Alasil commited on
Commit
64ef085
·
verified ·
1 Parent(s): c5018d7

[Run 20260607_160004] Log metrics

Browse files
metrics/UBC_NLP_MARBERT/20260607_160004_metrics.json ADDED
@@ -0,0 +1,320 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "run_id": "20260607_160004",
3
+ "hyperparameters": {
4
+ "model_architecture": "UBC-NLP/MARBERT",
5
+ "task": "sequence-classification",
6
+ "dataset": "Nexus-Analytics/multilingual-hate-detection-dataset_v2",
7
+ "learning_rate": 2e-05,
8
+ "batch_size": 16,
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": 0.21311009175933285,
19
+ "val_loss": 0.17298398119607628,
20
+ "val_accuracy": 0.9612831858407079,
21
+ "val_f1_weighted": 0.9612493806305092,
22
+ "val_f1_macro": 0.9612128344573214,
23
+ "val_precision_macro": 0.9624241523304999,
24
+ "val_recall_macro": 0.9608112475759534,
25
+ "hate_recall": 0.9394796380090498,
26
+ "hate_f1": 0.9595609474292317,
27
+ "classification_report": {
28
+ "Hate": {
29
+ "precision": 0.9805194805194806,
30
+ "recall": 0.9394796380090498,
31
+ "f1-score": 0.9595609474292317,
32
+ "support": 1768.0
33
+ },
34
+ "Normal": {
35
+ "precision": 0.9443288241415192,
36
+ "recall": 0.9821428571428571,
37
+ "f1-score": 0.9628647214854111,
38
+ "support": 1848.0
39
+ },
40
+ "accuracy": 0.9612831858407079,
41
+ "macro avg": {
42
+ "precision": 0.9624241523304999,
43
+ "recall": 0.9608112475759534,
44
+ "f1-score": 0.9612128344573214,
45
+ "support": 3616.0
46
+ },
47
+ "weighted avg": {
48
+ "precision": 0.9620238132112747,
49
+ "recall": 0.9612831858407079,
50
+ "f1-score": 0.9612493806305092,
51
+ "support": 3616.0
52
+ }
53
+ }
54
+ },
55
+ {
56
+ "epoch": 2,
57
+ "train_loss": 0.1010062225904647,
58
+ "val_loss": 0.1437140431204106,
59
+ "val_accuracy": 0.9698561946902655,
60
+ "val_f1_weighted": 0.9698482544191077,
61
+ "val_f1_macro": 0.969827763396765,
62
+ "val_precision_macro": 0.9701468186234363,
63
+ "val_recall_macro": 0.9696516718575543,
64
+ "hate_recall": 0.9604072398190046,
65
+ "hate_f1": 0.9689015691868759,
66
+ "classification_report": {
67
+ "Hate": {
68
+ "precision": 0.9775474956822107,
69
+ "recall": 0.9604072398190046,
70
+ "f1-score": 0.9689015691868759,
71
+ "support": 1768.0
72
+ },
73
+ "Normal": {
74
+ "precision": 0.9627461415646621,
75
+ "recall": 0.9788961038961039,
76
+ "f1-score": 0.9707539576066542,
77
+ "support": 1848.0
78
+ },
79
+ "accuracy": 0.9698561946902655,
80
+ "macro avg": {
81
+ "precision": 0.9701468186234363,
82
+ "recall": 0.9696516718575543,
83
+ "f1-score": 0.969827763396765,
84
+ "support": 3616.0
85
+ },
86
+ "weighted avg": {
87
+ "precision": 0.9699830868301006,
88
+ "recall": 0.9698561946902655,
89
+ "f1-score": 0.9698482544191077,
90
+ "support": 3616.0
91
+ }
92
+ }
93
+ },
94
+ {
95
+ "epoch": 3,
96
+ "train_loss": 0.07073115703309252,
97
+ "val_loss": 0.11939409197403032,
98
+ "val_accuracy": 0.9745575221238938,
99
+ "val_f1_weighted": 0.9745568679755232,
100
+ "val_f1_macro": 0.9745437850081111,
101
+ "val_precision_macro": 0.9745693813882644,
102
+ "val_recall_macro": 0.9745205774617539,
103
+ "hate_recall": 0.9728506787330317,
104
