Hugging Face's logo

Datasets:
project-telos
/
cognitive_map_probes_results

Dataset card Files
xet
 Community
gsarti commited on
Commit
f0d5a0f
·
verified ·
1 Parent(s): d42cb7c

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.json +144 -144
  2. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.txt +37 -37
  3. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.json +149 -149
  4. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.txt +38 -38
  5. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.json +138 -138
  6. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.txt +36 -36
  7. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.json +149 -149
  8. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.txt +38 -38
  9. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.json +147 -147
  10. layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.txt +38 -38
  11. layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.json +1147 -1146
  12. layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.txt +251 -251
  13. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.json +149 -149
  14. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.txt +38 -38
  15. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.json +149 -149
  16. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.txt +38 -38
  17. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.json +149 -149
  18. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.txt +38 -38
  19. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.json +149 -149
  20. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.txt +38 -38
  21. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.json +149 -149
  22. layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.txt +38 -38
  23. layer15/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.json +251 -250
  24. layer15/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.txt +24 -1
  25. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.json +146 -146
  26. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.txt +38 -38
  27. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.json +144 -144
  28. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.txt +37 -37
  29. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.json +137 -137
  30. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.txt +37 -37
  31. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.json +148 -148
  32. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.txt +38 -38
  33. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.json +148 -148
  34. layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.txt +38 -38
  35. layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.json +1142 -1141
  36. layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.txt +250 -250
  37. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.json +149 -149
  38. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.txt +38 -38
  39. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.json +149 -149
  40. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.txt +38 -38
  41. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.json +149 -149
  42. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.txt +38 -38
  43. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.json +149 -149
  44. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.txt +38 -38
  45. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.json +149 -149
  46. layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.txt +38 -38
  47. layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.json +1172 -1171
  48. layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.txt +255 -255
  49. layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.json +149 -149
  50. layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.txt +38 -38
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.json CHANGED
@@ -1,55 +1,55 @@
1
  {
2
  "global": {
3
- "accuracy": 0.03445860764812999,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.6332819722650231,
8
- "precision": 0.008270283322601416,
9
- "recall": 0.6332819722650231,
10
- "f1": 0.01632734134472142,
11
  "gt_support": 649,
12
- "predicted": 49696
13
  },
14
  "1": {
15
- "accuracy": 0.015939384724545133,
16
- "precision": 0.5775941230486685,
17
- "recall": 0.015939384724545133,
18
- "f1": 0.031022662819659198,
19
  "gt_support": 39462,
20
- "predicted": 1089
21
  },
22
  "2": {
23
- "accuracy": 0.31741140215716485,
24
- "precision": 0.008293075684380032,
25
- "recall": 0.31741140215716485,
26
- "f1": 0.01616383537996783,
27
  "gt_support": 649,
28
- "predicted": 24840
29
  },
30
  "3": {
31
- "accuracy": 0.03865604066827292,
32
- "precision": 0.5027548209366391,
33
- "recall": 0.03865604066827292,
34
- "f1": 0.0717920979519583,
35
  "gt_support": 37769,
36
- "predicted": 2904
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.03670393777345649,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.7352941176470589,
48
- "precision": 0.008264462809917356,
49
- "recall": 0.7352941176470589,
50
- "f1": 0.016345210853220007,
51
  "gt_support": 68,
52
- "predicted": 6050
53
  },
54
  "1": {
55
  "accuracy": 0.0,
@@ -60,215 +60,215 @@
60
  "predicted": 0
61
  },
62
  "2": {
63
- "accuracy": 0.22058823529411764,
64
- "precision": 0.008264462809917356,
65
- "recall": 0.22058823529411764,
66
- "f1": 0.015932023366967606,
67
  "gt_support": 68,
68
- "predicted": 1815
69
  },
70
  "3": {
71
- "accuracy": 0.04411764705882353,
72
- "precision": 0.6528925619834711,
73
- "recall": 0.04411764705882353,
74
- "f1": 0.0826503923278117,
75
  "gt_support": 5372,
76
- "predicted": 363
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
- "accuracy": 0.035877491492464754,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.7411764705882353,
87
- "precision": 0.008264462809917356,
88
- "recall": 0.7411764705882353,
89
- "f1": 0.01634665282823041,
90
  "gt_support": 85,
91
- "predicted": 7623
92
  },
93
  "1": {
94
- "accuracy": 0.0,
95
- "precision": 0.0,
96
- "recall": 0.0,
97
- "f1": 0.0,
98
  "gt_support": 3995,
99
- "predicted": 0
100
  },
101
  "2": {
102
- "accuracy": 0.21176470588235294,
103
- "precision": 0.008264462809917356,
104
- "recall": 0.21176470588235294,
105
- "f1": 0.015908086610693768,
106
  "gt_support": 85,
107
- "predicted": 2178
108
  },
109
  "3": {
110
- "accuracy": 0.047058823529411764,
111
- "precision": 0.5950413223140496,
112
- "recall": 0.047058823529411764,
113
- "f1": 0.0872198667474258,
114
  "gt_support": 6120,
115
- "predicted": 484
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
- "accuracy": 0.055146506386175805,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.6727272727272727,
126
- "precision": 0.008264462809917356,
127
- "recall": 0.6727272727272727,
128
- "f1": 0.016328331862312445,
129
  "gt_support": 55,
130
- "predicted": 4477
131
  },
132
  "1": {
133
- "accuracy": 0.01818181818181818,
134
- "precision": 0.4380165289256198,
135
- "recall": 0.01818181818181818,
136
- "f1": 0.03491436100131752,
137
  "gt_support": 2915,
138
- "predicted": 121
139
  },
140
  "2": {
141
- "accuracy": 0.23636363636363636,
142
- "precision": 0.008264462809917356,
143
- "recall": 0.23636363636363636,
144
- "f1": 0.01597051597051597,
145
  "gt_support": 55,
146
- "predicted": 1573
147
  },
148
  "3": {
149
- "accuracy": 0.07272727272727272,
150
- "precision": 0.5454545454545454,
151
- "recall": 0.07272727272727272,
152
- "f1": 0.1283422459893048,
153
  "gt_support": 3630,
154
- "predicted": 484
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
- "accuracy": 0.026879563507983633,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.6699029126213593,
165
- "precision": 0.008264462809917356,
166
- "recall": 0.6699029126213593,
167
- "f1": 0.01632749645054425,
168
  "gt_support": 103,
169
- "predicted": 8349
170
  },
171
  "1": {
172
- "accuracy": 0.0,
173
- "precision": 0.0,
174
- "recall": 0.0,
175
- "f1": 0.0,
176
  "gt_support": 6180,
177
- "predicted": 0
178
  },
179
  "2": {
180
- "accuracy": 0.2912621359223301,
181
- "precision": 0.008264462809917356,
182
- "recall": 0.2912621359223301,
183
- "f1": 0.01607286364854005,
184
  "gt_support": 103,
185
- "predicted": 3630
186
  },
187
  "3": {
188
- "accuracy": 0.038834951456310676,
189
- "precision": 0.48760330578512395,
190
- "recall": 0.038834951456310676,
191
- "f1": 0.07194025301021185,
192
  "gt_support": 6077,
193
- "predicted": 484
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
- "accuracy": 0.015803972741771786,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.7280701754385965,
204
- "precision": 0.00827022718214428,
205
- "recall": 0.7280701754385965,
206
- "f1": 0.016354679802955664,
207
  "gt_support": 114,
208
- "predicted": 10036
209
  },
210
  "1": {
211
- "accuracy": 0.0,
212
- "precision": 0.0,
213
- "recall": 0.0,
214
- "f1": 0.0,
215
  "gt_support": 7524,
216
- "predicted": 0
217
  },
218
  "2": {
219
- "accuracy": 0.2543859649122807,
220
- "precision": 0.008248009101251423,
221
- "recall": 0.2543859649122807,
222
- "f1": 0.01597796143250689,
223
  "gt_support": 114,
224
- "predicted": 3516
225
  },
226
  "3": {
227
- "accuracy": 0.017543859649122806,
228
- "precision": 0.4380165289256198,
229
- "recall": 0.017543859649122806,
230
- "f1": 0.03373647358370464,
231
  "gt_support": 6042,
232
- "predicted": 242
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
- "accuracy": 0.04113783943329398,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.48660714285714285,
243
- "precision": 0.008282045437276803,
244
- "recall": 0.48660714285714285,
245
- "f1": 0.016286888307807248,
246
  "gt_support": 224,
247
- "predicted": 13161
248
  },
249
  "1": {
250
- "accuracy": 0.03571428571428571,
251
- "precision": 0.5950413223140496,
252
- "recall": 0.03571428571428571,
253
- "f1": 0.06738418343472158,
254
  "gt_support": 16128,
255
- "predicted": 968
256
  },
257
  "2": {
258
- "accuracy": 0.45089285714285715,
259
- "precision": 0.008327836411609498,
260
- "recall": 0.45089285714285715,
261
- "f1": 0.01635362694300518,
262
  "gt_support": 224,
263
- "predicted": 12128
264
  },
265
  "3": {
266
- "accuracy": 0.03125,
267
- "precision": 0.3884297520661157,
268
- "recall": 0.03125,
269
- "f1": 0.057846153846153846,
270
  "gt_support": 10528,
271
- "predicted": 847
272
  }
273
  },
274
  "total_samples": 27104
@@ -278,13 +278,13 @@
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 11
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.31608705064371123,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.2588597842835131,
8
+ "precision": 0.00837070254110613,
9
+ "recall": 0.2588597842835131,
10
+ "f1": 0.016216998889907814,
11
  "gt_support": 649,
12
+ "predicted": 20070
13
  },
14
  "1": {
15
+ "accuracy": 0.4219502306015914,
16
+ "precision": 0.5620211293752321,
17
+ "recall": 0.4219502306015914,
18
+ "f1": 0.4820159504407359,
19
  "gt_support": 39462,
20
+ "predicted": 29627
21
  },
22
  "2": {
23
+ "accuracy": 0.18798151001540833,
24
+ "precision": 0.008472222222222223,
25
+ "recall": 0.18798151001540833,
26
+ "f1": 0.016213701907103466,
27
  "gt_support": 649,
28
+ "predicted": 14400
29
  },
30
  "3": {
31
+ "accuracy": 0.20866318938812253,
32
+ "precision": 0.5460781596452328,
33
+ "recall": 0.20866318938812253,
34
+ "f1": 0.30194823853949154,
35
  "gt_support": 37769,
36
+ "predicted": 14432
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.16079241614000972,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.16176470588235295,
48
+ "precision": 0.00975177304964539,
49
+ "recall": 0.16176470588235295,
50
+ "f1": 0.01839464882943144,
51
  "gt_support": 68,
52
+ "predicted": 1128
53
  },
54
  "1": {
55
  "accuracy": 0.0,
 
60
  "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.6176470588235294,
64
+ "precision": 0.008144269924374637,
65
+ "recall": 0.6176470588235294,
66
+ "f1": 0.016076555023923446,
67
  "gt_support": 68,
68
+ "predicted": 5157
69
  },
70
  "3": {
71
+ "accuracy": 0.23641102010424422,
72
+ "precision": 0.6536284096757591,
73
+ "recall": 0.23641102010424422,
74
+ "f1": 0.3472317156527683,
75
  "gt_support": 5372,
76
+ "predicted": 1943
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
+ "accuracy": 0.2031113271754983,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.21176470588235294,
87
+ "precision": 0.008399440037330844,
88
+ "recall": 0.21176470588235294,
89
+ "f1": 0.01615798922800718,
90
  "gt_support": 85,
91
+ "predicted": 2143
92
  },
93
  "1": {
94
+ "accuracy": 0.003254067584480601,
95
+ "precision": 0.5416666666666666,
96
+ "recall": 0.003254067584480601,
97
+ "f1": 0.006469270962926102,
98
  "gt_support": 3995,
99
+ "predicted": 24
100
  },
101
  "2": {
102
+ "accuracy": 0.5058823529411764,
103
+ "precision": 0.009102455546147333,
104
+ "recall": 0.5058823529411764,
105
+ "f1": 0.017883135787065917,
106
  "gt_support": 85,
107
+ "predicted": 4724
108
  },
109
  "3": {
110
+ "accuracy": 0.3292483660130719,
111
+ "precision": 0.5936947554507955,
112
+ "recall": 0.3292483660130719,
113
+ "f1": 0.42358629388269914,
114
  "gt_support": 6120,
115
+ "predicted": 3394
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
+ "accuracy": 0.4064613072877536,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.10909090909090909,
126
+ "precision": 0.008915304606240713,
127
+ "recall": 0.10909090909090909,
128
+ "f1": 0.016483516483516484,
129
  "gt_support": 55,
130
+ "predicted": 673
131
  },
132
  "1": {
133
+ "accuracy": 0.06380789022298457,
134
+ "precision": 0.43356643356643354,
135
+ "recall": 0.06380789022298457,
136
+ "f1": 0.11124401913875599,
137
  "gt_support": 2915,
138
+ "predicted": 429
139
  },
140
  "2": {
141
+ "accuracy": 0.14545454545454545,
142
+ "precision": 0.008421052631578947,
143
+ "recall": 0.14545454545454545,
144
+ "f1": 0.01592039800995025,
145
  "gt_support": 55,
146
+ "predicted": 950
147
  },
148
  "3": {
149
+ "accuracy": 0.6900826446280992,
150
+ "precision": 0.5442102976319791,
151
+ "recall": 0.6900826446280992,
152
+ "f1": 0.608526660998421,
153
  "gt_support": 3630,
154
+ "predicted": 4603
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
+ "accuracy": 0.3176602744122603,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.23300970873786409,
165
+ "precision": 0.008284432171211598,
166
+ "recall": 0.23300970873786409,
167
+ "f1": 0.016,
168
  "gt_support": 103,
169
+ "predicted": 2897
170
  },
171
  "1": {
172
+ "accuracy": 0.41828478964401294,
173
+ "precision": 0.49682875264270615,
174
+ "recall": 0.41828478964401294,
175
+ "f1": 0.454186066941931,
176
  "gt_support": 6180,
177
+ "predicted": 5203
178
  },
179
  "2": {
180
+ "accuracy": 0.13592233009708737,
181
+ "precision": 0.00850546780072904,
182
+ "recall": 0.13592233009708737,
183
+ "f1": 0.016009148084619784,
184
  "gt_support": 103,
185
+ "predicted": 1646
186
  },
187
  "3": {
188
+ "accuracy": 0.21984531841369095,
189
+ "precision": 0.4917188075082812,
190
+ "recall": 0.21984531841369095,
191
+ "f1": 0.30384352967932676,
192
  "gt_support": 6077,
193
+ "predicted": 2717
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
+ "accuracy": 0.3649412788168769,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.2631578947368421,
204
+ "precision": 0.008047210300429184,
205
+ "recall": 0.2631578947368421,
206
+ "f1": 0.01561686621551275,
207
  "gt_support": 114,
208
+ "predicted": 3728
209
  },
210
  "1": {
211
+ "accuracy": 0.5971557682083998,
212
+ "precision": 0.5453993687788298,
213
+ "recall": 0.5971557682083998,
214
+ "f1": 0.5701053165841897,
215
  "gt_support": 7524,
216
+ "predicted": 8238
217
  },
218
  "2": {
219
+ "accuracy": 0.05263157894736842,
220
+ "precision": 0.008708272859216255,
221
+ "recall": 0.05263157894736842,
222
+ "f1": 0.0149439601494396,
223
  "gt_support": 114,
224
+ "predicted": 689
225
  },
226
  "3": {
227
+ "accuracy": 0.08358159549817941,
228
+ "precision": 0.4433713784021071,
229
+ "recall": 0.08358159549817941,
230
+ "f1": 0.14064893468876202,
231
  "gt_support": 6042,
232
+ "predicted": 1139
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
+ "accuracy": 0.35832349468713104,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.35267857142857145,
243
+ "precision": 0.008314914219555837,
244
+ "recall": 0.35267857142857145,
245
+ "f1": 0.016246786632390747,
246
  "gt_support": 224,
247
+ "predicted": 9501
248
  },
249
  "1": {
250
+ "accuracy": 0.5812251984126984,
251
+ "precision": 0.5958177080022882,
252
+ "recall": 0.5812251984126984,
253
+ "f1": 0.5884309971438435,
254
  "gt_support": 16128,
255
+ "predicted": 15733
256
  },
257
  "2": {
258
+ "accuracy": 0.04017857142857143,
259
+ "precision": 0.007293354943273906,
260
+ "recall": 0.04017857142857143,
261
+ "f1": 0.01234567901234568,
262
  "gt_support": 224,
263
+ "predicted": 1234
264
  },
265
  "3": {
266
+ "accuracy": 0.023746200607902737,
267
+ "precision": 0.39308176100628933,
268
+ "recall": 0.023746200607902737,
269
+ "f1": 0.04478681476173415,
270
  "gt_support": 10528,
271
+ "predicted": 636
272
  }
273
  },
274
  "total_samples": 27104
 
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 11
290
  }
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 649 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.0345 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.6333 0.0083 0.6333 0.0163 649 49696
30
- # 0.0159 0.5776 0.0159 0.0310 39462 1089
31
- G 0.3174 0.0083 0.3174 0.0162 649 24840
32
- _ 0.0387 0.5028 0.0387 0.0718 37769 2904
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.0367 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.7353 0.0083 0.7353 0.0163 68 6050
49
  # 0.0000 0.0000 0.0000 0.0000 2720 0
50
- G 0.2206 0.0083 0.2206 0.0159 68 1815
51
- _ 0.0441 0.6529 0.0441 0.0827 5372 363
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.0359 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.7412 0.0083 0.7412 0.0163 85 7623
64
- # 0.0000 0.0000 0.0000 0.0000 3995 0
65
- G 0.2118 0.0083 0.2118 0.0159 85 2178
66
- _ 0.0471 0.5950 0.0471 0.0872 6120 484
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.0551 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.6727 0.0083 0.6727 0.0163 55 4477
79
- # 0.0182 0.4380 0.0182 0.0349 2915 121
80
- G 0.2364 0.0083 0.2364 0.0160 55 1573
81
- _ 0.0727 0.5455 0.0727 0.1283 3630 484
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0269 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.6699 0.0083 0.6699 0.0163 103 8349
94
- # 0.0000 0.0000 0.0000 0.0000 6180 0
95
- G 0.2913 0.0083 0.2913 0.0161 103 3630
96
- _ 0.0388 0.4876 0.0388 0.0719 6077 484
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0158 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.7281 0.0083 0.7281 0.0164 114 10036
109
- # 0.0000 0.0000 0.0000 0.0000 7524 0
110
- G 0.2544 0.0082 0.2544 0.0160 114 3516
111
- _ 0.0175 0.4380 0.0175 0.0337 6042 242
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.0411 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.4866 0.0083 0.4866 0.0163 224 13161
124
- # 0.0357 0.5950 0.0357 0.0674 16128 968
125
- G 0.4509 0.0083 0.4509 0.0164 224 12128
126
- _ 0.0312 0.3884 0.0312 0.0578 10528 847
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.3161 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.2589 0.0084 0.2589 0.0162 649 20070
30
+ # 0.4220 0.5620 0.4220 0.4820 39462 29627
31
+ G 0.1880 0.0085 0.1880 0.0162 649 14400
32
+ _ 0.2087 0.5461 0.2087 0.3019 37769 14432
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.1608 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.1618 0.0098 0.1618 0.0184 68 1128
49
  # 0.0000 0.0000 0.0000 0.0000 2720 0
50
+ G 0.6176 0.0081 0.6176 0.0161 68 5157
51
+ _ 0.2364 0.6536 0.2364 0.3472 5372 1943
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.2031 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.2118 0.0084 0.2118 0.0162 85 2143
64
+ # 0.0033 0.5417 0.0033 0.0065 3995 24
65
+ G 0.5059 0.0091 0.5059 0.0179 85 4724
66
+ _ 0.3292 0.5937 0.3292 0.4236 6120 3394
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.4065 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.1091 0.0089 0.1091 0.0165 55 673
79
+ # 0.0638 0.4336 0.0638 0.1112 2915 429
80
+ G 0.1455 0.0084 0.1455 0.0159 55 950
81
+ _ 0.6901 0.5442 0.6901 0.6085 3630 4603
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.3177 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.2330 0.0083 0.2330 0.0160 103 2897
94
+ # 0.4183 0.4968 0.4183 0.4542 6180 5203
95
+ G 0.1359 0.0085 0.1359 0.0160 103 1646
96
+ _ 0.2198 0.4917 0.2198 0.3038 6077 2717
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.3649 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2632 0.0080 0.2632 0.0156 114 3728
109
+ # 0.5972 0.5454 0.5972 0.5701 7524 8238
110
+ G 0.0526 0.0087 0.0526 0.0149 114 689
111
+ _ 0.0836 0.4434 0.0836 0.1406 6042 1139
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.3583 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.3527 0.0083 0.3527 0.0162 224 9501
124
+ # 0.5812 0.5958 0.5812 0.5884 16128 15733
125
+ G 0.0402 0.0073 0.0402 0.0123 224 1234
126
+ _ 0.0237 0.3931 0.0237 0.0448 10528 636
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size11.json
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.18671926364234057,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.5851851851851851,
8
- "precision": 0.005917085897611944,
9
- "recall": 0.5851851851851851,
10
- "f1": 0.011715708689150612,
11
  "gt_support": 810,
12
- "predicted": 80107
13
  },
14
  "1": {
15
- "accuracy": 0.21208021406452002,
16
- "precision": 0.515134918026801,
17
- "recall": 0.21208021406452002,
18
- "f1": 0.3004610846670713,
19
  "gt_support": 66522,
20
- "predicted": 27387
21
  },
22
  "2": {
23
- "accuracy": 0.06296296296296296,
24
- "precision": 0.005953770721456923,
25
- "recall": 0.06296296296296296,
26
- "f1": 0.010878839590443687,
27
  "gt_support": 810,
28
- "predicted": 8566
29
  },
30
  "3": {
31
- "accuracy": 0.15894280560889043,
32
- "precision": 0.5245799327892463,
33
- "recall": 0.15894280560889043,
34
- "f1": 0.24396615240349193,
35
  "gt_support": 68748,
36
- "predicted": 20830
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.15856139053254437,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.71875,
48
- "precision": 0.005917159763313609,
49
- "recall": 0.71875,
50
- "f1": 0.011737688185761673,
51
  "gt_support": 64,
52
- "predicted": 7774
53
  },
54
  "1": {
55
- "accuracy": 0.078125,
56
- "precision": 0.28402366863905326,
57
- "recall": 0.078125,
58
- "f1": 0.1225427623181006,
59
  "gt_support": 3072,
60
- "predicted": 845
61
  },
62
  "2": {
63
- "accuracy": 0.015625,
64
- "precision": 0.005917159763313609,
65
- "recall": 0.015625,
66
- "f1": 0.008583690987124463,
67
  "gt_support": 64,
68
- "predicted": 169
69
  },
70
  "3": {
71
- "accuracy": 0.1875,
72
- "precision": 0.7041420118343196,
73
- "recall": 0.1875,
74
- "f1": 0.29614267938614686,
75
  "gt_support": 7616,
76
- "predicted": 2028
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
- "accuracy": 0.1879175415155564,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.6451612903225806,
87
- "precision": 0.005925925925925926,
88
- "recall": 0.6451612903225806,
89
- "f1": 0.011743981209630066,
90
  "gt_support": 62,
91
- "predicted": 6750
92
  },
93
  "1": {
94
- "accuracy": 0.08064516129032258,
95
- "precision": 0.3431952662721893,
96
- "recall": 0.08064516129032258,
97
- "f1": 0.13060121594235533,
98
  "gt_support": 3596,
99
- "predicted": 845
100
  },
101
  "2": {
102
- "accuracy": 0.03225806451612903,
103
- "precision": 0.005917159763313609,
104
- "recall": 0.03225806451612903,
105
- "f1": 0.009999999999999998,
106
  "gt_support": 62,
107
- "predicted": 338
108
  },
109
  "3": {
110
- "accuracy": 0.24223142941698728,
111
- "precision": 0.6432220039292731,
112
- "recall": 0.24223142941698728,
113
- "f1": 0.35192948511232935,
114
  "gt_support": 6758,
115
- "predicted": 2545
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
- "accuracy": 0.18002366863905325,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.624,
126
- "precision": 0.005931107900539883,
127
- "recall": 0.624,
128
- "f1": 0.011750527267249173,
129
  "gt_support": 125,
130
- "predicted": 13151
131
  },
132
  "1": {
133
- "accuracy": 0.17,
134
- "precision": 0.4041958041958042,
135
- "recall": 0.17,
136
- "f1": 0.23933747412008285,
137
  "gt_support": 8500,
138
- "predicted": 3575
139
  },
140
  "2": {
141
- "accuracy": 0.024,
142
- "precision": 0.005917159763313609,
143
- "recall": 0.024,
144
- "f1": 0.00949367088607595,
145
  "gt_support": 125,
146
- "predicted": 507
147
  },
148
  "3": {
149
- "accuracy": 0.184,
150
- "precision": 0.5850462487153134,
151
- "recall": 0.184,
152
- "f1": 0.27995327964590894,
153
  "gt_support": 12375,
154
- "predicted": 3892
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
- "accuracy": 0.17815052080786348,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.628099173553719,
165
- "precision": 0.0059180812957483255,
166
- "recall": 0.628099173553719,
167
- "f1": 0.011725680783769189,
168
  "gt_support": 121,
169
- "predicted": 12842
170
  },
171
  "1": {
172
- "accuracy": 0.19855901674083493,
173
- "precision": 0.4618038442582553,
174
- "recall": 0.19855901674083493,
175
- "f1": 0.27771191464137523,
176
  "gt_support": 9438,
177
- "predicted": 4058
178
  },
179
  "2": {
180
- "accuracy": 0.01652892561983471,
181
- "precision": 0.005917159763313609,
182
- "recall": 0.01652892561983471,
183
- "f1": 0.008714596949891067,
184
  "gt_support": 121,
185
- "predicted": 338
186
  },
187
  "3": {
188
- "accuracy": 0.15702479338842976,
189
- "precision": 0.5266272189349113,
190
- "recall": 0.15702479338842976,
191
- "f1": 0.2419170243204578,
192
  "gt_support": 10769,
193
- "predicted": 3211
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
- "accuracy": 0.18406593406593408,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.5714285714285714,
204
- "precision": 0.005937471008442341,
205
- "recall": 0.5714285714285714,
206
- "f1": 0.011752823432191717,
207
  "gt_support": 112,
208
- "predicted": 10779
209
  },
210
  "1": {
211
- "accuracy": 0.20485553772070625,
212
- "precision": 0.5261530533367689,
213
- "recall": 0.20485553772070625,
214
- "f1": 0.2948949382626905,
215
  "gt_support": 9968,
216
- "predicted": 3881
217
  },
218
  "2": {
219
- "accuracy": 0.07142857142857142,
220
- "precision": 0.006158583525789068,
221
- "recall": 0.07142857142857142,
222
- "f1": 0.011339475549255846,
223
  "gt_support": 112,
224
- "predicted": 1299
225
  },
226
  "3": {
227
- "accuracy": 0.15682234432234432,
228
- "precision": 0.4614348265409229,
229
- "recall": 0.15682234432234432,
230
- "f1": 0.234087996582657,
231
  "gt_support": 8736,
232
- "predicted": 2969
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
- "accuracy": 0.19867862199150543,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.5214723926380368,
243
- "precision": 0.005900524105376418,
244
- "recall": 0.5214723926380368,
245
- "f1": 0.011669011909256271,
246
  "gt_support": 326,
247
- "predicted": 28811
248
  },
249
  "1": {
250
- "accuracy": 0.25719919869788405,
251
- "precision": 0.5793555665233026,
252
- "recall": 0.25719919869788405,
253
- "f1": 0.35624634193925986,
254
  "gt_support": 31948,
255
- "predicted": 14183
256
  },
257
  "2": {
258
- "accuracy": 0.10736196319018405,
259
- "precision": 0.005917159763313609,
260
- "recall": 0.10736196319018405,
261
- "f1": 0.011216151257811249,
262
  "gt_support": 326,
263
- "predicted": 5915
264
  },
265
  "3": {
266
- "accuracy": 0.11220769983106606,
267
- "precision": 0.40808407437348426,
268
- "recall": 0.11220769983106606,
269
- "f1": 0.1760172948847589,
270
  "gt_support": 22494,
271
- "predicted": 6185
272
  }
273
  },
274
  "total_samples": 55094
@@ -278,13 +278,13 @@
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 13
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.27560815253122944,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.2271604938271605,
8
+ "precision": 0.006057812602884046,
9
+ "recall": 0.2271604938271605,
10
+ "f1": 0.011800923550538737,
11
  "gt_support": 810,
12
+ "predicted": 30374
13
  },
14
  "1": {
15
+ "accuracy": 0.2682571179459427,
16
+ "precision": 0.5307536731901731,
17
+ "recall": 0.2682571179459427,
18
+ "f1": 0.35638680300367476,
19
  "gt_support": 66522,
20
+ "predicted": 33622
21
  },
22
  "2": {
23
+ "accuracy": 0.2962962962962963,
24
+ "precision": 0.00585480093676815,
25
+ "recall": 0.2962962962962963,
26
+ "f1": 0.011482704176833645,
27
  "gt_support": 810,
28
+ "predicted": 40992
29
  },
30
  "3": {
31
+ "accuracy": 0.28304823413044744,
32
+ "precision": 0.6099617578835183,
33
+ "recall": 0.28304823413044744,
34
+ "f1": 0.38666666666666666,
35
  "gt_support": 68748,
36
+ "predicted": 31902
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.5634245562130178,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.0625,
48
+ "precision": 0.013201320132013201,
49
+ "recall": 0.0625,
50
+ "f1": 0.021798365122615806,
51
  "gt_support": 64,
52
+ "predicted": 303
53
  },
54
  "1": {
55
+ "accuracy": 0.0,
56
+ "precision": 0.0,
57
+ "recall": 0.0,
58
+ "f1": 0.0,
59
  "gt_support": 3072,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.171875,
64
+ "precision": 0.005810882197569995,
65
+ "recall": 0.171875,
66
+ "f1": 0.011241696474195198,
67
  "gt_support": 64,
68
+ "predicted": 1893
69
  },
70
  "3": {
71
+ "accuracy": 0.798188025210084,
72
+ "precision": 0.7052204176334107,
73
+ "recall": 0.798188025210084,
74
+ "f1": 0.748829761024883,
75
  "gt_support": 7616,
76
+ "predicted": 8620
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
+ "accuracy": 0.3471082267608322,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.22580645161290322,
87
+ "precision": 0.005952380952380952,
88
+ "recall": 0.22580645161290322,
89
+ "f1": 0.011599005799502899,
90
  "gt_support": 62,
91
+ "predicted": 2352
92
  },
93
  "1": {
94
+ "accuracy": 0.0,
95
+ "precision": 0.0,
96
+ "recall": 0.0,
97
+ "f1": 0.0,
98
  "gt_support": 3596,
99
+ "predicted": 0
100
  },
101
  "2": {
102
+ "accuracy": 0.24193548387096775,
103
+ "precision": 0.005933544303797469,
104
+ "recall": 0.24193548387096775,
105
+ "f1": 0.011583011583011584,
106
  "gt_support": 62,
107
+ "predicted": 2528
108
  },
109
  "3": {
110
+ "accuracy": 0.5338857650192365,
111
+ "precision": 0.6445158985351911,
112
+ "recall": 0.5338857650192365,
113
+ "f1": 0.584007769504694,
114
  "gt_support": 6758,
115
+ "predicted": 5598
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
+ "accuracy": 0.35105325443786983,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.192,
126
+ "precision": 0.006083650190114068,
127
+ "recall": 0.192,
128
+ "f1": 0.011793611793611793,
129
  "gt_support": 125,
130
+ "predicted": 3945
131
  },
132
  "1": {
133
+ "accuracy": 0.11270588235294117,
134
+ "precision": 0.39310627821091504,
135
+ "recall": 0.11270588235294117,
136
+ "f1": 0.17518515132120321,
137
  "gt_support": 8500,
138
+ "predicted": 2437
139
  },
140
  "2": {
141
+ "accuracy": 0.176,
142
+ "precision": 0.005803218148245845,
143
+ "recall": 0.176,
144
+ "f1": 0.011235955056179773,
145
  "gt_support": 125,
146
+ "predicted": 3791
147
  },
148
  "3": {
149
+ "accuracy": 0.5181414141414141,
150
+ "precision": 0.585463842220599,
151
+ "recall": 0.5181414141414141,
152
+ "f1": 0.5497492176447892,
153
  "gt_support": 12375,
154
+ "predicted": 10952
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
+ "accuracy": 0.2639737884493129,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.2644628099173554,
165
+ "precision": 0.006286836935166994,
166
+ "recall": 0.2644628099173554,
167
+ "f1": 0.012281711763577046,
168
  "gt_support": 121,
169
+ "predicted": 5090
170
  },
171
  "1": {
172
+ "accuracy": 0.29211697393515573,
173
+ "precision": 0.46196380697050937,
174
+ "recall": 0.29211697393515573,
175
+ "f1": 0.35791250162274435,
176
  "gt_support": 9438,
177
+ "predicted": 5968
178
  },
179
  "2": {
180
+ "accuracy": 0.21487603305785125,
181
+ "precision": 0.005784204671857619,
182
+ "recall": 0.21487603305785125,
183
+ "f1": 0.011265164644714037,
184
  "gt_support": 121,
185
+ "predicted": 4495
186
  },
187
  "3": {
188
+ "accuracy": 0.23985513975299472,
189
+ "precision": 0.5275735294117647,
190
+ "recall": 0.23985513975299472,
191
+ "f1": 0.32977976380466006,
192
  "gt_support": 10769,
193
+ "predicted": 4896
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
+ "accuracy": 0.2580304311073542,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.15178571428571427,
204
+ "precision": 0.005582922824302135,
205
+ "recall": 0.15178571428571427,
206
+ "f1": 0.010769718086791258,
207
  "gt_support": 112,
208
+ "predicted": 3045
209
  },
210
  "1": {
211
+ "accuracy": 0.46097512038523275,
212
+ "precision": 0.5252629172382259,
213
+ "recall": 0.46097512038523275,
214
+ "f1": 0.49102372301773883,
215
  "gt_support": 9968,
216
+ "predicted": 8748
217
  },
218
  "2": {
219
+ "accuracy": 0.3482142857142857,
220
+ "precision": 0.005884127942063971,
221
+ "recall": 0.3482142857142857,
222
+ "f1": 0.011572700296735905,
223
  "gt_support": 112,
224
+ "predicted": 6628
225
  },
226
  "3": {
227
+ "accuracy": 0.02667124542124542,
228
+ "precision": 0.45956607495069035,
229
+ "recall": 0.02667124542124542,
230
+ "f1": 0.05041653142918965,
231
  "gt_support": 8736,
232
+ "predicted": 507
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
+ "accuracy": 0.18693505644897812,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.2852760736196319,
243
+ "precision": 0.005946671782083254,
244
+ "recall": 0.2852760736196319,
245
+ "f1": 0.011650485436893206,
246
  "gt_support": 326,
247
+ "predicted": 15639
248
  },
249
  "1": {
250
+ "accuracy": 0.2984537373231501,
251
+ "precision": 0.5789665432023803,
252
+ "recall": 0.2984537373231501,
253
+ "f1": 0.3938699217217093,
254
  "gt_support": 31948,
255
+ "predicted": 16469
256
  },
257
  "2": {
258
+ "accuracy": 0.3895705521472393,
259
+ "precision": 0.005864154776746548,
260
+ "recall": 0.3895705521472393,
261
+ "f1": 0.011554382932265841,
262
  "gt_support": 326,
263
+ "predicted": 21657
264
  },
265
  "3": {
266
+ "accuracy": 0.024184226904952432,
267
+ "precision": 0.40933032355154253,
268
+ "recall": 0.024184226904952432,
269
+ "f1": 0.04567015069470679,
270
  "gt_support": 22494,
271
+ "predicted": 1329
272
  }
273
  },
274
  "total_samples": 55094
 
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 13
290
  }
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 810 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.1867 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.5852 0.0059 0.5852 0.0117 810 80107
30
- # 0.2121 0.5151 0.2121 0.3005 66522 27387
31
- G 0.0630 0.0060 0.0630 0.0109 810 8566
32
- _ 0.1589 0.5246 0.1589 0.2440 68748 20830
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.1586 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.7188 0.0059 0.7188 0.0117 64 7774
49
- # 0.0781 0.2840 0.0781 0.1225 3072 845
50
- G 0.0156 0.0059 0.0156 0.0086 64 169
51
- _ 0.1875 0.7041 0.1875 0.2961 7616 2028
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.1879 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.6452 0.0059 0.6452 0.0117 62 6750
64
- # 0.0806 0.3432 0.0806 0.1306 3596 845
65
- G 0.0323 0.0059 0.0323 0.0100 62 338
66
- _ 0.2422 0.6432 0.2422 0.3519 6758 2545
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.1800 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.6240 0.0059 0.6240 0.0118 125 13151
79
- # 0.1700 0.4042 0.1700 0.2393 8500 3575
80
- G 0.0240 0.0059 0.0240 0.0095 125 507
81
- _ 0.1840 0.5850 0.1840 0.2800 12375 3892
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.1782 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.6281 0.0059 0.6281 0.0117 121 12842
94
- # 0.1986 0.4618 0.1986 0.2777 9438 4058
95
- G 0.0165 0.0059 0.0165 0.0087 121 338
96
- _ 0.1570 0.5266 0.1570 0.2419 10769 3211
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.1841 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.5714 0.0059 0.5714 0.0118 112 10779
109
- # 0.2049 0.5262 0.2049 0.2949 9968 3881
110
- G 0.0714 0.0062 0.0714 0.0113 112 1299
111
- _ 0.1568 0.4614 0.1568 0.2341 8736 2969
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.1987 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.5215 0.0059 0.5215 0.0117 326 28811
124
- # 0.2572 0.5794 0.2572 0.3562 31948 14183
125
- G 0.1074 0.0059 0.1074 0.0112 326 5915
126
- _ 0.1122 0.4081 0.1122 0.1760 22494 6185
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2756 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.2272 0.0061 0.2272 0.0118 810 30374
30
+ # 0.2683 0.5308 0.2683 0.3564 66522 33622
31
+ G 0.2963 0.0059 0.2963 0.0115 810 40992
32
+ _ 0.2830 0.6100 0.2830 0.3867 68748 31902
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.5634 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0625 0.0132 0.0625 0.0218 64 303
49
+ # 0.0000 0.0000 0.0000 0.0000 3072 0
50
+ G 0.1719 0.0058 0.1719 0.0112 64 1893
51
+ _ 0.7982 0.7052 0.7982 0.7488 7616 8620
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.3471 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.2258 0.0060 0.2258 0.0116 62 2352
64
+ # 0.0000 0.0000 0.0000 0.0000 3596 0
65
+ G 0.2419 0.0059 0.2419 0.0116 62 2528
66
+ _ 0.5339 0.6445 0.5339 0.5840 6758 5598
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.3511 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.1920 0.0061 0.1920 0.0118 125 3945
79
+ # 0.1127 0.3931 0.1127 0.1752 8500 2437
80
+ G 0.1760 0.0058 0.1760 0.0112 125 3791
81
+ _ 0.5181 0.5855 0.5181 0.5497 12375 10952
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.2640 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.2645 0.0063 0.2645 0.0123 121 5090
94
+ # 0.2921 0.4620 0.2921 0.3579 9438 5968
95
+ G 0.2149 0.0058 0.2149 0.0113 121 4495
96
+ _ 0.2399 0.5276 0.2399 0.3298 10769 4896
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2580 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.1518 0.0056 0.1518 0.0108 112 3045
109
+ # 0.4610 0.5253 0.4610 0.4910 9968 8748
110
+ G 0.3482 0.0059 0.3482 0.0116 112 6628
111
+ _ 0.0267 0.4596 0.0267 0.0504 8736 507
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.1869 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2853 0.0059 0.2853 0.0117 326 15639
124
+ # 0.2985 0.5790 0.2985 0.3939 31948 16469
125
+ G 0.3896 0.0059 0.3896 0.0116 326 21657
126
+ _ 0.0242 0.4093 0.0242 0.0457 22494 1329
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size13.json
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.json CHANGED
@@ -1,55 +1,55 @@
1
  {
2
  "global": {
3
- "accuracy": 0.04964499769479023,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.7904564315352697,
8
- "precision": 0.0044444444444444444,
9
- "recall": 0.7904564315352697,
10
- "f1": 0.008839189392972728,
11
  "gt_support": 964,
12
- "predicted": 171450
13
  },
14
  "1": {
15
- "accuracy": 0.0033663385885070634,
16
- "precision": 0.5051851851851852,
17
- "recall": 0.0033663385885070634,
18
- "f1": 0.006688110461695367,
19
  "gt_support": 101297,
20
- "predicted": 675
21
  },
22
  "2": {
23
- "accuracy": 0.11721991701244813,
24
- "precision": 0.0044444444444444444,
25
- "recall": 0.11721991701244813,
26
- "f1": 0.008564174466633824,
27
  "gt_support": 964,
28
- "predicted": 25425
29
  },
30
  "3": {
31
- "accuracy": 0.08402903012975588,
32
- "precision": 0.4936434108527132,
33
- "recall": 0.08402903012975588,
34
- "f1": 0.1436121029881601,
35
  "gt_support": 113675,
36
- "predicted": 19350
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.03875968992248062,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.7558139534883721,
48
- "precision": 0.0044444444444444444,
49
- "recall": 0.7558139534883721,
50
- "f1": 0.008836924750186935,
51
  "gt_support": 86,
52
- "predicted": 14625
53
  },
54
  "1": {
55
  "accuracy": 0.0,
@@ -60,35 +60,35 @@
60
  "predicted": 0
61
  },
62
  "2": {
63
- "accuracy": 0.19767441860465115,
64
- "precision": 0.0044444444444444444,
65
- "recall": 0.19767441860465115,
66
- "f1": 0.008693428790590641,
67
  "gt_support": 86,
68
- "predicted": 3825
69
  },
70
  "3": {
71
- "accuracy": 0.046511627906976744,
72
  "precision": 0.7422222222222222,
73
- "recall": 0.046511627906976744,
74
- "f1": 0.08753767527191718,
75
  "gt_support": 14362,
76
- "predicted": 900
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
- "accuracy": 0.043014048531289914,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.8620689655172413,
87
- "precision": 0.0044444444444444444,
88
- "recall": 0.8620689655172413,
89
- "f1": 0.008843296781039971,
90
  "gt_support": 87,
91
- "predicted": 16875
92
  },
93
  "1": {
94
  "accuracy": 0.0,
@@ -99,176 +99,176 @@
99
  "predicted": 0
100
  },
101
  "2": {
102
- "accuracy": 0.08045977011494253,
103
- "precision": 0.0044444444444444444,
104
- "recall": 0.08045977011494253,
105
- "f1": 0.00842358604091456,
106
  "gt_support": 87,
107
- "predicted": 1575
108
  },
109
  "3": {
110
- "accuracy": 0.05747126436781609,
111
- "precision": 0.6755555555555556,
112
- "recall": 0.05747126436781609,
113
- "f1": 0.10593072687992194,
114
  "gt_support": 13224,
115
- "predicted": 1125
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
- "accuracy": 0.056437908496732025,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.7426470588235294,
126
- "precision": 0.0044444444444444444,
127
- "recall": 0.7426470588235294,
128
- "f1": 0.00883600892349416,
129
  "gt_support": 136,
130
- "predicted": 22725
131
  },
132
  "1": {
133
- "accuracy": 0.007352941176470588,
134
- "precision": 0.37777777777777777,
135
- "recall": 0.007352941176470588,
136
- "f1": 0.014425116673737802,
137
  "gt_support": 11560,
138
- "predicted": 225
139
  },
140
  "2": {
141
- "accuracy": 0.16911764705882354,
142
- "precision": 0.0044444444444444444,
143
- "recall": 0.16911764705882354,
144
- "f1": 0.008661269064206363,
145
  "gt_support": 136,
146
- "predicted": 5175
147
  },
148
  "3": {
149
- "accuracy": 0.08088235294117647,
150
- "precision": 0.6133333333333333,
151
- "recall": 0.08088235294117647,
152
- "f1": 0.14291766699618697,
153
  "gt_support": 18768,
154
- "predicted": 2475
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
- "accuracy": 0.047822222222222224,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.78,
165
- "precision": 0.0044444444444444444,
166
- "recall": 0.78,
167
- "f1": 0.008838526912181304,
168
  "gt_support": 100,
169
- "predicted": 17550
170
  },
171
  "1": {
172
- "accuracy": 0.0,
173
- "precision": 0.0,
174
- "recall": 0.0,
175
- "f1": 0.0,
176
  "gt_support": 10000,
177
- "predicted": 0
178
  },
179
  "2": {
180
- "accuracy": 0.14,
181
- "precision": 0.0044444444444444444,
182
- "recall": 0.14,
183
- "f1": 0.008615384615384615,
184
  "gt_support": 100,
185
- "predicted": 3150
186
  },
187
  "3": {
188
- "accuracy": 0.08,
189
- "precision": 0.5466666666666666,
190
- "recall": 0.08,
191
- "f1": 0.1395744680851064,
192
  "gt_support": 12300,
193
- "predicted": 1800
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
- "accuracy": 0.02785907859078591,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.8841463414634146,
204
- "precision": 0.0044444444444444444,
205
- "recall": 0.8841463414634146,
206
- "f1": 0.008844429534295038,
207
  "gt_support": 164,
208
- "predicted": 32625
209
  },
210
  "1": {
211
- "accuracy": 0.0,
212
- "precision": 0.0,
213
- "recall": 0.0,
214
- "f1": 0.0,
215
  "gt_support": 18696,
216
- "predicted": 0
217
  },
218
  "2": {
219
- "accuracy": 0.06707317073170732,
220
- "precision": 0.0044444444444444444,
221
- "recall": 0.06707317073170732,
222
- "f1": 0.008336491095111784,
223
  "gt_support": 164,
224
- "predicted": 2475
225
  },
226
  "3": {
227
- "accuracy": 0.04878048780487805,
228
- "precision": 0.48444444444444446,
229
- "recall": 0.04878048780487805,
230
- "f1": 0.08863590160601748,
231
  "gt_support": 17876,
232
- "predicted": 1800
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
- "accuracy": 0.06075589656152316,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.7621483375959079,
243
- "precision": 0.0044444444444444444,
244
- "recall": 0.7621483375959079,
245
- "f1": 0.008837354131759612,
246
  "gt_support": 391,
247
- "predicted": 67050
248
  },
249
  "1": {
250
- "accuracy": 0.005115089514066497,
251
- "precision": 0.5688888888888889,
252
- "recall": 0.005115089514066497,
253
- "f1": 0.010139015406550756,
254
  "gt_support": 50048,
255
- "predicted": 450
256
  },
257
  "2": {
258
- "accuracy": 0.10485933503836317,
259
- "precision": 0.0044444444444444444,
260
- "recall": 0.10485933503836317,
261
- "f1": 0.008527454242928452,
262
  "gt_support": 391,
263
- "predicted": 9225
264
  },
265
  "3": {
266
- "accuracy": 0.1278772378516624,
267
- "precision": 0.4222222222222222,
268
- "recall": 0.1278772378516624,
269
- "f1": 0.19630127079243725,
270
  "gt_support": 37145,
271
- "predicted": 11250
272
  }
273
  },
274
  "total_samples": 87975
@@ -278,13 +278,13 @@
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 15
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.39108344859382205,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.1524896265560166,
8
+ "precision": 0.004301893418395716,
9
+ "recall": 0.1524896265560166,
10
+ "f1": 0.008367724491248044,
11
  "gt_support": 964,
12
+ "predicted": 34171
13
  },
14
  "1": {
15
+ "accuracy": 0.48209719932475786,
16
+ "precision": 0.5436864019950569,
17
+ "recall": 0.48209719932475786,
18
+ "f1": 0.5110428581145778,
19
  "gt_support": 101297,
20
+ "predicted": 89822
21
  },
22
  "2": {
23
+ "accuracy": 0.18568464730290457,
24
+ "precision": 0.004411148623672343,
25
+ "recall": 0.18568464730290457,
26
+ "f1": 0.008617576968442337,
27
  "gt_support": 964,
28
+ "predicted": 40579
29
  },
30
  "3": {
31
+ "accuracy": 0.31374532658895976,
32
+ "precision": 0.6815662742699893,
33
+ "recall": 0.31374532658895976,
34
+ "f1": 0.42969102968018646,
35
  "gt_support": 113675,
36
+ "predicted": 52328
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.7422222222222222,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.0,
48
+ "precision": 0.0,
49
+ "recall": 0.0,
50
+ "f1": 0.0,
51
  "gt_support": 86,
52
+ "predicted": 0
53
  },
54
  "1": {
55
  "accuracy": 0.0,
 
60
  "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.0,
64
+ "precision": 0.0,
65
+ "recall": 0.0,
66
+ "f1": 0.0,
67
  "gt_support": 86,
68
+ "predicted": 0
69
  },
70
  "3": {
71
+ "accuracy": 1.0,
72
  "precision": 0.7422222222222222,
73
+ "recall": 1.0,
74
+ "f1": 0.8520408163265306,
75
  "gt_support": 14362,
76
+ "predicted": 19350
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
+ "accuracy": 0.6208429118773946,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.034482758620689655,
87
+ "precision": 0.003667481662591687,
88
+ "recall": 0.034482758620689655,
89
+ "f1": 0.006629834254143646,
90
  "gt_support": 87,
91
+ "predicted": 818
92
  },
93
  "1": {
94
  "accuracy": 0.0,
 
99
  "predicted": 0
100
  },
101
  "2": {
102
+ "accuracy": 0.034482758620689655,
103
+ "precision": 0.0038510911424903724,
104
+ "recall": 0.034482758620689655,
105
+ "f1": 0.006928406466512702,
106
  "gt_support": 87,
107
+ "predicted": 779
108
  },
109
  "3": {
110
+ "accuracy": 0.918557168784029,
111
+ "precision": 0.6756591389476027,
112
+ "recall": 0.918557168784029,
113
+ "f1": 0.778603935645151,
114
  "gt_support": 13224,
115
+ "predicted": 17978
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
+ "accuracy": 0.3099673202614379,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.34558823529411764,
126
+ "precision": 0.004229281022226221,
127
+ "recall": 0.34558823529411764,
128
+ "f1": 0.008356298337630012,
129
  "gt_support": 136,
130
+ "predicted": 11113
131
  },
132
  "1": {
133
+ "accuracy": 0.05328719723183391,
134
+ "precision": 0.38213399503722084,
135
+ "recall": 0.05328719723183391,
136
+ "f1": 0.09353173398117218,
137
  "gt_support": 11560,
138
+ "predicted": 1612
139
  },
140
  "2": {
141
+ "accuracy": 0.11764705882352941,
142
+ "precision": 0.004567513559805881,
143
+ "recall": 0.11764705882352941,
144
+ "f1": 0.008793624622148942,
145
  "gt_support": 136,
146
+ "predicted": 3503
147
  },
148
  "3": {
149
+ "accuracy": 0.4692028985507246,
150
+ "precision": 0.6127191761758976,
151
+ "recall": 0.4692028985507246,
152
+ "f1": 0.5314423657211828,
153
  "gt_support": 18768,
154
+ "predicted": 14372
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
+ "accuracy": 0.1156,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.43,
165
+ "precision": 0.0043967280163599185,
166
+ "recall": 0.43,
167
+ "f1": 0.008704453441295546,
168
  "gt_support": 100,
169
+ "predicted": 9780
170
  },
171
  "1": {
172
+ "accuracy": 0.2177,
173
+ "precision": 0.4455587392550143,
174
+ "recall": 0.2177,
175
+ "f1": 0.2924895875319092,
176
  "gt_support": 10000,
177
+ "predicted": 4886
178
  },
179
  "2": {
180
+ "accuracy": 0.31,
181
+ "precision": 0.004301970580072162,
182
+ "recall": 0.31,
183
+ "f1": 0.008486175745962223,
184
  "gt_support": 100,
185
+ "predicted": 7206
186
  },
187
  "3": {
188
+ "accuracy": 0.028455284552845527,
189
+ "precision": 0.5573248407643312,
190
+ "recall": 0.028455284552845527,
191
+ "f1": 0.054146039603960396,
192
  "gt_support": 12300,
193
+ "predicted": 628
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
+ "accuracy": 0.3035230352303523,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.23170731707317074,
204
+ "precision": 0.004137630662020906,
205
+ "recall": 0.23170731707317074,
206
+ "f1": 0.008130081300813007,
207
  "gt_support": 164,
208
+ "predicted": 9184
209
  },
210
  "1": {
211
+ "accuracy": 0.5956889174154899,
212
+ "precision": 0.5068954530972646,
213
+ "recall": 0.5956889174154899,
214
+ "f1": 0.5477168219932623,
215
  "gt_support": 18696,
216
+ "predicted": 21971
217
  },
218
  "2": {
219
+ "accuracy": 0.1524390243902439,
220
+ "precision": 0.004351610095735422,
221
+ "recall": 0.1524390243902439,
222
+ "f1": 0.008461668641056016,
223
  "gt_support": 164,
224
+ "predicted": 5745
225
  },
226
  "3": {
227
+ "accuracy": 0.0,
228
+ "precision": 0.0,
229
+ "recall": 0.0,
230
+ "f1": 0.0,
231
  "gt_support": 17876,
232
+ "predicted": 0
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
+ "accuracy": 0.3981244671781756,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.04092071611253197,
243
+ "precision": 0.004884004884004884,
244
+ "recall": 0.04092071611253197,
245
+ "f1": 0.00872647941096264,
246
  "gt_support": 391,
247
+ "predicted": 3276
248
  },
249
  "1": {
250
+ "accuracy": 0.6974304667519181,
251
+ "precision": 0.5689208351669846,
252
+ "recall": 0.6974304667519181,
253
+ "f1": 0.6266550569564009,
254
  "gt_support": 50048,
255
+ "predicted": 61353
256
  },
257
  "2": {
258
+ "accuracy": 0.2659846547314578,
259
+ "precision": 0.0044547245780861815,
260
+ "recall": 0.2659846547314578,
261
+ "f1": 0.008762691157265029,
262
  "gt_support": 391,
263
+ "predicted": 23346
264
  },
265
  "3": {
266
+ "accuracy": 0.0,
267
+ "precision": 0.0,
268
+ "recall": 0.0,
269
+ "f1": 0.0,
270
  "gt_support": 37145,
271
+ "predicted": 0
272
  }
273
  },
274
  "total_samples": 87975
 
