Hugging Face's logo

PulpBio
/
FEMBA

PyTorch
English
femba
eeg
biosignal
mamba
state-space-model
bidirectional
foundation-model
self-supervised
masked-modeling
efficient
edge-ai
neuroscience
Eval Results (legacy)
Model card Files
xet
 Community
thorir commited on
Commit
8c84bf1
·
verified ·
1 Parent(s): c7437ad

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +293 -0
README.md CHANGED
@@ -2,6 +2,293 @@
2
  license: cc-by-nd-4.0
3
  language:
4
  - en
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  ---
6
 
7
  <div align="center">
@@ -211,6 +498,12 @@ If you use FEMBA in your research, please cite:
211
 
212
  ---
213
 
 
 
 
 
 
 
214
  ## 🗒️ Changelog
215
 
216
  - **v1.0:** Initial release of FEMBA model card with task-specific checkpoints and instructions.
 
2
  license: cc-by-nd-4.0
3
  language:
4
  - en
5
+ library_name: pytorch
6
+ tags:
7
+ - eeg
8
+ - biosignal
9
+ - mamba
10
+ - state-space-model
11
+ - bidirectional
12
+ - foundation-model
13
+ - self-supervised
14
+ - masked-modeling
15
+ - efficient
16
+ - edge-ai
17
+ - neuroscience
18
+ datasets:
19
+ - TUEG
20
+ - TUAB
21
+ - TUAR
22
+ - TUSL
23
+ metrics:
24
+ - balanced_accuracy
25
+ - roc_auc
26
+ - pr_auc
27
+ thumbnail: https://raw.githubusercontent.com/pulp-bio/BioFoundation/refs/heads/main/docs/model/logo/FEMBA_logo.png
28
+ model-index:
29
+ - name: FEMBA-Tiny
30
+ results:
31
+ - task:
32
+ type: time-series-classification
33
+ name: EEG Artifact Detection (Binary)
34
+ dataset:
35
+ type: TUAR
36
+ name: TUH EEG Artifact Corpus (TUAR)
37
+ config: binary
38
+ metrics:
39
+ - type: roc_auc
40
+ value: 0.937
41
+ name: AUROC
42
+ - type: pr_auc
43
+ value: 0.912
44
+ name: AUC-PR
45
+ - task:
46
+ type: time-series-classification
47
+ name: EEG Artifact Detection (Multilabel)
48
+ dataset:
49
+ type: TUAR
50
+ name: TUH EEG Artifact Corpus (TUAR)
51
+ config: multilabel
52
+ metrics:
53
+ - type: roc_auc
54
+ value: 0.887
55
+ name: AUROC
56
+ - type: pr_auc
57
+ value: 0.645
58
+ name: AUC-PR
59
+ - task:
60
+ type: time-series-classification
61
+ name: EEG Artifact Detection (Multiclass-Multioutput)
62
+ dataset:
63
+ type: TUAR
64
+ name: TUH EEG Artifact Corpus (TUAR)
65
+ config: multiclass-multioutput
66
+ metrics:
67
+ - type: roc_auc
68
+ value: 0.893
69
+ name: AUROC
70
+ - type: pr_auc
71
+ value: 0.504
72
+ name: AUC-PR
73
+ - task:
74
+ type: time-series-classification
75
+ name: EEG Artifact Detection (Multiclass)
76
+ dataset:
77
+ type: TUAR
78
+ name: TUH EEG Artifact Corpus (TUAR)
79
+ config: multiclass
80
+ metrics:
81
+ - type: roc_auc
82
+ value: 0.918
83
+ name: AUROC
84
+ - type: pr_auc
85
+ value: 0.518
86
+ name: AUC-PR
87
+ - task:
88
+ type: time-series-classification
89
+ name: EEG Slowing Classification
90
+ dataset:
91
+ type: TUSL
92
+ name: TUH EEG Slowing Corpus (TUSL)
93
+ metrics:
94
+ - type: roc_auc
95
+ value: 0.708
96
+ name: AUROC
97
+ - type: pr_auc
98
+ value: 0.277
99
+ name: AUC-PR
100
+ - name: FEMBA-Base
101
+ results:
102
+ - task:
103
+ type: time-series-classification
104
+ name: EEG Abnormality Detection
105
+ dataset:
106
+ type: TUAB
107
+ name: TUH EEG Abnormal Corpus (TUAB)
108
+ metrics:
109
+ - type: balanced_accuracy
110
+ value: 81.05
111
+ name: Balanced Accuracy (%)
112
+ - type: roc_auc
113
+ value: 0.8829
114
+ name: AUROC
115
+ - type: pr_auc
116
+ value: 0.8894
117
+ name: AUC-PR
118
+ - task:
119
+ type: time-series-classification
120
+ name: EEG Artifact Detection (Binary)
121
+ dataset:
122
+ type: TUAR
123
+ name: TUH EEG Artifact Corpus (TUAR)
124
+ config: binary
125
+ metrics:
126
+ - type: roc_auc
127
+ value: 0.949
128
+ name: AUROC
129
+ - type: pr_auc
130
+ value: 0.932
131
+ name: AUC-PR
132
+ - task:
133
+ type: time-series-classification
134
+ name: EEG Artifact Detection (Multilabel)
135
+ dataset:
136
+ type: TUAR
137
+ name: TUH EEG Artifact Corpus (TUAR)
138
+ config: multilabel
139
+ metrics:
140
+ - type: roc_auc
141
+ value: 0.909
142
+ name: AUROC
143
+ - type: pr_auc
144
+ value: 0.634
145
+ name: AUC-PR
146
+ - task:
147
+ type: time-series-classification
148
+ name: EEG Artifact Detection (Multiclass-Multioutput)
149
+ dataset:
150
