent(s): 0a472a1

[Run 20260607_195908] Log metrics

Browse files
metrics/UBC_NLP_MARBERT/20260607_195908_metrics.json ADDED
@@ -0,0 +1,242 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_id": "20260607_195908",
3
+ "hyperparameters": {
4
+ "model_architecture": "UBC-NLP/MARBERT",
5
+ "task": "sequence-classification",
6
+ "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
7
+ "learning_rate": 5e-05,
8
+ "batch_size": 128,
9
+ "num_epochs": 6,
10
+ "max_seq_length": 128,
11
+ "early_stopping_patience": 2,
12
+ "optimizer": "AdamW",
13
+ "seed": 42
14
+ },
15
+ "per_epoch_results": [
16
+ {
17
+ "epoch": 1,
18
+ "train_loss": 0.21657190260223366,
19
+ "val_loss": 0.08153876662254333,
20
+ "val_accuracy": 0.9715154867256637,
21
+ "val_f1_weighted": 0.971511385946218,
22
+ "val_f1_macro": 0.9714941195064465,
23
+ "val_precision_macro": 0.9716567704309111,
24
+ "val_recall_macro": 0.9713852324146441,
25
+ "hate_recall": 0.9654977375565611,
26
+ "hate_f1": 0.9707136764287745,
27
+ "classification_report": {
28
+ "Hate": {
29
+ "precision": 0.9759862778730704,
30
+ "recall": 0.9654977375565611,
31
+ "f1-score": 0.9707136764287745,
32
+ "support": 1768.0
33
+ },
34
+ "Normal": {
35
+ "precision": 0.967327262988752,
36
+ "recall": 0.9772727272727273,
37
+ "f1-score": 0.9722745625841185,
38
+ "support": 1848.0
39
+ },
40
+ "accuracy": 0.9715154867256637,
41
+ "macro avg": {
42
+ "precision": 0.9716567704309111,
43
+ "recall": 0.9713852324146441,
44
+ "f1-score": 0.9714941195064465,
45
+ "support": 3616.0
46
+ },
47
+ "weighted avg": {
48
+ "precision": 0.9715609848680317,
49
+ "recall": 0.9715154867256637,
50
+ "f1-score": 0.971511385946218,
51
+ "support": 3616.0
52
+ }
53
+ }
54
+ },
55
+ {
56
+ "epoch": 2,
57
+ "train_loss": 0.06782547577114945,
58
+ "val_loss": 0.0781356720574971,
59
+ "val_accuracy": 0.9695796460176991,
60
+ "val_f1_weighted": 0.9695820291241608,
61
+ "val_f1_macro": 0.9695701135918527,
62
+ "val_precision_macro": 0.9695155384103238,
63
+ "val_recall_macro": 0.9696626902509256,
64
+ "hate_recall": 0.9734162895927602,
65
+ "hate_f1": 0.9690315315315315,
66
+ "classification_report": {
67
+ "Hate": {
68
+ "precision": 0.9646860986547086,
69
+ "recall": 0.9734162895927602,
70
+ "f1-score": 0.9690315315315315,
71
+ "support": 1768.0
72
+ },
73
+ "Normal": {
74
+ "precision": 0.9743449781659389,
75
+ "recall": 0.9659090909090909,
76
+ "f1-score": 0.970108695652174,
77
+ "support": 1848.0
78
+ },
79
+ "accuracy": 0.9695796460176991,
80
+ "macro avg": {
81
+ "precision": 0.9695155384103238,
82
+ "recall": 0.9696626902509256,
83
+ "f1-score": 0.9695701135918527,
84
+ "support": 3616.0
85
+ },
86
+ "weighted avg": {
87
+ "precision": 0.9696223844226161,
88
+ "recall": 0.9695796460176991,
89
+ "f1-score": 0.9695820291241608,
90
+ "support": 3616.0
91
+ }
92
+ }
93
+ },
94
+ {
95
+ "epoch": 3,
96
+ "train_loss": 0.055689407923526946,
97
+ "val_loss": 0.11876626887583527,
98
+ "val_accuracy": 0.9601769911504425,
99
+ "val_f1_weighted": 0.9601742500619426,
100
+ "val_f1_macro": 0.9601766865850536,
101
+ "val_precision_macro": 0.9608353508672027,
102
+ "val_recall_macro": 0.9607084092378211,
103
+ "hate_recall": 0.9847285067873304,
104
+ "hate_f1": 0.9602868174296746,
105
+ "classification_report": {
106
+ "Hate": {
107
+ "precision": 0.9370290635091496,
108
+ "recall": 0.9847285067873304,
109
+ "f1-score": 0.9602868174296746,
110
+ "support": 1768.0
111
+ },
112
+ "Normal": {
113
+ "precision": 0.984641638225256,
114
+ "recall": 0.9366883116883117,
115
+ "f1-score": 0.9600665557404326,
116