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 15
290
  }
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 964 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.0496 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.7905 0.0044 0.7905 0.0088 964 171450
30
- # 0.0034 0.5052 0.0034 0.0067 101297 675
31
- G 0.1172 0.0044 0.1172 0.0086 964 25425
32
- _ 0.0840 0.4936 0.0840 0.1436 113675 19350
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.0388 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.7558 0.0044 0.7558 0.0088 86 14625
49
  # 0.0000 0.0000 0.0000 0.0000 4816 0
50
- G 0.1977 0.0044 0.1977 0.0087 86 3825
51
- _ 0.0465 0.7422 0.0465 0.0875 14362 900
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.0430 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.8621 0.0044 0.8621 0.0088 87 16875
64
  # 0.0000 0.0000 0.0000 0.0000 6177 0
65
- G 0.0805 0.0044 0.0805 0.0084 87 1575
66
- _ 0.0575 0.6756 0.0575 0.1059 13224 1125
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.0564 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.7426 0.0044 0.7426 0.0088 136 22725
79
- # 0.0074 0.3778 0.0074 0.0144 11560 225
80
- G 0.1691 0.0044 0.1691 0.0087 136 5175
81
- _ 0.0809 0.6133 0.0809 0.1429 18768 2475
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0478 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.7800 0.0044 0.7800 0.0088 100 17550
94
- # 0.0000 0.0000 0.0000 0.0000 10000 0
95
- G 0.1400 0.0044 0.1400 0.0086 100 3150
96
- _ 0.0800 0.5467 0.0800 0.1396 12300 1800
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0279 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.8841 0.0044 0.8841 0.0088 164 32625
109
- # 0.0000 0.0000 0.0000 0.0000 18696 0
110
- G 0.0671 0.0044 0.0671 0.0083 164 2475
111
- _ 0.0488 0.4844 0.0488 0.0886 17876 1800
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.0608 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.7621 0.0044 0.7621 0.0088 391 67050
124
- # 0.0051 0.5689 0.0051 0.0101 50048 450
125
- G 0.1049 0.0044 0.1049 0.0085 391 9225
126
- _ 0.1279 0.4222 0.1279 0.1963 37145 11250
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.3911 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.1525 0.0043 0.1525 0.0084 964 34171
30
+ # 0.4821 0.5437 0.4821 0.5110 101297 89822
31
+ G 0.1857 0.0044 0.1857 0.0086 964 40579
32
+ _ 0.3137 0.6816 0.3137 0.4297 113675 52328
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.7422 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0000 0.0000 0.0000 0.0000 86 0
49
  # 0.0000 0.0000 0.0000 0.0000 4816 0
50
+ G 0.0000 0.0000 0.0000 0.0000 86 0
51
+ _ 1.0000 0.7422 1.0000 0.8520 14362 19350
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.6208 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.0345 0.0037 0.0345 0.0066 87 818
64
  # 0.0000 0.0000 0.0000 0.0000 6177 0
65
+ G 0.0345 0.0039 0.0345 0.0069 87 779
66
+ _ 0.9186 0.6757 0.9186 0.7786 13224 17978
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.3100 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.3456 0.0042 0.3456 0.0084 136 11113
79
+ # 0.0533 0.3821 0.0533 0.0935 11560 1612
80
+ G 0.1176 0.0046 0.1176 0.0088 136 3503
81
+ _ 0.4692 0.6127 0.4692 0.5314 18768 14372
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.1156 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.4300 0.0044 0.4300 0.0087 100 9780
94
+ # 0.2177 0.4456 0.2177 0.2925 10000 4886
95
+ G 0.3100 0.0043 0.3100 0.0085 100 7206
96
+ _ 0.0285 0.5573 0.0285 0.0541 12300 628
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.3035 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2317 0.0041 0.2317 0.0081 164 9184
109
+ # 0.5957 0.5069 0.5957 0.5477 18696 21971
110
+ G 0.1524 0.0044 0.1524 0.0085 164 5745
111
+ _ 0.0000 0.0000 0.0000 0.0000 17876 0
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.3981 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.0409 0.0049 0.0409 0.0087 391 3276
124
+ # 0.6974 0.5689 0.6974 0.6267 50048 61353
125
+ G 0.2660 0.0045 0.2660 0.0088 391 23346
126
+ _ 0.0000 0.0000 0.0000 0.0000 37145 0
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size15.json
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.1399314970743542,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.3146853146853147,
8
- "precision": 0.020183897734918143,
9
- "recall": 0.3146853146853147,
10
- "f1": 0.037934668071654375,
11
  "gt_support": 286,
12
- "predicted": 4459
13
  },
14
  "1": {
15
- "accuracy": 0.0623784046692607,
16
- "precision": 0.5816326530612245,
17
- "recall": 0.0623784046692607,
18
- "f1": 0.11267296288161652,
19
  "gt_support": 8224,
20
- "predicted": 882
21
  },
22
  "2": {
23
- "accuracy": 0.36363636363636365,
24
- "precision": 0.020100502512562814,
25
- "recall": 0.36363636363636365,
26
- "f1": 0.0380952380952381,
27
  "gt_support": 286,
28
- "predicted": 5174
29
  },
30
  "3": {
31
- "accuracy": 0.24032196243771559,
32
- "precision": 0.3583881108888254,
33
- "recall": 0.24032196243771559,
34
- "f1": 0.2877136629574395,
35
  "gt_support": 5218,
36
- "predicted": 3499
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.09774436090225563,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.3157894736842105,
48
- "precision": 0.02054794520547945,
49
- "recall": 0.3157894736842105,
50
- "f1": 0.03858520900321544,
51
  "gt_support": 38,
52
- "predicted": 584
53
  },
54
  "1": {
55
- "accuracy": 0.05263157894736842,
56
- "precision": 0.4897959183673469,
57
- "recall": 0.05263157894736842,
58
- "f1": 0.09504950495049504,
59
  "gt_support": 912,
60
- "predicted": 98
61
  },
62
  "2": {
63
- "accuracy": 0.5,
64
- "precision": 0.01981230448383733,
65
- "recall": 0.5,
66
- "f1": 0.03811434302908726,
67
  "gt_support": 38,
68
- "predicted": 959
69
  },
70
  "3": {
71
- "accuracy": 0.11784897025171624,
72
- "precision": 0.4660633484162896,
73
- "recall": 0.11784897025171624,
74
- "f1": 0.18812785388127853,
75
  "gt_support": 874,
76
- "predicted": 221
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
- "accuracy": 0.12894248608534323,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.3409090909090909,
87
- "precision": 0.01984126984126984,
88
- "recall": 0.3409090909090909,
89
- "f1": 0.0375,
90
  "gt_support": 44,
91
- "predicted": 756
92
  },
93
  "1": {
94
- "accuracy": 0.06818181818181818,
95
- "precision": 0.5306122448979592,
96
- "recall": 0.06818181818181818,
97
- "f1": 0.12083656080557706,
98
  "gt_support": 1144,
99
- "predicted": 147
100
  },
101
  "2": {
102
- "accuracy": 0.38636363636363635,
103
- "precision": 0.019744483159117306,
104
- "recall": 0.38636363636363635,
105
- "f1": 0.03756906077348066,
106
  "gt_support": 44,
107
- "predicted": 861
108
  },
109
  "3": {
110
- "accuracy": 0.18181818181818182,
111
- "precision": 0.42857142857142855,
112
- "recall": 0.18181818181818182,
113
- "f1": 0.2553191489361702,
114
  "gt_support": 924,
115
- "predicted": 392
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
- "accuracy": 0.19761075161772026,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.3170731707317073,
126
- "precision": 0.019969278033794162,
127
- "recall": 0.3170731707317073,
128
- "f1": 0.03757225433526012,
129
  "gt_support": 41,
130
- "predicted": 651
131
  },
132
  "1": {
133
- "accuracy": 0.14634146341463414,
134
- "precision": 0.5714285714285714,
135
- "recall": 0.14634146341463414,
136
- "f1": 0.23300970873786406,
137
  "gt_support": 1148,
138
- "predicted": 294
139
  },
140
  "2": {
141
- "accuracy": 0.2682926829268293,
142
- "precision": 0.02040816326530612,
143
- "recall": 0.2682926829268293,
144
- "f1": 0.03793103448275861,
145
  "gt_support": 41,
146
- "predicted": 539
147
  },
148
  "3": {
149
- "accuracy": 0.2631578947368421,
150
- "precision": 0.3904761904761905,
151
- "recall": 0.2631578947368421,
152
- "f1": 0.31441717791411045,
153
  "gt_support": 779,
154
- "predicted": 525
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
- "accuracy": 0.12331741207121147,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.3191489361702128,
165
- "precision": 0.0195822454308094,
166
- "recall": 0.3191489361702128,
167
- "f1": 0.036900369003690044,
168
  "gt_support": 47,
169
- "predicted": 766
170
  },
171
  "1": {
172
- "accuracy": 0.02127659574468085,
173
- "precision": 0.5918367346938775,
174
- "recall": 0.02127659574468085,
175
- "f1": 0.04107648725212464,
176
  "gt_support": 1363,
177
- "predicted": 49
178
  },
179
  "2": {
180
- "accuracy": 0.3829787234042553,
181
- "precision": 0.020247469066366704,
182
- "recall": 0.3829787234042553,
183
- "f1": 0.038461538461538464,
184
  "gt_support": 47,
185
- "predicted": 889
186
  },
187
  "3": {
188
- "accuracy": 0.2624113475177305,
189
- "precision": 0.37061769616026713,
190
- "recall": 0.2624113475177305,
191
- "f1": 0.30726643598615916,
192
  "gt_support": 846,
193
- "predicted": 599
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
- "accuracy": 0.13172541743970315,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.34545454545454546,
204
- "precision": 0.020191285866099893,
205
- "recall": 0.34545454545454546,
206
- "f1": 0.038152610441767064,
207
  "gt_support": 55,
208
- "predicted": 941
209
  },
210
  "1": {
211
- "accuracy": 0.03636363636363636,
212
- "precision": 0.6326530612244898,
213
- "recall": 0.03636363636363636,
214
- "f1": 0.06877426511369938,
215
  "gt_support": 1705,
216
- "predicted": 98
217
  },
218
  "2": {
219
- "accuracy": 0.32727272727272727,
220
- "precision": 0.020594965675057208,
221
- "recall": 0.32727272727272727,
222
- "f1": 0.038751345532831,
223
  "gt_support": 55,
224
- "predicted": 874
225
  },
226
  "3": {
227
- "accuracy": 0.2909090909090909,
228
- "precision": 0.3273657289002558,
229
- "recall": 0.2909090909090909,
230
- "f1": 0.30806257521058966,
231
  "gt_support": 880,
232
- "predicted": 782
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
- "accuracy": 0.155570424891268,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.26229508196721313,
243
- "precision": 0.02102496714848883,
244
- "recall": 0.26229508196721313,
245
- "f1": 0.038929440389294405,
246
  "gt_support": 61,
247
- "predicted": 761
248
  },
249
  "1": {
250
- "accuracy": 0.06557377049180328,
251
- "precision": 0.6530612244897959,
252
- "recall": 0.06557377049180328,
253
- "f1": 0.11918063314711361,
254
  "gt_support": 1952,
255
- "predicted": 196
256
  },
257
  "2": {
258
- "accuracy": 0.3442622950819672,
259
- "precision": 0.019961977186311788,
260
- "recall": 0.3442622950819672,
261
- "f1": 0.03773584905660378,
262
  "gt_support": 61,
263
- "predicted": 1052
264
  },
265
  "3": {
266
- "accuracy": 0.32786885245901637,
267
- "precision": 0.30612244897959184,
268
- "recall": 0.32786885245901637,
269
- "f1": 0.31662269129287596,
270
  "gt_support": 915,
271
- "predicted": 980
272
  }
273
  },
274
  "total_samples": 2989
@@ -278,13 +278,13 @@
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 7
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.20343941772513202,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.34965034965034963,
8
+ "precision": 0.02002803925495694,
9
+ "recall": 0.34965034965034963,
10
+ "f1": 0.03788596325061565,
11
  "gt_support": 286,
12
+ "predicted": 4993
13
  },
14
  "1": {
15
+ "accuracy": 0.17935311284046693,
16
+ "precision": 0.6156093489148581,
17
+ "recall": 0.17935311284046693,
18
+ "f1": 0.2777777777777778,
19
  "gt_support": 8224,
20
+ "predicted": 2396
21
  },
22
  "2": {
23
+ "accuracy": 0.26223776223776224,
24
+ "precision": 0.020048115477145148,
25
+ "recall": 0.26223776223776224,
26
+ "f1": 0.037248572138068044,
27
  "gt_support": 286,
28
+ "predicted": 3741
29
  },
30
  "3": {
31
+ "accuracy": 0.23016481410502107,
32
+ "precision": 0.41643550624133147,
33
+ "recall": 0.23016481410502107,
34
+ "f1": 0.2964700074055789,
35
  "gt_support": 5218,
36
+ "predicted": 2884
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.26100966702470463,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.10526315789473684,
48
+ "precision": 0.02072538860103627,
49
+ "recall": 0.10526315789473684,
50
+ "f1": 0.03463203463203463,
51
  "gt_support": 38,
52
+ "predicted": 193
53
  },
54
  "1": {
55
+ "accuracy": 0.0,
56
+ "precision": 0.0,
57
+ "recall": 0.0,
58
+ "f1": 0.0,
59
  "gt_support": 912,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.42105263157894735,
64
+ "precision": 0.023323615160349854,
65
+ "recall": 0.42105263157894735,
66
+ "f1": 0.04419889502762431,
67
  "gt_support": 38,
68
+ "predicted": 686
69
  },
70
  "3": {
71
+ "accuracy": 0.5331807780320366,
72
+ "precision": 0.474059003051882,
73
+ "recall": 0.5331807780320366,
74
+ "f1": 0.5018847603661819,
75
  "gt_support": 874,
76
+ "predicted": 983
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
+ "accuracy": 0.19851576994434136,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.38636363636363635,
87
+ "precision": 0.01990632318501171,
88
+ "recall": 0.38636363636363635,
89
+ "f1": 0.0378619153674833,
90
  "gt_support": 44,
91
+ "predicted": 854
92
  },
93
  "1": {
94
+ "accuracy": 0.05856643356643357,
95
+ "precision": 0.5403225806451613,
96
+ "recall": 0.05856643356643357,
97
+ "f1": 0.10567823343848581,
98
  "gt_support": 1144,
99
+ "predicted": 124
100
  },
101
  "2": {
102
+ "accuracy": 0.18181818181818182,
103
+ "precision": 0.019851116625310174,
104
+ "recall": 0.18181818181818182,
105
+ "f1": 0.035794183445190156,
106
  "gt_support": 44,
107
+ "predicted": 403
108
  },
109
  "3": {
110
+ "accuracy": 0.36363636363636365,
111
+ "precision": 0.4335483870967742,
112
+ "recall": 0.36363636363636365,
113
+ "f1": 0.3955267804590936,
114
  "gt_support": 924,
115
+ "predicted": 775
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
+ "accuracy": 0.13091090094574415,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.5121951219512195,
126
+ "precision": 0.02001906577693041,
127
+ "recall": 0.5121951219512195,
128
+ "f1": 0.03853211009174312,
129
  "gt_support": 41,
130
+ "predicted": 1049
131
  },
132
  "1": {
133
+ "accuracy": 0.07317073170731707,
134
+ "precision": 0.5675675675675675,
135
+ "recall": 0.07317073170731707,
136
+ "f1": 0.12962962962962962,
137
  "gt_support": 1148,
138
+ "predicted": 148
139
  },
140
  "2": {
141
+ "accuracy": 0.1951219512195122,
142
+ "precision": 0.01937046004842615,
143
+ "recall": 0.1951219512195122,
144
+ "f1": 0.03524229074889868,
145
  "gt_support": 41,
146
+ "predicted": 413
147
  },
148
  "3": {
149
+ "accuracy": 0.1925545571245186,
150
+ "precision": 0.37593984962406013,
151
+ "recall": 0.1925545571245186,
152
+ "f1": 0.2546689303904923,
153
  "gt_support": 779,
154
+ "predicted": 399
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
+ "accuracy": 0.22839774207555363,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.48936170212765956,
165
+ "precision": 0.020390070921985817,
166
+ "recall": 0.48936170212765956,
167
+ "f1": 0.03914893617021277,
168
  "gt_support": 47,
169
+ "predicted": 1128
170
  },
171
  "1": {
172
+ "accuracy": 0.2920029347028613,
173
+ "precision": 0.5887573964497042,
174
+ "recall": 0.2920029347028613,
175
+ "f1": 0.39038744482589505,
176
  "gt_support": 1363,
177
+ "predicted": 676
178
  },
179
  "2": {
180
+ "accuracy": 0.06382978723404255,
181
+ "precision": 0.012711864406779662,
182
+ "recall": 0.06382978723404255,
183
+ "f1": 0.02120141342756184,
184
  "gt_support": 47,
185
+ "predicted": 236
186
  },
187
  "3": {
188
+ "accuracy": 0.12056737588652482,
189
+ "precision": 0.38783269961977185,
190
+ "recall": 0.12056737588652482,
191
+ "f1": 0.18394950405770966,
192
  "gt_support": 846,
193
+ "predicted": 263
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
+ "accuracy": 0.25009276437847866,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.3090909090909091,
204
+ "precision": 0.01990632318501171,
205
+ "recall": 0.3090909090909091,
206
+ "f1": 0.0374037403740374,
207
  "gt_support": 55,
208
+ "predicted": 854
209
  },
210
  "1": {
211
+ "accuracy": 0.35777126099706746,
212
+ "precision": 0.6308169596690796,
213
+ "recall": 0.35777126099706746,
214
+ "f1": 0.4565868263473054,
215
  "gt_support": 1705,
216
+ "predicted": 967
217
  },
218
  "2": {
219
+ "accuracy": 0.2727272727272727,
220
+ "precision": 0.019329896907216496,
221
+ "recall": 0.2727272727272727,
222
+ "f1": 0.03610108303249098,
223
  "gt_support": 55,
224
+ "predicted": 776
225
  },
226
  "3": {
227
+ "accuracy": 0.03636363636363636,
228
+ "precision": 0.32653061224489793,
229
+ "recall": 0.03636363636363636,
230
+ "f1": 0.06543967280163598,
231
  "gt_support": 880,
232
+ "predicted": 98
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
+ "accuracy": 0.15858146537303447,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.29508196721311475,
243
+ "precision": 0.019672131147540985,
244
+ "recall": 0.29508196721311475,
245
+ "f1": 0.03688524590163935,
246
  "gt_support": 61,
247
+ "predicted": 915
248
  },
249
  "1": {
250
+ "accuracy": 0.16188524590163936,
251
+ "precision": 0.656964656964657,
252
+ "recall": 0.16188524590163936,
253
+ "f1": 0.25976161117961366,
254
  "gt_support": 1952,
255
+ "predicted": 481
256
  },
257
  "2": {
258
+ "accuracy": 0.4098360655737705,
259
+ "precision": 0.020374898125509373,
260
+ "recall": 0.4098360655737705,
261
+ "f1": 0.03881987577639752,
262
  "gt_support": 61,
263
+ "predicted": 1227
264
  },
265
  "3": {
266
+ "accuracy": 0.12568306010928962,
267
+ "precision": 0.31420765027322406,
268
+ "recall": 0.12568306010928962,
269
+ "f1": 0.1795472287275566,
270
  "gt_support": 915,
271
+ "predicted": 366
272
  }
273
  },
274
  "total_samples": 2989
 
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 7
290
  }
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 286 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.1399 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.3147 0.0202 0.3147 0.0379 286 4459
30
- # 0.0624 0.5816 0.0624 0.1127 8224 882
31
- G 0.3636 0.0201 0.3636 0.0381 286 5174
32
- _ 0.2403 0.3584 0.2403 0.2877 5218 3499
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.0977 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.3158 0.0205 0.3158 0.0386 38 584
49
- # 0.0526 0.4898 0.0526 0.0950 912 98
50
- G 0.5000 0.0198 0.5000 0.0381 38 959
51
- _ 0.1178 0.4661 0.1178 0.1881 874 221
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.1289 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.3409 0.0198 0.3409 0.0375 44 756
64
- # 0.0682 0.5306 0.0682 0.1208 1144 147
65
- G 0.3864 0.0197 0.3864 0.0376 44 861
66
- _ 0.1818 0.4286 0.1818 0.2553 924 392
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.1976 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.3171 0.0200 0.3171 0.0376 41 651
79
- # 0.1463 0.5714 0.1463 0.2330 1148 294
80
- G 0.2683 0.0204 0.2683 0.0379 41 539
81
- _ 0.2632 0.3905 0.2632 0.3144 779 525
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.1233 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.3191 0.0196 0.3191 0.0369 47 766
94
- # 0.0213 0.5918 0.0213 0.0411 1363 49
95
- G 0.3830 0.0202 0.3830 0.0385 47 889
96
- _ 0.2624 0.3706 0.2624 0.3073 846 599
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.1317 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.3455 0.0202 0.3455 0.0382 55 941
109
- # 0.0364 0.6327 0.0364 0.0688 1705 98
110
- G 0.3273 0.0206 0.3273 0.0388 55 874
111
- _ 0.2909 0.3274 0.2909 0.3081 880 782
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.1556 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.2623 0.0210 0.2623 0.0389 61 761
124
- # 0.0656 0.6531 0.0656 0.1192 1952 196
125
- G 0.3443 0.0200 0.3443 0.0377 61 1052
126
- _ 0.3279 0.3061 0.3279 0.3166 915 980
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2034 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.3497 0.0200 0.3497 0.0379 286 4993
30
+ # 0.1794 0.6156 0.1794 0.2778 8224 2396
31
+ G 0.2622 0.0200 0.2622 0.0372 286 3741
32
+ _ 0.2302 0.4164 0.2302 0.2965 5218 2884
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.2610 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.1053 0.0207 0.1053 0.0346 38 193
49
+ # 0.0000 0.0000 0.0000 0.0000 912 0
50
+ G 0.4211 0.0233 0.4211 0.0442 38 686
51
+ _ 0.5332 0.4741 0.5332 0.5019 874 983
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.1985 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.3864 0.0199 0.3864 0.0379 44 854
64
+ # 0.0586 0.5403 0.0586 0.1057 1144 124
65
+ G 0.1818 0.0199 0.1818 0.0358 44 403
66
+ _ 0.3636 0.4335 0.3636 0.3955 924 775
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.1309 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.5122 0.0200 0.5122 0.0385 41 1049
79
+ # 0.0732 0.5676 0.0732 0.1296 1148 148
80
+ G 0.1951 0.0194 0.1951 0.0352 41 413
81
+ _ 0.1926 0.3759 0.1926 0.2547 779 399
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.2284 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.4894 0.0204 0.4894 0.0391 47 1128
94
+ # 0.2920 0.5888 0.2920 0.3904 1363 676
95
+ G 0.0638 0.0127 0.0638 0.0212 47 236
96
+ _ 0.1206 0.3878 0.1206 0.1839 846 263
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2501 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.3091 0.0199 0.3091 0.0374 55 854
109
+ # 0.3578 0.6308 0.3578 0.4566 1705 967
110
+ G 0.2727 0.0193 0.2727 0.0361 55 776
111
+ _ 0.0364 0.3265 0.0364 0.0654 880 98
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.1586 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2951 0.0197 0.2951 0.0369 61 915
124
+ # 0.1619 0.6570 0.1619 0.2598 1952 481
125
+ G 0.4098 0.0204 0.4098 0.0388 61 1227
126
+ _ 0.1257 0.3142 0.1257 0.1795 915 366
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size7.json
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.07625978767353606,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0019455252918287938,
8
- "precision": 0.012345679012345678,
9
- "recall": 0.0019455252918287938,
10
- "f1": 0.0033613445378151263,
11
  "gt_support": 514,
12
- "predicted": 81
13
  },
14
  "1": {
15
- "accuracy": 0.00141155712395236,
16
- "precision": 0.3950617283950617,
17
- "recall": 0.00141155712395236,
18
- "f1": 0.0028130631620588106,
19
  "gt_support": 22670,
20
- "predicted": 81
21
  },
22
  "2": {
23
- "accuracy": 0.8463035019455253,
24
- "precision": 0.012345679012345678,
25
- "recall": 0.8463035019455253,
26
- "f1": 0.024336345072589444,
27
  "gt_support": 514,
28
- "predicted": 35235
29
  },
30
  "3": {
31
- "accuracy": 0.15092551293487957,
32
- "precision": 0.4340227673561007,
33
- "recall": 0.15092551293487957,
34
- "f1": 0.22396889091134736,
35
  "gt_support": 17936,
36
- "predicted": 6237
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.11408321902149063,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
 
 
 
 
 
 
 
 
47
  "accuracy": 0.0,
48
  "precision": 0.0,
49
  "recall": 0.0,
50
  "f1": 0.0,
51
- "gt_support": 54,
52
- "predicted": 0
53
- },
54
- "1": {
55
- "accuracy": 0.018518518518518517,
56
- "precision": 0.3950617283950617,
57
- "recall": 0.018518518518518517,
58
- "f1": 0.035378662244333885,
59
  "gt_support": 1728,
60
- "predicted": 81
61
  },
62
  "2": {
63
- "accuracy": 0.8148148148148148,
64
- "precision": 0.012345679012345678,
65
- "recall": 0.8148148148148148,
66
- "f1": 0.024322830292979547,
67
  "gt_support": 54,
68
- "predicted": 3564
69
  },
70
  "3": {
71
- "accuracy": 0.16666666666666666,
72
- "precision": 0.5802469135802469,
73
- "recall": 0.16666666666666666,
74
- "f1": 0.25895316804407714,
75
  "gt_support": 2538,
76
- "predicted": 729
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
- "accuracy": 0.10060415024954032,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 47,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.0,
95
- "precision": 0.0,
96
- "recall": 0.0,
97
- "f1": 0.0,
98
  "gt_support": 1692,
99
- "predicted": 0
100
  },
101
  "2": {
102
- "accuracy": 0.8297872340425532,
103
  "precision": 0.012345679012345678,
104
- "recall": 0.8297872340425532,
105
- "f1": 0.024329382407985024,
106
  "gt_support": 47,
107
- "predicted": 3159
108
  },
109
  "3": {
110
- "accuracy": 0.1702127659574468,
111
- "precision": 0.5308641975308642,
112
- "recall": 0.1702127659574468,
113
- "f1": 0.25777444735856125,
114
  "gt_support": 2021,
115
- "predicted": 648
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
- "accuracy": 0.08236594803758983,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 67,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.0,
134
- "precision": 0.0,
135
- "recall": 0.0,
136
- "f1": 0.0,
137
  "gt_support": 2680,
138
- "predicted": 0
139
  },
140
  "2": {
141
- "accuracy": 0.8507462686567164,
142
- "precision": 0.012345679012345678,
143
- "recall": 0.8507462686567164,
144
- "f1": 0.024338172502134922,
145
  "gt_support": 67,
146
- "predicted": 4617
147
  },
148
  "3": {
149
- "accuracy": 0.14925373134328357,
150
- "precision": 0.48148148148148145,
151
- "recall": 0.14925373134328357,
152
- "f1": 0.22787028921998248,
153
  "gt_support": 2613,
154
- "predicted": 810
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
- "accuracy": 0.061930783242258654,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.01639344262295082,
165
- "precision": 0.012345679012345678,
166
- "recall": 0.01639344262295082,
167
- "f1": 0.014084507042253521,
168
  "gt_support": 61,
169
- "predicted": 81
170
  },
171
  "1": {
172
- "accuracy": 0.0,
173
- "precision": 0.0,
174
- "recall": 0.0,
175
- "f1": 0.0,
176
  "gt_support": 2623,
177
- "predicted": 0
178
  },
179
  "2": {
180
- "accuracy": 0.8688524590163934,
181
- "precision": 0.012345679012345678,
182
- "recall": 0.8688524590163934,
183
- "f1": 0.024345429490124023,
184
  "gt_support": 61,
185
- "predicted": 4293
186
  },
187
  "3": {
188
- "accuracy": 0.11475409836065574,
189
- "precision": 0.4444444444444444,
190
- "recall": 0.11475409836065574,
191
- "f1": 0.18241042345276873,
192
  "gt_support": 2196,
193
- "predicted": 567
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
- "accuracy": 0.07676323187874343,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 101,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.0,
212
- "precision": 0.0,
213
- "recall": 0.0,
214
- "f1": 0.0,
215
  "gt_support": 4747,
216
- "predicted": 0
217
  },
218
  "2": {
219
- "accuracy": 0.8316831683168316,
220
- "precision": 0.012345679012345678,
221
- "recall": 0.8316831683168316,
222
- "f1": 0.02433019551049964,
223
  "gt_support": 101,
224
- "predicted": 6804
225
  },
226
  "3": {
227
- "accuracy": 0.16831683168316833,
228
- "precision": 0.3950617283950617,
229
- "recall": 0.16831683168316833,
230
- "f1": 0.23605988283792578,
231
  "gt_support": 3232,
232
- "predicted": 1377
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
- "accuracy": 0.061191626409017714,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 184,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.0,
251
- "precision": 0.0,
252
- "recall": 0.0,
253
- "f1": 0.0,
254
  "gt_support": 9200,
255
- "predicted": 0
256
  },
257
  "2": {
258
- "accuracy": 0.8586956521739131,
259
- "precision": 0.012345679012345678,
260
- "recall": 0.8586956521739131,
261
- "f1": 0.024341395778770602,
262
  "gt_support": 184,
263
- "predicted": 12798
264
  },
265
  "3": {
266
- "accuracy": 0.14130434782608695,
267
- "precision": 0.35802469135802467,
268
- "recall": 0.14130434782608695,
269
- "f1": 0.2026337006181134,
270
  "gt_support": 5336,
271
- "predicted": 2106
272
  }
273
  },
274
  "total_samples": 14904
@@ -278,13 +278,13 @@
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 9
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.287073065283182,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.2821011673151751,
8
+ "precision": 0.01231945624468989,
9
+ "recall": 0.2821011673151751,
10
+ "f1": 0.023607945294692283,
11
  "gt_support": 514,
12
+ "predicted": 11770
13
  },
14
  "1": {
15
+ "accuracy": 0.32867225408028233,
16
+ "precision": 0.5981856133590238,
17
+ "recall": 0.32867225408028233,
18
+ "f1": 0.42424414963275064,
19
  "gt_support": 22670,
20
+ "predicted": 12456
21
  },
22
  "2": {
23
+ "accuracy": 0.2140077821011673,
24
+ "precision": 0.012109202994275649,
25
+ "recall": 0.2140077821011673,
26
+ "f1": 0.022921441967076475,
27
  "gt_support": 514,
28
+ "predicted": 9084
29
  },
30
  "3": {
31
+ "accuracy": 0.23673059768064228,
32
+ "precision": 0.5100913022585295,
33
+ "recall": 0.23673059768064228,
34
+ "f1": 0.32338156892612335,
35
  "gt_support": 17936,
36
+ "predicted": 8324
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.39094650205761317,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.018518518518518517,
48
+ "precision": 0.012345679012345678,
49
+ "recall": 0.018518518518518517,
50
+ "f1": 0.014814814814814814,
51
+ "gt_support": 54,
52
+ "predicted": 81
53
+ },
54
+ "1": {
55
  "accuracy": 0.0,
56
  "precision": 0.0,
57
  "recall": 0.0,
58
  "f1": 0.0,
 
 
 
 
 
 
 
 
59
  "gt_support": 1728,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.3148148148148148,
64
+ "precision": 0.012292118582791034,
65
+ "recall": 0.3148148148148148,
66
+ "f1": 0.02366040361864996,
67
  "gt_support": 54,
68
+ "predicted": 1383
69
  },
70
  "3": {
71
+ "accuracy": 0.6666666666666666,
72
+ "precision": 0.5814432989690722,
73
+ "recall": 0.6666666666666666,
74
+ "f1": 0.6211453744493393,
75
  "gt_support": 2538,
76
+ "predicted": 2910
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
+ "accuracy": 0.26214867349619125,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.3617021276595745,
87
+ "precision": 0.01223021582733813,
88
+ "recall": 0.3617021276595745,
89
+ "f1": 0.023660403618649968,
90
  "gt_support": 47,
91
+ "predicted": 1390
92
  },
93
  "1": {
94
+ "accuracy": 0.02127659574468085,
95
+ "precision": 0.4444444444444444,
96
+ "recall": 0.02127659574468085,
97
+ "f1": 0.04060913705583757,
98
  "gt_support": 1692,
99
+ "predicted": 81
100
  },
101
  "2": {
102
+ "accuracy": 0.14893617021276595,
103
  "precision": 0.012345679012345678,
104
+ "recall": 0.14893617021276595,
105
+ "f1": 0.02280130293159609,
106
  "gt_support": 47,
107
+ "predicted": 567
108
  },
109
  "3": {
110
+ "accuracy": 0.46412666996536367,
111
+ "precision": 0.5302430751837196,
112
+ "recall": 0.46412666996536367,
113
+ "f1": 0.49498680738786277,
114
  "gt_support": 2021,
115
+ "predicted": 1769
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
+ "accuracy": 0.20932375161230882,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.373134328358209,
126
+ "precision": 0.012236906510034264,
127
+ "recall": 0.373134328358209,
128
+ "f1": 0.023696682464454975,
129
  "gt_support": 67,
130
+ "predicted": 2043
131
  },
132
  "1": {
133
+ "accuracy": 0.045895522388059704,
134
+ "precision": 0.4522058823529412,
135
+ "recall": 0.045895522388059704,
136
+ "f1": 0.08333333333333333,
137
  "gt_support": 2680,
138
+ "predicted": 272
139
  },
140
  "2": {
141
+ "accuracy": 0.19402985074626866,
142
+ "precision": 0.011937557392102846,
143
+ "recall": 0.19402985074626866,
144
+ "f1": 0.02249134948096886,
145
  "gt_support": 67,
146
+ "predicted": 1089
147
  },
148
  "3": {
149
+ "accuracy": 0.373134328358209,
150
+ "precision": 0.4819574888779041,
151
+ "recall": 0.373134328358209,
152
+ "f1": 0.4206212251941329,
153
  "gt_support": 2613,
154
+ "predicted": 2023
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
+ "accuracy": 0.1600890507994333,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.5081967213114754,
165
+ "precision": 0.01245480112494978,
166
+ "recall": 0.5081967213114754,
167
+ "f1": 0.02431372549019608,
168
  "gt_support": 61,
169
+ "predicted": 2489
170
  },
171
  "1": {
172
+ "accuracy": 0.22378955394586353,
173
+ "precision": 0.5326678765880217,
174
+ "recall": 0.22378955394586353,
175
+ "f1": 0.31516778523489936,
176
  "gt_support": 2623,
177
+ "predicted": 1102
178
  },
179
  "2": {
180
+ "accuracy": 0.19672131147540983,
181
+ "precision": 0.012182741116751269,
182
+ "recall": 0.19672131147540983,
183
+ "f1": 0.02294455066921606,
184
  "gt_support": 61,
185
+ "predicted": 985
186
  },
187
  "3": {
188
+ "accuracy": 0.07331511839708561,
189
+ "precision": 0.4410958904109589,
190
+ "recall": 0.07331511839708561,
191
+ "f1": 0.12573213588442014,
192
  "gt_support": 2196,
193
+ "predicted": 365
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
+ "accuracy": 0.2421464368659088,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.25742574257425743,
204
+ "precision": 0.012524084778420038,
205
+ "recall": 0.25742574257425743,
206
+ "f1": 0.02388608176389527,
207
  "gt_support": 101,
208
+ "predicted": 2076
209
  },
210
  "1": {
211
+ "accuracy": 0.3353697071834843,
212
+ "precision": 0.582723279648609,
213
+ "recall": 0.3353697071834843,
214
+ "f1": 0.4257253643535232,
215
  "gt_support": 4747,
216
+ "predicted": 2732
217
  },
218
  "2": {
219
+ "accuracy": 0.3069306930693069,
220
+ "precision": 0.012277227722772278,
221
+ "recall": 0.3069306930693069,
222
+ "f1": 0.02361005331302361,
223
  "gt_support": 101,
224
+ "predicted": 2525
225
  },
226
  "3": {
227
+ "accuracy": 0.10272277227722772,
228
+ "precision": 0.3915094339622642,
229
+ "recall": 0.10272277227722772,
230
+ "f1": 0.1627450980392157,
231
  "gt_support": 3232,
232
+ "predicted": 848
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
+ "accuracy": 0.35802469135802467,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.24456521739130435,
243
+ "precision": 0.01219181793551883,
244
+ "recall": 0.24456521739130435,
245
+ "f1": 0.0232258064516129,
246
  "gt_support": 184,
247
+ "predicted": 3691
248
  },
249
  "1": {
250
+ "accuracy": 0.5557608695652174,
251
+ "precision": 0.6183335348893457,
252
+ "recall": 0.5557608695652174,
253
+ "f1": 0.5853798156734787,
254
  "gt_support": 9200,
255
+ "predicted": 8269
256
  },
257
  "2": {
258
+ "accuracy": 0.16304347826086957,
259
+ "precision": 0.011834319526627219,
260
+ "recall": 0.16304347826086957,
261
+ "f1": 0.022066936373666787,
262
  "gt_support": 184,
263
+ "predicted": 2535
264
  },
265
  "3": {
266
+ "accuracy": 0.027736131934032984,
267
+ "precision": 0.36185819070904646,
268
+ "recall": 0.027736131934032984,
269
+ "f1": 0.0515230635335074,
270
  "gt_support": 5336,
271
+ "predicted": 409
272
  }
273
  },
274
  "total_samples": 14904
 
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 9
290
  }
layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 514 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.0763 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0019 0.0123 0.0019 0.0034 514 81
30
- # 0.0014 0.3951 0.0014 0.0028 22670 81
31
- G 0.8463 0.0123 0.8463 0.0243 514 35235
32
- _ 0.1509 0.4340 0.1509 0.2240 17936 6237
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.1141 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 54 0
49
- # 0.0185 0.3951 0.0185 0.0354 1728 81
50
- G 0.8148 0.0123 0.8148 0.0243 54 3564
51
- _ 0.1667 0.5802 0.1667 0.2590 2538 729
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.1006 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 47 0
64
- # 0.0000 0.0000 0.0000 0.0000 1692 0
65
- G 0.8298 0.0123 0.8298 0.0243 47 3159
66
- _ 0.1702 0.5309 0.1702 0.2578 2021 648
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.0824 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 67 0
79
- # 0.0000 0.0000 0.0000 0.0000 2680 0
80
- G 0.8507 0.0123 0.8507 0.0243 67 4617
81
- _ 0.1493 0.4815 0.1493 0.2279 2613 810
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0619 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0164 0.0123 0.0164 0.0141 61 81
94
- # 0.0000 0.0000 0.0000 0.0000 2623 0
95
- G 0.8689 0.0123 0.8689 0.0243 61 4293
96
- _ 0.1148 0.4444 0.1148 0.1824 2196 567
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0768 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 101 0
109
- # 0.0000 0.0000 0.0000 0.0000 4747 0
110
- G 0.8317 0.0123 0.8317 0.0243 101 6804
111
- _ 0.1683 0.3951 0.1683 0.2361 3232 1377
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.0612 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 184 0
124
- # 0.0000 0.0000 0.0000 0.0000 9200 0
125
- G 0.8587 0.0123 0.8587 0.0243 184 12798
126
- _ 0.1413 0.3580 0.1413 0.2026 5336 2106
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2871 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.2821 0.0123 0.2821 0.0236 514 11770
30
+ # 0.3287 0.5982 0.3287 0.4242 22670 12456
31
+ G 0.2140 0.0121 0.2140 0.0229 514 9084
32
+ _ 0.2367 0.5101 0.2367 0.3234 17936 8324
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.3909 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0185 0.0123 0.0185 0.0148 54 81
49
+ # 0.0000 0.0000 0.0000 0.0000 1728 0
50
+ G 0.3148 0.0123 0.3148 0.0237 54 1383
51
+ _ 0.6667 0.5814 0.6667 0.6211 2538 2910
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.2621 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.3617 0.0122 0.3617 0.0237 47 1390
64
+ # 0.0213 0.4444 0.0213 0.0406 1692 81
65
+ G 0.1489 0.0123 0.1489 0.0228 47 567
66
+ _ 0.4641 0.5302 0.4641 0.4950 2021 1769
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.2093 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.3731 0.0122 0.3731 0.0237 67 2043
79
+ # 0.0459 0.4522 0.0459 0.0833 2680 272
80
+ G 0.1940 0.0119 0.1940 0.0225 67 1089
81
+ _ 0.3731 0.4820 0.3731 0.4206 2613 2023
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.1601 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.5082 0.0125 0.5082 0.0243 61 2489
94
+ # 0.2238 0.5327 0.2238 0.3152 2623 1102
95
+ G 0.1967 0.0122 0.1967 0.0229 61 985
96
+ _ 0.0733 0.4411 0.0733 0.1257 2196 365
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2421 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2574 0.0125 0.2574 0.0239 101 2076
109
+ # 0.3354 0.5827 0.3354 0.4257 4747 2732
110
+ G 0.3069 0.0123 0.3069 0.0236 101 2525
111
+ _ 0.1027 0.3915 0.1027 0.1627 3232 848
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.3580 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2446 0.0122 0.2446 0.0232 184 3691
124
+ # 0.5558 0.6183 0.5558 0.5854 9200 8269
125
+ G 0.1630 0.0118 0.1630 0.0221 184 2535
126
+ _ 0.0277 0.3619 0.0277 0.0515 5336 409
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_size9.json
layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.json CHANGED
@@ -1,631 +1,645 @@
1
  {
2
  "global": {
3
- "accuracy": 0.32904885027751923,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.08346261247285137,
8
- "precision": 0.004499456385380948,
9
- "recall": 0.08346261247285137,
10
- "f1": 0.008538598273235144,
11
  "gt_support": 3223,
12
- "predicted": 59785
13
  },
14
  "1": {
15
- "accuracy": 0.08120499632623071,
16
- "precision": 0.3945934917882281,
17
- "recall": 0.08120499632623071,
18
- "f1": 0.13469131933563147,
19
  "gt_support": 238175,
20
- "predicted": 49015
21
  },
22
  "2": {
23
- "accuracy": 0.10704312752094322,
24
- "precision": 0.004846730915118991,
25
- "recall": 0.10704312752094322,
26
- "f1": 0.009273570324574962,
27
  "gt_support": 3223,
28
- "predicted": 71182
29
  },
30
  "3": {
31
- "accuracy": 0.6624066144502067,
32
- "precision": 0.3527244161340228,
33
- "recall": 0.6624066144502067,
34
- "f1": 0.46032872463921254,
35
  "gt_support": 243346,
36
- "predicted": 456997
37
  },
38
  "4": {
39
- "accuracy": 0.2422726046339078,
40
- "precision": 0.6516055149893419,
41
- "recall": 0.2422726046339078,
42
- "f1": 0.35321630957210115,
43
  "gt_support": 237208,
44
- "predicted": 88196
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
- "by_size": {
50
- "7": {
51
- "accuracy": 0.31717171717171716,
52
- "baseline_accuracy": 0.7822222222222223,
53
  "per_class": {
54
  "0": {
55
- "accuracy": 0.09090909090909091,
56
- "precision": 0.0044812133746983795,
57
- "recall": 0.09090909090909091,
58
- "f1": 0.008541392904073587,
59
- "gt_support": 286,
60
- "predicted": 5802
61
  },
62
  "1": {
63
- "accuracy": 0.017874513618677042,
64
- "precision": 0.13183856502242153,
65
- "recall": 0.017874513618677042,
66
- "f1": 0.03148088660456152,
67
- "gt_support": 8224,
68
- "predicted": 1115
69
  },
70
  "2": {
71
- "accuracy": 0.013986013986013986,
72
- "precision": 0.0061162079510703364,
73
- "recall": 0.013986013986013986,
74
- "f1": 0.00851063829787234,
75
- "gt_support": 286,
76
- "predicted": 654
77
  },
78
  "3": {
79
- "accuracy": 0.874664622460713,
80
- "precision": 0.11136868304824186,
81
- "recall": 0.874664622460713,
82
- "f1": 0.19758003419987444,
83
- "gt_support": 5218,
84
- "predicted": 40981
85
  },
86
  "4": {
87
- "accuracy": 0.3112881436745073,
88
- "precision": 0.9918344094189138,
89
- "recall": 0.3112881436745073,
90
- "f1": 0.4738561103214685,
91
- "gt_support": 50336,
92
- "predicted": 15798
93
  }
94
  },
95
- "total_samples": 64350
96
  },
97
- "9": {
98
- "accuracy": 0.28072632944228276,
99
- "baseline_accuracy": 0.64,
100
  "per_class": {
101
  "0": {
102
- "accuracy": 0.09922178988326848,
103
- "precision": 0.004461942257217848,
104
- "recall": 0.09922178988326848,
105
- "f1": 0.00853985264567984,
106
- "gt_support": 514,
107
- "predicted": 11430
108
  },
109
  "1": {
110
- "accuracy": 0.04380238200264667,
111
- "precision": 0.21114182436742504,
112
- "recall": 0.04380238200264667,
113
- "f1": 0.07255324589924378,
114
- "gt_support": 22670,
115
- "predicted": 4703
116
  },
117
  "2": {
118
- "accuracy": 0.021400778210116732,
119
- "precision": 0.006719609040928528,
120
- "recall": 0.021400778210116732,
121
- "f1": 0.010227801022780102,
122
- "gt_support": 514,
123
- "predicted": 1637
124
  },
125
  "3": {
126
- "accuracy": 0.8281110615521855,
127
- "precision": 0.18438334057476258,
128
- "recall": 0.8281110615521855,
129
- "f1": 0.3016113147394178,
130
- "gt_support": 17936,
131
- "predicted": 80555
132
  },
133
  "4": {
134
- "accuracy": 0.22370838737570256,
135
- "precision": 0.9557287157287158,
136
- "recall": 0.22370838737570256,
137
- "f1": 0.3625535082821515,
138
- "gt_support": 74016,
139
- "predicted": 17325
140
  }
141
  },
142
- "total_samples": 115650
143
  },
144
- "11": {
145
- "accuracy": 0.2977709296353364,
146
- "baseline_accuracy": 0.4622222222222222,
147
  "per_class": {
148
  "0": {
149
- "accuracy": 0.08782742681047766,
150
- "precision": 0.004538939321548017,
151
- "recall": 0.08782742681047766,
152
- "f1": 0.008631786173998637,
153
- "gt_support": 649,
154
- "predicted": 12558
155
  },
156
  "1": {
157
- "accuracy": 0.08866757893669859,
158
- "precision": 0.29949499272447144,
159
- "recall": 0.08866757893669859,
160
- "f1": 0.1368266692736338,
161
- "gt_support": 39462,
162
- "predicted": 11683
163
  },
164
  "2": {
165
- "accuracy": 0.12172573189522343,
166
- "precision": 0.005811387376783875,
167
- "recall": 0.12172573189522343,
168
- "f1": 0.01109316857403637,
169
- "gt_support": 649,
170
- "predicted": 13594
171
  },
172
  "3": {
173
- "accuracy": 0.6618920278535306,
174
- "precision": 0.279100145137881,
175
- "recall": 0.6618920278535306,
176
- "f1": 0.39263697688846305,
177
- "gt_support": 37769,
178
- "predicted": 89570
179
  },
180
  "4": {
181
- "accuracy": 0.2199834064240844,
182
- "precision": 0.797422126745435,
183
- "recall": 0.2199834064240844,
184
- "f1": 0.34483719633982074,
185
- "gt_support": 67496,
186
- "predicted": 18620
187
  }
188
  },
189
- "total_samples": 146025
190
  },
191
- "13": {
192
- "accuracy": 0.33342112482853226,
193
- "baseline_accuracy": 0.37721810699588476,
194
  "per_class": {
195
  "0": {
196
- "accuracy": 0.07654320987654321,
197
- "precision": 0.004537802825148211,
198
- "recall": 0.07654320987654321,
199
- "f1": 0.008567677744766116,
200
- "gt_support": 810,
201
- "predicted": 13663
202
  },
203
  "1": {
204
- "accuracy": 0.09658458855716906,
205
- "precision": 0.41395528638618645,
206
- "recall": 0.09658458855716906,
207
- "f1": 0.15662518435454578,
208
- "gt_support": 66522,
209
- "predicted": 15521
210
  },
211
  "2": {
212
- "accuracy": 0.1580246913580247,
213
- "precision": 0.005012335043270548,
214
- "recall": 0.1580246913580247,
215
- "f1": 0.009716476259156639,
216
- "gt_support": 810,
217
- "predicted": 25537
218
  },
219
  "3": {
220
- "accuracy": 0.6364839704427765,
221
- "precision": 0.40177948360083743,
222
- "recall": 0.6364839704427765,
223
- "f1": 0.49260368352321343,
224
- "gt_support": 68748,
225
- "predicted": 108908
226
  },
227
  "4": {
228
- "accuracy": 0.22914462081128747,
229
- "precision": 0.5581869931797433,
230
- "recall": 0.22914462081128747,
231
- "f1": 0.3249089573467123,
232
- "gt_support": 45360,
233
- "predicted": 18621
234
  }
235
  },
236
- "total_samples": 182250
237
  },
238
- "15": {
239
- "accuracy": 0.37572153065929,
240
- "baseline_accuracy": 0.5240894421392347,
241
  "per_class": {
242
  "0": {
243
- "accuracy": 0.07572614107883817,
244
- "precision": 0.004469752632867989,
245
- "recall": 0.07572614107883817,
246
- "f1": 0.008441258094357078,
247
- "gt_support": 964,
248
- "predicted": 16332
249
  },
250
  "1": {
251
- "accuracy": 0.08171021846648963,
252
- "precision": 0.5175389232789346,
253
- "recall": 0.08171021846648963,
254
- "f1": 0.1411373518629039,
255
- "gt_support": 101297,
256
- "predicted": 15993
257
  },
258
  "2": {
259
- "accuracy": 0.1275933609958506,
260
- "precision": 0.004133064516129032,
261
- "recall": 0.1275933609958506,
262
- "f1": 0.008006769951829189,
263
- "gt_support": 964,
264
- "predicted": 29760
265
  },
266
  "3": {
267
- "accuracy": 0.6423663954255553,
268
- "precision": 0.5330661468941401,
269
- "recall": 0.6423663954255553,
270
- "f1": 0.58263450598026,
271
- "gt_support": 113675,
272
- "predicted": 136983
273
- },
274
- "4": {
275
  "accuracy": 0.0,
276
  "precision": 0.0,
277
  "recall": 0.0,
278
  "f1": 0.0,
279
- "gt_support": 0,
280
- "predicted": 17832
 
 
 
 
 
 
 
 
281
  }
282
  },
283
- "total_samples": 216900
284
- }
285
- },
286
- "by_complexity": {
287
- "0.0": {
288
- "accuracy": 0.40755555555555556,
289
- "baseline_accuracy": 0.4410322580645161,
290
  "per_class": {
291
  "0": {
292
- "accuracy": 0.054838709677419356,
293
- "precision": 0.004465458366167586,
294
- "recall": 0.054838709677419356,
295
- "f1": 0.008258440612096187,
296
- "gt_support": 310,
297
- "predicted": 3807
298
  },
299
  "1": {
300
- "accuracy": 0.014492753623188406,
301
- "precision": 0.17158176943699732,
302
- "recall": 0.014492753623188406,
303
- "f1": 0.026727918145750675,
304
- "gt_support": 13248,
305
- "predicted": 1119
306
  },
307
  "2": {
308
- "accuracy": 0.23225806451612904,
309
- "precision": 0.005293728402323359,
310
- "recall": 0.23225806451612904,
311
- "f1": 0.010351520379555747,
312
- "gt_support": 310,
313
- "predicted": 13601
314
  },
315
  "3": {
316
- "accuracy": 0.6046420908913595,
317
- "precision": 0.49771212972625833,
318
- "recall": 0.6046420908913595,
319
- "f1": 0.5459909295055259,
320
- "gt_support": 30762,
321
- "predicted": 37371
322
  },
323
  "4": {
324
- "accuracy": 0.38001592356687897,
325
- "precision": 0.6891423621137742,
326
- "recall": 0.38001592356687897,
327
- "f1": 0.4898901775633789,
328
- "gt_support": 25120,
329
- "predicted": 13852
330
  }
331
  },
332
- "total_samples": 69750
 
 
 
 
 
 
 
 
 
 
 
 
 
333
  },
334
- "0.2": {
335
- "accuracy": 0.40347350427350426,
336
- "baseline_accuracy": 0.3972239316239316,
 
 
 
