Alasil commited on
Commit
110eeae
·
verified ·
1 Parent(s): 4f88dcb

[Run 20260607_182656] Log metrics

Browse files
metrics/UBC_NLP_MARBERT/20260607_182656_metrics.json ADDED
@@ -0,0 +1,320 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_id": "20260607_182656",
3
+ "hyperparameters": {
4
+ "model_architecture": "UBC-NLP/MARBERT",
5
+ "task": "sequence-classification",
6
+ "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
7
+ "learning_rate": 3e-05,
8
+ "batch_size": 64,
9
+ "num_epochs": 5,
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": 0.1804913546815938,
19
+ "val_loss": 0.09567254954507869,
20
+ "val_accuracy": 0.9648783185840708,
21
+ "val_f1_weighted": 0.9648518196186497,
22
+ "val_f1_macro": 0.964820179062923,
23
+ "val_precision_macro": 0.9658552870911443,
24
+ "val_recall_macro": 0.9644509901862843,
25
+ "hate_recall": 0.9451357466063348,
26
+ "hate_f1": 0.9633900259440761,
27
+ "classification_report": {
28
+ "Hate": {
29
+ "precision": 0.982363315696649,
30
+ "recall": 0.9451357466063348,
31
+ "f1-score": 0.9633900259440761,
32
+ "support": 1768.0
33
+ },
34
+ "Normal": {
35
+ "precision": 0.9493472584856397,
36
+ "recall": 0.9837662337662337,
37
+ "f1-score": 0.9662503321817698,
38
+ "support": 1848.0
39
+ },
40
+ "accuracy": 0.9648783185840708,
41
+ "macro avg": {
42
+ "precision": 0.9658552870911443,
43
+ "recall": 0.9644509901862843,
44
+ "f1-score": 0.964820179062923,
45
+ "support": 3616.0
46
+ },
47
+ "weighted avg": {
48
+ "precision": 0.9654900652193411,
49
+ "recall": 0.9648783185840708,
50
+ "f1-score": 0.9648518196186497,
51
+ "support": 3616.0
52
+ }
53
+ }
54
+ },
55
+ {
56
+ "epoch": 2,
57
+ "train_loss": 0.08093556008657976,
58
+ "val_loss": 0.09877155760401174,
59
+ "val_accuracy": 0.9662610619469026,
60
+ "val_f1_weighted": 0.9662228318870574,
61
+ "val_f1_macro": 0.9661880772871978,
62
+ "val_precision_macro": 0.9678227075954349,
63
+ "val_recall_macro": 0.9657058627646863,
64
+ "hate_recall": 0.9406108597285068,
65
+ "hate_f1": 0.9646171693735499,
66
+ "classification_report": {
67
+ "Hate": {
68
+ "precision": 0.9898809523809524,
69
+ "recall": 0.9406108597285068,
70
+ "f1-score": 0.9646171693735499,
71
+ "support": 1768.0
72
+ },
73
+ "Normal": {
74
+ "precision": 0.9457644628099173,
75
+ "recall": 0.9908008658008658,
76
+ "f1-score": 0.9677589852008457,
77
+ "support": 1848.0
78
+ },
79
+ "accuracy": 0.9662610619469026,
80
+ "macro avg": {
81
+ "precision": 0.9678227075954349,
82
+ "recall": 0.9657058627646863,
83
+ "f1-score": 0.9661880772871978,
84
+ "support": 3616.0
85
+ },
86
+ "weighted avg": {
87
+ "precision": 0.9673346933302686,
88
+ "recall": 0.9662610619469026,
89
+ "f1-score": 0.9662228318870574,
90
+ "support": 3616.0
91
+ }
92
+ }
93
+ },
94
+ {
95
+ "epoch": 3,
96
+ "train_loss": 0.05505622040646179,
97
+ "val_loss": 0.12153002150201549,
98
+ "val_accuracy": 0.973174778761062,
99
+ "val_f1_weighted": 0.9731545393937718,
100
+ "val_f1_macro": 0.9731303729850671,
101
+ "val_precision_macro": 0.974180907253144,
102
+ "val_recall_macro": 0.9727515131926897,
103
+ "hate_recall": 0.9536199095022625,
104