+ "hate_f1": 0.9739524348810872,
105
+ "classification_report": {
106
+ "Hate": {
107
+ "precision": 0.9750566893424036,
108
+ "recall": 0.9728506787330317,
109
+ "f1-score": 0.9739524348810872,
110
+ "support": 1768.0
111
+ },
112
+ "Normal": {
113
+ "precision": 0.9740820734341252,
114
+ "recall": 0.9761904761904762,
115
+ "f1-score": 0.9751351351351352,
116
+ "support": 1848.0
117
+ },
118
+ "accuracy": 0.9745575221238938,
119
+ "macro avg": {
120
+ "precision": 0.9745693813882644,
121
+ "recall": 0.9745205774617539,
122
+ "f1-score": 0.9745437850081111,
123
+ "support": 3616.0
124
+ },
125
+ "weighted avg": {
126
+ "precision": 0.9745586002388366,
127
+ "recall": 0.9745575221238938,
128
+ "f1-score": 0.9745568679755232,
129
+ "support": 3616.0
130
+ }
131
+ }
132
+ },
133
+ {
134
+ "epoch": 4,
135
+ "train_loss": 0.050746984494552255,
136
+ "val_loss": 0.20435507836184147,
137
+ "val_accuracy": 0.9659845132743363,
138
+ "val_f1_weighted": 0.9659588489219992,
139
+ "val_f1_macro": 0.9659282049192088,
140
+ "val_precision_macro": 0.9669653697794109,
141
+ "val_recall_macro": 0.9655577265871383,
142
+ "hate_recall": 0.9462669683257918,
143
+ "hate_f1": 0.9645430959930816,
144
+ "classification_report": {
145
+ "Hate": {
146
+ "precision": 0.9835390946502057,
147
+ "recall": 0.9462669683257918,
148
+ "f1-score": 0.9645430959930816,
149
+ "support": 1768.0
150
+ },
151
+ "Normal": {
152
+ "precision": 0.9503916449086162,
153
+ "recall": 0.9848484848484849,
154
+ "f1-score": 0.9673133138453361,
155
+ "support": 1848.0
156
+ },
157
+ "accuracy": 0.9659845132743363,
158
+ "macro avg": {
159
+ "precision": 0.9669653697794109,
160
+ "recall": 0.9655577265871383,
161
+ "f1-score": 0.9659282049192088,
162
+ "support": 3616.0
163
+ },
164
+ "weighted avg": {
165
+ "precision": 0.9665986944504111,
166
+ "recall": 0.9659845132743363,
167
+ "f1-score": 0.9659588489219992,
168
+ "support": 3616.0
169
+ }
170
+ }
171
+ },
172
+ {
173
+ "epoch": 5,
174
+ "train_loss": 0.043556458314452796,
175
+ "val_loss": 0.10186352706988203,
176
+ "val_accuracy": 0.9734513274336283,
177
+ "val_f1_weighted": 0.9734381126209277,
178
+ "val_f1_macro": 0.9734169689206065,
179
+ "val_precision_macro": 0.9740754007343064,
180
+ "val_recall_macro": 0.9731322598969658,
181
+ "hate_recall": 0.958710407239819,
182
+ "hate_f1": 0.9724612736660929,
183
+ "classification_report": {
184
+ "Hate": {
185
+ "precision": 0.9866123399301513,
186
+ "recall": 0.958710407239819,
187
+ "f1-score": 0.9724612736660929,
188
+ "support": 1768.0
189
+ },
190
+ "Normal": {
191
+ "precision": 0.9615384615384616,
192
+ "recall": 0.9875541125541125,
193
+ "f1-score": 0.9743726641751201,
194
+ "support": 1848.0
195
+ },
196
+ "accuracy": 0.9734513274336283,
197
+ "macro avg": {
198
+ "precision": 0.9740754007343064,
199
+ "recall": 0.9731322598969658,
200
+ "f1-score": 0.9734169689206065,
201
+ "support": 3616.0
202
+ },
203
+ "weighted avg": {
204
+ "precision": 0.9737980348228938,
205
+ "recall": 0.9734513274336283,
206
+ "f1-score": 0.9734381126209277,
207
+ "support": 3616.0
208
+ }
209
+ }
210
+ }
211
+ ],
212
+ "best_val_macro_f1": 0.9745437850081111,