337
  "per_class": {
338
  "0": {
339
- "accuracy": 0.06769230769230769,
340
- "precision": 0.004450738417964798,
341
- "recall": 0.06769230769230769,
342
- "f1": 0.008352315869400152,
343
- "gt_support": 325,
344
- "predicted": 4943
345
  },
346
  "1": {
347
- "accuracy": 0.026921223801493616,
348
- "precision": 0.2514060742407199,
349
- "recall": 0.026921223801493616,
350
- "f1": 0.04863453378304863,
351
- "gt_support": 16604,
352
- "predicted": 1778
353
  },
354
  "2": {
355
- "accuracy": 0.09538461538461539,
356
- "precision": 0.0051856808297089324,
357
- "recall": 0.09538461538461539,
358
- "f1": 0.009836585752816119,
359
- "gt_support": 325,
360
- "predicted": 5978
361
  },
362
  "3": {
363
- "accuracy": 0.7061314421454884,
364
- "precision": 0.4311569830993021,
365
- "recall": 0.7061314421454884,
366
- "f1": 0.5354024458685183,
367
- "gt_support": 29047,
368
- "predicted": 47572
369
  },
370
  "4": {
371
- "accuracy": 0.31661944527288993,
372
- "precision": 0.6607281779990665,
373
- "recall": 0.31661944527288993,
374
- "f1": 0.4280961742023287,
375
- "gt_support": 26824,
376
- "predicted": 12854
377
  }
378
  },
379
- "total_samples": 73125
380
  },
381
- "0.4": {
382
- "accuracy": 0.37479035639413,
383
- "baseline_accuracy": 0.4000524109014675,
384
  "per_class": {
385
  "0": {
386
- "accuracy": 0.10141509433962265,
387
- "precision": 0.0044211392144766605,
388
- "recall": 0.10141509433962265,
389
- "f1": 0.008472906403940886,
390
- "gt_support": 424,
391
- "predicted": 9726
392
  },
393
  "1": {
394
- "accuracy": 0.03518262881020781,
395
- "precision": 0.2924031007751938,
396
- "recall": 0.03518262881020781,
397
- "f1": 0.06280804582389768,
398
- "gt_support": 26803,
399
- "predicted": 3225
400
  },
401
  "2": {
402
- "accuracy": 0.08254716981132075,
403
- "precision": 0.004425338222278417,
404
- "recall": 0.08254716981132075,
405
- "f1": 0.008400336013440538,
406
- "gt_support": 424,
407
- "predicted": 7909
408
  },
409
  "3": {
410
- "accuracy": 0.7054369186427355,
411
- "precision": 0.4352317366915081,
412
- "recall": 0.7054369186427355,
413
- "f1": 0.5383308006078541,
414
- "gt_support": 38165,
415
- "predicted": 61859
416
  },
417
  "4": {
418
- "accuracy": 0.2640278528934559,
419
- "precision": 0.6159608863654286,
420
- "recall": 0.2640278528934559,
421
- "f1": 0.36962025316455693,
422
- "gt_support": 29584,
423
- "predicted": 12681
424
  }
425
  },
426
- "total_samples": 95400
427
  },
428
- "0.6": {
429
- "accuracy": 0.3425411522633745,
430
- "baseline_accuracy": 0.3553909465020576,
431
  "per_class": {
432
  "0": {
433
- "accuracy": 0.06481481481481481,
434
- "precision": 0.004495103547921014,
435
- "recall": 0.06481481481481481,
436
- "f1": 0.008407146074163038,
437
- "gt_support": 432,
438
- "predicted": 6229
439
  },
440
  "1": {
441
- "accuracy": 0.050466153222537495,
442
- "precision": 0.3522754067436925,
443
- "recall": 0.050466153222537495,
444
- "f1": 0.08828482789185996,
445
- "gt_support": 29604,
446
- "predicted": 4241
447
  },
448
  "2": {
449
- "accuracy": 0.1087962962962963,
450
- "precision": 0.004967237370534771,
451
- "recall": 0.1087962962962963,
452
- "f1": 0.009500707499494644,
453
- "gt_support": 432,
454
- "predicted": 9462
455
  },
456
  "3": {
457
- "accuracy": 0.6893562818441655,
458
- "precision": 0.3479044826667085,
459
- "recall": 0.6893562818441655,
460
- "f1": 0.4624297935748747,
461
- "gt_support": 32188,
462
- "predicted": 63779
463
  },
464
  "4": {
465
- "accuracy": 0.27608267716535434,
466
- "precision": 0.7070205352509452,
467
- "recall": 0.27608267716535434,
468
- "f1": 0.3971019923802386,
469
- "gt_support": 34544,
470
- "predicted": 13489
471
  }
472
  },
473
- "total_samples": 97200
474
  },
475
- "0.8": {
476
- "accuracy": 0.3171998371998372,
477
- "baseline_accuracy": 0.3470899470899471,
478
  "per_class": {
479
  "0": {
480
- "accuracy": 0.06776556776556776,
481
- "precision": 0.004479961254389151,
482
- "recall": 0.06776556776556776,
483
- "f1": 0.008404315729699035,
484
- "gt_support": 546,
485
- "predicted": 8259
486
  },
487
  "1": {
488
- "accuracy": 0.07516416510318949,
489
- "precision": 0.3948016752894802,
490
- "recall": 0.07516416510318949,
491
- "f1": 0.12628551164348478,
492
- "gt_support": 42640,
493
- "predicted": 8118
494
  },
495
  "2": {
496
- "accuracy": 0.10622710622710622,
497
- "precision": 0.005224754526619224,
498
- "recall": 0.10622710622710622,
499
- "f1": 0.009959646260839703,
500
- "gt_support": 546,
501
- "predicted": 11101
502
  },
503
  "3": {
504
- "accuracy": 0.699015394658108,
505
- "precision": 0.3147195689444036,
506
- "recall": 0.699015394658108,
507
- "f1": 0.4340263117896408,
508
- "gt_support": 36766,
509
- "predicted": 81660
510
  },
511
  "4": {
512
- "accuracy": 0.2353607857952399,
513
- "precision": 0.7269544924154026,
514
- "recall": 0.2353607857952399,
515
- "f1": 0.35559360730593603,
516
- "gt_support": 42352,
517
- "predicted": 13712
518
  }
519
  },
520
- "total_samples": 122850
521
  },
522
- "1.0": {
523
- "accuracy": 0.27232152894884765,
524
- "baseline_accuracy": 0.4095034663668728,
525
  "per_class": {
526
  "0": {
527
- "accuracy": 0.10286677908937605,
528
- "precision": 0.004548674546064651,
529
- "recall": 0.10286677908937605,
530
- "f1": 0.008712107687363873,
531
- "gt_support": 1186,
532
- "predicted": 26821
533
  },
534
  "1": {
535
- "accuracy": 0.1195138914308723,
536
- "precision": 0.42771991877906596,
537
- "recall": 0.1195138914308723,
538
- "f1": 0.1868249767541664,
539
- "gt_support": 109276,
540
- "predicted": 30534
541
  },
542
  "2": {
543
- "accuracy": 0.08600337268128162,
544
- "precision": 0.004409666681077342,
545
- "recall": 0.08600337268128162,
546
- "f1": 0.008389192745815684,
547
- "gt_support": 1186,
548
- "predicted": 23131
549
  },
550
  "3": {
551
- "accuracy": 0.6185846266586407,
552
- "precision": 0.28691519580470515,
553
- "recall": 0.6185846266586407,
554
- "f1": 0.39200743032001795,
555
- "gt_support": 76418,
556
- "predicted": 164756
557
  },
558
  "4": {
559
- "accuracy": 0.15376218521527213,
560
- "precision": 0.560625694187338,
561
- "recall": 0.15376218521527213,
562
- "f1": 0.2413339708343294,
563
- "gt_support": 78784,
564
- "predicted": 21608
565
  }
566
  },
567
- "total_samples": 266850
568
  }
569
  },
570
  "by_size_complexity": {
571
  "7_0.0": {
572
- "accuracy": 0.4242105263157895,
573
  "baseline_accuracy": 0.7822222222222223,
574
  "per_class": {
575
  "0": {
576
  "accuracy": 0.07894736842105263,
577
- "precision": 0.0044444444444444444,
578
  "recall": 0.07894736842105263,
579
- "f1": 0.008415147265077139,
580
  "gt_support": 38,
581
- "predicted": 675
582
  },
583
  "1": {
584
- "accuracy": 0.02631578947368421,
585
- "precision": 0.10666666666666667,
586
- "recall": 0.02631578947368421,
587
- "f1": 0.04221635883905014,
588
  "gt_support": 912,
589
- "predicted": 225
590
  },
591
  "2": {
592
- "accuracy": 0.02631578947368421,
593
- "precision": 0.0044444444444444444,
594
- "recall": 0.02631578947368421,
595
- "f1": 0.0076045627376425855,
596
  "gt_support": 38,
597
- "predicted": 225
598
  },
599
  "3": {
600
- "accuracy": 0.8649885583524027,
601
- "precision": 0.16571679088119246,
602
- "recall": 0.8649885583524027,
603
- "f1": 0.2781456953642384,
604
  "gt_support": 874,
605
- "predicted": 4562
606
  },
607
  "4": {
608
- "accuracy": 0.42508971291866027,
609
- "precision": 0.9930143206426825,
610
- "recall": 0.42508971291866027,
611
- "f1": 0.5953303319024186,
612
  "gt_support": 6688,
613
- "predicted": 2863
614
  }
615
  },
616
  "total_samples": 8550
617
  },
618
  "7_0.2": {
619
- "accuracy": 0.31343434343434345,
620
  "baseline_accuracy": 0.7822222222222223,
621
  "per_class": {
622
  "0": {
623
- "accuracy": 0.06818181818181818,
624
- "precision": 0.004464285714285714,
625
- "recall": 0.06818181818181818,
626
- "f1": 0.008379888268156424,
627
  "gt_support": 44,
628
- "predicted": 672
629
  },
630
  "1": {
631
  "accuracy": 0.0,
@@ -636,90 +650,90 @@
636
  "predicted": 0
637
  },
638
  "2": {
 
 
 
 
 
 
 
 
639
  "accuracy": 0.0,
640
  "precision": 0.0,
641
  "recall": 0.0,
642
  "f1": 0.0,
643
- "gt_support": 44,
644
- "predicted": 0
645
- },
646
- "3": {
647
- "accuracy": 0.9318181818181818,
648
- "precision": 0.12324649298597194,
649
- "recall": 0.9318181818181818,
650
- "f1": 0.21769911504424777,
651
  "gt_support": 924,
652
- "predicted": 6986
653
  },
654
  "4": {
655
- "accuracy": 0.2891270661157025,
656
- "precision": 0.9986619090098127,
657
- "recall": 0.2891270661157025,
658
- "f1": 0.44842779891848594,
659
  "gt_support": 7744,
660
- "predicted": 2242
661
  }
662
  },
663
  "total_samples": 9900
664
  },
665
  "7_0.4": {
666
- "accuracy": 0.29897018970189704,
667
  "baseline_accuracy": 0.7822222222222223,
668
  "per_class": {
669
  "0": {
670
- "accuracy": 0.1951219512195122,
671
- "precision": 0.0044444444444444444,
672
- "recall": 0.1951219512195122,
673
- "f1": 0.008690928843020097,
674
  "gt_support": 41,
675
- "predicted": 1800
676
  },
677
  "1": {
678
- "accuracy": 0.024390243902439025,
679
- "precision": 0.13023255813953488,
680
- "recall": 0.024390243902439025,
681
- "f1": 0.04108584005869406,
682
  "gt_support": 1148,
683
- "predicted": 215
684
  },
685
  "2": {
686
- "accuracy": 0.024390243902439025,
687
- "precision": 0.005291005291005291,
688
- "recall": 0.024390243902439025,
689
- "f1": 0.008695652173913042,
690
  "gt_support": 41,
691
- "predicted": 189
692
  },
693
  "3": {
694
- "accuracy": 0.754813863928113,
695
- "precision": 0.12071443235475261,
696
- "recall": 0.754813863928113,
697
- "f1": 0.20814159292035397,
698
  "gt_support": 779,
699
- "predicted": 4871
700
  },
701
  "4": {
702
- "accuracy": 0.2955931263858093,
703
- "precision": 0.9920930232558139,
704
- "recall": 0.2955931263858093,
705
- "f1": 0.4554772581678412,
706
  "gt_support": 7216,
707
- "predicted": 2150
708
  }
709
  },
710
  "total_samples": 9225
711
  },
712
  "7_0.6": {
713
- "accuracy": 0.30921985815602837,
714
  "baseline_accuracy": 0.7822222222222223,
715
  "per_class": {
716
  "0": {
717
- "accuracy": 0.06382978723404255,
718
- "precision": 0.0044444444444444444,
719
- "recall": 0.06382978723404255,
720
- "f1": 0.008310249307479225,
721
  "gt_support": 47,
722
- "predicted": 675
723
  },
724
  "1": {
725
  "accuracy": 0.0,
@@ -730,278 +744,278 @@
730
  "predicted": 0
731
  },
732
  "2": {
 
 
 
 
 
 
 
 
733
  "accuracy": 0.0,
734
  "precision": 0.0,
735
  "recall": 0.0,
736
  "f1": 0.0,
737
- "gt_support": 47,
738
- "predicted": 0
739
- },
740
- "3": {
741
- "accuracy": 0.9349881796690307,
742
- "precision": 0.10670443814919736,
743
- "recall": 0.9349881796690307,
744
- "f1": 0.19154861363361184,
745
  "gt_support": 846,
746
- "predicted": 7413
747
  },
748
  "4": {
749
- "accuracy": 0.2993230174081238,
750
- "precision": 0.9955770004020909,
751
- "recall": 0.2993230174081238,
752
- "f1": 0.4602658239613347,
753
  "gt_support": 8272,
754
- "predicted": 2487
755
  }
756
  },
757
  "total_samples": 10575
758
  },
759
  "7_0.8": {
760
- "accuracy": 0.28185858585858586,
761
  "baseline_accuracy": 0.7822222222222223,
762
  "per_class": {
763
  "0": {
764
- "accuracy": 0.09090909090909091,
765
- "precision": 0.004629629629629629,
766
- "recall": 0.09090909090909091,
767
- "f1": 0.00881057268722467,
768
  "gt_support": 55,
769
- "predicted": 1080
770
  },
771
  "1": {
772
- "accuracy": 0.01818181818181818,
773
- "precision": 0.13777777777777778,
774
- "recall": 0.01818181818181818,
775
- "f1": 0.03212435233160622,
776
  "gt_support": 1705,
777
- "predicted": 225
778
  },
779
  "2": {
780
- "accuracy": 0.01818181818181818,
781
- "precision": 0.0044444444444444444,
782
- "recall": 0.01818181818181818,
783
- "f1": 0.0071428571428571435,
784
  "gt_support": 55,
785
- "predicted": 225
786
  },
787
  "3": {
788
- "accuracy": 0.8727272727272727,
789
- "precision": 0.09414072076489335,
790
- "recall": 0.8727272727272727,
791
- "f1": 0.169949103784023,
792
  "gt_support": 880,
793
- "predicted": 8158
794
  },
795
  "4": {
796
- "accuracy": 0.2771694214876033,
797
- "precision": 0.9985113509490138,
798
- "recall": 0.2771694214876033,
799
- "f1": 0.4338966604673729,
800
  "gt_support": 9680,
801
- "predicted": 2687
802
  }
803
  },
804
  "total_samples": 12375
805
  },
806
  "7_1.0": {
807
- "accuracy": 0.3033879781420765,
808
  "baseline_accuracy": 0.7822222222222223,
809
  "per_class": {
810
  "0": {
811
- "accuracy": 0.06557377049180328,
812
- "precision": 0.0044444444444444444,
813
- "recall": 0.06557377049180328,
814
- "f1": 0.008324661810613943,
815
  "gt_support": 61,
816
- "predicted": 900
817
  },
818
  "1": {
819
- "accuracy": 0.03278688524590164,
820
- "precision": 0.14222222222222222,
821
- "recall": 0.03278688524590164,
822
- "f1": 0.05328892589508742,
823
  "gt_support": 1952,
824
- "predicted": 450
825
  },
826
  "2": {
827
- "accuracy": 0.01639344262295082,
828
- "precision": 0.06666666666666667,
829
- "recall": 0.01639344262295082,
830
- "f1": 0.02631578947368421,
831
  "gt_support": 61,
832
- "predicted": 15
833
  },
834
  "3": {
835
- "accuracy": 0.8743169398907104,
836
- "precision": 0.08897786675564454,
837
- "recall": 0.8743169398907104,
838
- "f1": 0.16151827175449224,
839
  "gt_support": 915,
840
- "predicted": 8991
841
  },
842
  "4": {
843
- "accuracy": 0.3069113263785395,
844
- "precision": 0.9780350252300386,
845
- "recall": 0.3069113263785395,
846
- "f1": 0.46721020914569306,
847
  "gt_support": 10736,
848
- "predicted": 3369
849
  }
850
  },
851
  "total_samples": 13725
852
  },
853
  "9_0.0": {
854
- "accuracy": 0.40131687242798353,
855
  "baseline_accuracy": 0.64,
856
  "per_class": {
857
  "0": {
858
- "accuracy": 0.09259259259259259,
859
- "precision": 0.0044444444444444444,
860
- "recall": 0.09259259259259259,
861
- "f1": 0.008481764206955046,
862
  "gt_support": 54,
863
- "predicted": 1125
864
  },
865
  "1": {
866
- "accuracy": 0.018518518518518517,
867
- "precision": 0.1461187214611872,
868
- "recall": 0.018518518518518517,
869
- "f1": 0.03287108371854135,
870
  "gt_support": 1728,
871
- "predicted": 219
872
  },
873
  "2": {
874
- "accuracy": 0.037037037037037035,
875
- "precision": 0.005076142131979695,
876
- "recall": 0.037037037037037035,
877
- "f1": 0.008928571428571428,
878
  "gt_support": 54,
879
- "predicted": 394
880
  },
881
  "3": {
882
- "accuracy": 0.8266351457840819,
883
- "precision": 0.2796960405279296,
884
- "recall": 0.8266351457840819,
885
- "f1": 0.4179699173224425,
886
  "gt_support": 2538,
887
- "predicted": 7501
888
  },
889
  "4": {
890
- "accuracy": 0.35223765432098764,
891
- "precision": 0.9409137753349365,
892
- "recall": 0.35223765432098764,
893
- "f1": 0.5125853841115373,
894
  "gt_support": 7776,
895
- "predicted": 2911
896
  }
897
  },
898
  "total_samples": 12150
899
  },
900
  "9_0.2": {
901
- "accuracy": 0.3806146572104019,
902
  "baseline_accuracy": 0.64,
903
  "per_class": {
904
  "0": {
905
- "accuracy": 0.1276595744680851,
906
- "precision": 0.0044444444444444444,
907
- "recall": 0.1276595744680851,
908
- "f1": 0.008589835361488905,
909
  "gt_support": 47,
910
- "predicted": 1350
911
  },
912
  "1": {
913
- "accuracy": 0.02127659574468085,
914
- "precision": 0.16,
915
- "recall": 0.02127659574468085,
916
- "f1": 0.03755868544600939,
917
  "gt_support": 1692,
918
- "predicted": 225
919
  },
920
  "2": {
921
- "accuracy": 0.02127659574468085,
922
- "precision": 0.027777777777777776,
923
- "recall": 0.02127659574468085,
924
- "f1": 0.024096385542168672,
925
  "gt_support": 47,
926
- "predicted": 36
927
  },
928
  "3": {
929
- "accuracy": 0.8134586838198912,
930
- "precision": 0.25374286155270875,
931
- "recall": 0.8134586838198912,
932
- "f1": 0.38682352941176473,
933
  "gt_support": 2021,
934
- "predicted": 6479
935
  },
936
  "4": {
937
- "accuracy": 0.34544917257683216,
938
- "precision": 0.9408450704225352,
939
- "recall": 0.34544917257683216,
940
- "f1": 0.5053496163406463,
941
  "gt_support": 6768,
942
- "predicted": 2485
943
  }
944
  },
945
  "total_samples": 10575
946
  },
947
  "9_0.4": {
948
- "accuracy": 0.29180762852404646,
949
  "baseline_accuracy": 0.64,
950
  "per_class": {
951
  "0": {
952
- "accuracy": 0.13432835820895522,
953
- "precision": 0.0044444444444444444,
954
- "recall": 0.13432835820895522,
955
- "f1": 0.008604206500956023,
956
  "gt_support": 67,
957
- "predicted": 2025
958
  },
959
  "1": {
960
- "accuracy": 0.014925373134328358,
961
- "precision": 0.17777777777777778,
962
- "recall": 0.014925373134328358,
963
- "f1": 0.027538726333907058,
964
  "gt_support": 2680,
965
- "predicted": 225
966
  },
967
  "2": {
968
- "accuracy": 0.029850746268656716,
969
- "precision": 0.010256410256410256,
970
- "recall": 0.029850746268656716,
971
- "f1": 0.015267175572519085,
972
  "gt_support": 67,
973
- "predicted": 195
974
  },
975
  "3": {
976
- "accuracy": 0.8197474167623421,
977
- "precision": 0.20683661645422943,
978
- "recall": 0.8197474167623421,
979
- "f1": 0.3303261623872311,
980
  "gt_support": 2613,
981
- "predicted": 10356
982
  },
983
  "4": {
984
- "accuracy": 0.22864842454394693,
985
- "precision": 0.9700967458223395,
986
- "recall": 0.22864842454394693,
987
- "f1": 0.3700721355477269,
988
  "gt_support": 9648,
989
- "predicted": 2274
990
  }
991
  },
992
  "total_samples": 15075
993
  },
994
  "9_0.6": {
995
- "accuracy": 0.35001821493624774,
996
  "baseline_accuracy": 0.64,
997
  "per_class": {
998
  "0": {
999
- "accuracy": 0.06557377049180328,
1000
- "precision": 0.0044444444444444444,
1001
- "recall": 0.06557377049180328,
1002
- "f1": 0.008324661810613943,
1003
  "gt_support": 61,
1004
- "predicted": 900
1005
  },
1006
  "1": {
1007
  "accuracy": 0.0,
@@ -1012,410 +1026,410 @@
1012
  "predicted": 0
1013
  },
1014
  "2": {
1015
- "accuracy": 0.01639344262295082,
1016
- "precision": 0.004761904761904762,
1017
- "recall": 0.01639344262295082,
1018
- "f1": 0.007380073800738008,
1019
  "gt_support": 61,
1020
- "predicted": 210
1021
  },
1022
  "3": {
1023
- "accuracy": 0.9061930783242259,
1024
- "precision": 0.20517579131869265,
1025
- "recall": 0.9061930783242259,
1026
- "f1": 0.33459436738125264,
1027
  "gt_support": 2196,
1028
- "predicted": 9699
1029
  },
1030
  "4": {
1031
- "accuracy": 0.31978597449908924,
1032
- "precision": 0.9633058984910837,
1033
- "recall": 0.31978597449908924,
1034
- "f1": 0.48017094017094014,
1035
  "gt_support": 8784,
1036
- "predicted": 2916
1037
  }
1038
  },
1039
  "total_samples": 13725
1040
  },
1041
  "9_0.8": {
1042
- "accuracy": 0.2597139713971397,
1043
  "baseline_accuracy": 0.64,
1044
  "per_class": {
1045
  "0": {
1046
- "accuracy": 0.07920792079207921,
1047
- "precision": 0.0044444444444444444,
1048
- "recall": 0.07920792079207921,
1049
- "f1": 0.008416622830089426,
1050
  "gt_support": 101,
1051
- "predicted": 1800
1052
  },
1053
  "1": {
1054
- "accuracy": 0.04950495049504951,
1055
- "precision": 0.2088888888888889,
1056
- "recall": 0.04950495049504951,
1057
- "f1": 0.0800408719346049,
1058
  "gt_support": 4747,
1059
- "predicted": 1125
1060
  },
1061
  "2": {
1062
- "accuracy": 0.039603960396039604,
1063
- "precision": 0.0055248618784530384,
1064
- "recall": 0.039603960396039604,
1065
- "f1": 0.009696969696969697,
1066
  "gt_support": 101,
1067
- "predicted": 724
1068
  },
1069
  "3": {
1070
- "accuracy": 0.8146658415841584,
1071
- "precision": 0.16554542596667715,
1072
- "recall": 0.8146658415841584,
1073
- "f1": 0.2751737471913048,
1074
  "gt_support": 3232,
1075
- "predicted": 15905
1076
  },
1077
  "4": {
1078
- "accuracy": 0.20778327832783278,
1079
- "precision": 0.9530116682434563,
1080
- "recall": 0.20778327832783278,
1081
- "f1": 0.3411797911374541,
1082
  "gt_support": 14544,
1083
- "predicted": 3171
1084
  }
1085
  },
1086
  "total_samples": 22725
1087
  },
1088
  "9_1.0": {
1089
- "accuracy": 0.20434782608695654,
1090
  "baseline_accuracy": 0.64,
1091
  "per_class": {
1092
  "0": {
1093
- "accuracy": 0.10326086956521739,
1094
- "precision": 0.004491725768321513,
1095
- "recall": 0.10326086956521739,
1096
- "f1": 0.008608971454463073,
1097
  "gt_support": 184,
1098
- "predicted": 4230
1099
  },
1100
  "1": {
1101
- "accuracy": 0.07065217391304347,
1102
- "precision": 0.22344448264008251,
1103
- "recall": 0.07065217391304347,
1104
- "f1": 0.1073581633495747,
1105
  "gt_support": 9200,
1106
- "predicted": 2909
1107
  },
1108
  "2": {
1109
- "accuracy": 0.005434782608695652,
1110
- "precision": 0.01282051282051282,
1111
- "recall": 0.005434782608695652,
1112
- "f1": 0.007633587786259542,
1113
  "gt_support": 184,
1114
- "predicted": 78
1115
  },
1116
  "3": {
1117
- "accuracy": 0.8144677661169415,
1118
- "precision": 0.14195655724318146,
1119
- "recall": 0.8144677661169415,
1120
- "f1": 0.24177352507579764,
1121
  "gt_support": 5336,
1122
- "predicted": 30615
1123
  },
1124
  "4": {
1125
- "accuracy": 0.129981884057971,
1126
- "precision": 0.9652466367713004,
1127
- "recall": 0.129981884057971,
1128
- "f1": 0.22911122937732836,
1129
  "gt_support": 26496,
1130
- "predicted": 3568
1131
  }
1132
  },
1133
  "total_samples": 41400
1134
  },
1135
  "11_0.0": {
1136
- "accuracy": 0.3445751633986928,
1137
  "baseline_accuracy": 0.4622222222222222,
1138
  "per_class": {
1139
  "0": {
1140
- "accuracy": 0.11764705882352941,
1141
- "precision": 0.004489337822671156,
1142
- "recall": 0.11764705882352941,
1143
- "f1": 0.008648648648648649,
1144
  "gt_support": 68,
1145
- "predicted": 1782
1146
  },
1147
  "1": {
1148
- "accuracy": 0.029411764705882353,
1149
- "precision": 0.17777777777777778,
1150
- "recall": 0.029411764705882353,
1151
- "f1": 0.050473186119873815,
1152
  "gt_support": 2720,
1153
- "predicted": 450
1154
  },
1155
  "2": {
1156
- "accuracy": 0.3382352941176471,
1157
- "precision": 0.006027253668763103,
1158
- "recall": 0.3382352941176471,
1159
- "f1": 0.011843460350154478,
1160
  "gt_support": 68,
1161
- "predicted": 3816
1162
  },
1163
  "3": {
1164
- "accuracy": 0.46854058078927774,
1165
- "precision": 0.4236660494866184,
1166
- "recall": 0.46854058078927774,
1167
- "f1": 0.4449748077433041,
1168
  "gt_support": 5372,
1169
- "predicted": 5941
1170
  },
1171
  "4": {
1172
- "accuracy": 0.373868778280543,
1173
- "precision": 0.79855028692238,
1174
- "recall": 0.373868778280543,
1175
- "f1": 0.5092940383318887,
1176
  "gt_support": 7072,
1177
- "predicted": 3311
1178
  }
1179
  },
1180
  "total_samples": 15300
1181
  },
1182
  "11_0.2": {
1183
- "accuracy": 0.39388235294117646,
1184
  "baseline_accuracy": 0.4622222222222222,
1185
  "per_class": {
1186
  "0": {
1187
- "accuracy": 0.047058823529411764,
1188
- "precision": 0.004459308807134894,
1189
- "recall": 0.047058823529411764,
1190
- "f1": 0.008146639511201629,
1191
  "gt_support": 85,
1192
- "predicted": 897
1193
  },
1194
  "1": {
1195
- "accuracy": 0.03529411764705882,
1196
- "precision": 0.21559633027522937,
1197
- "recall": 0.03529411764705882,
1198
- "f1": 0.060658206065820605,
1199
  "gt_support": 3995,
1200
- "predicted": 654
1201
  },
1202
  "2": {
1203
- "accuracy": 0.09411764705882353,
1204
- "precision": 0.006130268199233717,
1205
- "recall": 0.09411764705882353,
1206
- "f1": 0.011510791366906473,
1207
  "gt_support": 85,
1208
- "predicted": 1305
1209
  },
1210
  "3": {
1211
- "accuracy": 0.7642156862745098,
1212
- "precision": 0.36399719822554283,
1213
- "recall": 0.7642156862745098,
1214
- "f1": 0.4931203542622173,
1215
  "gt_support": 6120,
1216
- "predicted": 12849
1217
  },
1218
  "4": {
1219
- "accuracy": 0.3057692307692308,
1220
- "precision": 0.7903508771929825,
1221
- "recall": 0.3057692307692308,
1222
- "f1": 0.44094616639477974,
1223
  "gt_support": 8840,
1224
- "predicted": 3420
1225
  }
1226
  },
1227
  "total_samples": 19125
1228
  },
1229
  "11_0.4": {
1230
- "accuracy": 0.3198383838383838,
1231
  "baseline_accuracy": 0.4622222222222222,
1232
  "per_class": {
1233
  "0": {
1234
- "accuracy": 0.16363636363636364,
1235
- "precision": 0.0045203415369161224,
1236
- "recall": 0.16363636363636364,
1237
- "f1": 0.00879765395894428,
1238
  "gt_support": 55,
1239
- "predicted": 1991
1240
  },
1241
  "1": {
1242
- "accuracy": 0.05454545454545454,
1243
- "precision": 0.23555555555555555,
1244
- "recall": 0.05454545454545454,
1245
- "f1": 0.08857938718662953,
1246
  "gt_support": 2915,
1247
- "predicted": 675
1248
  },
1249
  "2": {
1250
- "accuracy": 0.12727272727272726,
1251
- "precision": 0.0054858934169279,
1252
- "recall": 0.12727272727272726,
1253
- "f1": 0.010518407212622089,
1254
  "gt_support": 55,
1255
- "predicted": 1276
1256
  },
1257
  "3": {
1258
- "accuracy": 0.5479338842975207,
1259
- "precision": 0.3327199732351957,
1260
- "recall": 0.5479338842975207,
1261
- "f1": 0.414029975020816,
1262
  "gt_support": 3630,
1263
- "predicted": 5978
1264
  },
1265
  "4": {
1266
- "accuracy": 0.31363636363636366,
1267
- "precision": 0.7307535641547862,
1268
- "recall": 0.31363636363636366,
1269
- "f1": 0.43889908256880733,
1270
  "gt_support": 5720,
1271
- "predicted": 2455
1272
  }
1273
  },
1274
  "total_samples": 12375
1275
  },
1276
  "11_0.6": {
1277
- "accuracy": 0.3105070118662352,
1278
  "baseline_accuracy": 0.4622222222222222,
1279
  "per_class": {
1280
  "0": {
1281
- "accuracy": 0.06796116504854369,
1282
- "precision": 0.004707464694014795,
1283
- "recall": 0.06796116504854369,
1284
- "f1": 0.00880503144654088,
1285
  "gt_support": 103,
1286
- "predicted": 1487
1287
  },
1288
  "1": {
1289
- "accuracy": 0.06796116504854369,
1290
- "precision": 0.27149321266968324,
1291
- "recall": 0.06796116504854369,
1292
- "f1": 0.10870971916655882,
1293
  "gt_support": 6180,
1294
- "predicted": 1547
1295
  },
1296
  "2": {
1297
- "accuracy": 0.10679611650485436,
1298
- "precision": 0.0054753608760577405,
1299
- "recall": 0.10679611650485436,
1300
- "f1": 0.010416666666666668,
1301
  "gt_support": 103,
1302
- "predicted": 2009
1303
  },
1304
  "3": {
1305
- "accuracy": 0.7001810103669573,
1306
- "precision": 0.28455828261887245,
1307
- "recall": 0.7001810103669573,
1308
- "f1": 0.40466000951022346,
1309
  "gt_support": 6077,
1310
- "predicted": 14953
1311
  },
1312
  "4": {
1313
- "accuracy": 0.23366318147871545,
1314
- "precision": 0.7873545139981126,
1315
- "recall": 0.23366318147871545,
1316
- "f1": 0.3603772226621553,
1317
  "gt_support": 10712,
1318
- "predicted": 3179
1319
  }
1320
  },
1321
  "total_samples": 23175
1322
  },
1323
  "11_0.8": {
1324
- "accuracy": 0.3004288499025341,
1325
  "baseline_accuracy": 0.4622222222222222,
1326
  "per_class": {
1327
  "0": {
1328
- "accuracy": 0.043859649122807015,
1329
- "precision": 0.0044444444444444444,
1330
- "recall": 0.043859649122807015,
1331
- "f1": 0.008071025020177562,
1332
  "gt_support": 114,
1333
- "predicted": 1125
1334
  },
1335
  "1": {
1336
- "accuracy": 0.043859649122807015,
1337
- "precision": 0.29333333333333333,
1338
- "recall": 0.043859649122807015,
1339
- "f1": 0.07630939993062781,
1340
  "gt_support": 7524,
1341
- "predicted": 1125
1342
  },
1343
  "2": {
1344
- "accuracy": 0.13157894736842105,
1345
- "precision": 0.00641573994867408,
1346
- "recall": 0.13157894736842105,
1347
- "f1": 0.012234910277324634,
1348
  "gt_support": 114,
1349
- "predicted": 2338
1350
  },
1351
  "3": {
1352
- "accuracy": 0.7499172459450513,
1353
- "precision": 0.25734083035156474,
1354
- "recall": 0.7499172459450513,
1355
- "f1": 0.3831874497864604,
1356
  "gt_support": 6042,
1357
- "predicted": 17607
1358
  },
1359
  "4": {
1360
- "accuracy": 0.23827597840755735,
1361
- "precision": 0.8176555716353111,
1362
- "recall": 0.23827597840755735,
1363
- "f1": 0.3690157403174188,
1364
  "gt_support": 11856,
1365
- "predicted": 3455
1366
  }
1367
  },
1368
  "total_samples": 25650
1369
  },
1370
  "11_1.0": {
1371
- "accuracy": 0.2344642857142857,
1372
  "baseline_accuracy": 0.4622222222222222,
1373
  "per_class": {
1374
  "0": {
1375
- "accuracy": 0.10714285714285714,
1376
- "precision": 0.004548900682335102,
1377
- "recall": 0.10714285714285714,
1378
- "f1": 0.008727272727272728,
1379
  "gt_support": 224,
1380
- "predicted": 5276
1381
  },
1382
  "1": {
1383
- "accuracy": 0.14688740079365079,
1384
- "precision": 0.32757190265486724,
1385
- "recall": 0.14688740079365079,
1386
- "f1": 0.2028253424657534,
1387
  "gt_support": 16128,
1388
- "predicted": 7232
1389
  },
1390
  "2": {
1391
- "accuracy": 0.06696428571428571,
1392
- "precision": 0.005263157894736842,
1393
- "recall": 0.06696428571428571,
1394
- "f1": 0.009759271307742354,
1395
  "gt_support": 224,
1396
- "predicted": 2850
1397
  },
1398
  "3": {
1399
- "accuracy": 0.6677431610942249,
1400
- "precision": 0.21803858321444078,
1401
- "recall": 0.6677431610942249,
1402
- "f1": 0.3287350946925415,
1403
  "gt_support": 10528,
1404
- "predicted": 32242
1405
  },
1406
  "4": {
1407
- "accuracy": 0.10212053571428571,
1408
- "precision": 0.8496428571428571,
1409
- "recall": 0.10212053571428571,
1410
- "f1": 0.18232679337829552,
1411
  "gt_support": 23296,
1412
- "predicted": 2800
1413
  }
1414
  },
1415
  "total_samples": 50400
1416
  },
1417
  "13_0.0": {
1418
- "accuracy": 0.4103472222222222,
1419
  "baseline_accuracy": 0.5288888888888889,
1420
  "per_class": {
1421
  "0": {
@@ -1427,310 +1441,310 @@
1427
  "predicted": 0
1428
  },
1429
  "1": {
 
 
 
 
 
 
 
 
1430
  "accuracy": 0.0,
1431
  "precision": 0.0,
1432
  "recall": 0.0,
1433
  "f1": 0.0,
1434
- "gt_support": 3072,
1435
- "predicted": 0
1436
- },
1437
- "2": {
1438
- "accuracy": 0.34375,
1439
- "precision": 0.005208333333333333,
1440
- "recall": 0.34375,
1441
- "f1": 0.010261194029850746,
1442
  "gt_support": 64,
1443
- "predicted": 4224
1444
  },
1445
  "3": {
1446
- "accuracy": 0.5996586134453782,
1447
- "precision": 0.5758416340940613,
1448
- "recall": 0.5996586134453782,
1449
- "f1": 0.5875088441499967,
1450
  "gt_support": 7616,
1451
- "predicted": 7931
1452
  },
1453
  "4": {
1454
- "accuracy": 0.36830357142857145,
1455
- "precision": 0.5879732739420935,
1456
- "recall": 0.36830357142857145,
1457
- "f1": 0.4529078744209985,
1458
  "gt_support": 3584,
1459
- "predicted": 2245
1460
  }
1461
  },
1462
  "total_samples": 14400
1463
  },
1464
  "13_0.2": {
1465
- "accuracy": 0.4152688172043011,
1466
  "baseline_accuracy": 0.48444444444444446,
1467
  "per_class": {
1468
  "0": {
1469
- "accuracy": 0.06451612903225806,
1470
- "precision": 0.004449388209121246,
1471
- "recall": 0.06451612903225806,
1472
- "f1": 0.008324661810613945,
1473
  "gt_support": 62,
1474
- "predicted": 899
1475
  },
1476
  "1": {
1477
- "accuracy": 0.016129032258064516,
1478
- "precision": 0.2577777777777778,
1479
- "recall": 0.016129032258064516,
1480
- "f1": 0.030358544883538344,
1481
  "gt_support": 3596,
1482
- "predicted": 225
1483
  },
1484
  "2": {
1485
- "accuracy": 0.1935483870967742,
1486
- "precision": 0.005323868677905945,
1487
- "recall": 0.1935483870967742,
1488
- "f1": 0.010362694300518133,
1489
  "gt_support": 62,
1490
- "predicted": 2254
1491
  },
1492
  "3": {
1493
- "accuracy": 0.6667653151820065,
1494
- "precision": 0.534202726733847,
1495
- "recall": 0.6667653151820065,
1496
- "f1": 0.5931679062726256,
1497
  "gt_support": 6758,
1498
- "predicted": 8435
1499
  },
1500
  "4": {
1501
- "accuracy": 0.3493663594470046,
1502
- "precision": 0.5676181562938699,
1503
- "recall": 0.3493663594470046,
1504
- "f1": 0.4325191656266714,
1505
  "gt_support": 3472,
1506
- "predicted": 2137
1507
  }
1508
  },
1509
  "total_samples": 13950
1510
  },
1511
  "13_0.4": {
1512
- "accuracy": 0.39555555555555555,
1513
  "baseline_accuracy": 0.44,
1514
  "per_class": {
1515
  "0": {
1516
- "accuracy": 0.08,
1517
- "precision": 0.004456327985739751,
1518
- "recall": 0.08,
1519
- "f1": 0.008442380751371888,
1520
  "gt_support": 125,
1521
- "predicted": 2244
1522
  },
1523
  "1": {
1524
- "accuracy": 0.04,
1525
- "precision": 0.30384271671134944,
1526
- "recall": 0.04,
1527
- "f1": 0.07069341927435284,
1528
  "gt_support": 8500,
1529
- "predicted": 1119
1530
  },
1531
  "2": {
1532
- "accuracy": 0.072,
1533
- "precision": 0.0038910505836575876,
1534
- "recall": 0.072,
1535
- "f1": 0.007383100902379,
1536
  "gt_support": 125,
1537
- "predicted": 2313
1538
  },
1539
  "3": {
1540
- "accuracy": 0.7343838383838384,
1541
- "precision": 0.46557377049180326,
1542
- "recall": 0.7343838383838384,
1543
- "f1": 0.5698698855620004,
1544
  "gt_support": 12375,
1545
- "predicted": 19520
1546
  },
1547
  "4": {
1548
- "accuracy": 0.2397142857142857,
1549
- "precision": 0.5728917719358143,
1550
- "recall": 0.2397142857142857,
1551
- "f1": 0.33799979856984597,
1552
  "gt_support": 7000,
1553
- "predicted": 2929
1554
  }
1555
  },
1556
  "total_samples": 28125
1557
  },
1558
  "13_0.6": {
1559
- "accuracy": 0.3496786042240588,
1560
  "baseline_accuracy": 0.39555555555555555,
1561
  "per_class": {
1562
  "0": {
1563
- "accuracy": 0.04132231404958678,
1564
- "precision": 0.004336513443191674,
1565
- "recall": 0.04132231404958678,
1566
- "f1": 0.007849293563579277,
1567
  "gt_support": 121,
1568
- "predicted": 1153
1569
  },
1570
  "1": {
1571
- "accuracy": 0.04948082220809494,
1572
- "precision": 0.349812734082397,
1573
- "recall": 0.04948082220809494,
1574
- "f1": 0.08669822704910425,
1575
  "gt_support": 9438,
1576
- "predicted": 1335
1577
  },
1578
  "2": {
1579
- "accuracy": 0.1652892561983471,
1580
- "precision": 0.004822763443453099,
1581
- "recall": 0.1652892561983471,
1582
- "f1": 0.009372071227741332,
1583
  "gt_support": 121,
1584
- "predicted": 4147
1585
  },
1586
  "3": {
1587
- "accuracy": 0.6759216268920049,
1588
- "precision": 0.41629968544466683,
1589
- "recall": 0.6759216268920049,
1590
- "f1": 0.5152544772421603,
1591
  "gt_support": 10769,
1592
- "predicted": 17485
1593
  },
1594
  "4": {
1595
- "accuracy": 0.25811688311688313,
1596
- "precision": 0.5632850241545894,
1597
- "recall": 0.25811688311688313,
1598
- "f1": 0.3540127517457748,
1599
  "gt_support": 6776,
1600
- "predicted": 3105
1601
  }
1602
  },
1603
  "total_samples": 27225
1604
  },
1605
  "13_0.8": {
1606
- "accuracy": 0.3028968253968254,
1607
  "baseline_accuracy": 0.39555555555555555,
1608
  "per_class": {
1609
  "0": {
1610
- "accuracy": 0.08928571428571429,
1611
- "precision": 0.004486316733961417,
1612
- "recall": 0.08928571428571429,
1613
- "f1": 0.008543357539513028,
1614
  "gt_support": 112,
1615
- "predicted": 2229
1616
  },
1617
  "1": {
1618
- "accuracy": 0.09179373996789728,
1619
- "precision": 0.400262467191601,
1620
- "recall": 0.09179373996789728,
1621
- "f1": 0.14933899134976333,
1622
  "gt_support": 9968,
1623
- "predicted": 2286
1624
  },
1625
  "2": {
1626
- "accuracy": 0.17857142857142858,
1627
- "precision": 0.0053533190578158455,
1628
- "recall": 0.17857142857142858,
1629
- "f1": 0.010395010395010394,
1630
  "gt_support": 112,
1631
- "predicted": 3736
1632
  },
1633
  "3": {
1634
- "accuracy": 0.6009615384615384,
1635
- "precision": 0.36486204739731737,
1636
- "recall": 0.6009615384615384,
1637
- "f1": 0.454054054054054,
1638
  "gt_support": 8736,
1639
- "predicted": 14389
1640
  },
1641
  "4": {
1642
- "accuracy": 0.22927295918367346,
1643
- "precision": 0.56171875,
1644
- "recall": 0.22927295918367346,
1645
- "f1": 0.32563405797101447,
1646
  "gt_support": 6272,
1647
- "predicted": 2560
1648
  }
1649
  },
1650
  "total_samples": 25200
1651
  },
1652
  "13_1.0": {
1653
- "accuracy": 0.28338104976141787,
1654
  "baseline_accuracy": 0.43555555555555553,
1655
  "per_class": {
1656
  "0": {
1657
- "accuracy": 0.10122699386503067,
1658
- "precision": 0.004623143737741664,
1659
- "recall": 0.10122699386503067,
1660
- "f1": 0.008842443729903537,
1661
  "gt_support": 326,
1662
- "predicted": 7138
1663
  },
1664
  "1": {
1665
- "accuracy": 0.1453925128333542,
1666
- "precision": 0.44003410382720726,
1667
- "recall": 0.1453925128333542,
1668
- "f1": 0.21856766421983811,
1669
  "gt_support": 31948,
1670
- "predicted": 10556
1671
  },
1672
  "2": {
1673
- "accuracy": 0.13803680981595093,
1674
- "precision": 0.005077287600135394,
1675
- "recall": 0.13803680981595093,
1676
- "f1": 0.00979431929480901,
1677
  "gt_support": 326,
1678
- "predicted": 8863
1679
  },
1680
  "3": {
1681
- "accuracy": 0.5809104650128923,
1682
- "precision": 0.3175609993195295,
1683
- "recall": 0.5809104650128923,
1684
- "f1": 0.4106407718173533,
1685
  "gt_support": 22494,
1686
- "predicted": 41148
1687
  },
1688
  "4": {
1689
- "accuracy": 0.16411042944785276,
1690
- "precision": 0.5307351638618246,
1691
- "recall": 0.16411042944785276,
1692
- "f1": 0.25070080749759427,
1693
  "gt_support": 18256,
1694
- "predicted": 5645
1695
  }
1696
  },
1697
  "total_samples": 73350
1698
  },
1699
  "15_0.0": {
1700
- "accuracy": 0.4518346253229974,
1701
  "baseline_accuracy": 0.7422222222222222,
1702
  "per_class": {
1703
  "0": {
1704
- "accuracy": 0.011627906976744186,
1705
- "precision": 0.0044444444444444444,
1706
- "recall": 0.011627906976744186,
1707
- "f1": 0.006430868167202572,
1708
  "gt_support": 86,
1709
- "predicted": 225
1710
  },
1711
  "1": {
1712
- "accuracy": 0.011627906976744186,
1713
- "precision": 0.24888888888888888,
1714
- "recall": 0.011627906976744186,
1715
- "f1": 0.02221781392580837,
1716
  "gt_support": 4816,
1717
- "predicted": 225
1718
  },
1719
  "2": {
1720
- "accuracy": 0.27906976744186046,
1721
- "precision": 0.004856333468231485,
1722
- "recall": 0.27906976744186046,
1723
- "f1": 0.009546539379474939,
1724
  "gt_support": 86,
1725
- "predicted": 4942
1726
  },
1727
  "3": {
1728
- "accuracy": 0.603119342709929,
1729
- "precision": 0.7574326687653026,
1730
- "recall": 0.603119342709929,
1731
- "f1": 0.6715249244127451,
1732
  "gt_support": 14362,
1733
- "predicted": 11436
1734
  },
1735
  "4": {
1736
  "accuracy": 0.0,
@@ -1738,46 +1752,46 @@
1738
  "recall": 0.0,
1739
  "f1": 0.0,
1740
  "gt_support": 0,
1741
- "predicted": 2522
1742
  }
1743
  },
1744
  "total_samples": 19350
1745
  },
1746
  "15_0.2": {
1747
- "accuracy": 0.4623243933588761,
1748
  "baseline_accuracy": 0.6755555555555556,
1749
  "per_class": {
1750
  "0": {
1751
- "accuracy": 0.05747126436781609,
1752
- "precision": 0.0044444444444444444,
1753
- "recall": 0.05747126436781609,
1754
- "f1": 0.00825082508250825,
1755
  "gt_support": 87,
1756
- "predicted": 1125
1757
  },
1758
  "1": {
1759
- "accuracy": 0.03432086773514651,
1760
- "precision": 0.314540059347181,
1761
- "recall": 0.03432086773514651,
1762
- "f1": 0.061888775361261125,
1763
  "gt_support": 6177,
1764
- "predicted": 674
1765
  },
1766
  "2": {
1767
- "accuracy": 0.11494252873563218,
1768
- "precision": 0.00419639110365086,
1769
- "recall": 0.11494252873563218,
1770
- "f1": 0.008097165991902834,
1771
  "gt_support": 87,
1772
- "predicted": 2383
1773
  },
1774
  "3": {
1775
- "accuracy": 0.6671960072595281,
1776
- "precision": 0.6880605162598455,
1777
- "recall": 0.6671960072595281,
1778
- "f1": 0.6774676546243329,
1779
  "gt_support": 13224,
1780
- "predicted": 12823
1781
  },
1782
  "4": {
1783
  "accuracy": 0.0,
@@ -1785,46 +1799,46 @@
1785
  "recall": 0.0,
1786
  "f1": 0.0,
1787
  "gt_support": 0,
1788
- "predicted": 2570
1789
  }
1790
  },
1791
  "total_samples": 19575
1792
  },
1793
  "15_0.4": {
1794
- "accuracy": 0.44166666666666665,
1795
  "baseline_accuracy": 0.6133333333333333,
1796
  "per_class": {
1797
  "0": {
1798
- "accuracy": 0.051470588235294115,
1799
- "precision": 0.004201680672268907,
1800
- "recall": 0.051470588235294115,
1801
- "f1": 0.007769145394006659,
1802
  "gt_support": 136,
1803
- "predicted": 1666
1804
  },
1805
  "1": {
1806
- "accuracy": 0.032525951557093424,
1807
- "precision": 0.3794147325933401,
1808
- "recall": 0.032525951557093424,
1809
- "f1": 0.05991554457812126,
1810
  "gt_support": 11560,
1811
- "predicted": 991
1812
  },
1813
  "2": {
1814
- "accuracy": 0.11764705882352941,
1815
- "precision": 0.0040650406504065045,
1816
- "recall": 0.11764705882352941,
1817
- "f1": 0.007858546168958744,
1818
  "gt_support": 136,
1819
- "predicted": 3936
1820
  },
1821
  "3": {
1822
- "accuracy": 0.698849104859335,
1823
- "precision": 0.6206113371817924,
1824
- "recall": 0.698849104859335,
1825
- "f1": 0.6574106561074633,
1826
  "gt_support": 18768,
1827
- "predicted": 21134
1828
  },
1829
  "4": {
1830
  "accuracy": 0.0,
@@ -1832,46 +1846,46 @@
1832
  "recall": 0.0,
1833
  "f1": 0.0,
1834
  "gt_support": 0,
1835
- "predicted": 2873
1836
  }
1837
  },
1838
  "total_samples": 30600
1839
  },
1840
  "15_0.6": {
1841
- "accuracy": 0.378,
1842
  "baseline_accuracy": 0.5466666666666666,
1843
  "per_class": {
1844
  "0": {
1845
- "accuracy": 0.09,
1846
- "precision": 0.004468718967229395,
1847
- "recall": 0.09,
1848
- "f1": 0.008514664143803218,
1849
  "gt_support": 100,
1850
- "predicted": 2014
1851
  },
1852
  "1": {
1853
- "accuracy": 0.0607,
1854
- "precision": 0.44665194996320823,
1855
- "recall": 0.0607,
1856
- "f1": 0.10687560524694076,
1857
  "gt_support": 10000,
1858
- "predicted": 1359
1859
  },
1860
  "2": {
1861
- "accuracy": 0.15,
1862
- "precision": 0.0048449612403100775,
1863
- "recall": 0.15,
1864
- "f1": 0.009386733416770963,
1865
  "gt_support": 100,
1866
- "predicted": 3096
1867
  },
1868
  "3": {
1869
- "accuracy": 0.6401626016260162,
1870
- "precision": 0.5533769063180828,
1871
- "recall": 0.6401626016260162,
1872
- "f1": 0.5936145350371291,
1873
  "gt_support": 12300,
1874
- "predicted": 14229
1875
  },
1876
  "4": {
1877
  "accuracy": 0.0,
@@ -1879,46 +1893,46 @@
1879
  "recall": 0.0,
1880
  "f1": 0.0,
1881
  "gt_support": 0,
1882
- "predicted": 1802
1883
  }
1884
  },
1885
  "total_samples": 22500
1886
  },
1887
  "15_0.8": {
1888
- "accuracy": 0.3858807588075881,
1889
  "baseline_accuracy": 0.5066666666666667,
1890
  "per_class": {
1891
  "0": {
1892
- "accuracy": 0.054878048780487805,
1893
- "precision": 0.0044444444444444444,
1894
- "recall": 0.054878048780487805,
1895
- "f1": 0.008222932846048425,
1896
  "gt_support": 164,
1897
- "predicted": 2025
1898
  },
1899
  "1": {
1900
- "accuracy": 0.09060761660248182,
1901
- "precision": 0.504617217753947,
1902
- "recall": 0.09060761660248182,
1903
- "f1": 0.15362989162472226,
1904
  "gt_support": 18696,
1905
- "predicted": 3357
1906
  },
1907
  "2": {
1908
- "accuracy": 0.10975609756097561,
1909
- "precision": 0.004413928396272682,
1910
- "recall": 0.10975609756097561,
1911
- "f1": 0.008486562942008486,
1912
  "gt_support": 164,
1913
- "predicted": 4078
1914
  },
1915
  "3": {
1916
- "accuracy": 0.7002685164466323,
1917
- "precision": 0.48896527479395335,
1918
- "recall": 0.7002685164466323,
1919
- "f1": 0.5758446994962854,
1920
  "gt_support": 17876,
1921
- "predicted": 25601
1922
  },
1923
  "4": {
1924
  "accuracy": 0.0,
@@ -1926,46 +1940,46 @@
1926
  "recall": 0.0,
1927
  "f1": 0.0,
1928
  "gt_support": 0,
1929
- "predicted": 1839
1930
  }
1931
  },
1932
  "total_samples": 36900
1933
  },
1934
  "15_1.0": {
1935
- "accuracy": 0.3119295254333618,
1936
  "baseline_accuracy": 0.5688888888888889,
1937
  "per_class": {
1938
  "0": {
1939
- "accuracy": 0.10741687979539642,
1940
- "precision": 0.004527325644065969,
1941
- "recall": 0.10741687979539642,
1942
- "f1": 0.008688456764584195,
1943
  "gt_support": 391,
1944
- "predicted": 9277
1945
  },
1946
  "1": {
1947
- "accuracy": 0.10653772378516624,
1948
- "precision": 0.5680196015766485,
1949
- "recall": 0.10653772378516624,
1950
- "f1": 0.17942289896525615,
1951
  "gt_support": 50048,
1952
- "predicted": 9387
1953
  },
1954
  "2": {
1955
- "accuracy": 0.10230179028132992,
1956
- "precision": 0.003532008830022075,
1957
- "recall": 0.10230179028132992,
1958
- "f1": 0.006828269033799932,
1959
  "gt_support": 391,
1960
- "predicted": 11325
1961
  },
1962
  "3": {
1963
- "accuracy": 0.5930273253466146,
1964
- "precision": 0.42557959814528595,
1965
- "recall": 0.5930273253466146,
1966
- "f1": 0.4955401833417693,
1967
  "gt_support": 37145,
1968
- "predicted": 51760
1969
  },
1970
  "4": {
1971
  "accuracy": 0.0,
@@ -1973,23 +1987,10 @@
1973
  "recall": 0.0,
1974
  "f1": 0.0,
1975
  "gt_support": 0,
1976
- "predicted": 6226
1977
  }
1978
  },
1979
  "total_samples": 87975
1980
  }
1981
- },
1982
- "total_trajectories": 300,
1983
- "total_steps": 3223,
1984
- "config": {
1985
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_general.pt",
1986
- "trajectories_dir": "reveng/trajectories_test_full",
1987
- "activations_dir": "interp/activations_test_full",
1988
- "layers": "15",
1989
- "steps": "all",
1990
- "token_categories": {
1991
- "output": "all"
1992
- },
1993
- "pad_to_size": 15
1994
  }
1995
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.557248939911056,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.09587340986658392,
8
+ "precision": 0.00974025974025974,
9
+ "recall": 0.09587340986658392,
10
+ "f1": 0.01768392136664091,
11
  "gt_support": 3223,
12
+ "predicted": 31724
13
  },
14
  "1": {
15
+ "accuracy": 0.6826829012280886,
16
+ "precision": 0.472149788894761,
17
+ "recall": 0.6826829012280886,
18
+ "f1": 0.5582256035073204,
19
  "gt_support": 238175,
20
+ "predicted": 344378
21
  },
22
  "2": {
23
+ "accuracy": 0.20912193608439342,
24
+ "precision": 0.014440897306794077,
25
+ "recall": 0.20912193608439342,
26
+ "f1": 0.02701619368286035,
27
  "gt_support": 3223,
28
+ "predicted": 46673
29
  },
30
  "3": {
31
+ "accuracy": 0.20205797506431172,
32
+ "precision": 0.6346481491042387,
33
+ "recall": 0.20205797506431172,
34
+ "f1": 0.306525113614403,
35
  "gt_support": 243346,
36
+ "predicted": 77476
37
  },
38
  "4": {
39
+ "accuracy": 0.8066844288556878,
40
+ "precision": 0.8507406946346321,
41
+ "recall": 0.8066844288556878,
42
+ "f1": 0.8281270286411674,
43
  "gt_support": 237208,
44
+ "predicted": 224924
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
+ "by_complexity": {
50
+ "0.0": {
51
+ "accuracy": 0.6471971326164875,
52
+ "baseline_accuracy": 0.4410322580645161,
53
  "per_class": {
54
  "0": {
55
+ "accuracy": 0.14838709677419354,
56
+ "precision": 0.009457236842105263,
57
+ "recall": 0.14838709677419354,
58
+ "f1": 0.01778121376111326,
59
+ "gt_support": 310,
60
+ "predicted": 4864
61
  },
62
  "1": {
63
+ "accuracy": 0.00037741545893719806,
64
+ "precision": 0.8333333333333334,
65
+ "recall": 0.00037741545893719806,
66
+ "f1": 0.0007544892108042854,
67
+ "gt_support": 13248,
68
+ "predicted": 6
69
  },
70
  "2": {
71
+ "accuracy": 0.23870967741935484,
72
+ "precision": 0.014679627058123389,
73
+ "recall": 0.23870967741935484,
74
+ "f1": 0.027658381610913853,
75
+ "gt_support": 310,
76
+ "predicted": 5041
77
  },
78
  "3": {
79
+ "accuracy": 0.7939990897860998,
80
+ "precision": 0.6765365759078193,
81
+ "recall": 0.7939990897860998,
82
+ "f1": 0.7305765348089436,
83
+ "gt_support": 30762,
84
+ "predicted": 36103
85
  },
86
  "4": {
87
+ "accuracy": 0.8197452229299363,
88
+ "precision": 0.8675429726996967,
89
+ "recall": 0.8197452229299363,
90
+ "f1": 0.8429670869494024,
91
+ "gt_support": 25120,
92
+ "predicted": 23736
93
  }
94
  },
95
+ "total_samples": 69750
96
  },
97
+ "0.2": {
98
+ "accuracy": 0.5569641025641026,
99
+ "baseline_accuracy": 0.3972239316239316,
100
  "per_class": {
101
  "0": {
102
+ "accuracy": 0.2523076923076923,
103
+ "precision": 0.008888888888888889,
104
+ "recall": 0.2523076923076923,
105
+ "f1": 0.01717277486910995,
106
+ "gt_support": 325,
107
+ "predicted": 9225
108
  },
109
  "1": {
110
+ "accuracy": 0.12328354613346182,
111
+ "precision": 0.30178387144331414,
112
+ "recall": 0.12328354613346182,
113
+ "f1": 0.17505451746696882,
114
+ "gt_support": 16604,
115
+ "predicted": 6783
116
  },
117
  "2": {
118
+ "accuracy": 0.24,
119
+ "precision": 0.014473928372610873,
120
+ "recall": 0.24,
121
+ "f1": 0.027301365068253414,
122
+ "gt_support": 325,
123
+ "predicted": 5389
124
  },
125
  "3": {
126
+ "accuracy": 0.5553757703033015,
127
+ "precision": 0.6298610026550054,
128
+ "recall": 0.5553757703033015,
129
+ "f1": 0.5902779048281162,
130
+ "gt_support": 29047,
131
+ "predicted": 25612
132
  },
133
  "4": {
134
+ "accuracy": 0.8346629883686251,
135
+ "precision": 0.8572905498544954,
136
+ "recall": 0.8346629883686251,
137
+ "f1": 0.8458254627880619,
138
+ "gt_support": 26824,
139
+ "predicted": 26116
140
  }
141
  },
142
+ "total_samples": 73125
143
  },
144
+ "0.4": {
145
+ "accuracy": 0.4879035639412998,
146
+ "baseline_accuracy": 0.4000524109014675,
147
  "per_class": {
148
  "0": {
149
+ "accuracy": 0.13679245283018868,
150
+ "precision": 0.009281485037606016,
151
+ "recall": 0.13679245283018868,
152
+ "f1": 0.017383485688595834,
153
+ "gt_support": 424,
154
+ "predicted": 6249
155
  },
156
  "1": {
157
+ "accuracy": 0.5257993508189381,
158
+ "precision": 0.3598274013174692,
159
+ "recall": 0.5257993508189381,
160
+ "f1": 0.42726128939350294,
161
+ "gt_support": 26803,
162
+ "predicted": 39166
163
  },
164
  "2": {
165
+ "accuracy": 0.21462264150943397,
166
+ "precision": 0.014198782961460446,
167
+ "recall": 0.21462264150943397,
168
+ "f1": 0.02663544563149422,
169
+ "gt_support": 424,
170
+ "predicted": 6409
171
  },
172
  "3": {
173
+ "accuracy": 0.21462072579588629,
174
+ "precision": 0.5524008632317238,
175
+ "recall": 0.21462072579588629,
176
+ "f1": 0.3091351687958787,
177
+ "gt_support": 38165,
178
+ "predicted": 14828
179
  },
180
  "4": {
181
+ "accuracy": 0.8150689561925365,
182
+ "precision": 0.8387713927925421,
183
+ "recall": 0.8150689561925365,
184
+ "f1": 0.8267503257217308,
185
+ "gt_support": 29584,
186
+ "predicted": 28748
187
  }
188
  },
189
+ "total_samples": 95400
190
  },
191
+ "0.6": {
192
+ "accuracy": 0.5060699588477366,
193
+ "baseline_accuracy": 0.3553909465020576,
194
  "per_class": {
195
  "0": {
196
+ "accuracy": 0.13194444444444445,
197
+ "precision": 0.009435523919880815,
198
+ "recall": 0.13194444444444445,
199
+ "f1": 0.01761161748802719,
200
+ "gt_support": 432,
201
+ "predicted": 6041
202
  },
203
  "1": {
204
+ "accuracy": 0.7069990541818674,
205
+ "precision": 0.40778552780267313,
206
+ "recall": 0.7069990541818674,
207
+ "f1": 0.517237118497467,
208
+ "gt_support": 29604,
209
+ "predicted": 51326
210
  },
211
  "2": {
212
+ "accuracy": 0.2152777777777778,
213
+ "precision": 0.014884763124199745,
214
+ "recall": 0.2152777777777778,
215
+ "f1": 0.02784431137724551,
216
+ "gt_support": 432,
217
+ "predicted": 6248
218
  },
219
  "3": {
220
+ "accuracy": 0.013110475953771591,
221
+ "precision": 0.45230439442658094,
222
+ "recall": 0.013110475953771591,
223
+ "f1": 0.025482322393647533,
224
+ "gt_support": 32188,
225
+ "predicted": 933
226
  },
227
  "4": {
228
+ "accuracy": 0.801528485409912,
229
+ "precision": 0.847972559108171,
230
+ "recall": 0.801528485409912,
231
+ "f1": 0.8240966724209775,
232
+ "gt_support": 34544,
233
+ "predicted": 32652
234
  }
235
  },
236
+ "total_samples": 97200
237
  },
238
+ "0.8": {
239
+ "accuracy": 0.5557427757427758,
240
+ "baseline_accuracy": 0.3470899470899471,
241
  "per_class": {
242
  "0": {
243
+ "accuracy": 0.07326007326007326,
244
+ "precision": 0.011334655709832813,
245
+ "recall": 0.07326007326007326,
246
+ "f1": 0.019631901840490795,
247
+ "gt_support": 546,
248
+ "predicted": 3529
249
  },
250
  "1": {
251
+ "accuracy": 0.7725609756097561,
252
+ "precision": 0.47090272317918663,
253
+ "recall": 0.7725609756097561,
254
+ "f1": 0.5851414361206093,
255
+ "gt_support": 42640,
256
+ "predicted": 69955
257
  },
258
  "2": {
259
+ "accuracy": 0.2326007326007326,
260
+ "precision": 0.015101070154577883,
261
+ "recall": 0.2326007326007326,
262
+ "f1": 0.028360875390799463,
263
+ "gt_support": 546,
264
+ "predicted": 8410
265
  },
266
  "3": {
 