+ "hate_f1": 0.9720380513116171,
105
+ "classification_report": {
106
+ "Hate": {
107
+ "precision": 0.9911816578483245,
108
+ "recall": 0.9536199095022625,
109
+ "f1-score": 0.9720380513116171,
110
+ "support": 1768.0
111
+ },
112
+ "Normal": {
113
+ "precision": 0.9571801566579634,
114
+ "recall": 0.9918831168831169,
115
+ "f1-score": 0.9742226946585172,
116
+ "support": 1848.0
117
+ },
118
+ "accuracy": 0.973174778761062,
119
+ "macro avg": {
120
+ "precision": 0.974180907253144,
121
+ "recall": 0.9727515131926897,
122
+ "f1-score": 0.9731303729850671,
123
+ "support": 3616.0
124
+ },
125
+ "weighted avg": {
126
+ "precision": 0.9738047844523656,
127
+ "recall": 0.973174778761062,
128
+ "f1-score": 0.9731545393937718,
129
+ "support": 3616.0
130
+ }
131
+ }
132
+ },
133
+ {
134
+ "epoch": 4,
135
+ "train_loss": 0.03173725027910691,
136
+ "val_loss": 0.11768527087712612,
137
+ "val_accuracy": 0.9737278761061947,
138
+ "val_f1_weighted": 0.9737286299067954,
139
+ "val_f1_macro": 0.9737165690971827,
140
+ "val_precision_macro": 0.9736912492475454,
141
+ "val_recall_macro": 0.97374561712797,
142
+ "hate_recall": 0.9745475113122172,
143
+ "hate_f1": 0.9731714205026829,
144
+ "classification_report": {
145
+ "Hate": {
146
+ "precision": 0.9717992103778906,
147
+ "recall": 0.9745475113122172,
148
+ "f1-score": 0.9731714205026829,
149
+ "support": 1768.0
150
+ },
151
+ "Normal": {
152
+ "precision": 0.9755832881172002,
153
+ "recall": 0.9729437229437229,
154
+ "f1-score": 0.9742617176916825,
155
+ "support": 1848.0
156
+ },
157
+ "accuracy": 0.9737278761061947,
158
+ "macro avg": {
159
+ "precision": 0.9736912492475454,
160
+ "recall": 0.97374561712797,
161
+ "f1-score": 0.9737165690971827,
162
+ "support": 3616.0
163
+ },
164
+ "weighted avg": {
165
+ "precision": 0.9737331085145732,
166
+ "recall": 0.9737278761061947,
167
+ "f1-score": 0.9737286299067954,
168
+ "support": 3616.0
169
+ }
170
+ }
171
+ },
172
+ {
173
+ "epoch": 5,
174
+ "train_loss": 0.02443531528275123,
175
+ "val_loss": 0.11402751311400432,
176
+ "val_accuracy": 0.9764933628318584,
177
+ "val_f1_weighted": 0.9764947278348372,
178
+ "val_f1_macro": 0.9764848005404456,
179
+ "val_precision_macro": 0.976438507317727,
180
+ "val_recall_macro": 0.9765491861080096,
181
+ "hate_recall": 0.9790723981900452,
182
+ "hate_f1": 0.9760360868339442,
183
+ "classification_report": {
184
+ "Hate": {
185
+ "precision": 0.9730185497470489,
186
+ "recall": 0.9790723981900452,
187
+ "f1-score": 0.9760360868339442,
188
+ "support": 1768.0
189
+ },
190
+ "Normal": {
191
+ "precision": 0.979858464888405,
192
+ "recall": 0.974025974025974,
193
+ "f1-score": 0.9769335142469471,
194
+ "support": 1848.0
195
+ },
196
+ "accuracy": 0.9764933628318584,
197
+ "macro avg": {
198
+ "precision": 0.976438507317727,
199
+ "recall": 0.9765491861080096,
200
+ "f1-score": 0.9764848005404456,
201
+ "support": 3616.0
202
+ },
203
+ "weighted avg": {
204
+ "precision": 0.9765141700958393,
205
+ "recall": 0.9764933628318584,
206
+ "f1-score": 0.9764947278348372,
207
+ "support": 3616.0
208
+ }
209
+ }
210
+ }
211
+ ],
212
+ "best_val_macro_f1": 0.9764848005404456,