+ type: TUAR
151
+ name: TUH EEG Artifact Corpus (TUAR)
152
+ config: multiclass-multioutput
153
+ metrics:
154
+ - type: roc_auc
155
+ value: 0.888
156
+ name: AUROC
157
+ - type: pr_auc
158
+ value: 0.518
159
+ name: AUC-PR
160
+ - task:
161
+ type: time-series-classification
162
+ name: EEG Artifact Detection (Multiclass)
163
+ dataset:
164
+ type: TUAR
165
+ name: TUH EEG Artifact Corpus (TUAR)
166
+ config: multiclass
167
+ metrics:
168
+ - type: roc_auc
169
+ value: 0.900
170
+ name: AUROC
171
+ - type: pr_auc
172
+ value: 0.559
173
+ name: AUC-PR
174
+ - task:
175
+ type: time-series-classification
176
+ name: EEG Slowing Classification
177
+ dataset:
178
+ type: TUSL
179
+ name: TUH EEG Slowing Corpus (TUSL)
180
+ metrics:
181
+ - type: roc_auc
182
+ value: 0.731
183
+ name: AUROC
184
+ - type: pr_auc
185
+ value: 0.289
186
+ name: AUC-PR
187
+ - name: FEMBA-Large
188
+ results:
189
+ - task:
190
+ type: time-series-classification
191
+ name: EEG Abnormality Detection
192
+ dataset:
193
+ type: TUAB
194
+ name: TUH EEG Abnormal Corpus (TUAB)
195
+ metrics:
196
+ - type: balanced_accuracy
197
+ value: 81.47
198
+ name: Balanced Accuracy (%)
199
+ - type: roc_auc
200
+ value: 0.8856
201
+ name: AUROC
202
+ - type: pr_auc
203
+ value: 0.8992
204
+ name: AUC-PR
205
+ - task:
206
+ type: time-series-classification
207
+ name: EEG Artifact Detection (Binary)
208
+ dataset:
209
+ type: TUAR
210
+ name: TUH EEG Artifact Corpus (TUAR)
211
+ config: binary
212
+ metrics:
213
+ - type: roc_auc
214
+ value: 0.944
215
+ name: AUROC
216
+ - type: pr_auc
217
+ value: 0.913
218
+ name: AUC-PR
219
+ - task:
220
+ type: time-series-classification
221
+ name: EEG Artifact Detection (Multilabel)
222
+ dataset:
223
+ type: TUAR
224
+ name: TUH EEG Artifact Corpus (TUAR)
225
+ config: multilabel
226
+ metrics:
227
+ - type: roc_auc
228
+ value: 0.899
229
+ name: AUROC
230
+ - type: pr_auc
231
+ value: 0.608
232
+ name: AUC-PR
233
+ - task:
234
+ type: time-series-classification
235
+ name: EEG Artifact Detection (Multiclass-Multioutput)
236
+ dataset:
237
+ type: TUAR
238
+ name: TUH EEG Artifact Corpus (TUAR)
239
+ config: multiclass-multioutput
240
+ metrics:
241
+ - type: roc_auc
242
+ value: 0.878
243
+ name: AUROC
244
+ - type: pr_auc
245
+ value: 0.516
246
+ name: AUC-PR
247
+ - task:
248
+ type: time-series-classification
249
+ name: EEG Artifact Detection (Multiclass)
250
+ dataset:
251
+ type: TUAR
252
+ name: TUH EEG Artifact Corpus (TUAR)
253
+ config: multiclass
254
+ metrics:
255
+ - type: roc_auc
256
+ value: 0.915
257
+ name: AUROC
258
+ - type: pr_auc
259
+ value: 0.521
260
+ name: AUC-PR
261
+ - task:
262
+ type: time-series-classification
263
+ name: EEG Slowing Classification
264
+ dataset:
265
+ type: TUSL
266
+ name: TUH EEG Slowing Corpus (TUSL)
267
+ metrics:
268
+ - type: roc_auc
269
+ value: 0.714
270
+ name: AUROC
271
+ - type: pr_auc
272
+ value: 0.282
273
+ name: AUC-PR
274
+ - name: FEMBA-Huge
275
+ results:
276
+ - task:
277
+ type: time-series-classification
278
+ name: EEG Abnormality Detection
279
+ dataset:
280
+ type: TUAB
281
+ name: TUH EEG Abnormal Corpus (TUAB)
282
+ metrics:
283
+ - type: balanced_accuracy
284
+ value: 81.82
285
+ name: Balanced Accuracy (%)
286
+ - type: roc_auc
287
+ value: 0.8921
288
+ name: AUROC
289
+ - type: pr_auc
290
+ value: 0.9005
291
+ name: AUC-PR
292
  ---
293
 
294
  <div align="center">
 
498
 
499
  ---
500
 
501
+ ## 🔗 Related Models
502
+
503
+ - **[LuMamba](https://huggingface.co/PulpBio/LuMamba)** — Direct successor. Combines FEMBA's bi-Mamba encoder with LUNA's channel-unification cross-attention for topology-invariant EEG modeling, plus LeJEPA pre-training.
504
+ - **[LUNA](https://huggingface.co/PulpBio/LUNA)** — Transformer-based topology-agnostic EEG foundation model (NeurIPS 2025). Companion architecture focused on channel-heterogeneity rather than sequence-length efficiency.
505
+ - **[TinyMyo](https://huggingface.co/PulpBio/TinyMyo)** — Tiny foundation model for flexible EMG signal processing at the edge.
506
+ ---
507
  ## 🗒️ Changelog
508
 
509
  - **v1.0:** Initial release of FEMBA model card with task-specific checkpoints and instructions.