 
 
 
 
 
 
 
267
  "accuracy": 0.0,
268
  "precision": 0.0,
269
  "recall": 0.0,
270
  "f1": 0.0,
271
+ "gt_support": 36766,
272
+ "predicted": 0
273
+ },
274
+ "4": {
275
+ "accuracy": 0.8302795617680393,
276
+ "precision": 0.8585799394472117,
277
+ "recall": 0.8302795617680393,
278
+ "f1": 0.8441926345609065,
279
+ "gt_support": 42352,
280
+ "predicted": 40956
281
  }
282
  },
283
+ "total_samples": 122850
284
+ },
285
+ "1.0": {
286
+ "accuracy": 0.5779426644182125,
287
+ "baseline_accuracy": 0.4095034663668728,
 
 
288
  "per_class": {
289
  "0": {
290
+ "accuracy": 0.021922428330522766,
291
+ "precision": 0.014317180616740088,
292
+ "recall": 0.021922428330522766,
293
+ "f1": 0.0173217854763491,
294
+ "gt_support": 1186,
295
+ "predicted": 1816
296
  },
297
  "1": {
298
+ "accuracy": 0.8472217138255427,
299
+ "precision": 0.5226372063090627,
300
+ "recall": 0.8472217138255427,
301
+ "f1": 0.6464747327332779,
302
+ "gt_support": 109276,
303
+ "predicted": 177142
304
  },
305
  "2": {
306
+ "accuracy": 0.17790893760539628,
307
+ "precision": 0.013903531892461782,
308
+ "recall": 0.17790893760539628,
309
+ "f1": 0.025791468035692457,
310
+ "gt_support": 1186,
311
+ "predicted": 15176
312
  },
313
  "3": {
314
+ "accuracy": 0.0,
315
+ "precision": 0.0,
316
+ "recall": 0.0,
317
+ "f1": 0.0,
318
+ "gt_support": 76418,
319
+ "predicted": 0
320
  },
321
  "4": {
322
+ "accuracy": 0.7794222177091795,
323
+ "precision": 0.8444633918257329,
324
+ "recall": 0.7794222177091795,
325
+ "f1": 0.8106402640264027,
326
+ "gt_support": 78784,
327
+ "predicted": 72716
328
  }
329
  },
330
+ "total_samples": 266850
331
+ }
332
+ },
333
+ "total_trajectories": 300,
334
+ "total_steps": 3223,
335
+ "single_size_mode": false,
336
+ "config": {
337
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_general.pt",
338
+ "trajectories_dir": "reveng/trajectories_test_full",
339
+ "activations_dir": "interp/activations_test_full",
340
+ "layers": "15",
341
+ "steps": "all",
342
+ "token_categories": {
343
+ "prompt_suffix": "all"
344
  },
345
+ "pad_to_size": 15
346
+ },
347
+ "by_size": {
348
+ "7": {
349
+ "accuracy": 0.7535819735819735,
350
+ "baseline_accuracy": 0.7822222222222223,
351
  "per_class": {
352
  "0": {
353
+ "accuracy": 0.17132867132867133,
354
+ "precision": 0.02,
355
+ "recall": 0.17132867132867133,
356
+ "f1": 0.0358187134502924,
357
+ "gt_support": 286,
358
+ "predicted": 2450
359
  },
360
  "1": {
361
+ "accuracy": 0.0,
362
+ "precision": 0.0,
363
+ "recall": 0.0,
364
+ "f1": 0.0,
365
+ "gt_support": 8224,
366
+ "predicted": 0
367
  },
368
  "2": {
369
+ "accuracy": 0.8216783216783217,
370
+ "precision": 0.021261196055369583,
371
+ "recall": 0.8216783216783217,
372
+ "f1": 0.041449863303642295,
373
+ "gt_support": 286,
374
+ "predicted": 11053
375
  },
376
  "3": {
377
+ "accuracy": 0.0,
378
+ "precision": 0.0,
379
+ "recall": 0.0,
380
+ "f1": 0.0,
381
+ "gt_support": 5218,
382
+ "predicted": 0
383
  },
384
  "4": {
385
+ "accuracy": 0.9577439605848697,
386
+ "precision": 0.9481188664031309,
387
+ "recall": 0.9577439605848697,
388
+ "f1": 0.9529071089016928,
389
+ "gt_support": 50336,
390
+ "predicted": 50847
391
  }
392
  },
393
+ "total_samples": 64350
394
  },
395
+ "9": {
396
+ "accuracy": 0.5927194120190229,
397
+ "baseline_accuracy": 0.64,
398
  "per_class": {
399
  "0": {
400
+ "accuracy": 0.1245136186770428,
401
+ "precision": 0.0141280353200883,
402
+ "recall": 0.1245136186770428,
403
+ "f1": 0.025376685170499604,
404
+ "gt_support": 514,
405
+ "predicted": 4530
406
  },
407
  "1": {
408
+ "accuracy": 0.0,
409
+ "precision": 0.0,
410
+ "recall": 0.0,
411
+ "f1": 0.0,
412
+ "gt_support": 22670,
413
+ "predicted": 0
414
  },
415
  "2": {
416
+ "accuracy": 0.8424124513618677,
417
+ "precision": 0.012448609953138028,
418
+ "recall": 0.8424124513618677,
419
+ "f1": 0.02453466300252146,
420
+ "gt_support": 514,
421
+ "predicted": 34783
422
  },
423
  "3": {
424
+ "accuracy": 0.0022859054415700266,
425
+ "precision": 0.5256410256410257,
426
+ "recall": 0.0022859054415700266,
427
+ "f1": 0.004552015099367158,
428
+ "gt_support": 17936,
429
+ "predicted": 78
430
  },
431
  "4": {
432
+ "accuracy": 0.918855382619974,
433
+ "precision": 0.8918291611481923,
434
+ "recall": 0.918855382619974,
435
+ "f1": 0.9051405756113792,
436
+ "gt_support": 74016,
437
+ "predicted": 76259
438
  }
439
  },
440
+ "total_samples": 115650
441
  },
442
+ "11": {
443
+ "accuracy": 0.533799007019346,
444
+ "baseline_accuracy": 0.4622222222222222,
445
  "per_class": {
446
  "0": {
447
+ "accuracy": 0.300462249614792,
448
+ "precision": 0.007957559681697613,
449
+ "recall": 0.300462249614792,
450
+ "f1": 0.015504492327264054,
451
+ "gt_support": 649,
452
+ "predicted": 24505
453
  },
454
  "1": {
455
+ "accuracy": 0.5510110992853885,
456
+ "precision": 0.4409563789012593,
457
+ "recall": 0.5510110992853885,
458
+ "f1": 0.4898786793281741,
459
+ "gt_support": 39462,
460
+ "predicted": 49311
461
  },
462
  "2": {
463
+ "accuracy": 0.009244992295839754,
464
+ "precision": 0.00718562874251497,
465
+ "recall": 0.009244992295839754,
466
+ "f1": 0.008086253369272236,
467
+ "gt_support": 649,
468
+ "predicted": 835
469
  },
470
  "3": {
471
+ "accuracy": 0.10593343747517805,
472
+ "precision": 0.5283243100488578,
473
+ "recall": 0.10593343747517805,
474
+ "f1": 0.17648096687398,
475
+ "gt_support": 37769,
476
+ "predicted": 7573
477
  },
478
  "4": {
479
+ "accuracy": 0.7704456560388764,
480
+ "precision": 0.815065594583157,
481
+ "recall": 0.7704456560388764,
482
+ "f1": 0.7921277713885314,
483
+ "gt_support": 67496,
484
+ "predicted": 63801
485
  }
486
  },
487
+ "total_samples": 146025
488
  },
489
+ "13": {
490
+ "accuracy": 0.5025459533607681,
491
+ "baseline_accuracy": 0.37721810699588476,
492
  "per_class": {
493
  "0": {
494
+ "accuracy": 0.0012345679012345679,
495
+ "precision": 0.0041841004184100415,
496
+ "recall": 0.0012345679012345679,
497
+ "f1": 0.0019065776930409914,
498
+ "gt_support": 810,
499
+ "predicted": 239
500
  },
501
  "1": {
502
+ "accuracy": 0.7691139773308079,
503
+ "precision": 0.4337244197283871,
504
+ "recall": 0.7691139773308079,
505
+ "f1": 0.5546605667700181,
506
+ "gt_support": 66522,
507
+ "predicted": 117962
508
  },
509
  "2": {
510
+ "accuracy": 0.0,
511
+ "precision": 0.0,
512
+ "recall": 0.0,
513
+ "f1": 0.0,
514
+ "gt_support": 810,
515
+ "predicted": 2
516
  },
517
  "3": {
518
+ "accuracy": 0.25155640891371384,
519
+ "precision": 0.5758907758907759,
520
+ "recall": 0.25155640891371384,
521
+ "f1": 0.35015894227459554,
522
+ "gt_support": 68748,
523
+ "predicted": 30030
524
  },
525
  "4": {
526
+ "accuracy": 0.5099426807760141,
527
+ "precision": 0.6799835376429433,
528
+ "recall": 0.5099426807760141,
529
+ "f1": 0.5828136613880596,
530
+ "gt_support": 45360,
531
+ "predicted": 34017
532
  }
533
  },
534
+ "total_samples": 182250
535
  },
536
+ "15": {
537
+ "accuracy": 0.5418395573997233,
538
+ "baseline_accuracy": 0.5240894421392347,
539
  "per_class": {
540
  "0": {
541
+ "accuracy": 0.0,
542
+ "precision": 0.0,
543
+ "recall": 0.0,
544
+ "f1": 0.0,
545
+ "gt_support": 964,
546
+ "predicted": 0
547
  },
548
  "1": {
549
+ "accuracy": 0.8854260244627186,
550
+ "precision": 0.5064283899381723,
551
+ "recall": 0.8854260244627186,
552
+ "f1": 0.644327267763881,
553
+ "gt_support": 101297,
554
+ "predicted": 177105
555
  },
556
  "2": {
557
+ "accuracy": 0.0,
558
+ "precision": 0.0,
559
+ "recall": 0.0,
560
+ "f1": 0.0,
561
+ "gt_support": 964,
562
+ "predicted": 0
563
  },
564
  "3": {
565
+ "accuracy": 0.2448559489773477,
566
+ "precision": 0.699434602336977,
567
+ "recall": 0.2448559489773477,
568
+ "f1": 0.36272887209226556,
569
+ "gt_support": 113675,
570
+ "predicted": 39795
571
  },
572
  "4": {
573
+ "accuracy": 0.0,
574
+ "precision": 0.0,
575
+ "recall": 0.0,
576
+ "f1": 0.0,
577
+ "gt_support": 0,
578
+ "predicted": 0
579
  }
580
  },
581
+ "total_samples": 216900
582
  }
583
  },
584
  "by_size_complexity": {
585
  "7_0.0": {
586
+ "accuracy": 0.7502923976608187,
587
  "baseline_accuracy": 0.7822222222222223,
588
  "per_class": {
589
  "0": {
590
  "accuracy": 0.07894736842105263,
591
+ "precision": 0.018867924528301886,
592
  "recall": 0.07894736842105263,
593
+ "f1": 0.030456852791878174,
594
  "gt_support": 38,
595
+ "predicted": 159
596
  },
597
  "1": {
598
+ "accuracy": 0.0,
599
+ "precision": 0.0,
600
+ "recall": 0.0,
601
+ "f1": 0.0,
602
  "gt_support": 912,
603
+ "predicted": 0
604
  },
605
  "2": {
606
+ "accuracy": 0.9210526315789473,
607
+ "precision": 0.020710059171597635,
608
+ "recall": 0.9210526315789473,
609
+ "f1": 0.040509259259259266,
610
  "gt_support": 38,
611
+ "predicted": 1690
612
  },
613
  "3": {
614
+ "accuracy": 0.0,
615
+ "precision": 0.0,
616
+ "recall": 0.0,
617
+ "f1": 0.0,
618
  "gt_support": 874,
619
+ "predicted": 0
620
  },
621
  "4": {
622
+ "accuracy": 0.9534988038277512,
623
+ "precision": 0.9516490076108044,
624
+ "recall": 0.9534988038277512,
625
+ "f1": 0.9525730076928822,
626
  "gt_support": 6688,
627
+ "predicted": 6701
628
  }
629
  },
630
  "total_samples": 8550
631
  },
632
  "7_0.2": {
633
+ "accuracy": 0.7519191919191919,
634
  "baseline_accuracy": 0.7822222222222223,
635
  "per_class": {
636
  "0": {
637
+ "accuracy": 0.3181818181818182,
638
+ "precision": 0.019971469329529243,
639
+ "recall": 0.3181818181818182,
640
+ "f1": 0.03758389261744966,
641
  "gt_support": 44,
642
+ "predicted": 701
643
  },
644
  "1": {
645
  "accuracy": 0.0,
 
650
  "predicted": 0
651
  },
652
  "2": {
653
+ "accuracy": 0.6818181818181818,
654
+ "precision": 0.02127659574468085,
655
+ "recall": 0.6818181818181818,
656
+ "f1": 0.04126547455295736,
657
+ "gt_support": 44,
658
+ "predicted": 1410
659
+ },
660
+ "3": {
661
  "accuracy": 0.0,
662
  "precision": 0.0,
663
  "recall": 0.0,
664
  "f1": 0.0,
 
 
 
 
 
 
 
 
665
  "gt_support": 924,
666
+ "predicted": 0
667
  },
668
  "4": {
669
+ "accuracy": 0.9555785123966942,
670
+ "precision": 0.9500577737835408,
671
+ "recall": 0.9555785123966942,
672
+ "f1": 0.9528101461404751,
673
  "gt_support": 7744,
674
+ "predicted": 7789
675
  }
676
  },
677
  "total_samples": 9900
678
  },
679
  "7_0.4": {
680
+ "accuracy": 0.7538211382113821,
681
  "baseline_accuracy": 0.7822222222222223,
682
  "per_class": {
683
  "0": {
684
+ "accuracy": 0.21951219512195122,
685
+ "precision": 0.01882845188284519,
686
+ "recall": 0.21951219512195122,
687
+ "f1": 0.034682080924855495,
688
  "gt_support": 41,
689
+ "predicted": 478
690
  },
691
  "1": {
692
+ "accuracy": 0.0,
693
+ "precision": 0.0,
694
+ "recall": 0.0,
695
+ "f1": 0.0,
696
  "gt_support": 1148,
697
+ "predicted": 0
698
  },
699
  "2": {
700
+ "accuracy": 0.7317073170731707,
701
+ "precision": 0.02074688796680498,
702
+ "recall": 0.7317073170731707,
703
+ "f1": 0.04034969737726967,
704
  "gt_support": 41,
705
+ "predicted": 1446
706
  },
707
  "3": {
708
+ "accuracy": 0.0,
709
+ "precision": 0.0,
710
+ "recall": 0.0,
711
+ "f1": 0.0,
712
  "gt_support": 779,
713
+ "predicted": 0
714
  },
715
  "4": {
716
+ "accuracy": 0.9582871396895787,
717
+ "precision": 0.9471305300643748,
718
+ "recall": 0.9582871396895787,
719
+ "f1": 0.9526761727629675,
720
  "gt_support": 7216,
721
+ "predicted": 7301
722
  }
723
  },
724
  "total_samples": 9225
725
  },
726
  "7_0.6": {
727
+ "accuracy": 0.7540425531914894,
728
  "baseline_accuracy": 0.7822222222222223,
729
  "per_class": {
730
  "0": {
731
+ "accuracy": 0.19148936170212766,
732
+ "precision": 0.020689655172413793,
733
+ "recall": 0.19148936170212766,
734
+ "f1": 0.03734439834024897,
735
  "gt_support": 47,
736
+ "predicted": 435
737
  },
738
  "1": {
739
  "accuracy": 0.0,
 
744
  "predicted": 0
745
  },
746
  "2": {
747
+ "accuracy": 0.7872340425531915,
748
+ "precision": 0.020856820744081173,
749
+ "recall": 0.7872340425531915,
750
+ "f1": 0.040637012630422846,
751
+ "gt_support": 47,
752
+ "predicted": 1774
753
+ },
754
+ "3": {
755
  "accuracy": 0.0,
756
  "precision": 0.0,
757
  "recall": 0.0,
758
  "f1": 0.0,
 
 
 
 
 
 
 
 
759
  "gt_support": 846,
760
+ "predicted": 0
761
  },
762
  "4": {
763
+ "accuracy": 0.9584139264990329,
764
+ "precision": 0.947645230695673,
765
+ "recall": 0.9584139264990329,
766
+ "f1": 0.9529991585527107,
767
  "gt_support": 8272,
768
+ "predicted": 8366
769
  }
770
  },
771
  "total_samples": 10575
772
  },
773
  "7_0.8": {
774
+ "accuracy": 0.7545858585858586,
775
  "baseline_accuracy": 0.7822222222222223,
776
  "per_class": {
777
  "0": {
778
+ "accuracy": 0.10909090909090909,
779
+ "precision": 0.01948051948051948,
780
+ "recall": 0.10909090909090909,
781
+ "f1": 0.03305785123966942,
782
  "gt_support": 55,
783
+ "predicted": 308
784
  },
785
  "1": {
786
+ "accuracy": 0.0,
787
+ "precision": 0.0,
788
+ "recall": 0.0,
789
+ "f1": 0.0,
790
  "gt_support": 1705,
791
+ "predicted": 0
792
  },
793
  "2": {
794
+ "accuracy": 0.8909090909090909,
795
+ "precision": 0.021595416483032173,
796
+ "recall": 0.8909090909090909,
797
+ "f1": 0.04216867469879518,
798
  "gt_support": 55,
799
+ "predicted": 2269
800
  },
801
  "3": {
802
+ "accuracy": 0.0,
803
+ "precision": 0.0,
804
+ "recall": 0.0,
805
+ "f1": 0.0,
806
  "gt_support": 880,
807
+ "predicted": 0
808
  },
809
  "4": {
810
+ "accuracy": 0.9589876033057851,
811
+ "precision": 0.9474382527046336,
812
+ "recall": 0.9589876033057851,
813
+ "f1": 0.953177944347469,
814
  "gt_support": 9680,
815
+ "predicted": 9798
816
  }
817
  },
818
  "total_samples": 12375
819
  },
820
  "7_1.0": {
821
+ "accuracy": 0.7554098360655738,
822
  "baseline_accuracy": 0.7822222222222223,
823
  "per_class": {
824
  "0": {
825
+ "accuracy": 0.13114754098360656,
826
+ "precision": 0.02168021680216802,
827
+ "recall": 0.13114754098360656,
828
+ "f1": 0.037209302325581395,
829
  "gt_support": 61,
830
+ "predicted": 369
831
  },
832
  "1": {
833
+ "accuracy": 0.0,
834
+ "precision": 0.0,
835
+ "recall": 0.0,
836
+ "f1": 0.0,
837
  "gt_support": 1952,
838
+ "predicted": 0
839
  },
840
  "2": {
841
+ "accuracy": 0.8852459016393442,
842
+ "precision": 0.021915584415584416,
843
+ "recall": 0.8852459016393442,
844
+ "f1": 0.042772277227722776,
845
  "gt_support": 61,
846
+ "predicted": 2464
847
  },
848
  "3": {
849
+ "accuracy": 0.0,
850
+ "precision": 0.0,
851
+ "recall": 0.0,
852
+ "f1": 0.0,
853
  "gt_support": 915,
854
+ "predicted": 0
855
  },
856
  "4": {
857
+ "accuracy": 0.9599478390461997,
858
+ "precision": 0.9461990451707676,
859
+ "recall": 0.9599478390461997,
860
+ "f1": 0.9530238579619013,
861
  "gt_support": 10736,
862
+ "predicted": 10892
863
  }
864
  },
865
  "total_samples": 13725
866
  },
867
  "9_0.0": {
868
+ "accuracy": 0.5818930041152264,
869
  "baseline_accuracy": 0.64,
870
  "per_class": {
871
  "0": {
872
+ "accuracy": 0.3148148148148148,
873
+ "precision": 0.013229571984435798,
874
+ "recall": 0.3148148148148148,
875
+ "f1": 0.025392083644510826,
876
  "gt_support": 54,
877
+ "predicted": 1285
878
  },
879
  "1": {
880
+ "accuracy": 0.0,
881
+ "precision": 0.0,
882
+ "recall": 0.0,
883
+ "f1": 0.0,
884
  "gt_support": 1728,
885
+ "predicted": 0
886
  },
887
  "2": {
888
+ "accuracy": 0.6851851851851852,
889
+ "precision": 0.01198574667962423,
890
+ "recall": 0.6851851851851852,
891
+ "f1": 0.023559375994906078,
892
  "gt_support": 54,
893
+ "predicted": 3087
894
  },
895
  "3": {
896
+ "accuracy": 0.01615445232466509,
897
+ "precision": 0.5256410256410257,
898
+ "recall": 0.01615445232466509,
899
+ "f1": 0.03134556574923547,
900
  "gt_support": 2538,
901
+ "predicted": 78
902
  },
903
  "4": {
904
+ "accuracy": 0.8969907407407407,
905
+ "precision": 0.9058441558441559,
906
+ "recall": 0.8969907407407407,
907
+ "f1": 0.9013957094856553,
908
  "gt_support": 7776,
909
+ "predicted": 7700
910
  }
911
  },
912
  "total_samples": 12150
913
  },
914
  "9_0.2": {
915
+ "accuracy": 0.5888416075650118,
916
  "baseline_accuracy": 0.64,
917
  "per_class": {
918
  "0": {
919
+ "accuracy": 0.0425531914893617,
920
+ "precision": 0.013888888888888888,
921
+ "recall": 0.0425531914893617,
922
+ "f1": 0.020942408376963352,
923
  "gt_support": 47,
924
+ "predicted": 144
925
  },
926
  "1": {
927
+ "accuracy": 0.0,
928
+ "precision": 0.0,
929
+ "recall": 0.0,
930
+ "f1": 0.0,
931
  "gt_support": 1692,
932
+ "predicted": 0
933
  },
934
  "2": {
935
+ "accuracy": 0.9574468085106383,
936
+ "precision": 0.012704686617730096,
937
+ "recall": 0.9574468085106383,
938
+ "f1": 0.025076623014767346,
939
  "gt_support": 47,
940
+ "predicted": 3542
941
  },
942
  "3": {
943
+ "accuracy": 0.0,
944
+ "precision": 0.0,
945
+ "recall": 0.0,
946
+ "f1": 0.0,
947
  "gt_support": 2021,
948
+ "predicted": 0
949
  },
950
  "4": {
951
+ "accuracy": 0.9131205673758865,
952
+ "precision": 0.8970823051241109,
953
+ "recall": 0.9131205673758865,
954
+ "f1": 0.905030387347148,
955
  "gt_support": 6768,
956
+ "predicted": 6889
957
  }
958
  },
959
  "total_samples": 10575
960
  },
961
  "9_0.4": {
962
+ "accuracy": 0.5857379767827529,
963
  "baseline_accuracy": 0.64,
964
  "per_class": {
965
  "0": {
966
+ "accuracy": 0.07462686567164178,
967
+ "precision": 0.015060240963855422,
968
+ "recall": 0.07462686567164178,
969
+ "f1": 0.025062656641604012,
970
  "gt_support": 67,
971
+ "predicted": 332
972
  },
973
  "1": {
974
+ "accuracy": 0.0,
975
+ "precision": 0.0,
976
+ "recall": 0.0,
977
+ "f1": 0.0,
978
  "gt_support": 2680,
979
+ "predicted": 0
980
  },
981
  "2": {
982
+ "accuracy": 0.9104477611940298,
983
+ "precision": 0.012295908083047773,
984
+ "recall": 0.9104477611940298,
985
+ "f1": 0.02426412092283214,
986
  "gt_support": 67,
987
+ "predicted": 4961
988
  },
989
  "3": {
990
+ "accuracy": 0.0,
991
+ "precision": 0.0,
992
+ "recall": 0.0,
993
+ "f1": 0.0,
994
  "gt_support": 2613,
995
+ "predicted": 0
996
  },
997
  "4": {
998
+ "accuracy": 0.9083747927031509,
999
+ "precision": 0.8959313023921488,
1000
+ "recall": 0.9083747927031509,
1001
+ "f1": 0.9021101389603705,
1002
  "gt_support": 9648,
1003
+ "predicted": 9782
1004
  }
1005
  },
1006
  "total_samples": 15075
1007
  },
1008
  "9_0.6": {
1009
+ "accuracy": 0.5865209471766849,
1010
  "baseline_accuracy": 0.64,
1011
  "per_class": {
1012
  "0": {
1013
+ "accuracy": 0.09836065573770492,
1014
+ "precision": 0.013015184381778741,
1015
+ "recall": 0.09836065573770492,
1016
+ "f1": 0.022988505747126436,
1017
  "gt_support": 61,
1018
+ "predicted": 461
1019
  },
1020
  "1": {
1021
  "accuracy": 0.0,
 
1026
  "predicted": 0
1027
  },
1028
  "2": {
1029
+ "accuracy": 0.9016393442622951,
1030
+ "precision": 0.012672811059907835,
1031
+ "recall": 0.9016393442622951,
1032
+ "f1": 0.024994319472847082,
1033
  "gt_support": 61,
1034
+ "predicted": 4340
1035
  },
1036
  "3": {
1037
+ "accuracy": 0.0,
1038
+ "precision": 0.0,
1039
+ "recall": 0.0,
1040
+ "f1": 0.0,
1041
  "gt_support": 2196,
1042
+ "predicted": 0
1043
  },
1044
  "4": {
1045
+ "accuracy": 0.9094945355191257,
1046
+ "precision": 0.8952263558942178,
1047
+ "recall": 0.9094945355191257,
1048
+ "f1": 0.9023040433702281,
1049
  "gt_support": 8784,
1050
+ "predicted": 8924
1051
  }
1052
  },
1053
  "total_samples": 13725
1054
  },
1055
  "9_0.8": {
1056
+ "accuracy": 0.6004840484048405,
1057
  "baseline_accuracy": 0.64,
1058
  "per_class": {
1059
  "0": {
1060
+ "accuracy": 0.18811881188118812,
1061
+ "precision": 0.014878621769772905,
1062
+ "recall": 0.18811881188118812,
1063
+ "f1": 0.027576197387518143,
1064
  "gt_support": 101,
1065
+ "predicted": 1277
1066
  },
1067
  "1": {
1068
+ "accuracy": 0.0,
1069
+ "precision": 0.0,
1070
+ "recall": 0.0,
1071
+ "f1": 0.0,
1072
  "gt_support": 4747,
1073
+ "predicted": 0
1074
  },
1075
  "2": {
1076
+ "accuracy": 0.7722772277227723,
1077
+ "precision": 0.012701514411333659,
1078
+ "recall": 0.7722772277227723,
1079
+ "f1": 0.024991989746876005,
1080
  "gt_support": 101,
1081
+ "predicted": 6141
1082
  },
1083
  "3": {
1084
+ "accuracy": 0.0,
1085
+ "precision": 0.0,
1086
+ "recall": 0.0,
1087
+ "f1": 0.0,
1088
  "gt_support": 3232,
1089
+ "predicted": 0
1090
  },
1091
  "4": {
1092
+ "accuracy": 0.9315869086908691,
1093
+ "precision": 0.8851505846998106,
1094
+ "recall": 0.9315869086908691,
1095
+ "f1": 0.9077752839100868,
1096
  "gt_support": 14544,
1097
+ "predicted": 15307
1098
  }
1099
  },
1100
  "total_samples": 22725
1101
  },
1102
  "9_1.0": {
1103
+ "accuracy": 0.5972222222222222,
1104
  "baseline_accuracy": 0.64,
1105
  "per_class": {
1106
  "0": {
1107
+ "accuracy": 0.08152173913043478,
1108
+ "precision": 0.014548981571290009,
1109
+ "recall": 0.08152173913043478,
1110
+ "f1": 0.02469135802469136,
1111
  "gt_support": 184,
1112
+ "predicted": 1031
1113
  },
1114
  "1": {
1115
+ "accuracy": 0.0,
1116
+ "precision": 0.0,
1117
+ "recall": 0.0,
1118
+ "f1": 0.0,
1119
  "gt_support": 9200,
1120
+ "predicted": 0
1121
  },
1122
  "2": {
1123
+ "accuracy": 0.8532608695652174,
1124
+ "precision": 0.012350534927627438,
1125
+ "recall": 0.8532608695652174,
1126
+ "f1": 0.02434863523573201,
1127
  "gt_support": 184,
1128
+ "predicted": 12712
1129
  },
1130
  "3": {
1131
+ "accuracy": 0.0,
1132
+ "precision": 0.0,
1133
+ "recall": 0.0,
1134
+ "f1": 0.0,
1135
  "gt_support": 5336,
1136
+ "predicted": 0
1137
  },
1138
  "4": {
1139
+ "accuracy": 0.9266681763285024,
1140
+ "precision": 0.8877680153306577,
1141
+ "recall": 0.9266681763285024,
1142
+ "f1": 0.9068011005853784,
1143
  "gt_support": 26496,
1144
+ "predicted": 27657
1145
  }
1146
  },
1147
  "total_samples": 41400
1148
  },
1149
  "11_0.0": {
1150
+ "accuracy": 0.5342483660130719,
1151
  "baseline_accuracy": 0.4622222222222222,
1152
  "per_class": {
1153
  "0": {
1154
+ "accuracy": 0.38235294117647056,
1155
+ "precision": 0.00760233918128655,
1156
+ "recall": 0.38235294117647056,
1157
+ "f1": 0.014908256880733944,
1158
  "gt_support": 68,
1159
+ "predicted": 3420
1160
  },
1161
  "1": {
1162
+ "accuracy": 0.0,
1163
+ "precision": 0.0,
1164
+ "recall": 0.0,
1165
+ "f1": 0.0,
1166
  "gt_support": 2720,
1167
+ "predicted": 0
1168
  },
1169
  "2": {
1170
+ "accuracy": 0.029411764705882353,
1171
+ "precision": 0.007575757575757576,
1172
+ "recall": 0.029411764705882353,
1173
+ "f1": 0.012048192771084338,
1174
  "gt_support": 68,
1175
+ "predicted": 264
1176
  },
1177
  "3": {
1178
+ "accuracy": 0.5163812360387193,
1179
+ "precision": 0.5475720489538097,
1180
+ "recall": 0.5163812360387193,
1181
+ "f1": 0.5315194481701475,
1182
  "gt_support": 5372,
1183
+ "predicted": 5066
1184
  },
1185
  "4": {
1186
+ "accuracy": 0.7596153846153846,
1187
+ "precision": 0.8201526717557251,
1188
+ "recall": 0.7596153846153846,
1189
+ "f1": 0.7887241227426222,
1190
  "gt_support": 7072,
1191
+ "predicted": 6550
1192
  }
1193
  },
1194
  "total_samples": 15300
1195
  },
1196
  "11_0.2": {
1197
+ "accuracy": 0.4102483660130719,
1198
  "baseline_accuracy": 0.4622222222222222,
1199
  "per_class": {
1200
  "0": {
1201
+ "accuracy": 0.7764705882352941,
1202
+ "precision": 0.007875894988066826,
1203
+ "recall": 0.7764705882352941,
1204
+ "f1": 0.015593620791494388,
1205
  "gt_support": 85,
1206
+ "predicted": 8380
1207
  },
1208
  "1": {
1209
+ "accuracy": 0.00951188986232791,
1210
+ "precision": 0.2835820895522388,
1211
+ "recall": 0.00951188986232791,
1212
+ "f1": 0.01840639379995156,
1213
  "gt_support": 3995,
1214
+ "predicted": 134
1215
  },
1216
  "2": {
1217
+ "accuracy": 0.03529411764705882,
1218
+ "precision": 0.006864988558352402,
1219
+ "recall": 0.03529411764705882,
1220
+ "f1": 0.011494252873563218,
1221
  "gt_support": 85,
1222
+ "predicted": 437
1223
  },
1224
  "3": {
1225
+ "accuracy": 0.140359477124183,
1226
+ "precision": 0.5038123167155425,
1227
+ "recall": 0.140359477124183,
1228
+ "f1": 0.21955271565495205,
1229
  "gt_support": 6120,
1230
+ "predicted": 1705
1231
  },
1232
  "4": {
1233
+ "accuracy": 0.7782805429864253,
1234
+ "precision": 0.8123745424489314,
1235
+ "recall": 0.7782805429864253,
1236
+ "f1": 0.7949621584146975,
1237
  "gt_support": 8840,
1238
+ "predicted": 8469
1239
  }
1240
  },
1241
  "total_samples": 19125
1242
  },
1243
  "11_0.4": {
1244
+ "accuracy": 0.4176161616161616,
1245
  "baseline_accuracy": 0.4622222222222222,
1246
  "per_class": {
1247
  "0": {
1248
+ "accuracy": 0.7818181818181819,
1249
+ "precision": 0.00826923076923077,
1250
+ "recall": 0.7818181818181819,
1251
+ "f1": 0.01636536631779258,
1252
  "gt_support": 55,
1253
+ "predicted": 5200
1254
  },
1255
  "1": {
1256
+ "accuracy": 0.08610634648370498,
1257
+ "precision": 0.3189326556543837,
1258
+ "recall": 0.08610634648370498,
1259
+ "f1": 0.13560237709346298,
1260
  "gt_support": 2915,
1261
+ "predicted": 787
1262
  },
1263
  "2": {
1264
+ "accuracy": 0.0,
1265
+ "precision": 0.0,
1266
+ "recall": 0.0,
1267
+ "f1": 0.0,
1268
  "gt_support": 55,
1269
+ "predicted": 0
1270
  },
1271
  "3": {
1272
+ "accuracy": 0.10137741046831956,
1273
+ "precision": 0.45885286783042395,
1274
+ "recall": 0.10137741046831956,
1275
+ "f1": 0.16606498194945848,
1276
  "gt_support": 3630,
1277
+ "predicted": 802
1278
  },
1279
  "4": {
1280
+ "accuracy": 0.7877622377622377,
1281
+ "precision": 0.8066595059076263,
1282
+ "recall": 0.7877622377622377,
1283
+ "f1": 0.7970988855474969,
1284
  "gt_support": 5720,
1285
+ "predicted": 5586
1286
  }
1287
  },
1288
  "total_samples": 12375
1289
  },
1290
  "11_0.6": {
1291
+ "accuracy": 0.4853074433656958,
1292
  "baseline_accuracy": 0.4622222222222222,
1293
  "per_class": {
1294
  "0": {
1295
+ "accuracy": 0.4077669902912621,
1296
+ "precision": 0.00816326530612245,
1297
+ "recall": 0.4077669902912621,
1298
+ "f1": 0.01600609756097561,
1299
  "gt_support": 103,
1300
+ "predicted": 5145
1301
  },
1302
  "1": {
1303
+ "accuracy": 0.4854368932038835,
1304
+ "precision": 0.3817279552105866,
1305
+ "recall": 0.4854368932038835,
1306
+ "f1": 0.42738086758316124,
1307
  "gt_support": 6180,
1308
+ "predicted": 7859
1309
  },
1310
  "2": {
1311
+ "accuracy": 0.009708737864077669,
1312
+ "precision": 0.007462686567164179,
1313
+ "recall": 0.009708737864077669,
1314
+ "f1": 0.008438818565400843,
1315
  "gt_support": 103,
1316
+ "predicted": 134
1317
  },
1318
  "3": {
1319
+ "accuracy": 0.0,
1320
+ "precision": 0.0,
1321
+ "recall": 0.0,
1322
+ "f1": 0.0,
1323
  "gt_support": 6077,
1324
+ "predicted": 0
1325
  },
1326
  "4": {
1327
+ "accuracy": 0.7658700522778192,
1328
+ "precision": 0.8173757098734682,
1329
+ "recall": 0.7658700522778192,
1330
+ "f1": 0.7907850980770157,
1331
  "gt_support": 10712,
1332
+ "predicted": 10037
1333
  }
1334
  },
1335
  "total_samples": 23175
1336
  },
1337
  "11_0.8": {
1338
+ "accuracy": 0.5658089668615984,
1339
  "baseline_accuracy": 0.4622222222222222,
1340
  "per_class": {
1341
  "0": {
1342
+ "accuracy": 0.13157894736842105,
1343
+ "precision": 0.007716049382716049,
1344
+ "recall": 0.13157894736842105,
1345
+ "f1": 0.014577259475218658,
1346
  "gt_support": 114,
1347
+ "predicted": 1944
1348
  },
1349
  "1": {
1350
+ "accuracy": 0.6927166400850612,
1351
+ "precision": 0.42595619483491337,
1352
+ "recall": 0.6927166400850612,
1353
+ "f1": 0.5275303643724696,
1354
  "gt_support": 7524,
1355
+ "predicted": 12236
1356
  },
1357
  "2": {
1358
+ "accuracy": 0.0,
1359
+ "precision": 0.0,
1360
+ "recall": 0.0,
1361
+ "f1": 0.0,
1362
  "gt_support": 114,
1363
+ "predicted": 0
1364
  },
1365
  "3": {
1366
+ "accuracy": 0.0,
1367
+ "precision": 0.0,
1368
+ "recall": 0.0,
1369
+ "f1": 0.0,
1370
  "gt_support": 6042,
1371
+ "predicted": 0
1372
  },
1373
  "4": {
1374
+ "accuracy": 0.7832321187584346,
1375
+ "precision": 0.8095902353966871,
1376
+ "recall": 0.7832321187584346,
1377
+ "f1": 0.7961930892566236,
1378
  "gt_support": 11856,
1379
+ "predicted": 11470
1380
  }
1381
  },
1382
  "total_samples": 25650
1383
  },
1384
  "11_1.0": {
1385
+ "accuracy": 0.6150793650793651,
1386
  "baseline_accuracy": 0.4622222222222222,
1387
  "per_class": {
1388
  "0": {
1389
+ "accuracy": 0.013392857142857142,
1390
+ "precision": 0.007211538461538462,
1391
+ "recall": 0.013392857142857142,
1392
+ "f1": 0.009375,
1393
  "gt_support": 224,
1394
+ "predicted": 416
1395
  },
1396
  "1": {
1397
+ "accuracy": 0.8211185515873016,
1398
+ "precision": 0.46803322141721154,
1399
+ "recall": 0.8211185515873016,
1400
+ "f1": 0.5962226774418656,
1401
  "gt_support": 16128,
1402
+ "predicted": 28295
1403
  },
1404
  "2": {
1405
+ "accuracy": 0.0,
1406
+ "precision": 0.0,
1407
+ "recall": 0.0,
1408
+ "f1": 0.0,
1409
  "gt_support": 224,
1410
+ "predicted": 0
1411
  },
1412
  "3": {
1413
+ "accuracy": 0.0,
1414
+ "precision": 0.0,
1415
+ "recall": 0.0,
1416
+ "f1": 0.0,
1417
  "gt_support": 10528,
1418
+ "predicted": 0
1419
  },
1420
  "4": {
1421
+ "accuracy": 0.7621050824175825,
1422
+ "precision": 0.8185716261699479,
1423
+ "recall": 0.7621050824175825,
1424
+ "f1": 0.7893297765921975,
1425
  "gt_support": 23296,
1426
+ "predicted": 21689
1427
  }
1428
  },
1429
  "total_samples": 50400
1430
  },
1431
  "13_0.0": {
1432
+ "accuracy": 0.6334027777777778,
1433
  "baseline_accuracy": 0.5288888888888889,
1434
  "per_class": {
1435
  "0": {
 
1441
  "predicted": 0
1442
  },
1443
  "1": {
1444
+ "accuracy": 0.0016276041666666667,
1445
+ "precision": 0.8333333333333334,
1446
+ "recall": 0.0016276041666666667,
1447
+ "f1": 0.0032488628979857053,
1448
+ "gt_support": 3072,
1449
+ "predicted": 6
1450
+ },
1451
+ "2": {
1452
  "accuracy": 0.0,
1453
  "precision": 0.0,
1454
  "recall": 0.0,
1455
  "f1": 0.0,
 
 
 
 
 
 
 