213
+ "best_val_weighted_f1": 0.9764947278348372,
214
+ "best_epoch": 5,
215
+ "test_metrics": {
216
+ "overall": {
217
+ "accuracy": 0.9745575221238938,
218
+ "f1_macro": 0.97454567824682,
219
+ "f1_weighted": 0.9745578258130496,
220
+ "precision_macro": 0.9745346481321655,
221
+ "recall_macro": 0.9745573054396584,
222
+ "classification_report": {
223
+ "Hate": {
224
+ "precision": 0.9734463276836158,
225
+ "recall": 0.9745475113122172,
226
+ "f1-score": 0.9739966082532504,
227
+ "support": 1768.0
228
+ },
229
+ "Normal": {
230
+ "precision": 0.9756229685807151,
231
+ "recall": 0.9745670995670995,
232
+ "f1-score": 0.9750947482403898,
233
+ "support": 1848.0
234
+ },
235
+ "accuracy": 0.9745575221238938,
236
+ "macro avg": {
237
+ "precision": 0.9745346481321655,
238
+ "recall": 0.9745573054396584,
239
+ "f1-score": 0.97454567824682,
240
+ "support": 3616.0
241
+ },
242
+ "weighted avg": {
243
+ "precision": 0.9745587260181954,
244
+ "recall": 0.9745575221238938,
245
+ "f1-score": 0.9745578258130496,
246
+ "support": 3616.0
247
+ }
248
+ }
249
+ },
250
+ "arabic": {
251
+ "accuracy": 0.9766666666666667,
252
+ "f1_macro": 0.9766610192095866,
253
+ "f1_weighted": 0.9766682802258324,
254
+ "precision_macro": 0.9766143968391159,
255
+ "recall_macro": 0.9767462540571785,
256
+ "classification_report": {
257
+ "Hate": {
258
+ "precision": 0.9719101123595506,
259
+ "recall": 0.9807256235827665,
260
+ "f1-score": 0.9762979683972912,
261
+ "support": 882.0
262
+ },
263
+ "Normal": {
264
+ "precision": 0.9813186813186813,
265
+ "recall": 0.9727668845315904,
266
+ "f1-score": 0.9770240700218819,
267
+ "support": 918.0
268
+ },
269
+ "accuracy": 0.9766666666666667,
270
+ "macro avg": {
271
+ "precision": 0.9766143968391159,
272
+ "recall": 0.9767462540571785,
273
+ "f1-score": 0.9766610192095866,
274
+ "support": 1800.0
275
+ },
276
+ "weighted avg": {
277
+ "precision": 0.9767084825287073,
278
+ "recall": 0.9766666666666667,
279
+ "f1-score": 0.9766682802258324,
280
+ "support": 1800.0
281
+ }
282
+ }
283
+ },
284
+ "english": {
285
+ "accuracy": 0.9724669603524229,
286
+ "f1_macro": 0.9724460725942812,
287
+ "f1_weighted": 0.972464453821446,
288
+ "precision_macro": 0.972542735042735,
289
+ "recall_macro": 0.9723706886089468,
290
+ "classification_report": {
291
+ "Hate": {
292
+ "precision": 0.975,
293
+ "recall": 0.9683972911963883,
294
+ "f1-score": 0.9716874292185731,
295
+ "support": 886.0
296
+ },
297
+ "Normal": {
298
+ "precision": 0.9700854700854701,
299
+ "recall": 0.9763440860215054,
300
+ "f1-score": 0.9732047159699893,
301
+ "support": 930.0
302
+ },
303
+ "accuracy": 0.9724669603524229,
304
+ "macro avg": {
305
+ "precision": 0.972542735042735,
306
+ "recall": 0.9723706886089468,
307
+ "f1-score": 0.9724460725942812,
308
+ "support": 1816.0
309
+ },
310
+ "weighted avg": {
311
+ "precision": 0.9724831977860612,
312
+ "recall": 0.9724669603524229,
313
+ "f1-score": 0.972464453821446,
314
+ "support": 1816.0
315
+ }
316
+ }
317
+ }
318
+ },
319
+ "notes": "MARBERT sweep. LR=3e-05, Batch=64. Early stopping based on validation macro F1."
320
+ }