213
+ "best_val_weighted_f1": 0.9745568679755232,
214
+ "best_epoch": 3,
215
+ "test_metrics": {
216
+ "overall": {
217
+ "accuracy": 0.9759402654867256,
218
+ "f1_macro": 0.9759281867365222,
219
+ "f1_weighted": 0.9759401163663528,
220
+ "precision_macro": 0.9759341610188349,
221
+ "recall_macro": 0.9759223619517737,
222
+ "classification_report": {
223
+ "Hate": {
224
+ "precision": 0.9756649688737974,
225
+ "recall": 0.9751131221719457,
226
+ "f1-score": 0.9753889674681754,
227
+ "support": 1768.0
228
+ },
229
+ "Normal": {
230
+ "precision": 0.9762033531638724,
231
+ "recall": 0.9767316017316018,
232
+ "f1-score": 0.9764674060048688,
233
+ "support": 1848.0
234
+ },
235
+ "accuracy": 0.9759402654867256,
236
+ "macro avg": {
237
+ "precision": 0.9759341610188349,
238
+ "recall": 0.9759223619517737,
239
+ "f1-score": 0.9759281867365222,
240
+ "support": 3616.0
241
+ },
242
+ "weighted avg": {
243
+ "precision": 0.975940116597265,
244
+ "recall": 0.9759402654867256,
245
+ "f1-score": 0.9759401163663528,
246
+ "support": 3616.0
247
+ }
248
+ }
249
+ },
250
+ "arabic": {
251
+ "accuracy": 0.9788888888888889,
252
+ "f1_macro": 0.9788844833057762,
253
+ "f1_weighted": 0.9788905833439322,
254
+ "precision_macro": 0.9788378341004722,
255
+ "recall_macro": 0.9789915966386554,
256
+ "classification_report": {
257
+ "Hate": {
258
+ "precision": 0.9730941704035875,
259
+ "recall": 0.9841269841269841,
260
+ "f1-score": 0.9785794813979707,
261
+ "support": 882.0
262
+ },
263
+ "Normal": {
264
+ "precision": 0.9845814977973568,
265
+ "recall": 0.9738562091503268,
266
+ "f1-score": 0.9791894852135816,
267
+ "support": 918.0
268
+ },
269
+ "accuracy": 0.9788888888888889,
270
+ "macro avg": {
271
+ "precision": 0.9788378341004722,
272
+ "recall": 0.9789915966386554,
273
+ "f1-score": 0.9788844833057762,
274
+ "support": 1800.0
275
+ },
276
+ "weighted avg": {
277
+ "precision": 0.9789527073744098,
278
+ "recall": 0.9788888888888889,
279
+ "f1-score": 0.9788905833439322,
280
+ "support": 1800.0
281
+ }
282
+ }
283
+ },
284
+ "english": {
285
+ "accuracy": 0.9730176211453745,
286
+ "f1_macro": 0.9729928484951125,
287
+ "f1_weighted": 0.973012666615322,
288
+ "precision_macro": 0.9732023683011993,
289
+ "recall_macro": 0.9728549236631957,
290
+ "classification_report": {
291
+ "Hate": {
292
+ "precision": 0.9782857142857143,
293
+ "recall": 0.9661399548532731,
294
+ "f1-score": 0.9721749006246451,
295
+ "support": 886.0
296
+ },
297
+ "Normal": {
298
+ "precision": 0.9681190223166843,
299
+ "recall": 0.9795698924731183,
300
+ "f1-score": 0.9738107963655799,
301
+ "support": 930.0
302
+ },
303
+ "accuracy": 0.9730176211453745,
304
+ "macro avg": {
305
+ "precision": 0.9732023683011993,
306
+ "recall": 0.9728549236631957,
307
+ "f1-score": 0.9729928484951125,
308
+ "support": 1816.0
309
+ },
310
+ "weighted avg": {
311
+ "precision": 0.9730792035306494,
312
+ "recall": 0.9730176211453745,
313
+ "f1-score": 0.973012666615322,
314
+ "support": 1816.0
315
+ }
316
+ }
317
+ }
318
+ },
319
+ "notes": "MARBERT sweep. LR=2e-05, Batch=16. Early stopping based on validation macro F1."
320
+ }