 
1456
  "gt_support": 64,
1457
+ "predicted": 0
1458
  },
1459
  "3": {
1460
+ "accuracy": 0.9516806722689075,
1461
+ "precision": 0.6243431820139547,
1462
+ "recall": 0.9516806722689075,
1463
+ "f1": 0.7540182054616384,
1464
  "gt_support": 7616,
1465
+ "predicted": 11609
1466
  },
1467
  "4": {
1468
+ "accuracy": 0.5212053571428571,
1469
+ "precision": 0.6707360861759426,
1470
+ "recall": 0.5212053571428571,
1471
+ "f1": 0.5865913016172083,
1472
  "gt_support": 3584,
1473
+ "predicted": 2785
1474
  }
1475
  },
1476
  "total_samples": 14400
1477
  },
1478
  "13_0.2": {
1479
+ "accuracy": 0.5454480286738351,
1480
  "baseline_accuracy": 0.48444444444444446,
1481
  "per_class": {
1482
  "0": {
1483
+ "accuracy": 0.0,
1484
+ "precision": 0.0,
1485
+ "recall": 0.0,
1486
+ "f1": 0.0,
1487
  "gt_support": 62,
1488
+ "predicted": 0
1489
  },
1490
  "1": {
1491
+ "accuracy": 0.160734149054505,
1492
+ "precision": 0.272,
1493
+ "recall": 0.160734149054505,
1494
+ "f1": 0.20206257647264467,
1495
  "gt_support": 3596,
1496
+ "predicted": 2125
1497
  },
1498
  "2": {
1499
+ "accuracy": 0.0,
1500
+ "precision": 0.0,
1501
+ "recall": 0.0,
1502
+ "f1": 0.0,
1503
  "gt_support": 62,
1504
+ "predicted": 0
1505
  },
1506
  "3": {
1507
+ "accuracy": 0.7549570878958272,
1508
+ "precision": 0.5761065943992774,
1509
+ "recall": 0.7549570878958272,
1510
+ "f1": 0.6535160753170232,
1511
  "gt_support": 6758,
1512
+ "predicted": 8856
1513
  },
1514
  "4": {
1515
+ "accuracy": 0.5555875576036866,
1516
+ "precision": 0.6497137083192994,
1517
+ "recall": 0.5555875576036866,
1518
+ "f1": 0.5989753143921751,
1519
  "gt_support": 3472,
1520
+ "predicted": 2969
1521
  }
1522
  },
1523
  "total_samples": 13950
1524
  },
1525
  "13_0.4": {
1526
+ "accuracy": 0.4542933333333333,
1527
  "baseline_accuracy": 0.44,
1528
  "per_class": {
1529
  "0": {
1530
+ "accuracy": 0.008,
1531
+ "precision": 0.0041841004184100415,
1532
+ "recall": 0.008,
1533
+ "f1": 0.005494505494505494,
1534
  "gt_support": 125,
1535
+ "predicted": 239
1536
  },
1537
  "1": {
1538
+ "accuracy": 0.5089411764705882,
1539
+ "precision": 0.328398997950353,
1540
+ "recall": 0.5089411764705882,
1541
+ "f1": 0.3992063858256817,
1542
  "gt_support": 8500,
1543
+ "predicted": 13173
1544
  },
1545
  "2": {
1546
+ "accuracy": 0.0,
1547
+ "precision": 0.0,
1548
+ "recall": 0.0,
1549
+ "f1": 0.0,
1550
  "gt_support": 125,
1551
+ "predicted": 2
1552
  },
1553
  "3": {
1554
+ "accuracy": 0.3654141414141414,
1555
+ "precision": 0.5238646895273401,
1556
+ "recall": 0.3654141414141414,
1557
+ "f1": 0.43052315894701765,
1558
  "gt_support": 12375,
1559
+ "predicted": 8632
1560
  },
1561
  "4": {
1562
+ "accuracy": 0.5611428571428572,
1563
+ "precision": 0.6461589077150847,
1564
+ "recall": 0.5611428571428572,
1565
+ "f1": 0.600657542625583,
1566
  "gt_support": 7000,
1567
+ "predicted": 6079
1568
  }
1569
  },
1570
  "total_samples": 28125
1571
  },
1572
  "13_0.6": {
1573
+ "accuracy": 0.4377961432506887,
1574
  "baseline_accuracy": 0.39555555555555555,
1575
  "per_class": {
1576
  "0": {
1577
+ "accuracy": 0.0,
1578
+ "precision": 0.0,
1579
+ "recall": 0.0,
1580
+ "f1": 0.0,
1581
  "gt_support": 121,
1582
+ "predicted": 0
1583
  },
1584
  "1": {
1585
+ "accuracy": 0.8402203856749312,
1586
+ "precision": 0.3782133829350885,
1587
+ "recall": 0.8402203856749312,
1588
+ "f1": 0.5216247327742147,
1589
  "gt_support": 9438,
1590
+ "predicted": 20967
1591
  },
1592
  "2": {
1593
+ "accuracy": 0.0,
1594
+ "precision": 0.0,
1595
+ "recall": 0.0,
1596
+ "f1": 0.0,
1597
  "gt_support": 121,
1598
+ "predicted": 0
1599
  },
1600
  "3": {
1601
+ "accuracy": 0.039186553997585666,
1602
+ "precision": 0.45230439442658094,
1603
+ "recall": 0.039186553997585666,
1604
+ "f1": 0.07212442317552556,
1605
  "gt_support": 10769,
1606
+ "predicted": 933
1607
  },
1608
  "4": {
1609
+ "accuracy": 0.5264167650531287,
1610
+ "precision": 0.6698591549295775,
1611
+ "recall": 0.5264167650531287,
1612
+ "f1": 0.5895380547062227,
1613
  "gt_support": 6776,
1614
+ "predicted": 5325
1615
  }
1616
  },
1617
  "total_samples": 27225
1618
  },
1619
  "13_0.8": {
1620
+ "accuracy": 0.4793650793650794,
1621
  "baseline_accuracy": 0.39555555555555555,
1622
  "per_class": {
1623
  "0": {
1624
+ "accuracy": 0.0,
1625
+ "precision": 0.0,
1626
+ "recall": 0.0,
1627
+ "f1": 0.0,
1628
  "gt_support": 112,
1629
+ "predicted": 0
1630
  },
1631
  "1": {
1632
+ "accuracy": 0.9063001605136437,
1633
+ "precision": 0.4339305442144195,
1634
+ "recall": 0.9063001605136437,
1635
+ "f1": 0.5868710819501738,
1636
  "gt_support": 9968,
1637
+ "predicted": 20819
1638
  },
1639
  "2": {
1640
+ "accuracy": 0.0,
1641
+ "precision": 0.0,
1642
+ "recall": 0.0,
1643
+ "f1": 0.0,
1644
  "gt_support": 112,
1645
+ "predicted": 0
1646
  },
1647
  "3": {
1648
+ "accuracy": 0.0,
1649
+ "precision": 0.0,
1650
+ "recall": 0.0,
1651
+ "f1": 0.0,
1652
  "gt_support": 8736,
1653
+ "predicted": 0
1654
  },
1655
  "4": {
1656
+ "accuracy": 0.4856505102040816,
1657
+ "precision": 0.6952750513581374,
1658
+ "recall": 0.4856505102040816,
1659
+ "f1": 0.5718576926687319,
1660
  "gt_support": 6272,
1661
+ "predicted": 4381
1662
  }
1663
  },
1664
  "total_samples": 25200
1665
  },
1666
  "13_1.0": {
1667
+ "accuracy": 0.5191956373551465,
1668
  "baseline_accuracy": 0.43555555555555553,
1669
  "per_class": {
1670
  "0": {
1671
+ "accuracy": 0.0,
1672
+ "precision": 0.0,
1673
+ "recall": 0.0,
1674
+ "f1": 0.0,
1675
  "gt_support": 326,
1676
+ "predicted": 0
1677
  },
1678
  "1": {
1679
+ "accuracy": 0.9168023037435833,
1680
+ "precision": 0.48117361019844923,
1681
+ "recall": 0.9168023037435833,
1682
+ "f1": 0.6311139840551606,
1683
  "gt_support": 31948,
1684
+ "predicted": 60872
1685
  },
1686
  "2": {
1687
+ "accuracy": 0.0,
1688
+ "precision": 0.0,
1689
+ "recall": 0.0,
1690
+ "f1": 0.0,
1691
  "gt_support": 326,
1692
+ "predicted": 0
1693
  },
1694
  "3": {
1695
+ "accuracy": 0.0,
1696
+ "precision": 0.0,
1697
+ "recall": 0.0,
1698
+ "f1": 0.0,
1699
  "gt_support": 22494,
1700
+ "predicted": 0
1701
  },
1702
  "4": {
1703
+ "accuracy": 0.4816498685363716,
1704
+ "precision": 0.7046802372175028,
1705
+ "recall": 0.4816498685363716,
1706
+ "f1": 0.5722001691937268,
1707
  "gt_support": 18256,
1708
+ "predicted": 12478
1709
  }
1710
  },
1711
  "total_samples": 73350
1712
  },
1713
  "15_0.0": {
1714
+ "accuracy": 0.7422222222222222,
1715
  "baseline_accuracy": 0.7422222222222222,
1716
  "per_class": {
1717
  "0": {
1718
+ "accuracy": 0.0,
1719
+ "precision": 0.0,
1720
+ "recall": 0.0,
1721
+ "f1": 0.0,
1722
  "gt_support": 86,
1723
+ "predicted": 0
1724
  },
1725
  "1": {
1726
+ "accuracy": 0.0,
1727
+ "precision": 0.0,
1728
+ "recall": 0.0,
1729
+ "f1": 0.0,
1730
  "gt_support": 4816,
1731
+ "predicted": 0
1732
  },
1733
  "2": {
1734
+ "accuracy": 0.0,
1735
+ "precision": 0.0,
1736
+ "recall": 0.0,
1737
+ "f1": 0.0,
1738
  "gt_support": 86,
1739
+ "predicted": 0
1740
  },
1741
  "3": {
1742
+ "accuracy": 1.0,
1743
+ "precision": 0.7422222222222222,
1744
+ "recall": 1.0,
1745
+ "f1": 0.8520408163265306,
1746
  "gt_support": 14362,
1747
+ "predicted": 19350
1748
  },
1749
  "4": {
1750
  "accuracy": 0.0,
 
1752
  "recall": 0.0,
1753
  "f1": 0.0,
1754
  "gt_support": 0,
1755
+ "predicted": 0
1756
  }
1757
  },
1758
  "total_samples": 19350
1759
  },
1760
  "15_0.2": {
1761
+ "accuracy": 0.5926947637292465,
1762
  "baseline_accuracy": 0.6755555555555556,
1763
  "per_class": {
1764
  "0": {
1765
+ "accuracy": 0.0,
1766
+ "precision": 0.0,
1767
+ "recall": 0.0,
1768
+ "f1": 0.0,
1769
  "gt_support": 87,
1770
+ "predicted": 0
1771
  },
1772
  "1": {
1773
+ "accuracy": 0.23166585721223895,
1774
+ "precision": 0.3163129973474801,
1775
+ "recall": 0.23166585721223895,
1776
+ "f1": 0.2674516400336417,
1777
  "gt_support": 6177,
1778
+ "predicted": 4524
1779
  },
1780
  "2": {
1781
+ "accuracy": 0.0,
1782
+ "precision": 0.0,
1783
+ "recall": 0.0,
1784
+ "f1": 0.0,
1785
  "gt_support": 87,
1786
+ "predicted": 0
1787
  },
1788
  "3": {
1789
+ "accuracy": 0.7691318814277072,
1790
+ "precision": 0.6757690518902398,
1791
+ "recall": 0.7691318814277072,
1792
+ "f1": 0.7194341290893015,
1793
  "gt_support": 13224,
1794
+ "predicted": 15051
1795
  },
1796
  "4": {
1797
  "accuracy": 0.0,
 
1799
  "recall": 0.0,
1800
  "f1": 0.0,
1801
  "gt_support": 0,
1802
+ "predicted": 0
1803
  }
1804
  },
1805
  "total_samples": 19575
1806
  },
1807
  "15_0.4": {
1808
+ "accuracy": 0.4188562091503268,
1809
  "baseline_accuracy": 0.6133333333333333,
1810
  "per_class": {
1811
  "0": {
1812
+ "accuracy": 0.0,
1813
+ "precision": 0.0,
1814
+ "recall": 0.0,
1815
+ "f1": 0.0,
1816
  "gt_support": 136,
1817
+ "predicted": 0
1818
  },
1819
  "1": {
1820
+ "accuracy": 0.8231833910034602,
1821
+ "precision": 0.37752915972387524,
1822
+ "recall": 0.8231833910034602,
1823
+ "f1": 0.5176521786433117,
1824
  "gt_support": 11560,
1825
+ "predicted": 25206
1826
  },
1827
  "2": {
1828
+ "accuracy": 0.0,
1829
+ "precision": 0.0,
1830
+ "recall": 0.0,
1831
+ "f1": 0.0,
1832
  "gt_support": 136,
1833
+ "predicted": 0
1834
  },
1835
  "3": {
1836
+ "accuracy": 0.175884484228474,
1837
+ "precision": 0.6119762699295513,
1838
+ "recall": 0.175884484228474,
1839
+ "f1": 0.2732389702839169,
1840
  "gt_support": 18768,
1841
+ "predicted": 5394
1842
  },
1843
  "4": {
1844
  "accuracy": 0.0,
 
1846
  "recall": 0.0,
1847
  "f1": 0.0,
1848
  "gt_support": 0,
1849
+ "predicted": 0
1850
  }
1851
  },
1852
  "total_samples": 30600
1853
  },
1854
  "15_0.6": {
1855
+ "accuracy": 0.4444444444444444,
1856
  "baseline_accuracy": 0.5466666666666666,
1857
  "per_class": {
1858
  "0": {
1859
+ "accuracy": 0.0,
1860
+ "precision": 0.0,
1861
+ "recall": 0.0,
1862
+ "f1": 0.0,
1863
  "gt_support": 100,
1864
+ "predicted": 0
1865
  },
1866
  "1": {
1867
+ "accuracy": 1.0,
1868
+ "precision": 0.4444444444444444,
1869
+ "recall": 1.0,
1870
+ "f1": 0.6153846153846153,
1871
  "gt_support": 10000,
1872
+ "predicted": 22500
1873
  },
1874
  "2": {
1875
+ "accuracy": 0.0,
1876
+ "precision": 0.0,
1877
+ "recall": 0.0,
1878
+ "f1": 0.0,
1879
  "gt_support": 100,
1880
+ "predicted": 0
1881
  },
1882
  "3": {
1883
+ "accuracy": 0.0,
1884
+ "precision": 0.0,
1885
+ "recall": 0.0,
1886
+ "f1": 0.0,
1887
  "gt_support": 12300,
1888
+ "predicted": 0
1889
  },
1890
  "4": {
1891
  "accuracy": 0.0,
 
1893
  "recall": 0.0,
1894
  "f1": 0.0,
1895
  "gt_support": 0,
1896
+ "predicted": 0
1897
  }
1898
  },
1899
  "total_samples": 22500
1900
  },
1901
  "15_0.8": {
1902
+ "accuracy": 0.5066666666666667,
1903
  "baseline_accuracy": 0.5066666666666667,
1904
  "per_class": {
1905
  "0": {
1906
+ "accuracy": 0.0,
1907
+ "precision": 0.0,
1908
+ "recall": 0.0,
1909
+ "f1": 0.0,
1910
  "gt_support": 164,
1911
+ "predicted": 0
1912
  },
1913
  "1": {
1914
+ "accuracy": 1.0,
1915
+ "precision": 0.5066666666666667,
1916
+ "recall": 1.0,
1917
+ "f1": 0.6725663716814159,
1918
  "gt_support": 18696,
1919
+ "predicted": 36900
1920
  },
1921
  "2": {
1922
+ "accuracy": 0.0,
1923
+ "precision": 0.0,
1924
+ "recall": 0.0,
1925
+ "f1": 0.0,
1926
  "gt_support": 164,
1927
+ "predicted": 0
1928
  },
1929
  "3": {
1930
+ "accuracy": 0.0,
1931
+ "precision": 0.0,
1932
+ "recall": 0.0,
1933
+ "f1": 0.0,
1934
  "gt_support": 17876,
1935
+ "predicted": 0
1936
  },
1937
  "4": {
1938
  "accuracy": 0.0,
 
1940
  "recall": 0.0,
1941
  "f1": 0.0,
1942
  "gt_support": 0,
1943
+ "predicted": 0
1944
  }
1945
  },
1946
  "total_samples": 36900
1947
  },
1948
  "15_1.0": {
1949
+ "accuracy": 0.5688888888888889,
1950
  "baseline_accuracy": 0.5688888888888889,
1951
  "per_class": {
1952
  "0": {
1953
+ "accuracy": 0.0,
1954
+ "precision": 0.0,
1955
+ "recall": 0.0,
1956
+ "f1": 0.0,
1957
  "gt_support": 391,
1958
+ "predicted": 0
1959
  },
1960
  "1": {
1961
+ "accuracy": 1.0,
1962
+ "precision": 0.5688888888888889,
1963
+ "recall": 1.0,
1964
+ "f1": 0.725212464589235,
1965
  "gt_support": 50048,
1966
+ "predicted": 87975
1967
  },
1968
  "2": {
1969
+ "accuracy": 0.0,
1970
+ "precision": 0.0,
1971
+ "recall": 0.0,
1972
+ "f1": 0.0,
1973
  "gt_support": 391,
1974
+ "predicted": 0
1975
  },
1976
  "3": {
1977
+ "accuracy": 0.0,
1978
+ "precision": 0.0,
1979
+ "recall": 0.0,
1980
+ "f1": 0.0,
1981
  "gt_support": 37145,
1982
+ "predicted": 0
1983
  },
1984
  "4": {
1985
  "accuracy": 0.0,
 
1987
  "recall": 0.0,
1988
  "f1": 0.0,
1989
  "gt_support": 0,
1990
+ "predicted": 0
1991
  }
1992
  },
1993
  "total_samples": 87975
1994
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
1995
  }
1996
  }
layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_lr_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
@@ -27,17 +27,17 @@ Processed 300 trajectories, 3223 steps
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
- Overall Accuracy: 0.3290 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
- A 0.0835 0.0045 0.0835 0.0085 3223 59785
37
- # 0.0812 0.3946 0.0812 0.1347 238175 49015
38
- G 0.1070 0.0048 0.1070 0.0093 3223 71182
39
- _ 0.6624 0.3527 0.6624 0.4603 243346 456997
40
- + 0.2423 0.6516 0.2423 0.3532 237208 88196
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
@@ -47,81 +47,81 @@ METRICS BY SIZE
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
- Overall Accuracy: 0.3172 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
- A 0.0909 0.0045 0.0909 0.0085 286 5802
57
- # 0.0179 0.1318 0.0179 0.0315 8224 1115
58
- G 0.0140 0.0061 0.0140 0.0085 286 654
59
- _ 0.8747 0.1114 0.8747 0.1976 5218 40981
60
- + 0.3113 0.9918 0.3113 0.4739 50336 15798
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
- Overall Accuracy: 0.2807 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
- A 0.0992 0.0045 0.0992 0.0085 514 11430
73
- # 0.0438 0.2111 0.0438 0.0726 22670 4703
74
- G 0.0214 0.0067 0.0214 0.0102 514 1637
75
- _ 0.8281 0.1844 0.8281 0.3016 17936 80555
76
- + 0.2237 0.9557 0.2237 0.3626 74016 17325
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
- Overall Accuracy: 0.2978 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
- A 0.0878 0.0045 0.0878 0.0086 649 12558
89
- # 0.0887 0.2995 0.0887 0.1368 39462 11683
90
- G 0.1217 0.0058 0.1217 0.0111 649 13594
91
- _ 0.6619 0.2791 0.6619 0.3926 37769 89570
92
- + 0.2200 0.7974 0.2200 0.3448 67496 18620
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
- Overall Accuracy: 0.3334 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
- A 0.0765 0.0045 0.0765 0.0086 810 13663
105
- # 0.0966 0.4140 0.0966 0.1566 66522 15521
106
- G 0.1580 0.0050 0.1580 0.0097 810 25537
107
- _ 0.6365 0.4018 0.6365 0.4926 68748 108908
108
- + 0.2291 0.5582 0.2291 0.3249 45360 18621
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
- Overall Accuracy: 0.3757 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
- A 0.0757 0.0045 0.0757 0.0084 964 16332
121
- # 0.0817 0.5175 0.0817 0.1411 101297 15993
122
- G 0.1276 0.0041 0.1276 0.0080 964 29760
123
- _ 0.6424 0.5331 0.6424 0.5826 113675 136983
124
- + 0.0000 0.0000 0.0000 0.0000 0 17832
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
@@ -131,97 +131,97 @@ METRICS BY COMPLEXITY
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
- Overall Accuracy: 0.4076 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
- A 0.0548 0.0045 0.0548 0.0083 310 3807
141
- # 0.0145 0.1716 0.0145 0.0267 13248 1119
142
- G 0.2323 0.0053 0.2323 0.0104 310 13601
143
- _ 0.6046 0.4977 0.6046 0.5460 30762 37371
144
- + 0.3800 0.6891 0.3800 0.4899 25120 13852
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
- Overall Accuracy: 0.4035 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
- A 0.0677 0.0045 0.0677 0.0084 325 4943
157
- # 0.0269 0.2514 0.0269 0.0486 16604 1778
158
- G 0.0954 0.0052 0.0954 0.0098 325 5978
159
- _ 0.7061 0.4312 0.7061 0.5354 29047 47572
160
- + 0.3166 0.6607 0.3166 0.4281 26824 12854
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
- Overall Accuracy: 0.3748 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
- A 0.1014 0.0044 0.1014 0.0085 424 9726
173
- # 0.0352 0.2924 0.0352 0.0628 26803 3225
174
- G 0.0825 0.0044 0.0825 0.0084 424 7909
175
- _ 0.7054 0.4352 0.7054 0.5383 38165 61859
176
- + 0.2640 0.6160 0.2640 0.3696 29584 12681
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
- Overall Accuracy: 0.3425 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
- A 0.0648 0.0045 0.0648 0.0084 432 6229
189
- # 0.0505 0.3523 0.0505 0.0883 29604 4241
190
- G 0.1088 0.0050 0.1088 0.0095 432 9462
191
- _ 0.6894 0.3479 0.6894 0.4624 32188 63779
192
- + 0.2761 0.7070 0.2761 0.3971 34544 13489
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
- Overall Accuracy: 0.3172 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
- A 0.0678 0.0045 0.0678 0.0084 546 8259
205
- # 0.0752 0.3948 0.0752 0.1263 42640 8118
206
- G 0.1062 0.0052 0.1062 0.0100 546 11101
207
- _ 0.6990 0.3147 0.6990 0.4340 36766 81660
208
- + 0.2354 0.7270 0.2354 0.3556 42352 13712
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
- Overall Accuracy: 0.2723 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
- A 0.1029 0.0045 0.1029 0.0087 1186 26821
221
- # 0.1195 0.4277 0.1195 0.1868 109276 30534
222
- G 0.0860 0.0044 0.0860 0.0084 1186 23131
223
- _ 0.6186 0.2869 0.6186 0.3920 76418 164756
224
- + 0.1538 0.5606 0.1538 0.2413 78784 21608
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
@@ -231,481 +231,481 @@ METRICS BY SIZE-COMPLEXITY COMBINATION
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
- Overall Accuracy: 0.4242 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
- A 0.0789 0.0044 0.0789 0.0084 38 675
241
- # 0.0263 0.1067 0.0263 0.0422 912 225
242
- G 0.0263 0.0044 0.0263 0.0076 38 225
243
- _ 0.8650 0.1657 0.8650 0.2781 874 4562
244
- + 0.4251 0.9930 0.4251 0.5953 6688 2863
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
- Overall Accuracy: 0.3134 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
- A 0.0682 0.0045 0.0682 0.0084 44 672
257
  # 0.0000 0.0000 0.0000 0.0000 1144 0
258
- G 0.0000 0.0000 0.0000 0.0000 44 0
259
- _ 0.9318 0.1232 0.9318 0.2177 924 6986
260
- + 0.2891 0.9987 0.2891 0.4484 7744 2242
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
- Overall Accuracy: 0.2990 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
- A 0.1951 0.0044 0.1951 0.0087 41 1800
273
- # 0.0244 0.1302 0.0244 0.0411 1148 215
274
- G 0.0244 0.0053 0.0244 0.0087 41 189
275
- _ 0.7548 0.1207 0.7548 0.2081 779 4871
276
- + 0.2956 0.9921 0.2956 0.4555 7216 2150
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
- Overall Accuracy: 0.3092 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
- A 0.0638 0.0044 0.0638 0.0083 47 675
289
  # 0.0000 0.0000 0.0000 0.0000 1363 0
290
- G 0.0000 0.0000 0.0000 0.0000 47 0
291
- _ 0.9350 0.1067 0.9350 0.1915 846 7413
292
- + 0.2993 0.9956 0.2993 0.4603 8272 2487
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
- Overall Accuracy: 0.2819 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
- A 0.0909 0.0046 0.0909 0.0088 55 1080
305
- # 0.0182 0.1378 0.0182 0.0321 1705 225
306
- G 0.0182 0.0044 0.0182 0.0071 55 225
307
- _ 0.8727 0.0941 0.8727 0.1699 880 8158
308
- + 0.2772 0.9985 0.2772 0.4339 9680 2687
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
- Overall Accuracy: 0.3034 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
- A 0.0656 0.0044 0.0656 0.0083 61 900
321
- # 0.0328 0.1422 0.0328 0.0533 1952 450
322
- G 0.0164 0.0667 0.0164 0.0263 61 15
323
- _ 0.8743 0.0890 0.8743 0.1615 915 8991
324
- + 0.3069 0.9780 0.3069 0.4672 10736 3369
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
- Overall Accuracy: 0.4013 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
- A 0.0926 0.0044 0.0926 0.0085 54 1125
337
- # 0.0185 0.1461 0.0185 0.0329 1728 219
338
- G 0.0370 0.0051 0.0370 0.0089 54 394
339
- _ 0.8266 0.2797 0.8266 0.4180 2538 7501
340
- + 0.3522 0.9409 0.3522 0.5126 7776 2911
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
- Overall Accuracy: 0.3806 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
- A 0.1277 0.0044 0.1277 0.0086 47 1350
353
- # 0.0213 0.1600 0.0213 0.0376 1692 225
354
- G 0.0213 0.0278 0.0213 0.0241 47 36
355
- _ 0.8135 0.2537 0.8135 0.3868 2021 6479
356
- + 0.3454 0.9408 0.3454 0.5053 6768 2485
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
- Overall Accuracy: 0.2918 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
- A 0.1343 0.0044 0.1343 0.0086 67 2025
369
- # 0.0149 0.1778 0.0149 0.0275 2680 225
370
- G 0.0299 0.0103 0.0299 0.0153 67 195
371
- _ 0.8197 0.2068 0.8197 0.3303 2613 10356
372
- + 0.2286 0.9701 0.2286 0.3701 9648 2274
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
- Overall Accuracy: 0.3500 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
- A 0.0656 0.0044 0.0656 0.0083 61 900
385
  # 0.0000 0.0000 0.0000 0.0000 2623 0
386
- G 0.0164 0.0048 0.0164 0.0074 61 210
387
- _ 0.9062 0.2052 0.9062 0.3346 2196 9699
388
- + 0.3198 0.9633 0.3198 0.4802 8784 2916
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
- Overall Accuracy: 0.2597 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
- A 0.0792 0.0044 0.0792 0.0084 101 1800
401
- # 0.0495 0.2089 0.0495 0.0800 4747 1125
402
- G 0.0396 0.0055 0.0396 0.0097 101 724
403
- _ 0.8147 0.1655 0.8147 0.2752 3232 15905
404
- + 0.2078 0.9530 0.2078 0.3412 14544 3171
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
- Overall Accuracy: 0.2043 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
- A 0.1033 0.0045 0.1033 0.0086 184 4230
417
- # 0.0707 0.2234 0.0707 0.1074 9200 2909
418
- G 0.0054 0.0128 0.0054 0.0076 184 78
419
- _ 0.8145 0.1420 0.8145 0.2418 5336 30615
420
- + 0.1300 0.9652 0.1300 0.2291 26496 3568
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
- Overall Accuracy: 0.3446 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
- A 0.1176 0.0045 0.1176 0.0086 68 1782
433
- # 0.0294 0.1778 0.0294 0.0505 2720 450
434
- G 0.3382 0.0060 0.3382 0.0118 68 3816
435
- _ 0.4685 0.4237 0.4685 0.4450 5372 5941
436
- + 0.3739 0.7986 0.3739 0.5093 7072 3311
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
- Overall Accuracy: 0.3939 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
- A 0.0471 0.0045 0.0471 0.0081 85 897
449
- # 0.0353 0.2156 0.0353 0.0607 3995 654
450
- G 0.0941 0.0061 0.0941 0.0115 85 1305
451
- _ 0.7642 0.3640 0.7642 0.4931 6120 12849
452
- + 0.3058 0.7904 0.3058 0.4409 8840 3420
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
- Overall Accuracy: 0.3198 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
- A 0.1636 0.0045 0.1636 0.0088 55 1991
465
- # 0.0545 0.2356 0.0545 0.0886 2915 675
466
- G 0.1273 0.0055 0.1273 0.0105 55 1276
467
- _ 0.5479 0.3327 0.5479 0.4140 3630 5978
468
- + 0.3136 0.7308 0.3136 0.4389 5720 2455
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
- Overall Accuracy: 0.3105 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
- A 0.0680 0.0047 0.0680 0.0088 103 1487
481
- # 0.0680 0.2715 0.0680 0.1087 6180 1547
482
- G 0.1068 0.0055 0.1068 0.0104 103 2009
483
- _ 0.7002 0.2846 0.7002 0.4047 6077 14953
484
- + 0.2337 0.7874 0.2337 0.3604 10712 3179
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
- Overall Accuracy: 0.3004 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
- A 0.0439 0.0044 0.0439 0.0081 114 1125
497
- # 0.0439 0.2933 0.0439 0.0763 7524 1125
498
- G 0.1316 0.0064 0.1316 0.0122 114 2338
499
- _ 0.7499 0.2573 0.7499 0.3832 6042 17607
500
- + 0.2383 0.8177 0.2383 0.3690 11856 3455
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
- Overall Accuracy: 0.2345 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
- A 0.1071 0.0045 0.1071 0.0087 224 5276
513
- # 0.1469 0.3276 0.1469 0.2028 16128 7232
514
- G 0.0670 0.0053 0.0670 0.0098 224 2850
515
- _ 0.6677 0.2180 0.6677 0.3287 10528 32242
516
- + 0.1021 0.8496 0.1021 0.1823 23296 2800
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
- Overall Accuracy: 0.4103 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
  A 0.0000 0.0000 0.0000 0.0000 64 0
529
- # 0.0000 0.0000 0.0000 0.0000 3072 0
530
- G 0.3438 0.0052 0.3438 0.0103 64 4224
531
- _ 0.5997 0.5758 0.5997 0.5875 7616 7931
532
- + 0.3683 0.5880 0.3683 0.4529 3584 2245
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
- Overall Accuracy: 0.4153 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
- A 0.0645 0.0044 0.0645 0.0083 62 899
545
- # 0.0161 0.2578 0.0161 0.0304 3596 225
546
- G 0.1935 0.0053 0.1935 0.0104 62 2254
547
- _ 0.6668 0.5342 0.6668 0.5932 6758 8435
548
- + 0.3494 0.5676 0.3494 0.4325 3472 2137
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
- Overall Accuracy: 0.3956 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
- A 0.0800 0.0045 0.0800 0.0084 125 2244
561
- # 0.0400 0.3038 0.0400 0.0707 8500 1119
562
- G 0.0720 0.0039 0.0720 0.0074 125 2313
563
- _ 0.7344 0.4656 0.7344 0.5699 12375 19520
564
- + 0.2397 0.5729 0.2397 0.3380 7000 2929
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
- Overall Accuracy: 0.3497 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
- A 0.0413 0.0043 0.0413 0.0078 121 1153
577
- # 0.0495 0.3498 0.0495 0.0867 9438 1335
578
- G 0.1653 0.0048 0.1653 0.0094 121 4147
579
- _ 0.6759 0.4163 0.6759 0.5153 10769 17485
580
- + 0.2581 0.5633 0.2581 0.3540 6776 3105
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
- Overall Accuracy: 0.3029 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
- A 0.0893 0.0045 0.0893 0.0085 112 2229
593
- # 0.0918 0.4003 0.0918 0.1493 9968 2286
594
- G 0.1786 0.0054 0.1786 0.0104 112 3736
595
- _ 0.6010 0.3649 0.6010 0.4541 8736 14389
596
- + 0.2293 0.5617 0.2293 0.3256 6272 2560
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
- Overall Accuracy: 0.2834 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
- A 0.1012 0.0046 0.1012 0.0088 326 7138
609
- # 0.1454 0.4400 0.1454 0.2186 31948 10556
610
- G 0.1380 0.0051 0.1380 0.0098 326 8863
611
- _ 0.5809 0.3176 0.5809 0.4106 22494 41148
612
- + 0.1641 0.5307 0.1641 0.2507 18256 5645
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
- Overall Accuracy: 0.4518 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
- A 0.0116 0.0044 0.0116 0.0064 86 225
625
- # 0.0116 0.2489 0.0116 0.0222 4816 225
626
- G 0.2791 0.0049 0.2791 0.0095 86 4942
627
- _ 0.6031 0.7574 0.6031 0.6715 14362 11436
628
- + 0.0000 0.0000 0.0000 0.0000 0 2522
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
- Overall Accuracy: 0.4623 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
- A 0.0575 0.0044 0.0575 0.0083 87 1125
641
- # 0.0343 0.3145 0.0343 0.0619 6177 674
642
- G 0.1149 0.0042 0.1149 0.0081 87 2383
643
- _ 0.6672 0.6881 0.6672 0.6775 13224 12823
644
- + 0.0000 0.0000 0.0000 0.0000 0 2570
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
- Overall Accuracy: 0.4417 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
- A 0.0515 0.0042 0.0515 0.0078 136 1666
657
- # 0.0325 0.3794 0.0325 0.0599 11560 991
658
- G 0.1176 0.0041 0.1176 0.0079 136 3936
659
- _ 0.6988 0.6206 0.6988 0.6574 18768 21134
660
- + 0.0000 0.0000 0.0000 0.0000 0 2873
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
- Overall Accuracy: 0.3780 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
- A 0.0900 0.0045 0.0900 0.0085 100 2014
673
- # 0.0607 0.4467 0.0607 0.1069 10000 1359
674
- G 0.1500 0.0048 0.1500 0.0094 100 3096
675
- _ 0.6402 0.5534 0.6402 0.5936 12300 14229
676
- + 0.0000 0.0000 0.0000 0.0000 0 1802
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
- Overall Accuracy: 0.3859 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
- A 0.0549 0.0044 0.0549 0.0082 164 2025
689
- # 0.0906 0.5046 0.0906 0.1536 18696 3357
690
- G 0.1098 0.0044 0.1098 0.0085 164 4078
691
- _ 0.7003 0.4890 0.7003 0.5758 17876 25601
692
- + 0.0000 0.0000 0.0000 0.0000 0 1839
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
- Overall Accuracy: 0.3119 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
- A 0.1074 0.0045 0.1074 0.0087 391 9277
705
- # 0.1065 0.5680 0.1065 0.1794 50048 9387
706
- G 0.1023 0.0035 0.1023 0.0068 391 11325
707
- _ 0.5930 0.4256 0.5930 0.4955 37145 51760
708
- + 0.0000 0.0000 0.0000 0.0000 0 6226
709
  ---------------------------------------------------------------------------------------
710
 
711
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_lr_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer15_lr_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
 
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
+ Overall Accuracy: 0.5572 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
+ A 0.0959 0.0097 0.0959 0.0177 3223 31724
37
+ # 0.6827 0.4721 0.6827 0.5582 238175 344378
38
+ G 0.2091 0.0144 0.2091 0.0270 3223 46673
39
+ _ 0.2021 0.6346 0.2021 0.3065 243346 77476
40
+ + 0.8067 0.8507 0.8067 0.8281 237208 224924
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
 
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
+ Overall Accuracy: 0.7536 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
+ A 0.1713 0.0200 0.1713 0.0358 286 2450
57
+ # 0.0000 0.0000 0.0000 0.0000 8224 0
58
+ G 0.8217 0.0213 0.8217 0.0414 286 11053
59
+ _ 0.0000 0.0000 0.0000 0.0000 5218 0
60
+ + 0.9577 0.9481 0.9577 0.9529 50336 50847
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
+ Overall Accuracy: 0.5927 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
+ A 0.1245 0.0141 0.1245 0.0254 514 4530
73
+ # 0.0000 0.0000 0.0000 0.0000 22670 0
74
+ G 0.8424 0.0124 0.8424 0.0245 514 34783
75
+ _ 0.0023 0.5256 0.0023 0.0046 17936 78
76
+ + 0.9189 0.8918 0.9189 0.9051 74016 76259
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
+ Overall Accuracy: 0.5338 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
+ A 0.3005 0.0080 0.3005 0.0155 649 24505
89
+ # 0.5510 0.4410 0.5510 0.4899 39462 49311
90
+ G 0.0092 0.0072 0.0092 0.0081 649 835
91
+ _ 0.1059 0.5283 0.1059 0.1765 37769 7573
92
+ + 0.7704 0.8151 0.7704 0.7921 67496 63801
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
+ Overall Accuracy: 0.5025 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
+ A 0.0012 0.0042 0.0012 0.0019 810 239
105
+ # 0.7691 0.4337 0.7691 0.5547 66522 117962
106
+ G 0.0000 0.0000 0.0000 0.0000 810 2
107
+ _ 0.2516 0.5759 0.2516 0.3502 68748 30030
108
+ + 0.5099 0.6800 0.5099 0.5828 45360 34017
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
+ Overall Accuracy: 0.5418 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
+ A 0.0000 0.0000 0.0000 0.0000 964 0
121
+ # 0.8854 0.5064 0.8854 0.6443 101297 177105
122
+ G 0.0000 0.0000 0.0000 0.0000 964 0
123
+ _ 0.2449 0.6994 0.2449 0.3627 113675 39795
124
+ + 0.0000 0.0000 0.0000 0.0000 0 0
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
 
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
+ Overall Accuracy: 0.6472 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
+ A 0.1484 0.0095 0.1484 0.0178 310 4864
141
+ # 0.0004 0.8333 0.0004 0.0008 13248 6
142
+ G 0.2387 0.0147 0.2387 0.0277 310 5041
143
+ _ 0.7940 0.6765 0.7940 0.7306 30762 36103
144
+ + 0.8197 0.8675 0.8197 0.8430 25120 23736
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
+ Overall Accuracy: 0.5570 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
+ A 0.2523 0.0089 0.2523 0.0172 325 9225
157
+ # 0.1233 0.3018 0.1233 0.1751 16604 6783
158
+ G 0.2400 0.0145 0.2400 0.0273 325 5389
159
+ _ 0.5554 0.6299 0.5554 0.5903 29047 25612
160
+ + 0.8347 0.8573 0.8347 0.8458 26824 26116
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
+ Overall Accuracy: 0.4879 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
+ A 0.1368 0.0093 0.1368 0.0174 424 6249
173
+ # 0.5258 0.3598 0.5258 0.4273 26803 39166
174
+ G 0.2146 0.0142 0.2146 0.0266 424 6409
175
+ _ 0.2146 0.5524 0.2146 0.3091 38165 14828
176
+ + 0.8151 0.8388 0.8151 0.8268 29584 28748
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
+ Overall Accuracy: 0.5061 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
+ A 0.1319 0.0094 0.1319 0.0176 432 6041
189
+ # 0.7070 0.4078 0.7070 0.5172 29604 51326
190
+ G 0.2153 0.0149 0.2153 0.0278 432 6248
191
+ _ 0.0131 0.4523 0.0131 0.0255 32188 933
192
+ + 0.8015 0.8480 0.8015 0.8241 34544 32652
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
+ Overall Accuracy: 0.5557 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
+ A 0.0733 0.0113 0.0733 0.0196 546 3529
205
+ # 0.7726 0.4709 0.7726 0.5851 42640 69955
206
+ G 0.2326 0.0151 0.2326 0.0284 546 8410
207
+ _ 0.0000 0.0000 0.0000 0.0000 36766 0
208
+ + 0.8303 0.8586 0.8303 0.8442 42352 40956
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
+ Overall Accuracy: 0.5779 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
+ A 0.0219 0.0143 0.0219 0.0173 1186 1816
221
+ # 0.8472 0.5226 0.8472 0.6465 109276 177142
222
+ G 0.1779 0.0139 0.1779 0.0258 1186 15176
223
+ _ 0.0000 0.0000 0.0000 0.0000 76418 0
224
+ + 0.7794 0.8445 0.7794 0.8106 78784 72716
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
 
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
+ Overall Accuracy: 0.7503 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
+ A 0.0789 0.0189 0.0789 0.0305 38 159
241
+ # 0.0000 0.0000 0.0000 0.0000 912 0
242
+ G 0.9211 0.0207 0.9211 0.0405 38 1690
243
+ _ 0.0000 0.0000 0.0000 0.0000 874 0
244
+ + 0.9535 0.9516 0.9535 0.9526 6688 6701
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
+ Overall Accuracy: 0.7519 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
+ A 0.3182 0.0200 0.3182 0.0376 44 701
257
  # 0.0000 0.0000 0.0000 0.0000 1144 0
258
+ G 0.6818 0.0213 0.6818 0.0413 44 1410
259
+ _ 0.0000 0.0000 0.0000 0.0000 924 0
260
+ + 0.9556 0.9501 0.9556 0.9528 7744 7789
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
+ Overall Accuracy: 0.7538 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
+ A 0.2195 0.0188 0.2195 0.0347 41 478
273
+ # 0.0000 0.0000 0.0000 0.0000 1148 0
274
+ G 0.7317 0.0207 0.7317 0.0403 41 1446
275
+ _ 0.0000 0.0000 0.0000 0.0000 779 0
276
+ + 0.9583 0.9471 0.9583 0.9527 7216 7301
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
+ Overall Accuracy: 0.7540 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
+ A 0.1915 0.0207 0.1915 0.0373 47 435
289
  # 0.0000 0.0000 0.0000 0.0000 1363 0
290
+ G 0.7872 0.0209 0.7872 0.0406 47 1774
291
+ _ 0.0000 0.0000 0.0000 0.0000 846 0
292
+ + 0.9584 0.9476 0.9584 0.9530 8272 8366
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
+ Overall Accuracy: 0.7546 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
+ A 0.1091 0.0195 0.1091 0.0331 55 308
305
+ # 0.0000 0.0000 0.0000 0.0000 1705 0
306
+ G 0.8909 0.0216 0.8909 0.0422 55 2269
307
+ _ 0.0000 0.0000 0.0000 0.0000 880 0
308
+ + 0.9590 0.9474 0.9590 0.9532 9680 9798
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
+ Overall Accuracy: 0.7554 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
+ A 0.1311 0.0217 0.1311 0.0372 61 369
321
+ # 0.0000 0.0000 0.0000 0.0000 1952 0
322
+ G 0.8852 0.0219 0.8852 0.0428 61 2464
323
+ _ 0.0000 0.0000 0.0000 0.0000 915 0
324
+ + 0.9599 0.9462 0.9599 0.9530 10736 10892
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
+ Overall Accuracy: 0.5819 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
+ A 0.3148 0.0132 0.3148 0.0254 54 1285
337
+ # 0.0000 0.0000 0.0000 0.0000 1728 0
338
+ G 0.6852 0.0120 0.6852 0.0236 54 3087
339
+ _ 0.0162 0.5256 0.0162 0.0313 2538 78
340
+ + 0.8970 0.9058 0.8970 0.9014 7776 7700
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
+ Overall Accuracy: 0.5888 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
+ A 0.0426 0.0139 0.0426 0.0209 47 144
353
+ # 0.0000 0.0000 0.0000 0.0000 1692 0
354
+ G 0.9574 0.0127 0.9574 0.0251 47 3542
355
+ _ 0.0000 0.0000 0.0000 0.0000 2021 0
356
+ + 0.9131 0.8971 0.9131 0.9050 6768 6889
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
+ Overall Accuracy: 0.5857 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
+ A 0.0746 0.0151 0.0746 0.0251 67 332
369
+ # 0.0000 0.0000 0.0000 0.0000 2680 0
370
+ G 0.9104 0.0123 0.9104 0.0243 67 4961
371
+ _ 0.0000 0.0000 0.0000 0.0000 2613 0
372
+ + 0.9084 0.8959 0.9084 0.9021 9648 9782
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
+ Overall Accuracy: 0.5865 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
+ A 0.0984 0.0130 0.0984 0.0230 61 461
385
  # 0.0000 0.0000 0.0000 0.0000 2623 0
386
+ G 0.9016 0.0127 0.9016 0.0250 61 4340
387
+ _ 0.0000 0.0000 0.0000 0.0000 2196 0
388
+ + 0.9095 0.8952 0.9095 0.9023 8784 8924
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
+ Overall Accuracy: 0.6005 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
+ A 0.1881 0.0149 0.1881 0.0276 101 1277
401
+ # 0.0000 0.0000 0.0000 0.0000 4747 0
402
+ G 0.7723 0.0127 0.7723 0.0250 101 6141
403
+ _ 0.0000 0.0000 0.0000 0.0000 3232 0
404
+ + 0.9316 0.8852 0.9316 0.9078 14544 15307
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
+ Overall Accuracy: 0.5972 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
+ A 0.0815 0.0145 0.0815 0.0247 184 1031
417
+ # 0.0000 0.0000 0.0000 0.0000 9200 0
418
+ G 0.8533 0.0124 0.8533 0.0243 184 12712
419
+ _ 0.0000 0.0000 0.0000 0.0000 5336 0
420
+ + 0.9267 0.8878 0.9267 0.9068 26496 27657
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
+ Overall Accuracy: 0.5342 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
+ A 0.3824 0.0076 0.3824 0.0149 68 3420
433
+ # 0.0000 0.0000 0.0000 0.0000 2720 0
434
+ G 0.0294 0.0076 0.0294 0.0120 68 264
435
+ _ 0.5164 0.5476 0.5164 0.5315 5372 5066
436
+ + 0.7596 0.8202 0.7596 0.7887 7072 6550
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
+ Overall Accuracy: 0.4102 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
+ A 0.7765 0.0079 0.7765 0.0156 85 8380
449
+ # 0.0095 0.2836 0.0095 0.0184 3995 134
450
+ G 0.0353 0.0069 0.0353 0.0115 85 437
451
+ _ 0.1404 0.5038 0.1404 0.2196 6120 1705
452
+ + 0.7783 0.8124 0.7783 0.7950 8840 8469
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
+ Overall Accuracy: 0.4176 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
+ A 0.7818 0.0083 0.7818 0.0164 55 5200
465
+ # 0.0861 0.3189 0.0861 0.1356 2915 787
466
+ G 0.0000 0.0000 0.0000 0.0000 55 0
467
+ _ 0.1014 0.4589 0.1014 0.1661 3630 802
468
+ + 0.7878 0.8067 0.7878 0.7971 5720 5586
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
+ Overall Accuracy: 0.4853 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
+ A 0.4078 0.0082 0.4078 0.0160 103 5145
481
+ # 0.4854 0.3817 0.4854 0.4274 6180 7859
482
+ G 0.0097 0.0075 0.0097 0.0084 103 134
483
+ _ 0.0000 0.0000 0.0000 0.0000 6077 0
484
+ + 0.7659 0.8174 0.7659 0.7908 10712 10037
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
+ Overall Accuracy: 0.5658 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
+ A 0.1316 0.0077 0.1316 0.0146 114 1944
497
+ # 0.6927 0.4260 0.6927 0.5275 7524 12236
498
+ G 0.0000 0.0000 0.0000 0.0000 114 0
499
+ _ 0.0000 0.0000 0.0000 0.0000 6042 0
500
+ + 0.7832 0.8096 0.7832 0.7962 11856 11470
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
+ Overall Accuracy: 0.6151 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
+ A 0.0134 0.0072 0.0134 0.0094 224 416
513
+ # 0.8211 0.4680 0.8211 0.5962 16128 28295
514
+ G 0.0000 0.0000 0.0000 0.0000 224 0
515
+ _ 0.0000 0.0000 0.0000 0.0000 10528 0
516
+ + 0.7621 0.8186 0.7621 0.7893 23296 21689
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
+ Overall Accuracy: 0.6334 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
  A 0.0000 0.0000 0.0000 0.0000 64 0
529
+ # 0.0016 0.8333 0.0016 0.0032 3072 6
530
+ G 0.0000 0.0000 0.0000 0.0000 64 0
531
+ _ 0.9517 0.6243 0.9517 0.7540 7616 11609
532
+ + 0.5212 0.6707 0.5212 0.5866 3584 2785
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
+ Overall Accuracy: 0.5454 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
+ A 0.0000 0.0000 0.0000 0.0000 62 0
545
+ # 0.1607 0.2720 0.1607 0.2021 3596 2125
546
+ G 0.0000 0.0000 0.0000 0.0000 62 0
547
+ _ 0.7550 0.5761 0.7550 0.6535 6758 8856
548
+ + 0.5556 0.6497 0.5556 0.5990 3472 2969
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
+ Overall Accuracy: 0.4543 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
+ A 0.0080 0.0042 0.0080 0.0055 125 239
561
+ # 0.5089 0.3284 0.5089 0.3992 8500 13173
562
+ G 0.0000 0.0000 0.0000 0.0000 125 2
563
+ _ 0.3654 0.5239 0.3654 0.4305 12375 8632
564
+ + 0.5611 0.6462 0.5611 0.6007 7000 6079
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
+ Overall Accuracy: 0.4378 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
+ A 0.0000 0.0000 0.0000 0.0000 121 0
577
+ # 0.8402 0.3782 0.8402 0.5216 9438 20967
578
+ G 0.0000 0.0000 0.0000 0.0000 121 0
579
+ _ 0.0392 0.4523 0.0392 0.0721 10769 933
580
+ + 0.5264 0.6699 0.5264 0.5895 6776 5325
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
+ Overall Accuracy: 0.4794 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
+ A 0.0000 0.0000 0.0000 0.0000 112 0
593
+ # 0.9063 0.4339 0.9063 0.5869 9968 20819
594
+ G 0.0000 0.0000 0.0000 0.0000 112 0
595
+ _ 0.0000 0.0000 0.0000 0.0000 8736 0
596
+ + 0.4857 0.6953 0.4857 0.5719 6272 4381
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
+ Overall Accuracy: 0.5192 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
+ A 0.0000 0.0000 0.0000 0.0000 326 0
609
+ # 0.9168 0.4812 0.9168 0.6311 31948 60872
610
+ G 0.0000 0.0000 0.0000 0.0000 326 0
611
+ _ 0.0000 0.0000 0.0000 0.0000 22494 0
612
+ + 0.4816 0.7047 0.4816 0.5722 18256 12478
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
+ Overall Accuracy: 0.7422 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
+ A 0.0000 0.0000 0.0000 0.0000 86 0
625
+ # 0.0000 0.0000 0.0000 0.0000 4816 0
626
+ G 0.0000 0.0000 0.0000 0.0000 86 0
627
+ _ 1.0000 0.7422 1.0000 0.8520 14362 19350
628
+ + 0.0000 0.0000 0.0000 0.0000 0 0
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
+ Overall Accuracy: 0.5927 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
+ A 0.0000 0.0000 0.0000 0.0000 87 0
641
+ # 0.2317 0.3163 0.2317 0.2675 6177 4524
642
+ G 0.0000 0.0000 0.0000 0.0000 87 0
643
+ _ 0.7691 0.6758 0.7691 0.7194 13224 15051
644
+ + 0.0000 0.0000 0.0000 0.0000 0 0
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
+ Overall Accuracy: 0.4189 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
+ A 0.0000 0.0000 0.0000 0.0000 136 0
657
+ # 0.8232 0.3775 0.8232 0.5177 11560 25206
658
+ G 0.0000 0.0000 0.0000 0.0000 136 0
659
+ _ 0.1759 0.6120 0.1759 0.2732 18768 5394
660
+ + 0.0000 0.0000 0.0000 0.0000 0 0
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
+ Overall Accuracy: 0.4444 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
+ A 0.0000 0.0000 0.0000 0.0000 100 0
673
+ # 1.0000 0.4444 1.0000 0.6154 10000 22500
674
+ G 0.0000 0.0000 0.0000 0.0000 100 0
675
+ _ 0.0000 0.0000 0.0000 0.0000 12300 0
676
+ + 0.0000 0.0000 0.0000 0.0000 0 0
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
+ Overall Accuracy: 0.5067 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
+ A 0.0000 0.0000 0.0000 0.0000 164 0
689
+ # 1.0000 0.5067 1.0000 0.6726 18696 36900
690
+ G 0.0000 0.0000 0.0000 0.0000 164 0
691
+ _ 0.0000 0.0000 0.0000 0.0000 17876 0
692
+ + 0.0000 0.0000 0.0000 0.0000 0 0
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
+ Overall Accuracy: 0.5689 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
+ A 0.0000 0.0000 0.0000 0.0000 391 0
705
+ # 1.0000 0.5689 1.0000 0.7252 50048 87975
706
+ G 0.0000 0.0000 0.0000 0.0000 391 0
707
+ _ 0.0000 0.0000 0.0000 0.0000 37145 0
708
+ + 0.0000 0.0000 0.0000 0.0000 0 0
709
  ---------------------------------------------------------------------------------------
710
 