+ "support": 1848.0
117
+ },
118
+ "accuracy": 0.9601769911504425,
119
+ "macro avg": {
120
+ "precision": 0.9608353508672027,
121
+ "recall": 0.9607084092378211,
122
+ "f1-score": 0.9601766865850536,
123
+ "support": 3616.0
124
+ },
125
+ "weighted avg": {
126
+ "precision": 0.961362038640611,
127
+ "recall": 0.9601769911504425,
128
+ "f1-score": 0.9601742500619426,
129
+ "support": 3616.0
130
+ }
131
+ }
132
+ }
133
+ ],
134
+ "best_val_macro_f1": 0.9714941195064465,
135
+ "best_val_weighted_f1": 0.971511385946218,
136
+ "best_epoch": 1,
137
+ "test_metrics": {
138
+ "overall": {
139
+ "accuracy": 0.9679203539823009,
140
+ "f1_macro": 0.9678967739723365,
141
+ "f1_weighted": 0.9679160229600626,
142
+ "precision_macro": 0.9680471596998927,
143
+ "recall_macro": 0.9677944604415193,
144
+ "classification_report": {
145
+ "Hate": {
146
+ "precision": 0.972,
147
+ "recall": 0.9621040723981901,
148
+ "f1-score": 0.9670267197271177,
149
+ "support": 1768.0
150
+ },
151
+ "Normal": {
152
+ "precision": 0.9640943193997856,
153
+ "recall": 0.9734848484848485,
154
+ "f1-score": 0.9687668282175552,
155
+ "support": 1848.0
156
+ },
157
+ "accuracy": 0.9679203539823009,
158
+ "macro avg": {
159
+ "precision": 0.9680471596998927,
160
+ "recall": 0.9677944604415193,
161
+ "f1-score": 0.9678967739723365,
162
+ "support": 3616.0
163
+ },
164
+ "weighted avg": {
165
+ "precision": 0.9679597074808638,
166
+ "recall": 0.9679203539823009,
167
+ "f1-score": 0.9679160229600626,
168
+ "support": 3616.0
169
+ }
170
+ }
171
+ },
172
+ "arabic": {
173
+ "accuracy": 0.9733333333333334,
174
+ "f1_macro": 0.9733306663999733,
175
+ "f1_weighted": 0.9733360002666933,
176
+ "precision_macro": 0.9733333333333334,
177
+ "recall_macro": 0.9735227424303055,
178
+ "classification_report": {
179
+ "Hate": {
180
+ "precision": 0.9633333333333334,
181
+ "recall": 0.9829931972789115,
182
+ "f1-score": 0.9730639730639731,
183
+ "support": 882.0
184
+ },
185
+ "Normal": {
186
+ "precision": 0.9833333333333333,
187
+ "recall": 0.9640522875816994,
188
+ "f1-score": 0.9735973597359736,
189
+ "support": 918.0
190
+ },
191
+ "accuracy": 0.9733333333333334,
192
+ "macro avg": {
193
+ "precision": 0.9733333333333334,
194
+ "recall": 0.9735227424303055,
195
+ "f1-score": 0.9733306663999733,
196
+ "support": 1800.0
197
+ },
198
+ "weighted avg": {
199
+ "precision": 0.9735333333333334,
200
+ "recall": 0.9733333333333334,
201
+ "f1-score": 0.9733360002666933,
202
+ "support": 1800.0
203
+ }
204
+ }
205
+ },
206
+ "english": {
207
+ "accuracy": 0.9625550660792952,
208
+ "f1_macro": 0.962482257092302,
209
+ "f1_weighted": 0.9625223020351482,
210
+ "precision_macro": 0.9636731214224821,
211
+ "recall_macro": 0.962052477001869,
212
+ "classification_report": {
213
+ "Hate": {
214
+ "precision": 0.9811764705882353,
215
+ "recall": 0.9413092550790068,
216
+ "f1-score": 0.9608294930875576,
217
+ "support": 886.0
218
+ },
219
+ "Normal": {
220
+ "precision": 0.9461697722567288,
221
+ "recall": 0.9827956989247312,
222
+ "f1-score": 0.9641350210970464,
223
+ "support": 930.0
224
+ },
225
+ "accuracy": 0.9625550660792952,
226
+ "macro avg": {
227
+ "precision": 0.9636731214224821,
228
+ "recall": 0.962052477001869,
229
+ "f1-score": 0.962482257092302,
230
+ "support": 1816.0
231
+ },
232
+ "weighted avg": {
233
+ "precision": 0.9632490314647215,
234
+ "recall": 0.9625550660792952,
235
+ "f1-score": 0.9625223020351482,
236
+ "support": 1816.0
237
+ }
238
+ }
239
+ }
240
+ },
241
+ "notes": "MARBERT sweep. LR=5e-05, Batch=128. Early stopping based on validation macro F1."
242
+ }