711
+ Results saved to: reveng/cognitive_map_probes_results/layer15/lr_general/pre_reasoning/eval_cognitive_map_probe_layer15_lr_pre_reasoning_all_general.json
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5025149944606451,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 649,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.9735695099082662,
16
- "precision": 0.5028072609247601,
17
- "recall": 0.9735695099082662,
18
- "f1": 0.663134002468262,
19
  "gt_support": 39462,
20
- "predicted": 76409
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 649,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.027615240011649767,
32
- "precision": 0.49198113207547167,
33
- "recall": 0.027615240011649767,
34
- "f1": 0.05229511895510041,
35
  "gt_support": 37769,
36
- "predicted": 2120
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.33458920758386,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 68,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.9702205882352941,
56
- "precision": 0.3287244643746886,
57
- "recall": 0.9702205882352941,
58
- "f1": 0.4910681056940826,
59
  "gt_support": 2720,
60
- "predicted": 8028
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 68,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.02122114668652271,
72
- "precision": 0.57,
73
- "recall": 0.02122114668652271,
74
- "f1": 0.04091888011486001,
75
  "gt_support": 5372,
76
- "predicted": 200
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
- "accuracy": 0.3948468643655809,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 85,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.9642052565707134,
95
- "precision": 0.38795447678517475,
96
- "recall": 0.9642052565707134,
97
- "f1": 0.553289284688308,
98
  "gt_support": 3995,
99
- "predicted": 9929
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 85,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.03415032679738562,
111
- "precision": 0.5870786516853933,
112
- "recall": 0.03415032679738562,
113
- "f1": 0.06454601605929586,
114
  "gt_support": 6120,
115
- "predicted": 356
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
- "accuracy": 0.44522915101427496,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 55,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.9169811320754717,
134
- "precision": 0.4372648454114183,
135
- "recall": 0.9169811320754717,
136
- "f1": 0.5921577315019938,
137
  "gt_support": 2915,
138
- "predicted": 6113
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 55,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.07988980716253444,
150
- "precision": 0.5350553505535055,
151
- "recall": 0.07988980716253444,
152
- "f1": 0.13902205177372962,
153
  "gt_support": 3630,
154
- "predicted": 542
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
- "accuracy": 0.4968306186311482,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 103,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.9885113268608414,
173
- "precision": 0.4964245083699009,
174
- "recall": 0.9885113268608414,
175
- "f1": 0.6609325976414584,
176
  "gt_support": 6180,
177
- "predicted": 12306
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 103,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.013658054961329604,
189
- "precision": 0.5286624203821656,
190
- "recall": 0.013658054961329604,
191
- "f1": 0.026628168110362525,
192
  "gt_support": 6077,
193
- "predicted": 157
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
- "accuracy": 0.542844715093519,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 114,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.9803296119085593,
212
- "precision": 0.5451186165102357,
213
- "recall": 0.9803296119085593,
214
- "f1": 0.70064117786749,
215
  "gt_support": 7524,
216
- "predicted": 13531
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 114,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.018536908308507117,
228
- "precision": 0.42585551330798477,
229
- "recall": 0.018536908308507117,
230
- "f1": 0.035527359238699444,
231
  "gt_support": 6042,
232
- "predicted": 263
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
- "accuracy": 0.5905032467532467,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 224,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.9778025793650794,
251
- "precision": 0.5950494302316807,
252
- "recall": 0.9778025793650794,
253
- "f1": 0.7398545625146611,
254
  "gt_support": 16128,
255
- "predicted": 26502
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 224,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.022321428571428572,
267
- "precision": 0.3903654485049834,
268
- "recall": 0.022321428571428572,
269
- "f1": 0.042228212039532795,
270
  "gt_support": 10528,
271
- "predicted": 602
272
  }
273
  },
274
  "total_samples": 27104
@@ -278,13 +278,13 @@
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 11
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7178749251868737,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.5916795069337443,
8
+ "precision": 0.2428842504743833,
9
+ "recall": 0.5916795069337443,
10
+ "f1": 0.3443946188340807,
11
  "gt_support": 649,
12
+ "predicted": 1581
13
  },
14
  "1": {
15
+ "accuracy": 0.7041457604784349,
16
+ "precision": 0.7677239321434491,
17
+ "recall": 0.7041457604784349,
18
+ "f1": 0.7345617003277994,
19
  "gt_support": 39462,
20
+ "predicted": 36194
21
  },
22
  "2": {
23
+ "accuracy": 0.50231124807396,
24
+ "precision": 0.19509275882704968,
25
+ "recall": 0.50231124807396,
26
+ "f1": 0.2810344827586207,
27
  "gt_support": 649,
28
+ "predicted": 1671
29
  },
30
  "3": {
31
+ "accuracy": 0.7380920861023591,
32
+ "precision": 0.7132768722974183,
33
+ "recall": 0.7380920861023591,
34
+ "f1": 0.7254723364388695,
35
  "gt_support": 37769,
36
+ "predicted": 39083
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8361691784151677,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.7205882352941176,
48
+ "precision": 0.12128712871287128,
49
+ "recall": 0.7205882352941176,
50
+ "f1": 0.20762711864406777,
51
  "gt_support": 68,
52
+ "predicted": 404
53
  },
54
  "1": {
55
+ "accuracy": 0.8591911764705882,
56
+ "precision": 0.8208640674394099,
57
+ "recall": 0.8591911764705882,
58
+ "f1": 0.8395904436860068,
59
  "gt_support": 2720,
60
+ "predicted": 2847
61
  },
62
  "2": {
63
+ "accuracy": 0.4117647058823529,
64
+ "precision": 0.16091954022988506,
65
+ "recall": 0.4117647058823529,
66
+ "f1": 0.23140495867768596,
67
  "gt_support": 68,
68
+ "predicted": 174
69
  },
70
  "3": {
71
+ "accuracy": 0.8313477289650038,
72
+ "precision": 0.9298355194669998,
73
+ "recall": 0.8313477289650038,
74
+ "f1": 0.8778378378378379,
75
  "gt_support": 5372,
76
+ "predicted": 4803
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8218765192027224,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.6823529411764706,
87
+ "precision": 0.2761904761904762,
88
+ "recall": 0.6823529411764706,
89
+ "f1": 0.39322033898305087,
90
  "gt_support": 85,
91
+ "predicted": 210
92
  },
93
  "1": {
94
+ "accuracy": 0.8448060075093867,
95
+ "precision": 0.7859804378202142,
96
+ "recall": 0.8448060075093867,
97
+ "f1": 0.8143322475570033,
98
  "gt_support": 3995,
99
+ "predicted": 4294
100
  },
101
  "2": {
102
+ "accuracy": 0.36470588235294116,
103
+ "precision": 0.23846153846153847,
104
+ "recall": 0.36470588235294116,
105
+ "f1": 0.28837209302325584,
106
  "gt_support": 85,
107
+ "predicted": 130
108
  },
109
  "3": {
110
+ "accuracy": 0.8151960784313725,
111
+ "precision": 0.8828525924615113,
112
+ "recall": 0.8151960784313725,
113
+ "f1": 0.8476764930762042,
114
  "gt_support": 6120,
115
+ "predicted": 5651
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
+ "accuracy": 0.7610818933132982,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.45454545454545453,
126
+ "precision": 0.21008403361344538,
127
+ "recall": 0.45454545454545453,
128
+ "f1": 0.28735632183908044,
129
  "gt_support": 55,
130
+ "predicted": 119
131
  },
132
  "1": {
133
+ "accuracy": 0.7670668953687821,
134
+ "precision": 0.7515966386554622,
135
+ "recall": 0.7670668953687821,
136
+ "f1": 0.7592529711375212,
137
  "gt_support": 2915,
138
+ "predicted": 2975
139
  },
140
  "2": {
141
+ "accuracy": 0.36363636363636365,
142
+ "precision": 0.22988505747126436,
143
+ "recall": 0.36363636363636365,
144
+ "f1": 0.2816901408450704,
145
  "gt_support": 55,
146
+ "predicted": 87
147
  },
148
  "3": {
149
+ "accuracy": 0.7669421487603306,
150
+ "precision": 0.8013816925734024,
151
+ "recall": 0.7669421487603306,
152
+ "f1": 0.7837837837837839,
153
  "gt_support": 3630,
154
+ "predicted": 3474
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
+ "accuracy": 0.7170825643905961,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.6310679611650486,
165
+ "precision": 0.26859504132231404,
166
+ "recall": 0.6310679611650486,
167
+ "f1": 0.37681159420289856,
168
  "gt_support": 103,
169
+ "predicted": 242
170
  },
171
  "1": {
172
+ "accuracy": 0.6789644012944984,
173
+ "precision": 0.773598820058997,
174
+ "recall": 0.6789644012944984,
175
+ "f1": 0.7231988969320924,
176
  "gt_support": 6180,
177
+ "predicted": 5424
178
  },
179
  "2": {
180
+ "accuracy": 0.5048543689320388,
181
+ "precision": 0.22413793103448276,
182
+ "recall": 0.5048543689320388,
183
+ "f1": 0.3104477611940299,
184
  "gt_support": 103,
185
+ "predicted": 232
186
  },
187
  "3": {
188
+ "accuracy": 0.7609017607372058,
189
+ "precision": 0.7043412033511044,
190
+ "recall": 0.7609017607372058,
191
+ "f1": 0.7315298212308179,
192
  "gt_support": 6077,
193
+ "predicted": 6565
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
+ "accuracy": 0.6840655357401769,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.6578947368421053,
204
+ "precision": 0.35545023696682465,
205
+ "recall": 0.6578947368421053,
206
+ "f1": 0.46153846153846156,
207
  "gt_support": 114,
208
+ "predicted": 211
209
  },
210
  "1": {
211
+ "accuracy": 0.6299840510366826,
212
+ "precision": 0.783082768874938,
213
+ "recall": 0.6299840510366826,
214
+ "f1": 0.6982396700301982,
215
  "gt_support": 7524,
216
+ "predicted": 6053
217
  },
218
  "2": {
219
+ "accuracy": 0.6052631578947368,
220
+ "precision": 0.25274725274725274,
221
+ "recall": 0.6052631578947368,
222
+ "f1": 0.35658914728682173,
223
  "gt_support": 114,
224
+ "predicted": 273
225
  },
226
  "3": {
227
+ "accuracy": 0.7533929162528964,
228
+ "precision": 0.627256442055946,
229
+ "recall": 0.7533929162528964,
230
+ "f1": 0.6845627490788782,
231
  "gt_support": 6042,
232
+ "predicted": 7257
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
+ "accuracy": 0.6494613341204251,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.5,
243
+ "precision": 0.28354430379746837,
244
+ "recall": 0.5,
245
+ "f1": 0.3618739903069467,
246
  "gt_support": 224,
247
+ "predicted": 395
248
  },
249
  "1": {
250
+ "accuracy": 0.6760292658730159,
251
+ "precision": 0.7467296760495856,
252
+ "recall": 0.6760292658730159,
253
+ "f1": 0.7096228318526474,
254
  "gt_support": 16128,
255
+ "predicted": 14601
256
  },
257
  "2": {
258
+ "accuracy": 0.5625,
259
+ "precision": 0.1625806451612903,
260
+ "recall": 0.5625,
261
+ "f1": 0.25225225225225223,
262
  "gt_support": 224,
263
+ "predicted": 775
264
  },
265
  "3": {
266
+ "accuracy": 0.6137917933130699,
267
+ "precision": 0.5701932409776758,
268
+ "recall": 0.6137917933130699,
269
+ "f1": 0.5911897900370523,
270
  "gt_support": 10528,
271
+ "predicted": 11333
272
  }
273
  },
274
  "total_samples": 27104
 
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 11
290
  }
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 649 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5025 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 649 0
30
- # 0.9736 0.5028 0.9736 0.6631 39462 76409
31
- G 0.0000 0.0000 0.0000 0.0000 649 0
32
- _ 0.0276 0.4920 0.0276 0.0523 37769 2120
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.3346 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 68 0
49
- # 0.9702 0.3287 0.9702 0.4911 2720 8028
50
- G 0.0000 0.0000 0.0000 0.0000 68 0
51
- _ 0.0212 0.5700 0.0212 0.0409 5372 200
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.3948 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 85 0
64
- # 0.9642 0.3880 0.9642 0.5533 3995 9929
65
- G 0.0000 0.0000 0.0000 0.0000 85 0
66
- _ 0.0342 0.5871 0.0342 0.0645 6120 356
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.4452 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 55 0
79
- # 0.9170 0.4373 0.9170 0.5922 2915 6113
80
- G 0.0000 0.0000 0.0000 0.0000 55 0
81
- _ 0.0799 0.5351 0.0799 0.1390 3630 542
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.4968 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 103 0
94
- # 0.9885 0.4964 0.9885 0.6609 6180 12306
95
- G 0.0000 0.0000 0.0000 0.0000 103 0
96
- _ 0.0137 0.5287 0.0137 0.0266 6077 157
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.5428 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 114 0
109
- # 0.9803 0.5451 0.9803 0.7006 7524 13531
110
- G 0.0000 0.0000 0.0000 0.0000 114 0
111
- _ 0.0185 0.4259 0.0185 0.0355 6042 263
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.5905 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 224 0
124
- # 0.9778 0.5950 0.9778 0.7399 16128 26502
125
- G 0.0000 0.0000 0.0000 0.0000 224 0
126
- _ 0.0223 0.3904 0.0223 0.0422 10528 602
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.7179 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.5917 0.2429 0.5917 0.3444 649 1581
30
+ # 0.7041 0.7677 0.7041 0.7346 39462 36194
31
+ G 0.5023 0.1951 0.5023 0.2810 649 1671
32
+ _ 0.7381 0.7133 0.7381 0.7255 37769 39083
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8362 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.7206 0.1213 0.7206 0.2076 68 404
49
+ # 0.8592 0.8209 0.8592 0.8396 2720 2847
50
+ G 0.4118 0.1609 0.4118 0.2314 68 174
51
+ _ 0.8313 0.9298 0.8313 0.8778 5372 4803
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8219 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.6824 0.2762 0.6824 0.3932 85 210
64
+ # 0.8448 0.7860 0.8448 0.8143 3995 4294
65
+ G 0.3647 0.2385 0.3647 0.2884 85 130
66
+ _ 0.8152 0.8829 0.8152 0.8477 6120 5651
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7611 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.4545 0.2101 0.4545 0.2874 55 119
79
+ # 0.7671 0.7516 0.7671 0.7593 2915 2975
80
+ G 0.3636 0.2299 0.3636 0.2817 55 87
81
+ _ 0.7669 0.8014 0.7669 0.7838 3630 3474
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.7171 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.6311 0.2686 0.6311 0.3768 103 242
94
+ # 0.6790 0.7736 0.6790 0.7232 6180 5424
95
+ G 0.5049 0.2241 0.5049 0.3104 103 232
96
+ _ 0.7609 0.7043 0.7609 0.7315 6077 6565
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6841 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.6579 0.3555 0.6579 0.4615 114 211
109
+ # 0.6300 0.7831 0.6300 0.6982 7524 6053
110
+ G 0.6053 0.2527 0.6053 0.3566 114 273
111
+ _ 0.7534 0.6273 0.7534 0.6846 6042 7257
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.6495 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.5000 0.2835 0.5000 0.3619 224 395
124
+ # 0.6760 0.7467 0.6760 0.7096 16128 14601
125
+ G 0.5625 0.1626 0.5625 0.2523 224 775
126
+ _ 0.6138 0.5702 0.6138 0.5912 10528 11333
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size11.json
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5150485791511432,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 810,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.8139262198971768,
16
- "precision": 0.5023846196670811,
17
- "recall": 0.8139262198971768,
18
- "f1": 0.621287923991371,
19
  "gt_support": 66522,
20
- "predicted": 107774
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 810,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.237985105021237,
32
- "precision": 0.5619247149333699,
33
- "recall": 0.237985105021237,
34
- "f1": 0.33436197171585064,
35
  "gt_support": 68748,
36
- "predicted": 29116
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.5134060650887574,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 64,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.58203125,
56
- "precision": 0.3136291878617786,
57
- "recall": 0.58203125,
58
- "f1": 0.4076142710589308,
59
  "gt_support": 3072,
60
- "predicted": 5701
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 64,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.4943539915966387,
72
- "precision": 0.7360703812316716,
73
- "recall": 0.4943539915966387,
74
- "f1": 0.5914696410336974,
75
  "gt_support": 7616,
76
- "predicted": 5115
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
- "accuracy": 0.47356365718648596,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 62,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.6954949944382648,
95
- "precision": 0.3640465793304221,
96
- "recall": 0.6954949944382648,
97
- "f1": 0.4779285304796484,
98
  "gt_support": 3596,
99
- "predicted": 6870
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 62,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.36416099437703464,
111
- "precision": 0.6820953436807096,
112
- "recall": 0.36416099437703464,
113
- "f1": 0.474821531931314,
114
  "gt_support": 6758,
115
- "predicted": 3608
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
- "accuracy": 0.4404733727810651,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 125,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.8356470588235294,
134
- "precision": 0.40628038666132815,
135
- "recall": 0.8356470588235294,
136
- "f1": 0.5467421006042412,
137
  "gt_support": 8500,
138
- "predicted": 17483
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 125,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.17793939393939395,
150
- "precision": 0.6046128500823723,
151
- "recall": 0.17793939393939395,
152
- "f1": 0.27495785727664357,
153
  "gt_support": 12375,
154
- "predicted": 3642
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
- "accuracy": 0.48858134872120884,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 121,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.7682771773680864,
173
- "precision": 0.46895615056266976,
174
- "recall": 0.7682771773680864,
175
- "f1": 0.5824096385542169,
176
  "gt_support": 9438,
177
- "predicted": 15462
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 121,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.2544340235862197,
189
- "precision": 0.5494285141367555,
190
- "recall": 0.2544340235862197,
191
- "f1": 0.34780401117034776,
192
  "gt_support": 10769,
193
- "predicted": 4987
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
- "accuracy": 0.5190722738799662,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 112,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.7900280898876404,
212
- "precision": 0.5309108069844266,
213
- "recall": 0.7900280898876404,
214
- "f1": 0.6350550381033022,
215
  "gt_support": 9968,
216
- "predicted": 14833
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 112,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.22321428571428573,
228
- "precision": 0.47619047619047616,
229
- "recall": 0.22321428571428573,
230
- "f1": 0.303951367781155,
231
  "gt_support": 8736,
232
- "predicted": 4095
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
- "accuracy": 0.5602969470359749,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 326,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.8647176662075873,
251
- "precision": 0.5825197680548234,
252
- "recall": 0.8647176662075873,
253
- "f1": 0.6961057286482809,
254
  "gt_support": 31948,
255
- "predicted": 47425
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 326,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.1441717791411043,
267
- "precision": 0.42287130003911855,
268
- "recall": 0.1441717791411043,
269
- "f1": 0.2150316613068992,
270
  "gt_support": 22494,
271
- "predicted": 7669
272
  }
273
  },
274
  "total_samples": 55094
@@ -278,13 +278,13 @@
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 13
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7073635765943458,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.5469135802469136,
8
+ "precision": 0.21536217792902285,
9
+ "recall": 0.5469135802469136,
10
+ "f1": 0.3090338332752005,
11
  "gt_support": 810,
12
+ "predicted": 2057
13
  },
14
  "1": {
15
+ "accuracy": 0.6564895823937945,
16
+ "precision": 0.7505155702206641,
17
+ "recall": 0.6564895823937945,
18
+ "f1": 0.7003608371421698,
19
  "gt_support": 66522,
20
+ "predicted": 58188
21
  },
22
  "2": {
23
+ "accuracy": 0.3987654320987654,
24
+ "precision": 0.14104803493449783,
25
+ "recall": 0.3987654320987654,
26
+ "f1": 0.20838709677419356,
27
  "gt_support": 810,
28
+ "predicted": 2290
29
  },
30
  "3": {
31
+ "accuracy": 0.7621167161226509,
32
+ "precision": 0.7046466276645821,
33
+ "recall": 0.7621167161226509,
34
+ "f1": 0.7322557877892147,
35
  "gt_support": 68748,
36
+ "predicted": 74355
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8831360946745562,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.546875,
48
+ "precision": 0.22727272727272727,
49
+ "recall": 0.546875,
50
+ "f1": 0.3211009174311926,
51
  "gt_support": 64,
52
+ "predicted": 154
53
  },
54
  "1": {
55
+ "accuracy": 0.7473958333333334,
56
+ "precision": 0.9184,
57
+ "recall": 0.7473958333333334,
58
+ "f1": 0.8241206030150754,
59
  "gt_support": 3072,
60
+ "predicted": 2500
61
  },
62
  "2": {
63
+ "accuracy": 0.484375,
64
+ "precision": 0.19135802469135801,
65
+ "recall": 0.484375,
66
+ "f1": 0.2743362831858407,
67
  "gt_support": 64,
68
+ "predicted": 162
69
  },
70
  "3": {
71
+ "accuracy": 0.9440651260504201,
72
+ "precision": 0.89875,
73
+ "recall": 0.9440651260504201,
74
+ "f1": 0.9208504098360657,
75
  "gt_support": 7616,
76
+ "predicted": 8000
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8632372590188967,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.5645161290322581,
87
+ "precision": 0.23809523809523808,
88
+ "recall": 0.5645161290322581,
89
+ "f1": 0.33492822966507174,
90
  "gt_support": 62,
91
+ "predicted": 147
92
  },
93
  "1": {
94
+ "accuracy": 0.7975528364849833,
95
+ "precision": 0.8625563909774436,
96
+ "recall": 0.7975528364849833,
97
+ "f1": 0.8287819679237104,
98
  "gt_support": 3596,
99
+ "predicted": 3325
100
  },
101
  "2": {
102
+ "accuracy": 0.2903225806451613,
103
+ "precision": 0.16363636363636364,
104
+ "recall": 0.2903225806451613,
105
+ "f1": 0.20930232558139533,
106
  "gt_support": 62,
107
+ "predicted": 110
108
  },
109
  "3": {
110
+ "accuracy": 0.9061852619118083,
111
+ "precision": 0.8880510440835266,
112
+ "recall": 0.9061852619118083,
113
+ "f1": 0.8970265123773252,
114
  "gt_support": 6758,
115
+ "predicted": 6896
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
+ "accuracy": 0.7996686390532545,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.504,
126
+ "precision": 0.17119565217391305,
127
+ "recall": 0.504,
128
+ "f1": 0.25557809330628806,
129
  "gt_support": 125,
130
+ "predicted": 368
131
  },
132
  "1": {
133
+ "accuracy": 0.6688235294117647,
134
+ "precision": 0.8608419139915203,
135
+ "recall": 0.6688235294117647,
136
+ "f1": 0.7527807203389831,
137
  "gt_support": 8500,
138
+ "predicted": 6604
139
  },
140
  "2": {
141
+ "accuracy": 0.376,
142
+ "precision": 0.1910569105691057,
143
+ "recall": 0.376,
144
+ "f1": 0.2533692722371968,
145
  "gt_support": 125,
146
+ "predicted": 246
147
  },
148
  "3": {
149
+ "accuracy": 0.8968080808080808,
150
+ "precision": 0.7980153879341338,
151
+ "recall": 0.8968080808080808,
152
+ "f1": 0.8445323795753747,
153
  "gt_support": 12375,
154
+ "predicted": 13907
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
+ "accuracy": 0.7433615335713237,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.5619834710743802,
165
+ "precision": 0.20298507462686566,
166
+ "recall": 0.5619834710743802,
167
+ "f1": 0.2982456140350877,
168
  "gt_support": 121,
169
+ "predicted": 335
170
  },
171
  "1": {
172
+ "accuracy": 0.5969485060394152,
173
+ "precision": 0.8510574018126889,
174
+ "recall": 0.5969485060394152,
175
+ "f1": 0.7017063146095405,
176
  "gt_support": 9438,
177
+ "predicted": 6620
178
  },
179
  "2": {
180
+ "accuracy": 0.45454545454545453,
181
+ "precision": 0.16717325227963525,
182
+ "recall": 0.45454545454545453,
183
+ "f1": 0.24444444444444444,
184
  "gt_support": 121,
185
+ "predicted": 329
186
  },
187
  "3": {
188
+ "accuracy": 0.8769616491781966,
189
+ "precision": 0.7173566274211925,
190
+ "recall": 0.8769616491781966,
191
+ "f1": 0.7891702180997744,
192
  "gt_support": 10769,
193
+ "predicted": 13165
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
+ "accuracy": 0.6336115807269653,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.5089285714285714,
204
+ "precision": 0.21509433962264152,
205
+ "recall": 0.5089285714285714,
206
+ "f1": 0.3023872679045093,
207
  "gt_support": 112,
208
+ "predicted": 265
209
  },
210
  "1": {
211
+ "accuracy": 0.6079454253611557,
212
+ "precision": 0.6905195989061076,
213
+ "recall": 0.6079454253611557,
214
+ "f1": 0.646606914212548,
215
  "gt_support": 9968,
216
+ "predicted": 8776
217
  },
218
  "2": {
219
+ "accuracy": 0.375,
220
+ "precision": 0.16470588235294117,
221
+ "recall": 0.375,
222
+ "f1": 0.22888283378746596,
223
  "gt_support": 112,
224
+ "predicted": 255
225
  },
226
  "3": {
227
+ "accuracy": 0.6678113553113553,
228
+ "precision": 0.6056893687707641,
229
+ "recall": 0.6678113553113553,
230
+ "f1": 0.6352351916376305,
231
  "gt_support": 8736,
232
+ "predicted": 9632
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
+ "accuracy": 0.6197952590118706,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.5674846625766872,
243
+ "precision": 0.23477157360406092,
244
+ "recall": 0.5674846625766872,
245
+ "f1": 0.33213644524236985,
246
  "gt_support": 326,
247
+ "predicted": 788
248
  },
249
  "1": {
250
+ "accuracy": 0.6613246525604106,
251
+ "precision": 0.6958469189474031,
252
+ "recall": 0.6613246525604106,
253
+ "f1": 0.6781467156681806,
254
  "gt_support": 31948,
255
+ "predicted": 30363
256
  },
257
  "2": {
258
+ "accuracy": 0.3987730061349693,
259
+ "precision": 0.10942760942760943,
260
+ "recall": 0.3987730061349693,
261
+ "f1": 0.17173051519154559,
262
  "gt_support": 326,
263
+ "predicted": 1188
264
  },
265
  "3": {
266
+ "accuracy": 0.5647728283097715,
267
+ "precision": 0.5582948802460997,
268
+ "recall": 0.5647728283097715,
269
+ "f1": 0.5615151716060024,
270
  "gt_support": 22494,
271
+ "predicted": 22755
272
  }
273
  },
274
  "total_samples": 55094
 
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 13
290
  }
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 810 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5150 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 810 0
30
- # 0.8139 0.5024 0.8139 0.6213 66522 107774
31
- G 0.0000 0.0000 0.0000 0.0000 810 0
32
- _ 0.2380 0.5619 0.2380 0.3344 68748 29116
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.5134 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 64 0
49
- # 0.5820 0.3136 0.5820 0.4076 3072 5701
50
- G 0.0000 0.0000 0.0000 0.0000 64 0
51
- _ 0.4944 0.7361 0.4944 0.5915 7616 5115
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.4736 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 62 0
64
- # 0.6955 0.3640 0.6955 0.4779 3596 6870
65
- G 0.0000 0.0000 0.0000 0.0000 62 0
66
- _ 0.3642 0.6821 0.3642 0.4748 6758 3608
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.4405 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 125 0
79
- # 0.8356 0.4063 0.8356 0.5467 8500 17483
80
- G 0.0000 0.0000 0.0000 0.0000 125 0
81
- _ 0.1779 0.6046 0.1779 0.2750 12375 3642
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.4886 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 121 0
94
- # 0.7683 0.4690 0.7683 0.5824 9438 15462
95
- G 0.0000 0.0000 0.0000 0.0000 121 0
96
- _ 0.2544 0.5494 0.2544 0.3478 10769 4987
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.5191 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 112 0
109
- # 0.7900 0.5309 0.7900 0.6351 9968 14833
110
- G 0.0000 0.0000 0.0000 0.0000 112 0
111
- _ 0.2232 0.4762 0.2232 0.3040 8736 4095
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.5603 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 326 0
124
- # 0.8647 0.5825 0.8647 0.6961 31948 47425
125
- G 0.0000 0.0000 0.0000 0.0000 326 0
126
- _ 0.1442 0.4229 0.1442 0.2150 22494 7669
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.7074 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.5469 0.2154 0.5469 0.3090 810 2057
30
+ # 0.6565 0.7505 0.6565 0.7004 66522 58188
31
+ G 0.3988 0.1410 0.3988 0.2084 810 2290
32
+ _ 0.7621 0.7046 0.7621 0.7323 68748 74355
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8831 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.5469 0.2273 0.5469 0.3211 64 154
49
+ # 0.7474 0.9184 0.7474 0.8241 3072 2500
50
+ G 0.4844 0.1914 0.4844 0.2743 64 162
51
+ _ 0.9441 0.8988 0.9441 0.9209 7616 8000
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8632 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.5645 0.2381 0.5645 0.3349 62 147
64
+ # 0.7976 0.8626 0.7976 0.8288 3596 3325
65
+ G 0.2903 0.1636 0.2903 0.2093 62 110
66
+ _ 0.9062 0.8881 0.9062 0.8970 6758 6896
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7997 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.5040 0.1712 0.5040 0.2556 125 368
79
+ # 0.6688 0.8608 0.6688 0.7528 8500 6604
80
+ G 0.3760 0.1911 0.3760 0.2534 125 246
81
+ _ 0.8968 0.7980 0.8968 0.8445 12375 13907
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.7434 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.5620 0.2030 0.5620 0.2982 121 335
94
+ # 0.5969 0.8511 0.5969 0.7017 9438 6620
95
+ G 0.4545 0.1672 0.4545 0.2444 121 329
96
+ _ 0.8770 0.7174 0.8770 0.7892 10769 13165
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6336 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.5089 0.2151 0.5089 0.3024 112 265
109
+ # 0.6079 0.6905 0.6079 0.6466 9968 8776
110
+ G 0.3750 0.1647 0.3750 0.2289 112 255
111
+ _ 0.6678 0.6057 0.6678 0.6352 8736 9632
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.6198 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.5675 0.2348 0.5675 0.3321 326 788
124
+ # 0.6613 0.6958 0.6613 0.6781 31948 30363
125
+ G 0.3988 0.1094 0.3988 0.1717 326 1188
126
+ _ 0.5648 0.5583 0.5648 0.5615 22494 22755
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size13.json
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.49412171507607194,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.06431535269709543,
8
- "precision": 0.004942602040816327,
9
- "recall": 0.06431535269709543,
10
- "f1": 0.009179745336097129,
11
  "gt_support": 964,
12
- "predicted": 12544
13
  },
14
  "1": {
15
- "accuracy": 0.09531378026989941,
16
- "precision": 0.4871587870225541,
17
- "recall": 0.09531378026989941,
18
- "f1": 0.15943393110736812,
19
  "gt_support": 101297,
20
- "predicted": 19819
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 964,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.8573389047723774,
32
- "precision": 0.5281217316852447,
33
- "recall": 0.8573389047723774,
34
- "f1": 0.6536155486700737,
35
  "gt_support": 113675,
36
- "predicted": 184537
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.6402583979328166,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.03488372093023256,
48
- "precision": 0.005791505791505791,
49
- "recall": 0.03488372093023256,
50
- "f1": 0.009933774834437085,
51
  "gt_support": 86,
52
- "predicted": 518
53
  },
54
  "1": {
55
- "accuracy": 0.29277408637873753,
56
- "precision": 0.31814079422382674,
57
- "recall": 0.29277408637873753,
58
- "f1": 0.3049307958477509,
59
  "gt_support": 4816,
60
- "predicted": 4432
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 86,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.7642389639325999,
72
- "precision": 0.7622222222222222,
73
- "recall": 0.7642389639325999,
74
- "f1": 0.7632292608302622,
75
  "gt_support": 14362,
76
- "predicted": 14400
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
- "accuracy": 0.641941251596424,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.022988505747126436,
87
- "precision": 0.0036101083032490976,
88
- "recall": 0.022988505747126436,
89
- "f1": 0.0062402496099844,
90
  "gt_support": 87,
91
- "predicted": 554
92
  },
93
  "1": {
94
- "accuracy": 0.08224056985591711,
95
- "precision": 0.38572513287775245,
96
- "recall": 0.08224056985591711,
97
- "f1": 0.13557512676808114,
98
  "gt_support": 6177,
99
- "predicted": 1317
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 87,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.91167574107683,
111
- "precision": 0.6809760506100316,
112
- "recall": 0.91167574107683,
113
- "f1": 0.7796171753750647,
114
  "gt_support": 13224,
115
- "predicted": 17704
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
- "accuracy": 0.5788562091503268,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.051470588235294115,
126
- "precision": 0.0049469964664310955,
127
- "recall": 0.051470588235294115,
128
- "f1": 0.009026434558349452,
129
  "gt_support": 136,
130
- "predicted": 1415
131
  },
132
  "1": {
133
- "accuracy": 0.0685121107266436,
134
- "precision": 0.4469525959367946,
135
- "recall": 0.0685121107266436,
136
- "f1": 0.11881188118811882,
137
  "gt_support": 11560,
138
- "predicted": 1772
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 136,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.9012148337595908,
150
- "precision": 0.6170065297486594,
151
- "recall": 0.9012148337595908,
152
- "f1": 0.7325090405144974,
153
  "gt_support": 18768,
154
- "predicted": 27413
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
- "accuracy": 0.5193777777777778,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.07,
165
- "precision": 0.006535947712418301,
166
- "recall": 0.07,
167
- "f1": 0.011955593509820665,
168
  "gt_support": 100,
169
- "predicted": 1071
170
  },
171
  "1": {
172
- "accuracy": 0.0858,
173
- "precision": 0.4793296089385475,
174
- "recall": 0.0858,
175
- "f1": 0.14554707379134862,
176
  "gt_support": 10000,
177
- "predicted": 1790
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 100,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.8797560975609756,
189
- "precision": 0.5509954682010285,
190
- "recall": 0.8797560975609756,
191
- "f1": 0.6776041829737937,
192
  "gt_support": 12300,
193
- "predicted": 19639
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
- "accuracy": 0.45848238482384823,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.07317073170731707,
204
- "precision": 0.004987531172069825,
205
- "recall": 0.07317073170731707,
206
- "f1": 0.009338521400778212,
207
  "gt_support": 164,
208
- "predicted": 2406
209
  },
210
  "1": {
211
- "accuracy": 0.04535729567821994,
212
- "precision": 0.5535248041775457,
213
- "recall": 0.04535729567821994,
214
- "f1": 0.08384417638916354,
215
  "gt_support": 18696,
216
- "predicted": 1532
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 164,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.8982993958379951,
228
- "precision": 0.4871670408349008,
229
- "recall": 0.8982993958379951,
230
- "f1": 0.6317321688500728,
231
  "gt_support": 17876,
232
- "predicted": 32962
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
- "accuracy": 0.4081045751633987,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0792838874680307,
243
- "precision": 0.004711246200607902,
244
- "recall": 0.0792838874680307,
245
- "f1": 0.008893989384593314,
246
  "gt_support": 391,
247
- "predicted": 6580
248
  },
249
  "1": {
250
- "accuracy": 0.10467950767263427,
251
- "precision": 0.5836675579322638,
252
- "recall": 0.10467950767263427,
253
- "f1": 0.17752100840336135,
254
  "gt_support": 50048,
255
- "predicted": 8976
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 391,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.8246870372863104,
267
- "precision": 0.42299672737817423,
268
- "recall": 0.8246870372863104,
269
- "f1": 0.5591800226351721,
270
  "gt_support": 37145,
271
- "predicted": 72419
272
  }
273
  },
274
  "total_samples": 87975
@@ -278,13 +278,13 @@
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 15
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.6981512217611803,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.4387966804979253,
8
+ "precision": 0.13428571428571429,
9
+ "recall": 0.4387966804979253,
10
+ "f1": 0.20563928050559066,
11
  "gt_support": 964,
12
+ "predicted": 3150
13
  },
14
  "1": {
15
+ "accuracy": 0.6459125146845415,
16
+ "precision": 0.7222939780316829,
17
+ "recall": 0.6459125146845415,
18
+ "f1": 0.6819712114737181,
19
  "gt_support": 101297,
20
+ "predicted": 90585
21
  },
22
  "2": {
23
+ "accuracy": 0.4066390041493776,
24
+ "precision": 0.10486891385767791,
25
+ "recall": 0.4066390041493776,
26
+ "f1": 0.16673755848575073,
27
  "gt_support": 964,
28
+ "predicted": 3738
29
  },
30
  "3": {
31
+ "accuracy": 0.7493732131075435,
32
+ "precision": 0.7132809163756939,
33
+ "recall": 0.7493732131075435,
34
+ "f1": 0.7308817599162598,
35
  "gt_support": 113675,
36
+ "predicted": 119427
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.9166925064599484,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.6744186046511628,
48
+ "precision": 0.13302752293577982,
49
+ "recall": 0.6744186046511628,
50
+ "f1": 0.22222222222222224,
51
  "gt_support": 86,
52
+ "predicted": 436
53
  },
54
  "1": {
55
+ "accuracy": 0.8608803986710963,
56
+ "precision": 0.9219479653102068,
57
+ "recall": 0.8608803986710963,
58
+ "f1": 0.890368302373027,
59
  "gt_support": 4816,
60
+ "predicted": 4497
61
  },
62
  "2": {
63
+ "accuracy": 0.4186046511627907,
64
+ "precision": 0.14285714285714285,
65
+ "recall": 0.4186046511627907,
66
+ "f1": 0.21301775147928995,
67
  "gt_support": 86,
68
+ "predicted": 252
69
  },
70
  "3": {
71
+ "accuracy": 0.939841247737084,
72
+ "precision": 0.9529121073067419,
73
+ "recall": 0.939841247737084,
74
+ "f1": 0.9463315455533355,
75
  "gt_support": 14362,
76
+ "predicted": 14165
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8518007662835249,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.5172413793103449,
87
+ "precision": 0.1347305389221557,
88
+ "recall": 0.5172413793103449,
89
+ "f1": 0.2137767220902613,
90
  "gt_support": 87,
91
+ "predicted": 334
92
  },
93
  "1": {
94
+ "accuracy": 0.6980734984620366,
95
+ "precision": 0.890725056806445,
96
+ "recall": 0.6980734984620366,
97
+ "f1": 0.7827191867852605,
98
  "gt_support": 6177,
99
+ "predicted": 4841
100
  },
101
  "2": {
102
+ "accuracy": 0.2413793103448276,
103
+ "precision": 0.08823529411764706,
104
+ "recall": 0.2413793103448276,
105
+ "f1": 0.12923076923076926,
106
  "gt_support": 87,
107
+ "predicted": 238
108
  },
109
  "3": {
110
+ "accuracy": 0.9298245614035088,
111
+ "precision": 0.8682389493009461,
112
+ "recall": 0.9298245614035088,
113
+ "f1": 0.8979770685751843,
114
  "gt_support": 13224,
115
+ "predicted": 14162
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
+ "accuracy": 0.7923856209150327,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.38235294117647056,
126
+ "precision": 0.12935323383084577,
127
+ "recall": 0.38235294117647056,
128
+ "f1": 0.19330855018587362,
129
  "gt_support": 136,
130
+ "predicted": 402
131
  },
132
  "1": {
133
+ "accuracy": 0.5698096885813149,
134
+ "precision": 0.9190735314636529,
135
+ "recall": 0.5698096885813149,
136
+ "f1": 0.70347626421744,
137
  "gt_support": 11560,
138
+ "predicted": 7167
139
  },
140
  "2": {
141
+ "accuracy": 0.34558823529411764,
142
+ "precision": 0.09038461538461538,
143
+ "recall": 0.34558823529411764,
144
+ "f1": 0.14329268292682926,
145
  "gt_support": 136,
146
+ "predicted": 520
147
  },
148
  "3": {
149
+ "accuracy": 0.9356884057971014,
150
+ "precision": 0.780107502998534,
151
+ "recall": 0.9356884057971014,
152
+ "f1": 0.8508442549480365,
153
  "gt_support": 18768,
154
+ "predicted": 22511
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
+ "accuracy": 0.7148444444444444,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.48,
165
+ "precision": 0.1518987341772152,
166
+ "recall": 0.48,
167
+ "f1": 0.23076923076923078,
168
  "gt_support": 100,
169
+ "predicted": 316
170
  },
171
  "1": {
172
+ "accuracy": 0.4902,
173
+ "precision": 0.8621174815335912,
174
+ "recall": 0.4902,
175
+ "f1": 0.6250159377789112,
176
  "gt_support": 10000,
177
+ "predicted": 5686
178
  },
179
  "2": {
180
+ "accuracy": 0.46,
181
+ "precision": 0.10952380952380952,
182
+ "recall": 0.46,
183
+ "f1": 0.1769230769230769,
184
  "gt_support": 100,
185
+ "predicted": 420
186
  },
187
  "3": {
188
+ "accuracy": 0.9014634146341464,
189
+ "precision": 0.6896380146784425,
190
+ "recall": 0.9014634146341464,
191
+ "f1": 0.7814504193389245,
192
  "gt_support": 12300,
193
+ "predicted": 16078
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
+ "accuracy": 0.6268563685636857,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.4146341463414634,
204
+ "precision": 0.13991769547325103,
205
+ "recall": 0.4146341463414634,
206
+ "f1": 0.20923076923076925,
207
  "gt_support": 164,
208
+ "predicted": 486
209
  },
210
  "1": {
211
+ "accuracy": 0.5424154899443732,
212
+ "precision": 0.689441838330274,
213
+ "recall": 0.5424154899443732,
214
+ "f1": 0.6071546175722197,
215
  "gt_support": 18696,
216
+ "predicted": 14709
217
  },
218
  "2": {
219
+ "accuracy": 0.39634146341463417,
220
+ "precision": 0.11035653650254669,
221
+ "recall": 0.39634146341463417,
222
+ "f1": 0.17264276228419653,
223
  "gt_support": 164,
224
+ "predicted": 589
225
  },
226
  "3": {
227
+ "accuracy": 0.7192324904900426,
228
+ "precision": 0.6088747868914567,
229
+ "recall": 0.7192324904900426,
230
+ "f1": 0.6594686089454247,
231
  "gt_support": 17876,
232
+ "predicted": 21116
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
+ "accuracy": 0.6087524865018471,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.3887468030690537,
243
+ "precision": 0.1292517006802721,
244
+ "recall": 0.3887468030690537,
245
+ "f1": 0.19400127632418634,
246
  "gt_support": 391,
247
+ "predicted": 1176
248
  },
249
  "1": {
250
+ "accuracy": 0.7061421035805626,
251
+ "precision": 0.6583030641706249,
252
+ "recall": 0.7061421035805626,
253
+ "f1": 0.6813839376090539,
254
  "gt_support": 50048,
255
+ "predicted": 53685
256
  },
257
  "2": {
258
+ "accuracy": 0.45268542199488493,
259
+ "precision": 0.10296684118673648,
260
+ "recall": 0.45268542199488493,
261
+ "f1": 0.16777251184834124,
262
  "gt_support": 391,
263
+ "predicted": 1719
264
  },
265
  "3": {
266
+ "accuracy": 0.4814914524162068,
267
+ "precision": 0.5696767001114827,
268
+ "recall": 0.4814914524162068,
269
+ "f1": 0.5218850306390429,
270
  "gt_support": 37145,
271
+ "predicted": 31395
272
  }
273
  },
274
  "total_samples": 87975
 
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 15
290
  }
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 964 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.4941 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0643 0.0049 0.0643 0.0092 964 12544
30
- # 0.0953 0.4872 0.0953 0.1594 101297 19819
31
- G 0.0000 0.0000 0.0000 0.0000 964 0
32
- _ 0.8573 0.5281 0.8573 0.6536 113675 184537
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.6403 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0349 0.0058 0.0349 0.0099 86 518
49
- # 0.2928 0.3181 0.2928 0.3049 4816 4432
50
- G 0.0000 0.0000 0.0000 0.0000 86 0
51
- _ 0.7642 0.7622 0.7642 0.7632 14362 14400
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.6419 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0230 0.0036 0.0230 0.0062 87 554
64
- # 0.0822 0.3857 0.0822 0.1356 6177 1317
65
- G 0.0000 0.0000 0.0000 0.0000 87 0
66
- _ 0.9117 0.6810 0.9117 0.7796 13224 17704
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5789 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0515 0.0049 0.0515 0.0090 136 1415
79
- # 0.0685 0.4470 0.0685 0.1188 11560 1772
80
- G 0.0000 0.0000 0.0000 0.0000 136 0
81
- _ 0.9012 0.6170 0.9012 0.7325 18768 27413
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5194 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0700 0.0065 0.0700 0.0120 100 1071
94
- # 0.0858 0.4793 0.0858 0.1455 10000 1790
95
- G 0.0000 0.0000 0.0000 0.0000 100 0
96
- _ 0.8798 0.5510 0.8798 0.6776 12300 19639
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.4585 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0732 0.0050 0.0732 0.0093 164 2406
109
- # 0.0454 0.5535 0.0454 0.0838 18696 1532
110
- G 0.0000 0.0000 0.0000 0.0000 164 0
111
- _ 0.8983 0.4872 0.8983 0.6317 17876 32962
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.4081 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0793 0.0047 0.0793 0.0089 391 6580
124
- # 0.1047 0.5837 0.1047 0.1775 50048 8976
125
- G 0.0000 0.0000 0.0000 0.0000 391 0
126
- _ 0.8247 0.4230 0.8247 0.5592 37145 72419
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.6982 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.4388 0.1343 0.4388 0.2056 964 3150
30
+ # 0.6459 0.7223 0.6459 0.6820 101297 90585
31
+ G 0.4066 0.1049 0.4066 0.1667 964 3738
32
+ _ 0.7494 0.7133 0.7494 0.7309 113675 119427
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.9167 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.6744 0.1330 0.6744 0.2222 86 436
49
+ # 0.8609 0.9219 0.8609 0.8904 4816 4497
50
+ G 0.4186 0.1429 0.4186 0.2130 86 252
51
+ _ 0.9398 0.9529 0.9398 0.9463 14362 14165
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8518 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.5172 0.1347 0.5172 0.2138 87 334
64
+ # 0.6981 0.8907 0.6981 0.7827 6177 4841
65
+ G 0.2414 0.0882 0.2414 0.1292 87 238
66
+ _ 0.9298 0.8682 0.9298 0.8980 13224 14162
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7924 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.3824 0.1294 0.3824 0.1933 136 402
79
+ # 0.5698 0.9191 0.5698 0.7035 11560 7167
80
+ G 0.3456 0.0904 0.3456 0.1433 136 520
81
+ _ 0.9357 0.7801 0.9357 0.8508 18768 22511
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.7148 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.4800 0.1519 0.4800 0.2308 100 316
94
+ # 0.4902 0.8621 0.4902 0.6250 10000 5686
95
+ G 0.4600 0.1095 0.4600 0.1769 100 420
96
+ _ 0.9015 0.6896 0.9015 0.7815 12300 16078
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6269 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.4146 0.1399 0.4146 0.2092 164 486
109
+ # 0.5424 0.6894 0.5424 0.6072 18696 14709
110
+ G 0.3963 0.1104 0.3963 0.1726 164 589
111
+ _ 0.7192 0.6089 0.7192 0.6595 17876 21116
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.6088 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.3887 0.1293 0.3887 0.1940 391 1176
124
+ # 0.7061 0.6583 0.7061 0.6814 50048 53685
125
+ G 0.4527 0.1030 0.4527 0.1678 391 1719
126
+ _ 0.4815 0.5697 0.4815 0.5219 37145 31395
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size15.json
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5722848579991437,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 286,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.8966439688715954,
16
- "precision": 0.5924319113039287,
17
- "recall": 0.8966439688715954,
18
- "f1": 0.713463306080983,
19
  "gt_support": 8224,
20
- "predicted": 12447
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 286,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.1238022230739747,
32
- "precision": 0.412252712188896,
33
- "recall": 0.1238022230739747,
34
- "f1": 0.19042004421518052,
35
  "gt_support": 5218,
36
- "predicted": 1567
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.4940923737916219,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 38,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.7708333333333334,
56
- "precision": 0.4964689265536723,
57
- "recall": 0.7708333333333334,
58
- "f1": 0.6039518900343643,
59
  "gt_support": 912,
60
- "predicted": 1416
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 38,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.2482837528604119,
72
- "precision": 0.48654708520179374,
73
- "recall": 0.2482837528604119,
74
- "f1": 0.3287878787878788,
75
  "gt_support": 874,
76
- "predicted": 446
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
- "accuracy": 0.5236549165120594,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 44,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.8592657342657343,
95
- "precision": 0.5359869138495092,
96
- "recall": 0.8592657342657343,
97
- "f1": 0.6601746138347885,
98
  "gt_support": 1144,
99
- "predicted": 1834
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 44,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.15800865800865802,
111
- "precision": 0.453416149068323,
112
- "recall": 0.15800865800865802,
113
- "f1": 0.23434991974317818,
114
  "gt_support": 924,
115
- "predicted": 322
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
- "accuracy": 0.5440517670482827,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 41,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.8327526132404182,
134
- "precision": 0.5745192307692307,
135
- "recall": 0.8327526132404182,
136
- "f1": 0.6799431009957325,
137
  "gt_support": 1148,
138
- "predicted": 1664
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 41,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.17586649550706032,
150
- "precision": 0.39710144927536234,
151
- "recall": 0.17586649550706032,
152
- "f1": 0.24377224199288255,
153
  "gt_support": 779,
154
- "predicted": 345
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
- "accuracy": 0.5796786799826313,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 47,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.9574468085106383,
173
- "precision": 0.5899638336347197,
174
- "recall": 0.9574468085106383,
175
- "f1": 0.7300699300699302,
176
  "gt_support": 1363,
177
- "predicted": 2212
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 47,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.03546099290780142,
189
- "precision": 0.32967032967032966,
190
- "recall": 0.03546099290780142,
191
- "f1": 0.06403415154749201,
192
  "gt_support": 846,
193
- "predicted": 91
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
- "accuracy": 0.6103896103896104,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 55,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.9237536656891495,
212
- "precision": 0.6338028169014085,
213
- "recall": 0.9237536656891495,
214
- "f1": 0.7517899761336515,
215
  "gt_support": 1705,
216
- "predicted": 2485
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 55,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.07954545454545454,
228
- "precision": 0.3333333333333333,
229
- "recall": 0.07954545454545454,
230
- "f1": 0.12844036697247707,
231
  "gt_support": 880,
232
- "predicted": 210
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
- "accuracy": 0.6349949815991971,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 61,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.9487704918032787,
251
- "precision": 0.6530324400564175,
252
- "recall": 0.9487704918032787,
253
- "f1": 0.7736006683375104,
254
  "gt_support": 1952,
255
- "predicted": 2836
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 61,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.05027322404371585,
267
- "precision": 0.3006535947712418,
268
- "recall": 0.05027322404371585,
269
- "f1": 0.08614232209737827,
270
  "gt_support": 915,
271
- "predicted": 153
272
  }
273
  },
274
  "total_samples": 2989
@@ -278,13 +278,13 @@
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 7
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.8861138861138861,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.993006993006993,
8
+ "precision": 0.8114285714285714,
9
+ "recall": 0.993006993006993,
10
+ "f1": 0.8930817610062893,
11
  "gt_support": 286,
12
+ "predicted": 350
13
  },
14
  "1": {
15
+ "accuracy": 0.8503161478599222,
16
+ "precision": 0.9611050027487631,
17
+ "recall": 0.8503161478599222,
18
+ "f1": 0.9023225806451612,
19
  "gt_support": 8224,
20
+ "predicted": 7276
21
  },
22
  "2": {
23
+ "accuracy": 0.986013986013986,
24
+ "precision": 0.7943661971830986,
25
+ "recall": 0.986013986013986,
26
+ "f1": 0.8798751950078004,
27
  "gt_support": 286,
28
+ "predicted": 355
29
  },
30
  "3": {
31
+ "accuracy": 0.9311996933691069,
32
+ "precision": 0.8054036134593071,
33
+ "recall": 0.9311996933691069,
34
+ "f1": 0.8637454448493467,
35
  "gt_support": 5218,
36
+ "predicted": 6033
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.985499462943072,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 1.0,
48
+ "precision": 0.8837209302325582,
49
+ "recall": 1.0,
50
+ "f1": 0.9382716049382717,
51
  "gt_support": 38,
52
+ "predicted": 43
53
  },
54
  "1": {
55
+ "accuracy": 0.9912280701754386,
56
+ "precision": 0.9890590809628009,
57
+ "recall": 0.9912280701754386,
58
+ "f1": 0.9901423877327493,
59
  "gt_support": 912,
60
+ "predicted": 914
61
  },
62
  "2": {
63
+ "accuracy": 1.0,
64
+ "precision": 0.8444444444444444,
65
+ "recall": 1.0,
66
+ "f1": 0.9156626506024096,
67
  "gt_support": 38,
68
+ "predicted": 45
69
  },
70
  "3": {
71
+ "accuracy": 0.9782608695652174,
72
+ "precision": 0.9941860465116279,
73
+ "recall": 0.9782608695652174,
74
+ "f1": 0.986159169550173,
75
  "gt_support": 874,
76
+ "predicted": 860
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
+ "accuracy": 0.9392393320964749,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 1.0,
87
+ "precision": 0.8,
88
+ "recall": 1.0,
89
+ "f1": 0.888888888888889,
90
  "gt_support": 44,
91
+ "predicted": 55
92
  },
93
  "1": {
94
+ "accuracy": 0.9213286713286714,
95
+ "precision": 0.975925925925926,
96
+ "recall": 0.9213286713286714,
97
+ "f1": 0.947841726618705,
98
  "gt_support": 1144,
99
+ "predicted": 1080
100
  },
101
  "2": {
102
+ "accuracy": 1.0,
103
+ "precision": 0.8148148148148148,
104
+ "recall": 1.0,
105
+ "f1": 0.8979591836734693,
106
  "gt_support": 44,
107
+ "predicted": 54
108
  },
109
  "3": {
110
+ "accuracy": 0.9556277056277056,
111
+ "precision": 0.9131334022750776,
112
+ "recall": 0.9556277056277056,
113
+ "f1": 0.933897408778424,
114
  "gt_support": 924,
115
+ "predicted": 967
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
+ "accuracy": 0.8989547038327527,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 1.0,
126
+ "precision": 0.8723404255319149,
127
+ "recall": 1.0,
128
+ "f1": 0.9318181818181819,
129
  "gt_support": 41,
130
+ "predicted": 47
131
  },
132
  "1": {
133
+ "accuracy": 0.8554006968641115,
134
+ "precision": 0.9771144278606965,
135
+ "recall": 0.8554006968641115,
136
+ "f1": 0.9122155132373432,
137
  "gt_support": 1148,
138
+ "predicted": 1005
139
  },
140
  "2": {
141
+ "accuracy": 1.0,
142
+ "precision": 0.7192982456140351,
143
+ "recall": 1.0,
144
+ "f1": 0.8367346938775511,
145
  "gt_support": 41,
146
+ "predicted": 57
147
  },
148
  "3": {
149
+ "accuracy": 0.9525032092426188,
150
+ "precision": 0.8244444444444444,
151
+ "recall": 0.9525032092426188,
152
+ "f1": 0.8838594401429422,
153
  "gt_support": 779,
154
+ "predicted": 900
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
+ "accuracy": 0.8758141554494138,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.9787234042553191,
165
+ "precision": 0.6865671641791045,
166
+ "recall": 0.9787234042553191,
167
+ "f1": 0.807017543859649,
168
  "gt_support": 47,
169
+ "predicted": 67
170
  },
171
  "1": {
172
+ "accuracy": 0.8341892883345561,
173
+ "precision": 0.9635593220338983,
174
+ "recall": 0.8341892883345561,
175
+ "f1": 0.8942194258749508,
176
  "gt_support": 1363,
177
+ "predicted": 1180
178
  },
179
  "2": {
180
+ "accuracy": 0.9574468085106383,
181
+ "precision": 0.7758620689655172,
182
+ "recall": 0.9574468085106383,
183
+ "f1": 0.8571428571428572,
184
  "gt_support": 47,
185
+ "predicted": 58
186
  },
187
  "3": {
188
+ "accuracy": 0.9326241134751773,
189
+ "precision": 0.7905811623246493,
190
+ "recall": 0.9326241134751773,
191
+ "f1": 0.8557483731019522,
192
  "gt_support": 846,
193
+ "predicted": 998
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
+ "accuracy": 0.8564007421150278,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 1.0,
204
+ "precision": 0.8461538461538461,
205
+ "recall": 1.0,
206
+ "f1": 0.9166666666666666,
207
  "gt_support": 55,
208
+ "predicted": 65
209
  },
210
  "1": {
211
+ "accuracy": 0.8087976539589443,
212
+ "precision": 0.9629888268156425,
213
+ "recall": 0.8087976539589443,
214
+ "f1": 0.8791839336946127,
215
  "gt_support": 1705,
216
+ "predicted": 1432
217
  },
218
  "2": {
219
+ "accuracy": 1.0,
220
+ "precision": 0.8461538461538461,
221
+ "recall": 1.0,
222
+ "f1": 0.9166666666666666,
223
  "gt_support": 55,
224
+ "predicted": 65
225
  },
226
  "3": {
227
+ "accuracy": 0.9306818181818182,
228
+ "precision": 0.7228596646072374,
229
+ "recall": 0.9306818181818182,
230
+ "f1": 0.8137108792846497,
231
  "gt_support": 880,
232
+ "predicted": 1133
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
+ "accuracy": 0.81197724991636,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.9836065573770492,
243
+ "precision": 0.821917808219178,
244
+ "recall": 0.9836065573770492,
245
+ "f1": 0.8955223880597014,
246
  "gt_support": 61,
247
+ "predicted": 73
248
  },
249
  "1": {
250
+ "accuracy": 0.7873975409836066,
251
+ "precision": 0.9231231231231232,
252
+ "recall": 0.7873975409836066,
253
+ "f1": 0.8498755875034559,
254
  "gt_support": 1952,
255
+ "predicted": 1665
256
  },
257
  "2": {
258
+ "accuracy": 0.9672131147540983,
259
+ "precision": 0.7763157894736842,
260
+ "recall": 0.9672131147540983,
261
+ "f1": 0.8613138686131387,
262
  "gt_support": 61,
263
+ "predicted": 76
264
  },
265
  "3": {
266
+ "accuracy": 0.8426229508196721,
267
+ "precision": 0.6561702127659574,
268
+ "recall": 0.8426229508196721,
269
+ "f1": 0.7377990430622009,
270
  "gt_support": 915,
271
+ "predicted": 1175
272
  }
273
  },
274
  "total_samples": 2989
 
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 7
290
  }
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 286 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5723 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 286 0
30
- # 0.8966 0.5924 0.8966 0.7135 8224 12447
31
- G 0.0000 0.0000 0.0000 0.0000 286 0
32
- _ 0.1238 0.4123 0.1238 0.1904 5218 1567
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.4941 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 38 0
49
- # 0.7708 0.4965 0.7708 0.6040 912 1416
50
- G 0.0000 0.0000 0.0000 0.0000 38 0
51
- _ 0.2483 0.4865 0.2483 0.3288 874 446
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.5237 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 44 0
64
- # 0.8593 0.5360 0.8593 0.6602 1144 1834
65
- G 0.0000 0.0000 0.0000 0.0000 44 0
66
- _ 0.1580 0.4534 0.1580 0.2343 924 322
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5441 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 41 0
79
- # 0.8328 0.5745 0.8328 0.6799 1148 1664
80
- G 0.0000 0.0000 0.0000 0.0000 41 0
81
- _ 0.1759 0.3971 0.1759 0.2438 779 345
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5797 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 47 0
94
- # 0.9574 0.5900 0.9574 0.7301 1363 2212
95
- G 0.0000 0.0000 0.0000 0.0000 47 0
96
- _ 0.0355 0.3297 0.0355 0.0640 846 91
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.6104 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 55 0
109
- # 0.9238 0.6338 0.9238 0.7518 1705 2485
110
- G 0.0000 0.0000 0.0000 0.0000 55 0
111
- _ 0.0795 0.3333 0.0795 0.1284 880 210
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.6350 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 61 0
124
- # 0.9488 0.6530 0.9488 0.7736 1952 2836
125
- G 0.0000 0.0000 0.0000 0.0000 61 0
126
- _ 0.0503 0.3007 0.0503 0.0861 915 153
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.8861 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.9930 0.8114 0.9930 0.8931 286 350
30
+ # 0.8503 0.9611 0.8503 0.9023 8224 7276
31
+ G 0.9860 0.7944 0.9860 0.8799 286 355
32
+ _ 0.9312 0.8054 0.9312 0.8637 5218 6033
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.9855 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 1.0000 0.8837 1.0000 0.9383 38 43
49
+ # 0.9912 0.9891 0.9912 0.9901 912 914
50
+ G 1.0000 0.8444 1.0000 0.9157 38 45
51
+ _ 0.9783 0.9942 0.9783 0.9862 874 860
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.9392 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 1.0000 0.8000 1.0000 0.8889 44 55
64
+ # 0.9213 0.9759 0.9213 0.9478 1144 1080
65
+ G 1.0000 0.8148 1.0000 0.8980 44 54
66
+ _ 0.9556 0.9131 0.9556 0.9339 924 967
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.8990 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 1.0000 0.8723 1.0000 0.9318 41 47
79
+ # 0.8554 0.9771 0.8554 0.9122 1148 1005
80
+ G 1.0000 0.7193 1.0000 0.8367 41 57
81
+ _ 0.9525 0.8244 0.9525 0.8839 779 900
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.8758 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.9787 0.6866 0.9787 0.8070 47 67
94
+ # 0.8342 0.9636 0.8342 0.8942 1363 1180
95
+ G 0.9574 0.7759 0.9574 0.8571 47 58
96
+ _ 0.9326 0.7906 0.9326 0.8557 846 998
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.8564 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 1.0000 0.8462 1.0000 0.9167 55 65
109
+ # 0.8088 0.9630 0.8088 0.8792 1705 1432
110
+ G 1.0000 0.8462 1.0000 0.9167 55 65
111
+ _ 0.9307 0.7229 0.9307 0.8137 880 1133
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.8120 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.9836 0.8219 0.9836 0.8955 61 73
124
+ # 0.7874 0.9231 0.7874 0.8499 1952 1665
125
+ G 0.9672 0.7763 0.9672 0.8613 61 76
126
+ _ 0.8426 0.6562 0.8426 0.7378 915 1175
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size7.json
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5443867992506125,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 514,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.9962505513895016,
16
- "precision": 0.544676232967563,
17
- "recall": 0.9962505513895016,
18
- "f1": 0.7042956264130351,
19
  "gt_support": 22670,
20
- "predicted": 41465
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 514,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.0044603033006244425,
32
- "precision": 0.47337278106508873,
33
- "recall": 0.0044603033006244425,
34
- "f1": 0.00883733775200221,
35
  "gt_support": 17936,
36
- "predicted": 169
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.39871970736168266,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 54,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.9820601851851852,
56
- "precision": 0.3952026082906381,
57
- "recall": 0.9820601851851852,
58
- "f1": 0.5636001328462306,
59
  "gt_support": 1728,
60
- "predicted": 4294
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 54,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.018518518518518517,
72
- "precision": 0.5875,
73
- "recall": 0.018518518518518517,
74
- "f1": 0.03590527119938885,
75
  "gt_support": 2538,
76
- "predicted": 80
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
- "accuracy": 0.4444444444444444,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 47,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 1.0,
95
- "precision": 0.4444444444444444,
96
- "recall": 1.0,
97
- "f1": 0.6153846153846153,
98
  "gt_support": 1692,
99
- "predicted": 3807
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 47,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.0,
111
- "precision": 0.0,
112
- "recall": 0.0,
113
- "f1": 0.0,
114
  "gt_support": 2021,
115
- "predicted": 0
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
- "accuracy": 0.49382716049382713,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 67,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 1.0,
134
- "precision": 0.49382716049382713,
135
- "recall": 1.0,
136
- "f1": 0.6611570247933884,
137
  "gt_support": 2680,
138
- "predicted": 5427
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 67,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.0,
150
- "precision": 0.0,
151
- "recall": 0.0,
152
- "f1": 0.0,
153
  "gt_support": 2613,
154
- "predicted": 0
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
- "accuracy": 0.5308641975308642,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 61,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 1.0,
173
- "precision": 0.5308641975308642,
174
- "recall": 1.0,
175
- "f1": 0.6935483870967741,
176
  "gt_support": 2623,
177
- "predicted": 4941
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 61,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.0,
189
- "precision": 0.0,
190
- "recall": 0.0,
191
- "f1": 0.0,
192
  "gt_support": 2196,
193
- "predicted": 0
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
- "accuracy": 0.5781689280039115,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 101,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.9896776911733727,
212
- "precision": 0.5801432452457397,
213
- "recall": 0.9896776911733727,
214
- "f1": 0.7314908524717788,
215
  "gt_support": 4747,
216
- "predicted": 8098
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 101,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.009900990099009901,
228
- "precision": 0.3855421686746988,
229
- "recall": 0.009900990099009901,
230
- "f1": 0.019306184012066366,
231
  "gt_support": 3232,
232
- "predicted": 83
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
- "accuracy": 0.6170155662909286,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 184,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.9994565217391305,
251
- "precision": 0.617196939186468,
252
- "recall": 0.9994565217391305,
253
- "f1": 0.7631338700307079,
254
  "gt_support": 9200,
255
- "predicted": 14898
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 184,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.0001874062968515742,
267
- "precision": 0.16666666666666666,
268
- "recall": 0.0001874062968515742,
269
- "f1": 0.00037439161362785476,
270
  "gt_support": 5336,
271
- "predicted": 6
272
  }
273
  },
274
  "total_samples": 14904
@@ -278,13 +278,13 @@
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 9
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7913003794975261,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.9688715953307393,
8
+ "precision": 0.6140567200986436,
9
+ "recall": 0.9688715953307393,
10
+ "f1": 0.7516981132075471,
11
  "gt_support": 514,
12
+ "predicted": 811
13
  },
14
  "1": {
15
+ "accuracy": 0.7635641817379797,
16
+ "precision": 0.8560407497156421,
17
+ "recall": 0.7635641817379797,
18
+ "f1": 0.8071623417500176,
19
  "gt_support": 22670,
20
+ "predicted": 20221
21
  },
22
  "2": {
23
+ "accuracy": 0.9319066147859922,
24
+ "precision": 0.6017587939698492,
25
+ "recall": 0.9319066147859922,
26
+ "f1": 0.731297709923664,
27
  "gt_support": 514,
28
+ "predicted": 796
29
  },
30
  "3": {
31
+ "accuracy": 0.8172390722569135,
32
+ "precision": 0.7400787640109058,
33
+ "recall": 0.8172390722569135,
34
+ "f1": 0.7767473901754014,
35
  "gt_support": 17936,
36
+ "predicted": 19806
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8701417466849566,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.9629629629629629,
48
+ "precision": 0.27956989247311825,
49
+ "recall": 0.9629629629629629,
50
+ "f1": 0.43333333333333324,
51
  "gt_support": 54,
52
+ "predicted": 186
53
  },
54
  "1": {
55
+ "accuracy": 0.7980324074074074,
56
+ "precision": 0.9503790489317712,
57
+ "recall": 0.7980324074074074,
58
+ "f1": 0.8675684177414281,
59
  "gt_support": 1728,
60
+ "predicted": 1451
61
  },
62
  "2": {
63
+ "accuracy": 0.9259259259259259,
64
+ "precision": 0.4098360655737705,
65
+ "recall": 0.9259259259259259,
66
+ "f1": 0.5681818181818182,
67
  "gt_support": 54,
68
+ "predicted": 122
69
  },
70
  "3": {
71
+ "accuracy": 0.9160756501182034,
72
+ "precision": 0.8891013384321224,
73
+ "recall": 0.9160756501182034,
74
+ "f1": 0.9023869590529788,
75
  "gt_support": 2538,
76
+ "predicted": 2615
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
+ "accuracy": 0.899658523771999,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.9574468085106383,
87
+ "precision": 0.7258064516129032,
88
+ "recall": 0.9574468085106383,
89
+ "f1": 0.8256880733944956,
90
  "gt_support": 47,
91
+ "predicted": 62
92
  },
93
  "1": {
94
+ "accuracy": 0.8859338061465721,
95
+ "precision": 0.9019253910950662,
96
+ "recall": 0.8859338061465721,
97
+ "f1": 0.8938580799045917,
98
  "gt_support": 1692,
99
+ "predicted": 1662
100
  },
101
  "2": {
102
+ "accuracy": 0.9574468085106383,
103
+ "precision": 0.8035714285714286,
104
+ "recall": 0.9574468085106383,
105
+ "f1": 0.8737864077669905,
106
  "gt_support": 47,
107
+ "predicted": 56
108
  },
109
  "3": {
110
+ "accuracy": 0.9084611578426521,
111
+ "precision": 0.9057720769610261,
112
+ "recall": 0.9084611578426521,
113
+ "f1": 0.9071146245059288,
114
  "gt_support": 2021,
115
+ "predicted": 2027
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
+ "accuracy": 0.8380320619126589,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.9701492537313433,
126
+ "precision": 0.8333333333333334,
127
+ "recall": 0.9701492537313433,
128
+ "f1": 0.896551724137931,
129
  "gt_support": 67,
130
+ "predicted": 78
131
  },
132
  "1": {
133
+ "accuracy": 0.7985074626865671,
134
+ "precision": 0.8734693877551021,
135
+ "recall": 0.7985074626865671,
136
+ "f1": 0.834307992202729,
137
  "gt_support": 2680,
138
+ "predicted": 2450
139
  },
140
  "2": {
141
+ "accuracy": 0.8656716417910447,
142
+ "precision": 0.7341772151898734,
143
+ "recall": 0.8656716417910447,
144
+ "f1": 0.7945205479452054,
145
  "gt_support": 67,
146
+ "predicted": 79
147
  },
148
  "3": {
149
+ "accuracy": 0.8744737849215461,
150
+ "precision": 0.8102836879432624,
151
+ "recall": 0.8744737849215461,
152
+ "f1": 0.8411558991349164,
153
  "gt_support": 2613,
154
+ "predicted": 2820
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
+ "accuracy": 0.810159886662619,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.9508196721311475,
165
+ "precision": 0.7631578947368421,
166
+ "recall": 0.9508196721311475,
167
+ "f1": 0.8467153284671532,
168
  "gt_support": 61,
169
+ "predicted": 76
170
  },
171
  "1": {
172
+ "accuracy": 0.7460922607701106,
173
+ "precision": 0.8927919708029197,
174
+ "recall": 0.7460922607701106,
175
+ "f1": 0.8128764278296989,
176
  "gt_support": 2623,
177
+ "predicted": 2192
178
  },
179
  "2": {
180
+ "accuracy": 0.9180327868852459,
181
+ "precision": 0.6829268292682927,
182
+ "recall": 0.9180327868852459,
183
+ "f1": 0.7832167832167832,
184
  "gt_support": 61,
185
+ "predicted": 82
186
  },
187
  "3": {
188
+ "accuracy": 0.8797814207650273,
189
+ "precision": 0.7456580470860672,
190
+ "recall": 0.8797814207650273,
191
+ "f1": 0.8071861290996449,
192
  "gt_support": 2196,
193
+ "predicted": 2591
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
+ "accuracy": 0.7731328688424398,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.9801980198019802,
204
+ "precision": 0.7557251908396947,
205
+ "recall": 0.9801980198019802,
206
+ "f1": 0.853448275862069,
207
  "gt_support": 101,
208
+ "predicted": 131
209
  },
210
  "1": {
211
+ "accuracy": 0.717295133768696,
212
+ "precision": 0.8793904958677686,
213
+ "recall": 0.717295133768696,
214
+ "f1": 0.7901148625130525,
215
  "gt_support": 4747,
216
+ "predicted": 3872
217
  },
218
  "2": {
219
+ "accuracy": 0.9504950495049505,
220
+ "precision": 0.6114649681528662,
221
+ "recall": 0.9504950495049505,
222
+ "f1": 0.7441860465116279,
223
  "gt_support": 101,
224
+ "predicted": 157
225
  },
226
  "3": {
227
+ "accuracy": 0.8431311881188119,
228
+ "precision": 0.677692116388958,
229
+ "recall": 0.8431311881188119,
230
+ "f1": 0.7514132083275886,
231
  "gt_support": 3232,
232
+ "predicted": 4021
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
+ "accuracy": 0.727187332259796,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.9728260869565217,
243
+ "precision": 0.6438848920863309,
244
+ "recall": 0.9728260869565217,
245
+ "f1": 0.7748917748917749,
246
  "gt_support": 184,
247
+ "predicted": 278
248
  },
249
  "1": {
250
+ "accuracy": 0.7532608695652174,
251
+ "precision": 0.8063765417733302,
252
+ "recall": 0.7532608695652174,
253
+ "f1": 0.7789142407553108,
254
  "gt_support": 9200,
255
+ "predicted": 8594
256
  },
257
  "2": {
258
+ "accuracy": 0.9456521739130435,
259
+ "precision": 0.58,
260
+ "recall": 0.9456521739130435,
261
+ "f1": 0.7190082644628099,
262
  "gt_support": 184,
263
+ "predicted": 300
264
  },
265
  "3": {
266
+ "accuracy": 0.6662293853073463,
267
+ "precision": 0.6202023726448012,
268
+ "recall": 0.6662293853073463,
269
+ "f1": 0.6423924828333936,
270
  "gt_support": 5336,
271
+ "predicted": 5732
272
  }
273
  },
274
  "total_samples": 14904
 
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "15",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 9
290
  }
layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 514 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5444 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 514 0
30
- # 0.9963 0.5447 0.9963 0.7043 22670 41465
31
- G 0.0000 0.0000 0.0000 0.0000 514 0
32
- _ 0.0045 0.4734 0.0045 0.0088 17936 169
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.3987 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 54 0
49
- # 0.9821 0.3952 0.9821 0.5636 1728 4294
50
- G 0.0000 0.0000 0.0000 0.0000 54 0
51
- _ 0.0185 0.5875 0.0185 0.0359 2538 80
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.4444 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 47 0
64
- # 1.0000 0.4444 1.0000 0.6154 1692 3807
65
- G 0.0000 0.0000 0.0000 0.0000 47 0
66
- _ 0.0000 0.0000 0.0000 0.0000 2021 0
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.4938 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 67 0
79
- # 1.0000 0.4938 1.0000 0.6612 2680 5427
80
- G 0.0000 0.0000 0.0000 0.0000 67 0
81
- _ 0.0000 0.0000 0.0000 0.0000 2613 0
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5309 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 61 0
94
- # 1.0000 0.5309 1.0000 0.6935 2623 4941
95
- G 0.0000 0.0000 0.0000 0.0000 61 0
96
- _ 0.0000 0.0000 0.0000 0.0000 2196 0
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.5782 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 101 0
109
- # 0.9897 0.5801 0.9897 0.7315 4747 8098
110
- G 0.0000 0.0000 0.0000 0.0000 101 0
111
- _ 0.0099 0.3855 0.0099 0.0193 3232 83
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.6170 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 184 0
124
- # 0.9995 0.6172 0.9995 0.7631 9200 14898
125
- G 0.0000 0.0000 0.0000 0.0000 184 0
126
- _ 0.0002 0.1667 0.0002 0.0004 5336 6
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.7913 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.9689 0.6141 0.9689 0.7517 514 811
30
+ # 0.7636 0.8560 0.7636 0.8072 22670 20221
31
+ G 0.9319 0.6018 0.9319 0.7313 514 796
32
+ _ 0.8172 0.7401 0.8172 0.7767 17936 19806
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8701 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.9630 0.2796 0.9630 0.4333 54 186
49
+ # 0.7980 0.9504 0.7980 0.8676 1728 1451
50
+ G 0.9259 0.4098 0.9259 0.5682 54 122
51
+ _ 0.9161 0.8891 0.9161 0.9024 2538 2615
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8997 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.9574 0.7258 0.9574 0.8257 47 62
64
+ # 0.8859 0.9019 0.8859 0.8939 1692 1662
65
+ G 0.9574 0.8036 0.9574 0.8738 47 56
66
+ _ 0.9085 0.9058 0.9085 0.9071 2021 2027
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.8380 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.9701 0.8333 0.9701 0.8966 67 78
79
+ # 0.7985 0.8735 0.7985 0.8343 2680 2450
80
+ G 0.8657 0.7342 0.8657 0.7945 67 79
81
+ _ 0.8745 0.8103 0.8745 0.8412 2613 2820
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.8102 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.9508 0.7632 0.9508 0.8467 61 76
94
+ # 0.7461 0.8928 0.7461 0.8129 2623 2192
95
+ G 0.9180 0.6829 0.9180 0.7832 61 82
96
+ _ 0.8798 0.7457 0.8798 0.8072 2196 2591
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.7731 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.9802 0.7557 0.9802 0.8534 101 131
109
+ # 0.7173 0.8794 0.7173 0.7901 4747 3872
110
+ G 0.9505 0.6115 0.9505 0.7442 101 157
111
+ _ 0.8431 0.6777 0.8431 0.7514 3232 4021
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.7272 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.9728 0.6439 0.9728 0.7749 184 278
124
+ # 0.7533 0.8064 0.7533 0.7789 9200 8594
125
+ G 0.9457 0.5800 0.9457 0.7190 184 300
126
+ _ 0.6662 0.6202 0.6662 0.6424 5336 5732
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_size9.json
layer15/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.json CHANGED
@@ -46,243 +46,6 @@
46
  },
47
  "total_samples": 725175
48
  },
49
- "by_size": {
50
- "7": {
51
- "accuracy": 0.9397047397047397,
52
- "baseline_accuracy": 0.7822222222222223,
53
- "per_class": {
54
- "0": {
55
- "accuracy": 0.9965034965034965,
56
- "precision": 0.19507186858316222,
57
- "recall": 0.9965034965034965,
58
- "f1": 0.32627361190612475,
59
- "gt_support": 286,
60
- "predicted": 1461
61
- },
62
- "1": {
63
- "accuracy": 0.8340223735408561,
64
- "precision": 0.8297846600532301,
65
- "recall": 0.8340223735408561,
66
- "f1": 0.8318981200727715,
67
- "gt_support": 8224,
68
- "predicted": 8266
69
- },
70
- "2": {
71
- "accuracy": 0.986013986013986,
72
- "precision": 0.29559748427672955,
73
- "recall": 0.986013986013986,
74
- "f1": 0.45483870967741935,
75
- "gt_support": 286,
76
- "predicted": 954
77
- },
78
- "3": {
79
- "accuracy": 0.5199310080490609,
80
- "precision": 0.8130056937368895,
81
- "recall": 0.5199310080490609,
82
- "f1": 0.634248977206312,
83
- "gt_support": 5218,
84
- "predicted": 3337
85
- },
86
- "4": {
87
- "accuracy": 0.9999006675143038,
88
- "precision": 0.9999801319240245,
89
- "recall": 0.9999006675143038,
90
- "f1": 0.9999403981404219,
91
- "gt_support": 50336,
92
- "predicted": 50332
93
- }
94
- },
95
- "total_samples": 64350
96
- },
97
- "9": {
98
- "accuracy": 0.8894076956333765,
99
- "baseline_accuracy": 0.64,
100
- "per_class": {
101
- "0": {
102
- "accuracy": 0.9941634241245136,
103
- "precision": 0.12567634038366945,
104
- "recall": 0.9941634241245136,
105
- "f1": 0.22314410480349348,
106
- "gt_support": 514,
107
- "predicted": 4066
108
- },
109
- "1": {
110
- "accuracy": 0.7103661226290251,
111
- "precision": 0.8476235591346913,
112
- "recall": 0.7103661226290251,
113
- "f1": 0.7729487148719671,
114
- "gt_support": 22670,
115
- "predicted": 18999
116
- },
117
- "2": {
118
- "accuracy": 0.9260700389105059,
119
- "precision": 0.17734724292101342,
120
- "recall": 0.9260700389105059,
121
- "f1": 0.29768605378361473,
122
- "gt_support": 514,
123
- "predicted": 2684
124
- },
125
- "3": {
126
- "accuracy": 0.657671721677074,
127
- "precision": 0.7413738922757841,
128
- "recall": 0.657671721677074,
129
- "f1": 0.6970189381629096,
130
- "gt_support": 17936,
131
- "predicted": 15911
132
- },
133
- "4": {
134
- "accuracy": 0.9994190445309122,
135
- "precision": 0.9997702392215164,
136
- "recall": 0.9994190445309122,
137
- "f1": 0.9995946110292827,
138
- "gt_support": 74016,
139
- "predicted": 73990
140
- }
141
- },
142
- "total_samples": 115650
143
- },
144
- "11": {
145
- "accuracy": 0.8638178394110597,
146
- "baseline_accuracy": 0.4622222222222222,
147
- "per_class": {
148
- "0": {
149
- "accuracy": 0.8505392912172574,
150
- "precision": 0.20744081172491544,
151
- "recall": 0.8505392912172574,
152
- "f1": 0.33353474320241694,
153
- "gt_support": 649,
154
- "predicted": 2661
155
- },
156
- "1": {
157
- "accuracy": 0.6834676397547007,
158
- "precision": 0.8738659927423535,
159
- "recall": 0.6834676397547007,
160
- "f1": 0.7670278417654921,
161
- "gt_support": 39462,
162
- "predicted": 30864
163
- },
164
- "2": {
165
- "accuracy": 0.8335901386748844,
166
- "precision": 0.14781420765027323,
167
- "recall": 0.8335901386748844,
168
- "f1": 0.2511023439313066,
169
- "gt_support": 649,
170
- "predicted": 3660
171
- },
172
- "3": {
173
- "accuracy": 0.8096851915592153,
174
- "precision": 0.7397256960402506,
175
- "recall": 0.8096851915592153,
176
- "f1": 0.7731260270509418,
177
- "gt_support": 37769,
178
- "predicted": 41341
179
- },
180
- "4": {
181
- "accuracy": 0.9999703686144364,
182
- "precision": 0.9999259248285159,
183
- "recall": 0.9999703686144364,
184
- "f1": 0.9999481462276382,
185
- "gt_support": 67496,
186
- "predicted": 67499
187
- }
188
- },
189
- "total_samples": 146025
190
- },
191
- "13": {
192
- "accuracy": 0.7796707818930041,
193
- "baseline_accuracy": 0.37721810699588476,
194
- "per_class": {
195
- "0": {
196
- "accuracy": 0.7098765432098766,
197
- "precision": 0.18087448883296633,
198
- "recall": 0.7098765432098766,
199
- "f1": 0.2882928052143395,
200
- "gt_support": 810,
201
- "predicted": 3179
202
- },
203
- "1": {
204
- "accuracy": 0.6853071164426806,
205
- "precision": 0.7796952231096821,
206
- "recall": 0.6853071164426806,
207
- "f1": 0.7294605211575232,
208
- "gt_support": 66522,
209
- "predicted": 58469
210
- },
211
- "2": {
212
- "accuracy": 0.9444444444444444,
213
- "precision": 0.08850069412309117,
214
- "recall": 0.9444444444444444,
215
- "f1": 0.1618362597842183,
216
- "gt_support": 810,
217
- "predicted": 8644
218
- },
219
- "3": {
220
- "accuracy": 0.7244865305172514,
221
- "precision": 0.749045026618943,
222
- "recall": 0.7244865305172514,
223
- "f1": 0.7365611274604043,
224
- "gt_support": 68748,
225
- "predicted": 66494
226
- },
227
- "4": {
228
- "accuracy": 1.0,
229
- "precision": 0.9977124758050325,
230
- "recall": 1.0,
231
- "f1": 0.9988549282128071,
232
- "gt_support": 45360,
233
- "predicted": 45464
234
- }
235
- },
236
- "total_samples": 182250
237
- },
238
- "15": {
239
- "accuracy": 0.6900737667127709,
240
- "baseline_accuracy": 0.5240894421392347,
241
- "per_class": {
242
- "0": {
243
- "accuracy": 0.720954356846473,
244
- "precision": 0.12246696035242291,
245
- "recall": 0.720954356846473,
246
- "f1": 0.20936888085555053,
247
- "gt_support": 964,
248
- "predicted": 5675
249
- },
250
- "1": {
251
- "accuracy": 0.6680948103102757,
252
- "precision": 0.7417035640699663,
253
- "recall": 0.6680948103102757,
254
- "f1": 0.7029775476392042,
255
- "gt_support": 101297,
256
- "predicted": 91244
257
- },
258
- "2": {
259
- "accuracy": 0.8672199170124482,
260
- "precision": 0.07033484772000673,
261
- "recall": 0.8672199170124482,
262
- "f1": 0.13011673151750971,
263
- "gt_support": 964,
264
- "predicted": 11886
265
- },
266
- "3": {
267
- "accuracy": 0.7078953155926985,
268
- "precision": 0.7444377630787733,
269
- "recall": 0.7078953155926985,
270
- "f1": 0.7257068133651982,
271
- "gt_support": 113675,
272
- "predicted": 108095
273
- },
274
- "4": {
275
- "accuracy": 0.0,
276
- "precision": 0.0,
277
- "recall": 0.0,
278
- "f1": 0.0,
279
- "gt_support": 0,
280
- "predicted": 0
281
- }
282
- },
283
- "total_samples": 216900
284
- }
285
- },
286
  "by_complexity": {
287
  "0.0": {
288
  "accuracy": 0.9509390681003584,
@@ -567,6 +330,257 @@
567
  "total_samples": 266850
568
  }
569
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
570
  "by_size_complexity": {
571
  "7_0.0": {
572
  "accuracy": 0.9653801169590643,
@@ -1978,18 +1992,5 @@
1978
  },
1979
  "total_samples": 87975
1980
  }
1981
- },
1982
- "total_trajectories": 300,
1983
- "total_steps": 3223,
1984
- "config": {
1985
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_l15_s0_suffix_-3--1_mlp_1024_full_upsample_normalize.pt",
1986
- "trajectories_dir": "reveng/trajectories_test_full",
1987
- "activations_dir": "interp/activations_test_full",
1988
- "layers": "15",
1989
- "steps": "all",
1990
- "token_categories": {
1991
- "prompt_suffix": "-3:-1"
1992
- },
1993
- "pad_to_size": 15
1994
  }
1995
  }
 
46
  },
47
  "total_samples": 725175
48
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
  "by_complexity": {
50
  "0.0": {
51
  "accuracy": 0.9509390681003584,
 
330
  "total_samples": 266850
331
  }
332
  },
333
+ "total_trajectories": 300,
334
+ "total_steps": 3223,
335
+ "single_size_mode": false,
336
+ "config": {
337
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.pt",
338
+ "trajectories_dir": "reveng/trajectories_test_full",
339
+ "activations_dir": "interp/activations_test_full",
340
+ "layers": "15",
341
+ "steps": "all",
342
+ "token_categories": {
343
+ "prompt_suffix": "all"
344
+ },
345
+ "pad_to_size": 15
346
+ },
347
+ "by_size": {
348
+ "7": {
349
+ "accuracy": 0.9397047397047397,
350
+ "baseline_accuracy": 0.7822222222222223,
351
+ "per_class": {
352
+ "0": {
353
+ "accuracy": 0.9965034965034965,
354
+ "precision": 0.19507186858316222,
355
+ "recall": 0.9965034965034965,
356
+ "f1": 0.32627361190612475,
357
+ "gt_support": 286,
358
+ "predicted": 1461
359
+ },
360
+ "1": {
361
+ "accuracy": 0.8340223735408561,
362
+ "precision": 0.8297846600532301,
363
+ "recall": 0.8340223735408561,
364
+ "f1": 0.8318981200727715,
365
+ "gt_support": 8224,
366
+ "predicted": 8266
367
+ },
368
+ "2": {
369
+ "accuracy": 0.986013986013986,
370
+ "precision": 0.29559748427672955,
371
+ "recall": 0.986013986013986,
372
+ "f1": 0.45483870967741935,
373
+ "gt_support": 286,
374
+ "predicted": 954
375
+ },
376
+ "3": {
377
+ "accuracy": 0.5199310080490609,
378
+ "precision": 0.8130056937368895,
379
+ "recall": 0.5199310080490609,
380
+ "f1": 0.634248977206312,
381
+ "gt_support": 5218,
382
+ "predicted": 3337
383
+ },
384
+ "4": {
385
+ "accuracy": 0.9999006675143038,
386
+ "precision": 0.9999801319240245,
387
+ "recall": 0.9999006675143038,
388
+ "f1": 0.9999403981404219,
389
+ "gt_support": 50336,
390
+ "predicted": 50332
391
+ }
392
+ },
393
+ "total_samples": 64350
394
+ },
395
+ "9": {
396
+ "accuracy": 0.8894076956333765,
397
+ "baseline_accuracy": 0.64,
398
+ "per_class": {
399
+ "0": {
400
+ "accuracy": 0.9941634241245136,
401
+ "precision": 0.12567634038366945,
402
+ "recall": 0.9941634241245136,
403
+ "f1": 0.22314410480349348,
404
+ "gt_support": 514,
405
+ "predicted": 4066
406
+ },
407
+ "1": {
408
+ "accuracy": 0.7103661226290251,
409
+ "precision": 0.8476235591346913,
410
+ "recall": 0.7103661226290251,
411
+ "f1": 0.7729487148719671,
412
+ "gt_support": 22670,
413
+ "predicted": 18999
414
+ },
415
+ "2": {
416
+ "accuracy": 0.9260700389105059,
417
+ "precision": 0.17734724292101342,
418
+ "recall": 0.9260700389105059,
419
+ "f1": 0.29768605378361473,
420
+ "gt_support": 514,
421
+ "predicted": 2684
422
+ },
423
+ "3": {
424
+ "accuracy": 0.657671721677074,
425
+ "precision": 0.7413738922757841,
426
+ "recall": 0.657671721677074,
427
+ "f1": 0.6970189381629096,
428
+ "gt_support": 17936,
429
+ "predicted": 15911
430
+ },
431
+ "4": {
432
+ "accuracy": 0.9994190445309122,
433
+ "precision": 0.9997702392215164,
434
+ "recall": 0.9994190445309122,
435
+ "f1": 0.9995946110292827,
436
+ "gt_support": 74016,
437
+ "predicted": 73990
438
+ }
439
+ },
440
+ "total_samples": 115650
441
+ },
442
+ "11": {
443
+ "accuracy": 0.8638178394110597,
444
+ "baseline_accuracy": 0.4622222222222222,
445
+ "per_class": {
446
+ "0": {
447
+ "accuracy": 0.8505392912172574,
448
+ "precision": 0.20744081172491544,
449
+ "recall": 0.8505392912172574,
450
+ "f1": 0.33353474320241694,
451
+ "gt_support": 649,
452
+ "predicted": 2661
453
+ },
454
+ "1": {
455
+ "accuracy": 0.6834676397547007,
456
+ "precision": 0.8738659927423535,
457
+ "recall": 0.6834676397547007,
458
+ "f1": 0.7670278417654921,
459
+ "gt_support": 39462,
460
+ "predicted": 30864
461
+ },
462
+ "2": {
463
+ "accuracy": 0.8335901386748844,
464
+ "precision": 0.14781420765027323,
465
+ "recall": 0.8335901386748844,
466
+ "f1": 0.2511023439313066,
467
+ "gt_support": 649,
468
+ "predicted": 3660
469
+ },
470
+ "3": {
471
+ "accuracy": 0.8096851915592153,
472
+ "precision": 0.7397256960402506,
473
+ "recall": 0.8096851915592153,
474
+ "f1": 0.7731260270509418,
475
+ "gt_support": 37769,
476
+ "predicted": 41341
477
+ },
478
+ "4": {
479
+ "accuracy": 0.9999703686144364,
480
+ "precision": 0.9999259248285159,
481
+ "recall": 0.9999703686144364,
482
+ "f1": 0.9999481462276382,
483
+ "gt_support": 67496,
484
+ "predicted": 67499
485
+ }
486
+ },
487
+ "total_samples": 146025
488
+ },
489
+ "13": {
490
+ "accuracy": 0.7796707818930041,
491
+ "baseline_accuracy": 0.37721810699588476,
492
+ "per_class": {
493
+ "0": {
494
+ "accuracy": 0.7098765432098766,
495
+ "precision": 0.18087448883296633,
496
+ "recall": 0.7098765432098766,
497
+ "f1": 0.2882928052143395,
498
+ "gt_support": 810,
499
+ "predicted": 3179
500
+ },
501
+ "1": {
502
+ "accuracy": 0.6853071164426806,
503
+ "precision": 0.7796952231096821,
504
+ "recall": 0.6853071164426806,
505
+ "f1": 0.7294605211575232,
506
+ "gt_support": 66522,
507
+ "predicted": 58469
508
+ },
509
+ "2": {
510
+ "accuracy": 0.9444444444444444,
511
+ "precision": 0.08850069412309117,
512
+ "recall": 0.9444444444444444,
513
+ "f1": 0.1618362597842183,
514
+ "gt_support": 810,
515
+ "predicted": 8644
516
+ },
517
+ "3": {
518
+ "accuracy": 0.7244865305172514,
519
+ "precision": 0.749045026618943,
520
+ "recall": 0.7244865305172514,
521
+ "f1": 0.7365611274604043,
522
+ "gt_support": 68748,
523
+ "predicted": 66494
524
+ },
525
+ "4": {
526
+ "accuracy": 1.0,
527
+ "precision": 0.9977124758050325,
528
+ "recall": 1.0,
529
+ "f1": 0.9988549282128071,
530
+ "gt_support": 45360,
531
+ "predicted": 45464
532
+ }
533
+ },
534
+ "total_samples": 182250
535
+ },
536
+ "15": {
537
+ "accuracy": 0.6900737667127709,
538
+ "baseline_accuracy": 0.5240894421392347,
539
+ "per_class": {
540
+ "0": {
541
+ "accuracy": 0.720954356846473,
542
+ "precision": 0.12246696035242291,
543
+ "recall": 0.720954356846473,
544
+ "f1": 0.20936888085555053,
545
+ "gt_support": 964,
546
+ "predicted": 5675
547
+ },
548
+ "1": {
549
+ "accuracy": 0.6680948103102757,
550
+ "precision": 0.7417035640699663,
551
+ "recall": 0.6680948103102757,
552
+ "f1": 0.7029775476392042,
553
+ "gt_support": 101297,
554
+ "predicted": 91244
555
+ },
556
+ "2": {
557
+ "accuracy": 0.8672199170124482,
558
+ "precision": 0.07033484772000673,
559
+ "recall": 0.8672199170124482,
560
+ "f1": 0.13011673151750971,
561
+ "gt_support": 964,
562
+ "predicted": 11886
563
+ },
564
+ "3": {
565
+ "accuracy": 0.7078953155926985,
566
+ "precision": 0.7444377630787733,
567
+ "recall": 0.7078953155926985,
568
+ "f1": 0.7257068133651982,
569
+ "gt_support": 113675,
570
+ "predicted": 108095
571
+ },
572
+ "4": {
573
+ "accuracy": 0.0,
574
+ "precision": 0.0,
575
+ "recall": 0.0,
576
+ "f1": 0.0,
577
+ "gt_support": 0,
578
+ "predicted": 0
579
+ }
580
+ },
581
+ "total_samples": 216900
582
+ }
583
+ },
584
  "by_size_complexity": {
585
  "7_0.0": {
586
  "accuracy": 0.9653801169590643,
 
1992
  },
1993
  "total_samples": 87975
1994
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
1995
  }
1996
  }
layer15/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.txt CHANGED
@@ -1,3 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  =======================================================================================
2
  EVALUATION COMPLETE
3
  =======================================================================================
@@ -685,4 +706,6 @@ A 0.6573 0.1211 0.6573 0.2045 391 2123
685
  G 0.8721 0.0644 0.8721 0.1200 391 5294
686
  _ 0.4680 0.5748 0.4680 0.5159 37145 30243
687
  + 0.0000 0.0000 0.0000 0.0000 0 0
688
- ---------------------------------------------------------------------------------------
 
 
 
1
+ Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.pt
3
+ Loaded probe: cognitive_map_probe_layer15_mlp_pre_reasoning_all_general
4
+ Input dimension: 8642
5
+ Number of classes: 5
6
+ Normalized: True
7
+
8
+ Token categories: {'prompt_suffix': 'all'}
9
+
10
+ Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
+
12
+ Processing size11: 60 trajectories
13
+
14
+ Processing size13: 60 trajectories
15
+
16
+ Processing size15: 60 trajectories
17
+
18
+ Processing size7: 60 trajectories
19
+
20
+ Processing size9: 60 trajectories
21
+
22
  =======================================================================================
23
  EVALUATION COMPLETE
24
  =======================================================================================
 
706
  G 0.8721 0.0644 0.8721 0.1200 391 5294
707
  _ 0.4680 0.5748 0.4680 0.5159 37145 30243
708
  + 0.0000 0.0000 0.0000 0.0000 0 0
709
+ ---------------------------------------------------------------------------------------
710
+
711
+ Results saved to: reveng/cognitive_map_probes_results/layer15/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer15_mlp_pre_reasoning_all_general.json
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.json CHANGED
@@ -1,272 +1,272 @@
1
  {
2
  "global": {
3
- "accuracy": 0.02735295241248456,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.22496147919876733,
8
- "precision": 0.008264462809917356,
9
- "recall": 0.22496147919876733,
10
- "f1": 0.015943215943215943,
11
  "gt_support": 649,
12
- "predicted": 17666
13
  },
14
  "1": {
15
- "accuracy": 0.017383812275100097,
16
- "precision": 0.5154019534184824,
17
- "recall": 0.017383812275100097,
18
- "f1": 0.033633221386022115,
19
  "gt_support": 39462,
20
- "predicted": 1331
21
  },
22
  "2": {
23
- "accuracy": 0.7380585516178737,
24
- "precision": 0.008264462809917356,
25
- "recall": 0.7380585516178737,
26
- "f1": 0.016345891345891348,
27
  "gt_support": 649,
28
- "predicted": 57959
29
  },
30
  "3": {
31
- "accuracy": 0.02216103153379756,
32
- "precision": 0.5321042593769867,
33
- "recall": 0.02216103153379756,
34
- "f1": 0.04254994662193076,
35
  "gt_support": 37769,
36
- "predicted": 1573
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.0319640252795333,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
  "accuracy": 0.19117647058823528,
48
- "precision": 0.008264462809917356,
49
  "recall": 0.19117647058823528,
50
- "f1": 0.015843997562461916,
51
  "gt_support": 68,
52
- "predicted": 1573
53
  },
54
  "1": {
55
- "accuracy": 0.014705882352941176,
56
- "precision": 0.3305785123966942,
57
- "recall": 0.014705882352941176,
58
- "f1": 0.028159098908834918,
59
  "gt_support": 2720,
60
- "predicted": 121
61
  },
62
  "2": {
63
- "accuracy": 0.7647058823529411,
64
- "precision": 0.008264462809917356,
65
- "recall": 0.7647058823529411,
66
- "f1": 0.016352201257861635,
67
  "gt_support": 68,
68
- "predicted": 6292
69
  },
70
  "3": {
71
- "accuracy": 0.029411764705882353,
72
- "precision": 0.6528925619834711,
73
- "recall": 0.029411764705882353,
74
- "f1": 0.056287851799073745,
75
  "gt_support": 5372,
76
- "predicted": 242
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
- "accuracy": 0.04482255712202236,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.17647058823529413,
87
- "precision": 0.008264462809917356,
88
- "recall": 0.17647058823529413,
89
- "f1": 0.015789473684210527,
90
  "gt_support": 85,
91
- "predicted": 1815
92
  },
93
  "1": {
94
- "accuracy": 0.023529411764705882,
95
- "precision": 0.3884297520661157,
96
- "recall": 0.023529411764705882,
97
- "f1": 0.04437101722917159,
98
  "gt_support": 3995,
99
- "predicted": 242
100
  },
101
  "2": {
102
- "accuracy": 0.7529411764705882,
103
- "precision": 0.008264462809917356,
104
- "recall": 0.7529411764705882,
105
- "f1": 0.016349469919529954,
106
  "gt_support": 85,
107
- "predicted": 7744
108
  },
109
  "3": {
110
- "accuracy": 0.047058823529411764,
111
- "precision": 0.5950413223140496,
112
- "recall": 0.047058823529411764,
113
- "f1": 0.0872198667474258,
114
  "gt_support": 6120,
115
- "predicted": 484
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
- "accuracy": 0.02779864763335838,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.2909090909090909,
126
- "precision": 0.008264462809917356,
127
- "recall": 0.2909090909090909,
128
- "f1": 0.01607232546459066,
129
  "gt_support": 55,
130
- "predicted": 1936
131
  },
132
  "1": {
133
- "accuracy": 0.0,
134
- "precision": 0.0,
135
- "recall": 0.0,
136
- "f1": 0.0,
137
  "gt_support": 2915,
138
- "predicted": 0
139
  },
140
  "2": {
141
- "accuracy": 0.6727272727272727,
142
- "precision": 0.008264462809917356,
143
- "recall": 0.6727272727272727,
144
- "f1": 0.016328331862312445,
145
  "gt_support": 55,
146
- "predicted": 4477
147
  },
148
  "3": {
149
- "accuracy": 0.03636363636363636,
150
- "precision": 0.5454545454545454,
151
- "recall": 0.03636363636363636,
152
- "f1": 0.06818181818181818,
153
  "gt_support": 3630,
154
- "predicted": 242
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
- "accuracy": 0.0176522506619594,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.18446601941747573,
165
- "precision": 0.008264462809917356,
166
- "recall": 0.18446601941747573,
167
- "f1": 0.01582014987510408,
168
  "gt_support": 103,
169
- "predicted": 2299
170
  },
171
  "1": {
172
- "accuracy": 0.009708737864077669,
173
- "precision": 0.49586776859504134,
174
- "recall": 0.009708737864077669,
175
- "f1": 0.0190445960958578,
176
  "gt_support": 6180,
177
- "predicted": 121
178
  },
179
  "2": {
180
- "accuracy": 0.7961165048543689,
181
- "precision": 0.008264462809917356,
182
- "recall": 0.7961165048543689,
183
- "f1": 0.016359102244389027,
184
  "gt_support": 103,
185
- "predicted": 9922
186
  },
187
  "3": {
188
- "accuracy": 0.009708737864077669,
189
- "precision": 0.48760330578512395,
190
- "recall": 0.009708737864077669,
191
- "f1": 0.01903839948370442,
192
  "gt_support": 6077,
193
- "predicted": 121
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
- "accuracy": 0.025228360156589823,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.14912280701754385,
204
- "precision": 0.008264462809917356,
205
- "recall": 0.14912280701754385,
206
- "f1": 0.015660985720865963,
207
  "gt_support": 114,
208
- "predicted": 2057
209
  },
210
  "1": {
211
- "accuracy": 0.017543859649122806,
212
- "precision": 0.5454545454545454,
213
- "recall": 0.017543859649122806,
214
- "f1": 0.03399433427762039,
215
  "gt_support": 7524,
216
- "predicted": 242
217
  },
218
  "2": {
219
- "accuracy": 0.8157894736842105,
220
- "precision": 0.008264462809917356,
221
- "recall": 0.8157894736842105,
222
- "f1": 0.016363156505674322,
223
  "gt_support": 114,
224
- "predicted": 11253
225
  },
226
  "3": {
227
- "accuracy": 0.017543859649122806,
228
- "precision": 0.4380165289256198,
229
- "recall": 0.017543859649122806,
230
- "f1": 0.03373647358370464,
231
  "gt_support": 6042,
232
- "predicted": 242
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
- "accuracy": 0.024756493506493508,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.29464285714285715,
243
- "precision": 0.008264462809917356,
244
- "recall": 0.29464285714285715,
245
- "f1": 0.016077953714981726,
246
  "gt_support": 224,
247
- "predicted": 7986
248
  },
249
  "1": {
250
- "accuracy": 0.022321428571428572,
251
- "precision": 0.5950413223140496,
252
- "recall": 0.022321428571428572,
253
- "f1": 0.04302874559254168,
254
  "gt_support": 16128,
255
- "predicted": 605
256
  },
257
  "2": {
258
- "accuracy": 0.6741071428571429,
259
- "precision": 0.008264462809917356,
260
- "recall": 0.6741071428571429,
261
- "f1": 0.01632873749662071,
262
  "gt_support": 224,
263
- "predicted": 18271
264
  },
265
  "3": {
266
- "accuracy": 0.008928571428571428,
267
- "precision": 0.3884297520661157,
268
- "recall": 0.008928571428571428,
269
- "f1": 0.01745589600742804,
270
  "gt_support": 10528,
271
  "predicted": 242
272
  }
@@ -278,13 +278,13 @@
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 11
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.21320785951686638,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.2835130970724191,
8
+ "precision": 0.008383451795152179,
9
+ "recall": 0.2835130970724191,
10
+ "f1": 0.01628534761251494,
11
  "gt_support": 649,
12
+ "predicted": 21948
13
  },
14
  "1": {
15
+ "accuracy": 0.27109624448836855,
16
+ "precision": 0.5742042831839408,
17
+ "recall": 0.27109624448836855,
18
+ "f1": 0.3683059921160897,
19
  "gt_support": 39462,
20
+ "predicted": 18631
21
  },
22
  "2": {
23
+ "accuracy": 0.34822804314329736,
24
+ "precision": 0.008254803126598,
25
+ "recall": 0.34822804314329736,
26
+ "f1": 0.01612730581225247,
27
  "gt_support": 649,
28
+ "predicted": 27378
29
  },
30
  "3": {
31
+ "accuracy": 0.14919643093542323,
32
+ "precision": 0.5330117290957246,
33
+ "recall": 0.14919643093542323,
34
+ "f1": 0.23313543368982853,
35
  "gt_support": 37769,
36
+ "predicted": 10572
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.055177442877977635,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
  "accuracy": 0.19117647058823528,
48
+ "precision": 0.009084556254367574,
49
  "recall": 0.19117647058823528,
50
+ "f1": 0.01734489659773182,
51
  "gt_support": 68,
52
+ "predicted": 1431
53
  },
54
  "1": {
55
+ "accuracy": 0.0,
56
+ "precision": 0.0,
57
+ "recall": 0.0,
58
+ "f1": 0.0,
59
  "gt_support": 2720,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.7205882352941176,
64
+ "precision": 0.00789685737308622,
65
+ "recall": 0.7205882352941176,
66
+ "f1": 0.015622509166268133,
67
  "gt_support": 68,
68
+ "predicted": 6205
69
  },
70
  "3": {
71
+ "accuracy": 0.07297096053611318,
72
+ "precision": 0.6621621621621622,
73
+ "recall": 0.07297096053611318,
74
+ "f1": 0.13145539906103287,
75
  "gt_support": 5372,
76
+ "predicted": 592
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
+ "accuracy": 0.1545940690325717,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.21176470588235294,
87
+ "precision": 0.0081374321880651,
88
+ "recall": 0.21176470588235294,
89
+ "f1": 0.01567261645624728,
90
  "gt_support": 85,
91
+ "predicted": 2212
92
  },
93
  "1": {
94
+ "accuracy": 0.008010012515644555,
95
+ "precision": 0.5,
96
+ "recall": 0.008010012515644555,
97
+ "f1": 0.01576743040157674,
98
  "gt_support": 3995,
99
+ "predicted": 64
100
  },
101
  "2": {
102
+ "accuracy": 0.5529411764705883,
103
+ "precision": 0.008542348237004725,
104
+ "recall": 0.5529411764705883,
105
+ "f1": 0.01682477179165921,
106
  "gt_support": 85,
107
+ "predicted": 5502
108
  },
109
  "3": {
110
+ "accuracy": 0.24395424836601307,
111
+ "precision": 0.5955325089748703,
112
+ "recall": 0.24395424836601307,
113
+ "f1": 0.3461226382288165,
114
  "gt_support": 6120,
115
+ "predicted": 2507
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
+ "accuracy": 0.29962434259954923,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.09090909090909091,
126
+ "precision": 0.00865051903114187,
127
+ "recall": 0.09090909090909091,
128
+ "f1": 0.015797788309636653,
129
  "gt_support": 55,
130
+ "predicted": 578
131
  },
132
  "1": {
133
+ "accuracy": 0.01818181818181818,
134
+ "precision": 0.4380165289256198,
135
+ "recall": 0.01818181818181818,
136
+ "f1": 0.03491436100131752,
137
  "gt_support": 2915,
138
+ "predicted": 121
139
  },
140
  "2": {
141
+ "accuracy": 0.36363636363636365,
142
+ "precision": 0.008173273395995096,
143
+ "recall": 0.36363636363636365,
144
+ "f1": 0.01598721023181455,
145
  "gt_support": 55,
146
+ "predicted": 2447
147
  },
148
  "3": {
149
+ "accuracy": 0.5278236914600551,
150
+ "precision": 0.5460245084069536,
151
+ "recall": 0.5278236914600551,
152
+ "f1": 0.53676985572209,
153
  "gt_support": 3630,
154
+ "predicted": 3509
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
+ "accuracy": 0.16978255636684586,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.22330097087378642,
165
+ "precision": 0.00892510671323244,
166
+ "recall": 0.22330097087378642,
167
+ "f1": 0.017164179104477612,
168
  "gt_support": 103,
169
+ "predicted": 2577
170
  },
171
  "1": {
172
+ "accuracy": 0.148705501618123,
173
+ "precision": 0.4978331527627302,
174
+ "recall": 0.148705501618123,
175
+ "f1": 0.22900573137303762,
176
  "gt_support": 6180,
177
+ "predicted": 1846
178
  },
179
  "2": {
180
+ "accuracy": 0.47572815533980584,
181
+ "precision": 0.008560447239692523,
182
+ "recall": 0.47572815533980584,
183
+ "f1": 0.016818259824952806,
184
  "gt_support": 103,
185
+ "predicted": 5724
186
  },
187
  "3": {
188
+ "accuracy": 0.18512423893368438,
189
+ "precision": 0.48575129533678757,
190
+ "recall": 0.18512423893368438,
191
+ "f1": 0.2680805433099011,
192
  "gt_support": 6077,
193
+ "predicted": 2316
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
+ "accuracy": 0.19059011164274323,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.2982456140350877,
204
+ "precision": 0.007942069609904227,
205
+ "recall": 0.2982456140350877,
206
+ "f1": 0.015472127417519906,
207
  "gt_support": 114,
208
+ "predicted": 4281
209
  },
210
  "1": {
211
+ "accuracy": 0.2578415736310473,
212
+ "precision": 0.5440269209197981,
213
+ "recall": 0.2578415736310473,
214
+ "f1": 0.34986474301172227,
215
  "gt_support": 7524,
216
+ "predicted": 3566
217
  },
218
  "2": {
219
+ "accuracy": 0.32456140350877194,
220
+ "precision": 0.008147985025324819,
221
+ "recall": 0.32456140350877194,
222
+ "f1": 0.0158968850698174,
223
  "gt_support": 114,
224
+ "predicted": 4541
225
  },
226
  "3": {
227
+ "accuracy": 0.10228401191658391,
228
+ "precision": 0.4395448079658606,
229
+ "recall": 0.10228401191658391,
230
+ "f1": 0.16595059076262086,
231
  "gt_support": 6042,
232
+ "predicted": 1406
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
+ "accuracy": 0.2936835891381346,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.40625,
243
+ "precision": 0.008372435366639066,
244
+ "recall": 0.40625,
245
+ "f1": 0.016406742991075453,
246
  "gt_support": 224,
247
+ "predicted": 10869
248
  },
249
  "1": {
250
+ "accuracy": 0.4807787698412698,
251
+ "precision": 0.5949056314255026,
252
+ "recall": 0.4807787698412698,
253
+ "f1": 0.5317879432137713,
254
  "gt_support": 16128,
255
+ "predicted": 13034
256
  },
257
  "2": {
258
+ "accuracy": 0.10714285714285714,
259
+ "precision": 0.008110848259547145,
260
+ "recall": 0.10714285714285714,
261
+ "f1": 0.015080113100848254,
262
  "gt_support": 224,
263
+ "predicted": 2959
264
  },
265
  "3": {
266
+ "accuracy": 0.008643617021276596,
267
+ "precision": 0.3760330578512397,
268
+ "recall": 0.008643617021276596,
269
+ "f1": 0.01689879294336119,
270
  "gt_support": 10528,
271
  "predicted": 242
272
  }
 
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 11
290
  }
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 649 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.0274 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.2250 0.0083 0.2250 0.0159 649 17666
30
- # 0.0174 0.5154 0.0174 0.0336 39462 1331
31
- G 0.7381 0.0083 0.7381 0.0163 649 57959
32
- _ 0.0222 0.5321 0.0222 0.0425 37769 1573
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.0320 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.1912 0.0083 0.1912 0.0158 68 1573
49
- # 0.0147 0.3306 0.0147 0.0282 2720 121
50
- G 0.7647 0.0083 0.7647 0.0164 68 6292
51
- _ 0.0294 0.6529 0.0294 0.0563 5372 242
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.0448 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.1765 0.0083 0.1765 0.0158 85 1815
64
- # 0.0235 0.3884 0.0235 0.0444 3995 242
65
- G 0.7529 0.0083 0.7529 0.0163 85 7744
66
- _ 0.0471 0.5950 0.0471 0.0872 6120 484
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.0278 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.2909 0.0083 0.2909 0.0161 55 1936
79
- # 0.0000 0.0000 0.0000 0.0000 2915 0
80
- G 0.6727 0.0083 0.6727 0.0163 55 4477
81
- _ 0.0364 0.5455 0.0364 0.0682 3630 242
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0177 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.1845 0.0083 0.1845 0.0158 103 2299
94
- # 0.0097 0.4959 0.0097 0.0190 6180 121
95
- G 0.7961 0.0083 0.7961 0.0164 103 9922
96
- _ 0.0097 0.4876 0.0097 0.0190 6077 121
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0252 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.1491 0.0083 0.1491 0.0157 114 2057
109
- # 0.0175 0.5455 0.0175 0.0340 7524 242
110
- G 0.8158 0.0083 0.8158 0.0164 114 11253
111
- _ 0.0175 0.4380 0.0175 0.0337 6042 242
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.0248 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.2946 0.0083 0.2946 0.0161 224 7986
124
- # 0.0223 0.5950 0.0223 0.0430 16128 605
125
- G 0.6741 0.0083 0.6741 0.0163 224 18271
126
- _ 0.0089 0.3884 0.0089 0.0175 10528 242
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2132 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.2835 0.0084 0.2835 0.0163 649 21948
30
+ # 0.2711 0.5742 0.2711 0.3683 39462 18631
31
+ G 0.3482 0.0083 0.3482 0.0161 649 27378
32
+ _ 0.1492 0.5330 0.1492 0.2331 37769 10572
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.0552 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.1912 0.0091 0.1912 0.0173 68 1431
49
+ # 0.0000 0.0000 0.0000 0.0000 2720 0
50
+ G 0.7206 0.0079 0.7206 0.0156 68 6205
51
+ _ 0.0730 0.6622 0.0730 0.1315 5372 592
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.1546 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.2118 0.0081 0.2118 0.0157 85 2212
64
+ # 0.0080 0.5000 0.0080 0.0158 3995 64
65
+ G 0.5529 0.0085 0.5529 0.0168 85 5502
66
+ _ 0.2440 0.5955 0.2440 0.3461 6120 2507
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.2996 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.0909 0.0087 0.0909 0.0158 55 578
79
+ # 0.0182 0.4380 0.0182 0.0349 2915 121
80
+ G 0.3636 0.0082 0.3636 0.0160 55 2447
81
+ _ 0.5278 0.5460 0.5278 0.5368 3630 3509
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.1698 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.2233 0.0089 0.2233 0.0172 103 2577
94
+ # 0.1487 0.4978 0.1487 0.2290 6180 1846
95
+ G 0.4757 0.0086 0.4757 0.0168 103 5724
96
+ _ 0.1851 0.4858 0.1851 0.2681 6077 2316
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.1906 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2982 0.0079 0.2982 0.0155 114 4281
109
+ # 0.2578 0.5440 0.2578 0.3499 7524 3566
110
+ G 0.3246 0.0081 0.3246 0.0159 114 4541
111
+ _ 0.1023 0.4395 0.1023 0.1660 6042 1406
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.2937 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.4062 0.0084 0.4062 0.0164 224 10869
124
+ # 0.4808 0.5949 0.4808 0.5318 16128 13034
125
+ G 0.1071 0.0081 0.1071 0.0151 224 2959
126
+ _ 0.0086 0.3760 0.0086 0.0169 10528 242
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size11.json
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.json CHANGED
@@ -1,227 +1,227 @@
1
  {
2
  "global": {
3
- "accuracy": 0.05182993644532106,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.5049382716049383,
8
- "precision": 0.005914934848945001,
9
- "recall": 0.5049382716049383,
10
- "f1": 0.011692897065340137,
11
  "gt_support": 810,
12
- "predicted": 69147
13
  },
14
  "1": {
15
- "accuracy": 0.09019572472264814,
16
- "precision": 0.4671441918405481,
17
- "recall": 0.09019572472264814,
18
- "f1": 0.1511982461003452,
19
  "gt_support": 66522,
20
- "predicted": 12844
21
  },
22
  "2": {
23
- "accuracy": 0.3962962962962963,
24
- "precision": 0.005938505938505939,
25
- "recall": 0.3962962962962963,
26
- "f1": 0.011701662292213475,
27
  "gt_support": 810,
28
- "predicted": 54054
29
  },
30
  "3": {
31
- "accuracy": 0.005309245359865014,
32
- "precision": 0.4319526627218935,
33
- "recall": 0.005309245359865014,
34
- "f1": 0.010489560731682786,
35
  "gt_support": 68748,
36
- "predicted": 845
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.04937130177514793,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.125,
48
- "precision": 0.005532503457814661,
49
- "recall": 0.125,
50
- "f1": 0.010596026490066225,
51
  "gt_support": 64,
52
- "predicted": 1446
53
  },
54
  "1": {
55
- "accuracy": 0.15625,
56
- "precision": 0.28402366863905326,
57
- "recall": 0.15625,
58
- "f1": 0.2015959680806384,
59
  "gt_support": 3072,
60
- "predicted": 1690
61
  },
62
  "2": {
63
- "accuracy": 0.71875,
64
- "precision": 0.005989583333333334,
65
- "recall": 0.71875,
66
- "f1": 0.011880165289256199,
67
  "gt_support": 64,
68
- "predicted": 7680
69
  },
70
  "3": {
71
- "accuracy": 0.0,
72
- "precision": 0.0,
73
- "recall": 0.0,
74
- "f1": 0.0,
75
  "gt_support": 7616,
76
- "predicted": 0
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
- "accuracy": 0.0548768849016988,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.3064516129032258,
87
- "precision": 0.005915317559153176,
88
- "recall": 0.3064516129032258,
89
- "f1": 0.011606597434331093,
90
  "gt_support": 62,
91
- "predicted": 3212
92
  },
93
  "1": {
94
- "accuracy": 0.14516129032258066,
95
- "precision": 0.3431952662721893,
96
- "recall": 0.14516129032258066,
97
- "f1": 0.2040257963650577,
98
  "gt_support": 3596,
99
- "predicted": 1521
100
  },
101
  "2": {
102
- "accuracy": 0.5483870967741935,
103
- "precision": 0.005918189730200174,
104
- "recall": 0.5483870967741935,
105
- "f1": 0.011710005166178749,
106
  "gt_support": 62,
107
- "predicted": 5745
108
  },
109
  "3": {
110
- "accuracy": 0.0,
111
- "precision": 0.0,
112
- "recall": 0.0,
113
- "f1": 0.0,
114
  "gt_support": 6758,
115
- "predicted": 0
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
- "accuracy": 0.03768047337278106,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.528,
126
- "precision": 0.005941663665826431,
127
- "recall": 0.528,
128
- "f1": 0.011751090536811181,
129
  "gt_support": 125,
130
- "predicted": 11108
131
  },
132
  "1": {
133
- "accuracy": 0.08,
134
- "precision": 0.40236686390532544,
135
- "recall": 0.08,
136
- "f1": 0.13346418056918546,
137
  "gt_support": 8500,
138
- "predicted": 1690
139
  },
140
  "2": {
141
- "accuracy": 0.4,
142
- "precision": 0.006004563468235859,
143
- "recall": 0.4,
144
- "f1": 0.011831519167061051,
145
  "gt_support": 125,
146
- "predicted": 8327
147
  },
148
  "3": {
149
- "accuracy": 0.0,
150
- "precision": 0.0,
151
- "recall": 0.0,
152
- "f1": 0.0,
153
  "gt_support": 12375,
154
- "predicted": 0
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
- "accuracy": 0.04410973641742873,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.47107438016528924,
165
- "precision": 0.005917159763313609,
166
- "recall": 0.47107438016528924,
167
- "f1": 0.01168751281525528,
168
  "gt_support": 121,
169
- "predicted": 9633
170
  },
171
  "1": {
172
- "accuracy": 0.0743801652892562,
173
- "precision": 0.46153846153846156,
174
- "recall": 0.0743801652892562,
175
- "f1": 0.12811387900355872,
176
  "gt_support": 9438,
177
- "predicted": 1521
178
  },
179
  "2": {
180
- "accuracy": 0.4462809917355372,
181
- "precision": 0.005917159763313609,
182
- "recall": 0.4462809917355372,
183
- "f1": 0.011679463609819401,
184
  "gt_support": 121,
185
- "predicted": 9126
186
  },
187
  "3": {
188
- "accuracy": 0.008264462809917356,
189
- "precision": 0.5266272189349113,
190
- "recall": 0.008264462809917356,
191
- "f1": 0.016273541780947157,
192
  "gt_support": 10769,
193
- "predicted": 169
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
- "accuracy": 0.06170752324598478,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.5089285714285714,
204
- "precision": 0.005917159763313609,
205
- "recall": 0.5089285714285714,
206
- "f1": 0.011698306824012312,
207
  "gt_support": 112,
208
- "predicted": 9633
209
  },
210
  "1": {
211
- "accuracy": 0.10714285714285714,
212
- "precision": 0.5266272189349113,
213
- "recall": 0.10714285714285714,
214
- "f1": 0.1780593531177059,
215
  "gt_support": 9968,
216
- "predicted": 2028
217
  },
218
  "2": {
219
- "accuracy": 0.38392857142857145,
220
- "precision": 0.005917159763313609,
221
- "recall": 0.38392857142857145,
222
- "f1": 0.011654695758232823,
223
  "gt_support": 112,
224
- "predicted": 7267
225
  },
226
  "3": {
227
  "accuracy": 0.0,
@@ -235,40 +235,40 @@
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
- "accuracy": 0.056630486078338836,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.6196319018404908,
243
- "precision": 0.0059211490546680345,
244
- "recall": 0.6196319018404908,
245
- "f1": 0.011730205278592375,
246
  "gt_support": 326,
247
- "predicted": 34115
248
  },
249
  "1": {
250
- "accuracy": 0.07975460122699386,
251
- "precision": 0.5798816568047337,
252
- "recall": 0.07975460122699386,
253
- "f1": 0.14022343294260084,
254
  "gt_support": 31948,
255
- "predicted": 4394
256
  },
257
  "2": {
258
- "accuracy": 0.2883435582822086,
259
- "precision": 0.005908605192029669,
260
- "recall": 0.2883435582822086,
261
- "f1": 0.011579919926085617,
262
  "gt_support": 326,
263
- "predicted": 15909
264
  },
265
  "3": {
266
- "accuracy": 0.012269938650306749,
267
- "precision": 0.40828402366863903,
268
- "recall": 0.012269938650306749,
269
- "f1": 0.023823910228744066,
270
  "gt_support": 22494,
271
- "predicted": 676
272
  }
273
  },
274
  "total_samples": 55094
@@ -278,13 +278,13 @@
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 13
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.2376068376068376,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.3296296296296296,
8
+ "precision": 0.005884297520661157,
9
+ "recall": 0.3296296296296296,
10
+ "f1": 0.011562195518025332,
11
  "gt_support": 810,
12
+ "predicted": 45375
13
  },
14
  "1": {
15
+ "accuracy": 0.3233366405099065,
16
+ "precision": 0.5356493587349023,
17
+ "recall": 0.3233366405099065,
18
+ "f1": 0.40325468470241943,
19
  "gt_support": 66522,
20
+ "predicted": 40155
21
  },
22
  "2": {
23
+ "accuracy": 0.26296296296296295,
24
+ "precision": 0.006029211956521739,
25
+ "recall": 0.26296296296296295,
26
+ "f1": 0.011788145442470529,
27
  "gt_support": 810,
28
+ "predicted": 35328
29
  },
30
  "3": {
31
+ "accuracy": 0.15326991330656892,
32
+ "precision": 0.657248003992016,
33
+ "recall": 0.15326991330656892,
34
+ "f1": 0.2485727765982543,
35
  "gt_support": 68748,
36
+ "predicted": 16032
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.520987426035503,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.109375,
48
+ "precision": 0.005131964809384164,
49
+ "recall": 0.109375,
50
+ "f1": 0.00980392156862745,
51
  "gt_support": 64,
52
+ "predicted": 1364
53
  },
54
  "1": {
55
+ "accuracy": 0.0026041666666666665,
56
+ "precision": 0.5714285714285714,
57
+ "recall": 0.0026041666666666665,
58
+ "f1": 0.005184705119896306,
59
  "gt_support": 3072,
60
+ "predicted": 14
61
  },
62
  "2": {
63
+ "accuracy": 0.15625,
64
+ "precision": 0.0068073519400953025,
65
+ "recall": 0.15625,
66
+ "f1": 0.01304631441617743,
67
  "gt_support": 64,
68
+ "predicted": 1469
69
  },
70
  "3": {
71
+ "accuracy": 0.7366071428571429,
72
+ "precision": 0.7039779144183712,
73
+ "recall": 0.7366071428571429,
74
+ "f1": 0.7199230028873916,
75
  "gt_support": 7616,
76
+ "predicted": 7969
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
+ "accuracy": 0.26980339759496086,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.3709677419354839,
87
+ "precision": 0.005683222139856684,
88
+ "recall": 0.3709677419354839,
89
+ "f1": 0.011194937941104892,
90
  "gt_support": 62,
91
+ "predicted": 4047
92
  },
93
  "1": {
94
+ "accuracy": 0.004171301446051168,
95
+ "precision": 0.5357142857142857,
96
+ "recall": 0.004171301446051168,
97
+ "f1": 0.008278145695364239,
98
  "gt_support": 3596,
99
+ "predicted": 28
100
  },
101
  "2": {
102
+ "accuracy": 0.16129032258064516,
103
+ "precision": 0.004728132387706856,
104
+ "recall": 0.16129032258064516,
105
+ "f1": 0.009186954524575103,
106
  "gt_support": 62,
107
+ "predicted": 2115
108
  },
109
  "3": {
110
+ "accuracy": 0.4112163361941403,
111
+ "precision": 0.6480876865671642,
112
+ "recall": 0.4112163361941403,
113
+ "f1": 0.5031685678073511,
114
  "gt_support": 6758,
115
+ "predicted": 4288
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
+ "accuracy": 0.11498224852071005,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.52,
126
+ "precision": 0.00590426015078572,
127
+ "recall": 0.52,
128
+ "f1": 0.011675947548051015,
129
  "gt_support": 125,
130
+ "predicted": 11009
131
  },
132
  "1": {
133
+ "accuracy": 0.06552941176470588,
134
+ "precision": 0.4128984432913269,
135
+ "recall": 0.06552941176470588,
136
+ "f1": 0.1131079297390598,
137
  "gt_support": 8500,
138
+ "predicted": 1349
139
  },
140
  "2": {
141
+ "accuracy": 0.264,
142
+ "precision": 0.005741127348643006,
143
+ "recall": 0.264,
144
+ "f1": 0.011237868210454622,
145
  "gt_support": 125,
146
+ "predicted": 5748
147
  },
148
  "3": {
149
+ "accuracy": 0.14335353535353534,
150
+ "precision": 0.5876117919841007,
151
+ "recall": 0.14335353535353534,
152
+ "f1": 0.23047940756138754,
153
  "gt_support": 12375,
154
+ "predicted": 3019
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
+ "accuracy": 0.20944789476258008,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.38016528925619836,
165
+ "precision": 0.006089489012443739,
166
+ "recall": 0.38016528925619836,
167
+ "f1": 0.011986970684039089,
168
  "gt_support": 121,
169
+ "predicted": 7554
170
  },
171
  "1": {
172
+ "accuracy": 0.418732782369146,
173
+ "precision": 0.46200607902735563,
174
+ "recall": 0.418732782369146,
175
+ "f1": 0.43930635838150284,
176
  "gt_support": 9438,
177
+ "predicted": 8554
178
  },
179
  "2": {
180
+ "accuracy": 0.19834710743801653,
181
+ "precision": 0.00622245268343272,
182
+ "recall": 0.19834710743801653,
183
+ "f1": 0.012066365007541479,
184
  "gt_support": 121,
185
+ "predicted": 3857
186
  },
187
  "3": {
188
+ "accuracy": 0.024236233633577863,
189
+ "precision": 0.5392561983471075,
190
+ "recall": 0.024236233633577863,
191
+ "f1": 0.046387629965342576,
192
  "gt_support": 10769,
193
+ "predicted": 484
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
+ "accuracy": 0.28883136094674555,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.29464285714285715,
204
+ "precision": 0.006238185255198488,
205
+ "recall": 0.29464285714285715,
206
+ "f1": 0.012217697149203998,
207
  "gt_support": 112,
208
+ "predicted": 5290
209
  },
210
  "1": {
211
+ "accuracy": 0.5429373996789727,
212
+ "precision": 0.527177089421391,
213
+ "recall": 0.5429373996789727,
214
+ "f1": 0.534941188099239,
215
  "gt_support": 9968,
216
+ "predicted": 10266
217
  },
218
  "2": {
219
+ "accuracy": 0.19642857142857142,
220
+ "precision": 0.006524317912218268,
221
+ "recall": 0.19642857142857142,
222
+ "f1": 0.012629161882893225,
223
  "gt_support": 112,
224
+ "predicted": 3372
225
  },
226
  "3": {
227
  "accuracy": 0.0,
 
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
+ "accuracy": 0.21572222020546702,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.2852760736196319,
243
+ "precision": 0.005772453603128298,
244
+ "recall": 0.2852760736196319,
245
+ "f1": 0.011315933564519074,
246
  "gt_support": 326,
247
+ "predicted": 16111
248
  },
249
  "1": {
250
+ "accuracy": 0.36199449104795295,
251
+ "precision": 0.5798736462093863,
252
+ "recall": 0.36199449104795295,
253
+ "f1": 0.4457334463886534,
254
  "gt_support": 31948,
255
+ "predicted": 19944
256
  },
257
  "2": {
258
+ "accuracy": 0.3496932515337423,
259
+ "precision": 0.006074492460169447,
260
+ "recall": 0.3496932515337423,
261
+ "f1": 0.011941549258890695,
262
  "gt_support": 326,
263
+ "predicted": 18767
264
  },
265
  "3": {
266
+ "accuracy": 0.005023561838712546,
267
+ "precision": 0.41544117647058826,
268
+ "recall": 0.005023561838712546,
269
+ "f1": 0.009927084248440658,
270
  "gt_support": 22494,
271
+ "predicted": 272
272
  }
273
  },
274
  "total_samples": 55094
 
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 13
290
  }
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 810 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.0518 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.5049 0.0059 0.5049 0.0117 810 69147
30
- # 0.0902 0.4671 0.0902 0.1512 66522 12844
31
- G 0.3963 0.0059 0.3963 0.0117 810 54054
32
- _ 0.0053 0.4320 0.0053 0.0105 68748 845
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.0494 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.1250 0.0055 0.1250 0.0106 64 1446
49
- # 0.1562 0.2840 0.1562 0.2016 3072 1690
50
- G 0.7188 0.0060 0.7188 0.0119 64 7680
51
- _ 0.0000 0.0000 0.0000 0.0000 7616 0
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.0549 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.3065 0.0059 0.3065 0.0116 62 3212
64
- # 0.1452 0.3432 0.1452 0.2040 3596 1521
65
- G 0.5484 0.0059 0.5484 0.0117 62 5745
66
- _ 0.0000 0.0000 0.0000 0.0000 6758 0
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.0377 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.5280 0.0059 0.5280 0.0118 125 11108
79
- # 0.0800 0.4024 0.0800 0.1335 8500 1690
80
- G 0.4000 0.0060 0.4000 0.0118 125 8327
81
- _ 0.0000 0.0000 0.0000 0.0000 12375 0
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0441 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.4711 0.0059 0.4711 0.0117 121 9633
94
- # 0.0744 0.4615 0.0744 0.1281 9438 1521
95
- G 0.4463 0.0059 0.4463 0.0117 121 9126
96
- _ 0.0083 0.5266 0.0083 0.0163 10769 169
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0617 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.5089 0.0059 0.5089 0.0117 112 9633
109
- # 0.1071 0.5266 0.1071 0.1781 9968 2028
110
- G 0.3839 0.0059 0.3839 0.0117 112 7267
111
  _ 0.0000 0.0000 0.0000 0.0000 8736 0
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.0566 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.6196 0.0059 0.6196 0.0117 326 34115
124
- # 0.0798 0.5799 0.0798 0.1402 31948 4394
125
- G 0.2883 0.0059 0.2883 0.0116 326 15909
126
- _ 0.0123 0.4083 0.0123 0.0238 22494 676
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2376 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.3296 0.0059 0.3296 0.0116 810 45375
30
+ # 0.3233 0.5356 0.3233 0.4033 66522 40155
31
+ G 0.2630 0.0060 0.2630 0.0118 810 35328
32
+ _ 0.1533 0.6572 0.1533 0.2486 68748 16032
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.5210 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.1094 0.0051 0.1094 0.0098 64 1364
49
+ # 0.0026 0.5714 0.0026 0.0052 3072 14
50
+ G 0.1562 0.0068 0.1562 0.0130 64 1469
51
+ _ 0.7366 0.7040 0.7366 0.7199 7616 7969
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.2698 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.3710 0.0057 0.3710 0.0112 62 4047
64
+ # 0.0042 0.5357 0.0042 0.0083 3596 28
65
+ G 0.1613 0.0047 0.1613 0.0092 62 2115
66
+ _ 0.4112 0.6481 0.4112 0.5032 6758 4288
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.1150 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.5200 0.0059 0.5200 0.0117 125 11009
79
+ # 0.0655 0.4129 0.0655 0.1131 8500 1349
80
+ G 0.2640 0.0057 0.2640 0.0112 125 5748
81
+ _ 0.1434 0.5876 0.1434 0.2305 12375 3019
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.2094 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.3802 0.0061 0.3802 0.0120 121 7554
94
+ # 0.4187 0.4620 0.4187 0.4393 9438 8554
95
+ G 0.1983 0.0062 0.1983 0.0121 121 3857
96
+ _ 0.0242 0.5393 0.0242 0.0464 10769 484
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2888 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2946 0.0062 0.2946 0.0122 112 5290
109
+ # 0.5429 0.5272 0.5429 0.5349 9968 10266
110
+ G 0.1964 0.0065 0.1964 0.0126 112 3372
111
  _ 0.0000 0.0000 0.0000 0.0000 8736 0
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.2157 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2853 0.0058 0.2853 0.0113 326 16111
124
+ # 0.3620 0.5799 0.3620 0.4457 31948 19944
125
+ G 0.3497 0.0061 0.3497 0.0119 326 18767
126
+ _ 0.0050 0.4154 0.0050 0.0099 22494 272
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size13.json
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.json CHANGED
@@ -1,63 +1,63 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5144213923467036,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.014522821576763486,
8
- "precision": 0.0044444444444444444,
9
- "recall": 0.014522821576763486,
10
- "f1": 0.0068060281964025275,
11
  "gt_support": 964,
12
- "predicted": 3150
13
  },
14
  "1": {
15
- "accuracy": 0.057109292476578775,
16
- "precision": 0.4851153039832285,
17
- "recall": 0.057109292476578775,
18
- "f1": 0.10218862058610519,
19
  "gt_support": 101297,
20
- "predicted": 11925
21
  },
22
  "2": {
23
- "accuracy": 0.004149377593360996,
24
- "precision": 0.0044444444444444444,
25
- "recall": 0.004149377593360996,
26
- "f1": 0.004291845493562232,
27
  "gt_support": 964,
28
- "predicted": 900
29
  },
30
  "3": {
31
- "accuracy": 0.9305036287662195,
32
- "precision": 0.5264402140102028,
33
- "recall": 0.9305036287662195,
34
- "f1": 0.6724411951684679,
35
  "gt_support": 113675,
36
- "predicted": 200925
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.7192764857881137,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
 
 
 
 
 
 
 
 
47
  "accuracy": 0.0,
48
  "precision": 0.0,
49
  "recall": 0.0,
50
  "f1": 0.0,
51
- "gt_support": 86,
52
- "predicted": 0
53
- },
54
- "1": {
55
- "accuracy": 0.046511627906976744,
56
- "precision": 0.24888888888888888,
57
- "recall": 0.046511627906976744,
58
- "f1": 0.07837648705388384,
59
  "gt_support": 4816,
60
- "predicted": 900
61
  },
62
  "2": {
63
  "accuracy": 0.0,
@@ -65,38 +65,38 @@
65
  "recall": 0.0,
66
  "f1": 0.0,
67
  "gt_support": 86,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.9534883720930233,
72
- "precision": 0.7422222222222222,
73
- "recall": 0.9534883720930233,
74
- "f1": 0.8346946239180787,
75
  "gt_support": 14362,
76
- "predicted": 18450
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
- "accuracy": 0.6631417624521073,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
 
 
 
 
 
 
 
 
86
  "accuracy": 0.0,
87
  "precision": 0.0,
88
  "recall": 0.0,
89
  "f1": 0.0,
90
- "gt_support": 87,
91
- "predicted": 0
92
- },
93
- "1": {
94
- "accuracy": 0.034482758620689655,
95
- "precision": 0.31555555555555553,
96
- "recall": 0.034482758620689655,
97
- "f1": 0.062171628721541146,
98
  "gt_support": 6177,
99
- "predicted": 675
100
  },
101
  "2": {
102
  "accuracy": 0.0,
@@ -107,168 +107,168 @@
107
  "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.9655172413793104,
111
- "precision": 0.6755555555555556,
112
- "recall": 0.9655172413793104,
113
- "f1": 0.7949196862159134,
114
  "gt_support": 13224,
115
- "predicted": 18900
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
- "accuracy": 0.5905228758169935,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.007352941176470588,
126
- "precision": 0.0044444444444444444,
127
- "recall": 0.007352941176470588,
128
- "f1": 0.00554016620498615,
129
  "gt_support": 136,
130
- "predicted": 225
131
  },
132
  "1": {
133
- "accuracy": 0.058823529411764705,
134
- "precision": 0.37777777777777777,
135
- "recall": 0.058823529411764705,
136
- "f1": 0.10179640718562873,
137
  "gt_support": 11560,
138
- "predicted": 1800
139
  },
140
  "2": {
141
- "accuracy": 0.007352941176470588,
142
- "precision": 0.0044444444444444444,
143
- "recall": 0.007352941176470588,
144
- "f1": 0.00554016620498615,
145
  "gt_support": 136,
146
- "predicted": 225
147
  },
148
  "3": {
149
- "accuracy": 0.9264705882352942,
150
- "precision": 0.6133333333333333,
151
- "recall": 0.9264705882352942,
152
- "f1": 0.738061887176875,
153
  "gt_support": 18768,
154
- "predicted": 28350
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
- "accuracy": 0.5381777777777778,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.01,
165
- "precision": 0.0044444444444444444,
166
- "recall": 0.01,
167
- "f1": 0.006153846153846155,
168
  "gt_support": 100,
169
- "predicted": 225
170
  },
171
  "1": {
172
- "accuracy": 0.03,
173
- "precision": 0.4444444444444444,
174
- "recall": 0.03,
175
- "f1": 0.05620608899297423,
176
  "gt_support": 10000,
177
- "predicted": 675
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 100,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.96,
189
- "precision": 0.5466666666666666,
190
- "recall": 0.96,
191
- "f1": 0.6966371681415928,
192
  "gt_support": 12300,
193
- "predicted": 21600
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
- "accuracy": 0.4767479674796748,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.012195121951219513,
204
- "precision": 0.0044444444444444444,
205
- "recall": 0.012195121951219513,
206
- "f1": 0.0065146579804560255,
207
  "gt_support": 164,
208
- "predicted": 450
209
  },
210
  "1": {
211
- "accuracy": 0.04878048780487805,
212
- "precision": 0.5066666666666667,
213
- "recall": 0.04878048780487805,
214
- "f1": 0.08899297423887588,
215
  "gt_support": 18696,
216
- "predicted": 1800
217
  },
218
  "2": {
219
- "accuracy": 0.006097560975609756,
220
- "precision": 0.0044444444444444444,
221
- "recall": 0.006097560975609756,
222
- "f1": 0.0051413881748071984,
223
  "gt_support": 164,
224
- "predicted": 225
225
  },
226
  "3": {
227
- "accuracy": 0.9329268292682927,
228
- "precision": 0.48444444444444446,
229
- "recall": 0.9329268292682927,
230
- "f1": 0.6377315921301696,
231
  "gt_support": 17876,
232
- "predicted": 34425
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
- "accuracy": 0.4195282750781472,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.02557544757033248,
243
- "precision": 0.0044444444444444444,
244
- "recall": 0.02557544757033248,
245
- "f1": 0.0075728890571753115,
246
  "gt_support": 391,
247
- "predicted": 2250
248
  },
249
  "1": {
250
- "accuracy": 0.06905370843989769,
251
- "precision": 0.5688888888888889,
252
- "recall": 0.06905370843989769,
253
- "f1": 0.12315806353901253,
254
  "gt_support": 50048,
255
- "predicted": 6075
256
  },
257
  "2": {
258
- "accuracy": 0.005115089514066497,
259
- "precision": 0.0044444444444444444,
260
- "recall": 0.005115089514066497,
261
- "f1": 0.0047562425683709865,
262
  "gt_support": 391,
263
- "predicted": 450
264
  },
265
  "3": {
266
- "accuracy": 0.9002557544757033,
267
  "precision": 0.4222222222222222,
268
- "recall": 0.9002557544757033,
269
- "f1": 0.5748420645494005,
270
  "gt_support": 37145,
271
- "predicted": 79200
272
  }
273
  },
274
  "total_samples": 87975
@@ -278,13 +278,13 @@
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 15
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.4284923928077455,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.1037344398340249,
8
+ "precision": 0.004375410194705754,
9
+ "recall": 0.1037344398340249,
10
+ "f1": 0.008396658130064235,
11
  "gt_support": 964,
12
+ "predicted": 22855
13
  },
14
  "1": {
15
+ "accuracy": 0.529334531131228,
16
+ "precision": 0.5327583808596467,
17
+ "recall": 0.529334531131228,
18
+ "f1": 0.531040937294187,
19
  "gt_support": 101297,
20
+ "predicted": 100646
21
  },
22
  "2": {
23
+ "accuracy": 0.16390041493775934,
24
+ "precision": 0.004505660592579918,
25
+ "recall": 0.16390041493775934,
26
+ "f1": 0.008770225639033055,
27
  "gt_support": 964,
28
+ "predicted": 35067
29
  },
30
  "3": {
31
+ "accuracy": 0.34362876621948535,
32
+ "precision": 0.6696495919906741,
33
+ "recall": 0.34362876621948535,
34
+ "f1": 0.454190817815554,
35
  "gt_support": 113675,
36
+ "predicted": 58332
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.7227390180878553,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.03488372093023256,
48
+ "precision": 0.006198347107438017,
49
+ "recall": 0.03488372093023256,
50
+ "f1": 0.010526315789473684,
51
+ "gt_support": 86,
52
+ "predicted": 484
53
+ },
54
+ "1": {
55
  "accuracy": 0.0,
56
  "precision": 0.0,
57
  "recall": 0.0,
58
  "f1": 0.0,
 
 
 
 
 
 
 
 
59
  "gt_support": 4816,
60
+ "predicted": 0
61
  },
62
  "2": {
63
  "accuracy": 0.0,
 
65
  "recall": 0.0,
66
  "f1": 0.0,
67
  "gt_support": 86,
68
+ "predicted": 37
69
  },
70
  "3": {
71
+ "accuracy": 0.9735412895139952,
72
+ "precision": 0.7425779382866854,
73
+ "recall": 0.9735412895139952,
74
+ "f1": 0.8425175499382364,
75
  "gt_support": 14362,
76
+ "predicted": 18829
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
+ "accuracy": 0.6161430395913154,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.09195402298850575,
87
+ "precision": 0.004618937644341801,
88
+ "recall": 0.09195402298850575,
89
+ "f1": 0.008796041781198461,
90
+ "gt_support": 87,
91
+ "predicted": 1732
92
+ },
93
+ "1": {
94
  "accuracy": 0.0,
95
  "precision": 0.0,
96
  "recall": 0.0,
97
  "f1": 0.0,
 
 
 
 
 
 
 
 
98
  "gt_support": 6177,
99
+ "predicted": 0
100
  },
101
  "2": {
102
  "accuracy": 0.0,
 
107
  "predicted": 0
108
  },
109
  "3": {
110
+ "accuracy": 0.9114488808227466,
111
+ "precision": 0.6755029983747127,
112
+ "recall": 0.9114488808227466,
113
+ "f1": 0.7759358805163035,
114
  "gt_support": 13224,
115
+ "predicted": 17843
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
+ "accuracy": 0.3942483660130719,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.18382352941176472,
126
+ "precision": 0.004426345609065156,
127
+ "recall": 0.18382352941176472,
128
+ "f1": 0.008644536652835409,
129
  "gt_support": 136,
130
+ "predicted": 5648
131
  },
132
  "1": {
133
+ "accuracy": 0.05519031141868512,
134
+ "precision": 0.37976190476190474,
135
+ "recall": 0.05519031141868512,
136
+ "f1": 0.09637462235649545,
137
  "gt_support": 11560,
138
+ "predicted": 1680
139
  },
140
  "2": {
141
+ "accuracy": 0.15441176470588236,
142
+ "precision": 0.0044472681067344345,
143
+ "recall": 0.15441176470588236,
144
+ "f1": 0.008645533141210375,
145
  "gt_support": 136,
146
+ "predicted": 4722
147
  },
148
  "3": {
149
+ "accuracy": 0.6063512361466326,
150
+ "precision": 0.613477088948787,
151
+ "recall": 0.6063512361466326,
152
+ "f1": 0.6098933490540758,
153
  "gt_support": 18768,
154
+ "predicted": 18550
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
+ "accuracy": 0.33466666666666667,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.08,
165
+ "precision": 0.0039662865642042635,
166
+ "recall": 0.08,
167
+ "f1": 0.007557864903164855,
168
  "gt_support": 100,
169
+ "predicted": 2017
170
  },
171
  "1": {
172
+ "accuracy": 0.6047,
173
+ "precision": 0.4450577758151174,
174
+ "recall": 0.6047,
175
+ "f1": 0.512740068681901,
176
  "gt_support": 10000,
177
+ "predicted": 13587
178
  },
179
  "2": {
180
+ "accuracy": 0.18,
181
+ "precision": 0.00424929178470255,
182
+ "recall": 0.18,
183
+ "f1": 0.008302583025830259,
184
  "gt_support": 100,
185
+ "predicted": 4236
186
  },
187
  "3": {
188
+ "accuracy": 0.11845528455284553,
189
+ "precision": 0.5477443609022556,
190
+ "recall": 0.11845528455284553,
191
+ "f1": 0.1947860962566845,
192
  "gt_support": 12300,
193
+ "predicted": 2660
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
+ "accuracy": 0.36241192411924117,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.10975609756097561,
204
+ "precision": 0.004266413842142688,
205
+ "recall": 0.10975609756097561,
206
+ "f1": 0.008213552361396304,
207
  "gt_support": 164,
208
+ "predicted": 4219
209
  },
210
  "1": {
211
+ "accuracy": 0.7128262729995721,
212
+ "precision": 0.5065566916264396,
213
+ "recall": 0.7128262729995721,
214
+ "f1": 0.5922453060771025,
215
  "gt_support": 18696,
216
+ "predicted": 26309
217
  },
218
  "2": {
219
+ "accuracy": 0.17073170731707318,
220
+ "precision": 0.004394224733207784,
221
+ "recall": 0.17073170731707318,
222
+ "f1": 0.008567931456548349,
223
  "gt_support": 164,
224
+ "predicted": 6372
225
  },
226
  "3": {
227
+ "accuracy": 0.0,
228
+ "precision": 0.0,
229
+ "recall": 0.0,
230
+ "f1": 0.0,
231
  "gt_support": 17876,
232
+ "predicted": 0
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
+ "accuracy": 0.3856436487638534,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.09718670076726342,
243
+ "precision": 0.004340376927470017,
244
+ "recall": 0.09718670076726342,
245
+ "f1": 0.008309643560026243,
246
  "gt_support": 391,
247
+ "predicted": 8755
248
  },
249
  "1": {
250
+ "accuracy": 0.6715153452685422,
251
+ "precision": 0.5689520907398002,
252
+ "recall": 0.6715153452685422,
253
+ "f1": 0.6159936948990999,
254
  "gt_support": 50048,
255
+ "predicted": 59070
256
  },
257
  "2": {
258
+ "accuracy": 0.23273657289002558,
259
+ "precision": 0.004619289340101523,
260
+ "recall": 0.23273657289002558,
261
+ "f1": 0.009058782539445523,
262
  "gt_support": 391,
263
+ "predicted": 19700
264
  },
265
  "3": {
266
+ "accuracy": 0.005115089514066497,
267
  "precision": 0.4222222222222222,
268
+ "recall": 0.005115089514066497,
269
+ "f1": 0.010107727091368535,
270
  "gt_support": 37145,
271
+ "predicted": 450
272
  }
273
  },
274
  "total_samples": 87975
 
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 15
290
  }
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 964 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5144 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0145 0.0044 0.0145 0.0068 964 3150
30
- # 0.0571 0.4851 0.0571 0.1022 101297 11925
31
- G 0.0041 0.0044 0.0041 0.0043 964 900
32
- _ 0.9305 0.5264 0.9305 0.6724 113675 200925
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.7193 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 86 0
49
- # 0.0465 0.2489 0.0465 0.0784 4816 900
50
- G 0.0000 0.0000 0.0000 0.0000 86 0
51
- _ 0.9535 0.7422 0.9535 0.8347 14362 18450
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.6631 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 87 0
64
- # 0.0345 0.3156 0.0345 0.0622 6177 675
65
  G 0.0000 0.0000 0.0000 0.0000 87 0
66
- _ 0.9655 0.6756 0.9655 0.7949 13224 18900
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5905 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0074 0.0044 0.0074 0.0055 136 225
79
- # 0.0588 0.3778 0.0588 0.1018 11560 1800
80
- G 0.0074 0.0044 0.0074 0.0055 136 225
81
- _ 0.9265 0.6133 0.9265 0.7381 18768 28350
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5382 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0100 0.0044 0.0100 0.0062 100 225
94
- # 0.0300 0.4444 0.0300 0.0562 10000 675
95
- G 0.0000 0.0000 0.0000 0.0000 100 0
96
- _ 0.9600 0.5467 0.9600 0.6966 12300 21600
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.4767 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0122 0.0044 0.0122 0.0065 164 450
109
- # 0.0488 0.5067 0.0488 0.0890 18696 1800
110
- G 0.0061 0.0044 0.0061 0.0051 164 225
111
- _ 0.9329 0.4844 0.9329 0.6377 17876 34425
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.4195 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0256 0.0044 0.0256 0.0076 391 2250
124
- # 0.0691 0.5689 0.0691 0.1232 50048 6075
125
- G 0.0051 0.0044 0.0051 0.0048 391 450
126
- _ 0.9003 0.4222 0.9003 0.5748 37145 79200
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.4285 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.1037 0.0044 0.1037 0.0084 964 22855
30
+ # 0.5293 0.5328 0.5293 0.5310 101297 100646
31
+ G 0.1639 0.0045 0.1639 0.0088 964 35067
32
+ _ 0.3436 0.6696 0.3436 0.4542 113675 58332
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.7227 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0349 0.0062 0.0349 0.0105 86 484
49
+ # 0.0000 0.0000 0.0000 0.0000 4816 0
50
+ G 0.0000 0.0000 0.0000 0.0000 86 37
51
+ _ 0.9735 0.7426 0.9735 0.8425 14362 18829
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.6161 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.0920 0.0046 0.0920 0.0088 87 1732
64
+ # 0.0000 0.0000 0.0000 0.0000 6177 0
65
  G 0.0000 0.0000 0.0000 0.0000 87 0
66
+ _ 0.9114 0.6755 0.9114 0.7759 13224 17843
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.3942 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.1838 0.0044 0.1838 0.0086 136 5648
79
+ # 0.0552 0.3798 0.0552 0.0964 11560 1680
80
+ G 0.1544 0.0044 0.1544 0.0086 136 4722
81
+ _ 0.6064 0.6135 0.6064 0.6099 18768 18550
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.3347 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.0800 0.0040 0.0800 0.0076 100 2017
94
+ # 0.6047 0.4451 0.6047 0.5127 10000 13587
95
+ G 0.1800 0.0042 0.1800 0.0083 100 4236
96
+ _ 0.1185 0.5477 0.1185 0.1948 12300 2660
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.3624 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.1098 0.0043 0.1098 0.0082 164 4219
109
+ # 0.7128 0.5066 0.7128 0.5922 18696 26309
110
+ G 0.1707 0.0044 0.1707 0.0086 164 6372
111
+ _ 0.0000 0.0000 0.0000 0.0000 17876 0
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.3856 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.0972 0.0043 0.0972 0.0083 391 8755
124
+ # 0.6715 0.5690 0.6715 0.6160 50048 59070
125
+ G 0.2327 0.0046 0.2327 0.0091 391 19700
126
+ _ 0.0051 0.4222 0.0051 0.0101 37145 450
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size15.json
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5793492222063651,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0034965034965034965,
8
- "precision": 0.02040816326530612,
9
- "recall": 0.0034965034965034965,
10
- "f1": 0.005970149253731343,
11
  "gt_support": 286,
12
- "predicted": 49
13
  },
14
  "1": {
15
- "accuracy": 0.9867461089494164,
16
- "precision": 0.5872774641771602,
17
- "recall": 0.9867461089494164,
18
- "f1": 0.7363215679157971,
19
  "gt_support": 8224,
20
- "predicted": 13818
21
  },
22
  "2": {
23
- "accuracy": 0.01048951048951049,
24
- "precision": 0.02040816326530612,
25
- "recall": 0.01048951048951049,
26
- "f1": 0.013856812933025403,
27
  "gt_support": 286,
28
- "predicted": 147
29
  },
30
  "3": {
31
- "accuracy": 0.0,
32
- "precision": 0.0,
33
- "recall": 0.0,
34
- "f1": 0.0,
35
  "gt_support": 5218,
36
- "predicted": 0
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.4774436090225564,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
 
 
 
 
 
 
 
 
47
  "accuracy": 0.0,
48
  "precision": 0.0,
49
  "recall": 0.0,
50
  "f1": 0.0,
51
- "gt_support": 38,
52
- "predicted": 0
53
- },
54
- "1": {
55
- "accuracy": 0.9736842105263158,
56
- "precision": 0.4897959183673469,
57
- "recall": 0.9736842105263158,
58
- "f1": 0.6517431192660551,
59
  "gt_support": 912,
60
- "predicted": 1813
61
  },
62
  "2": {
63
- "accuracy": 0.02631578947368421,
64
- "precision": 0.02040816326530612,
65
- "recall": 0.02631578947368421,
66
- "f1": 0.022988505747126436,
67
  "gt_support": 38,
68
- "predicted": 49
69
  },
70
  "3": {
71
- "accuracy": 0.0,
72
- "precision": 0.0,
73
- "recall": 0.0,
74
- "f1": 0.0,
75
  "gt_support": 874,
76
- "predicted": 0
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
- "accuracy": 0.5190166975881262,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 44,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.9772727272727273,
95
- "precision": 0.5306122448979592,
96
- "recall": 0.9772727272727273,
97
- "f1": 0.6877883728083667,
98
  "gt_support": 1144,
99
- "predicted": 2107
100
  },
101
  "2": {
102
- "accuracy": 0.022727272727272728,
103
- "precision": 0.02040816326530612,
104
- "recall": 0.022727272727272728,
105
- "f1": 0.02150537634408602,
106
  "gt_support": 44,
107
- "predicted": 49
108
  },
109
  "3": {
110
- "accuracy": 0.0,
111
- "precision": 0.0,
112
- "recall": 0.0,
113
- "f1": 0.0,
114
  "gt_support": 924,
115
- "predicted": 0
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
- "accuracy": 0.5579890492782479,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.024390243902439025,
126
- "precision": 0.02040816326530612,
127
- "recall": 0.024390243902439025,
128
- "f1": 0.02222222222222222,
129
  "gt_support": 41,
130
- "predicted": 49
131
  },
132
  "1": {
133
- "accuracy": 0.975609756097561,
134
- "precision": 0.5714285714285714,
135
- "recall": 0.975609756097561,
136
- "f1": 0.7207207207207207,
137
  "gt_support": 1148,
138
- "predicted": 1960
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 41,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.0,
150
- "precision": 0.0,
151
- "recall": 0.0,
152
- "f1": 0.0,
153
  "gt_support": 779,
154
- "predicted": 0
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
- "accuracy": 0.5918367346938775,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 47,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 1.0,
173
- "precision": 0.5918367346938775,
174
- "recall": 1.0,
175
- "f1": 0.7435897435897436,
176
  "gt_support": 1363,
177
- "predicted": 2303
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 47,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.0,
189
- "precision": 0.0,
190
- "recall": 0.0,
191
- "f1": 0.0,
192
  "gt_support": 846,
193
- "predicted": 0
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
- "accuracy": 0.62152133580705,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 55,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.9818181818181818,
212
- "precision": 0.6326530612244898,
213
- "recall": 0.9818181818181818,
214
- "f1": 0.7694782808549759,
215
  "gt_support": 1705,
216
- "predicted": 2646
217
  },
218
  "2": {
219
- "accuracy": 0.01818181818181818,
220
- "precision": 0.02040816326530612,
221
- "recall": 0.01818181818181818,
222
- "f1": 0.01923076923076923,
223
  "gt_support": 55,
224
- "predicted": 49
225
  },
226
  "3": {
227
- "accuracy": 0.0,
228
- "precision": 0.0,
229
- "recall": 0.0,
230
- "f1": 0.0,
231
  "gt_support": 880,
232
- "predicted": 0
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
- "accuracy": 0.6530612244897959,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 61,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 1.0,
251
- "precision": 0.6530612244897959,
252
- "recall": 1.0,
253
- "f1": 0.7901234567901235,
254
  "gt_support": 1952,
255
- "predicted": 2989
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 61,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.0,
267
- "precision": 0.0,
268
- "recall": 0.0,
269
- "f1": 0.0,
270
  "gt_support": 915,
271
- "predicted": 0
272
  }
273
  },
274
  "total_samples": 2989
@@ -278,13 +278,13 @@
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 7
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.29577565291851005,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.10139860139860139,
8
+ "precision": 0.021739130434782608,
9
+ "recall": 0.10139860139860139,
10
+ "f1": 0.03580246913580247,
11
  "gt_support": 286,
12
+ "predicted": 1334
13
  },
14
  "1": {
15
+ "accuracy": 0.2587548638132296,
16
+ "precision": 0.6211325160537069,
17
+ "recall": 0.2587548638132296,
18
+ "f1": 0.36532188841201724,
19
  "gt_support": 8224,
20
+ "predicted": 3426
21
  },
22
  "2": {
23
+ "accuracy": 0.3146853146853147,
24
+ "precision": 0.019867549668874173,
25
+ "recall": 0.3146853146853147,
26
+ "f1": 0.03737541528239203,
27
  "gt_support": 286,
28
+ "predicted": 4530
29
  },
30
  "3": {
31
+ "accuracy": 0.3637408968953622,
32
+ "precision": 0.40177815410668927,
33
+ "recall": 0.3637408968953622,
34
+ "f1": 0.3818145242405955,
35
  "gt_support": 5218,
36
+ "predicted": 4724
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.3066595059076262,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.07894736842105263,
48
+ "precision": 0.024793388429752067,
49
+ "recall": 0.07894736842105263,
50
+ "f1": 0.03773584905660377,
51
+ "gt_support": 38,
52
+ "predicted": 121
53
+ },
54
+ "1": {
55
  "accuracy": 0.0,
56
  "precision": 0.0,
57
  "recall": 0.0,
58
  "f1": 0.0,
 
 
 
 
 
 
 
 
59
  "gt_support": 912,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.2894736842105263,
64
+ "precision": 0.020257826887661142,
65
+ "recall": 0.2894736842105263,
66
+ "f1": 0.0378657487091222,
67
  "gt_support": 38,
68
+ "predicted": 543
69
  },
70
  "3": {
71
+ "accuracy": 0.6372997711670481,
72
+ "precision": 0.46494156928213687,
73
+ "recall": 0.6372997711670481,
74
+ "f1": 0.5376447876447877,
75
  "gt_support": 874,
76
+ "predicted": 1198
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
+ "accuracy": 0.2792207792207792,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.09090909090909091,
87
+ "precision": 0.016597510373443983,
88
+ "recall": 0.09090909090909091,
89
+ "f1": 0.028070175438596492,
90
  "gt_support": 44,
91
+ "predicted": 241
92
  },
93
  "1": {
94
+ "accuracy": 0.033216783216783216,
95
+ "precision": 0.5507246376811594,
96
+ "recall": 0.033216783216783216,
97
+ "f1": 0.06265457543281121,
98
  "gt_support": 1144,
99
+ "predicted": 69
100
  },
101
  "2": {
102
+ "accuracy": 0.25,
103
+ "precision": 0.019400352733686066,
104
+ "recall": 0.25,
105
+ "f1": 0.03600654664484452,
106
  "gt_support": 44,
107
+ "predicted": 567
108
  },
109
  "3": {
110
+ "accuracy": 0.5941558441558441,
111
+ "precision": 0.4292415949960907,
112
+ "recall": 0.5941558441558441,
113
+ "f1": 0.49841125737630493,
114
  "gt_support": 924,
115
+ "predicted": 1279
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
+ "accuracy": 0.28322548531607766,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.04878048780487805,
126
+ "precision": 0.02631578947368421,
127
+ "recall": 0.04878048780487805,
128
+ "f1": 0.034188034188034185,
129
  "gt_support": 41,
130
+ "predicted": 76
131
  },
132
  "1": {
133
+ "accuracy": 0.21341463414634146,
134
+ "precision": 0.5658198614318707,
135
+ "recall": 0.21341463414634146,
136
+ "f1": 0.3099304237824162,
137
  "gt_support": 1148,
138
+ "predicted": 433
139
  },
140
  "2": {
141
+ "accuracy": 0.34146341463414637,
142
+ "precision": 0.020086083213773313,
143
+ "recall": 0.34146341463414637,
144
+ "f1": 0.037940379403794036,
145
  "gt_support": 41,
146
+ "predicted": 697
147
  },
148
  "3": {
149
+ "accuracy": 0.39537869062901154,
150
+ "precision": 0.3835616438356164,
151
+ "recall": 0.39537869062901154,
152
+ "f1": 0.3893805309734513,
153
  "gt_support": 779,
154
+ "predicted": 803
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
+ "accuracy": 0.3009118541033435,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.1702127659574468,
165
+ "precision": 0.02,
166
+ "recall": 0.1702127659574468,
167
+ "f1": 0.035794183445190156,
168
  "gt_support": 47,
169
+ "predicted": 400
170
  },
171
  "1": {
172
+ "accuracy": 0.3734409391049156,
173
+ "precision": 0.594626168224299,
174
+ "recall": 0.3734409391049156,
175
+ "f1": 0.4587652095538531,
176
  "gt_support": 1363,
177
+ "predicted": 856
178
  },
179
  "2": {
180
+ "accuracy": 0.23404255319148937,
181
+ "precision": 0.018456375838926176,
182
+ "recall": 0.23404255319148937,
183
+ "f1": 0.034214618973561435,
184
  "gt_support": 47,
185
+ "predicted": 596
186
  },
187
  "3": {
188
+ "accuracy": 0.1950354609929078,
189
+ "precision": 0.36585365853658536,
190
+ "recall": 0.1950354609929078,
191
+ "f1": 0.25443330763299926,
192
  "gt_support": 846,
193
+ "predicted": 451
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
+ "accuracy": 0.2894248608534323,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.14545454545454545,
204
+ "precision": 0.024242424242424242,
205
+ "recall": 0.14545454545454545,
206
+ "f1": 0.04155844155844156,
207
  "gt_support": 55,
208
+ "predicted": 330
209
  },
210
  "1": {
211
+ "accuracy": 0.3255131964809384,
212
+ "precision": 0.6364678899082569,
213
+ "recall": 0.3255131964809384,
214
+ "f1": 0.4307334109429569,
215
  "gt_support": 1705,
216
+ "predicted": 872
217
  },
218
  "2": {
219
+ "accuracy": 0.3090909090909091,
220
+ "precision": 0.01925254813137033,
221
+ "recall": 0.3090909090909091,
222
+ "f1": 0.036247334754797446,
223
  "gt_support": 55,
224
+ "predicted": 883
225
  },
226
  "3": {
227
+ "accuracy": 0.22727272727272727,
228
+ "precision": 0.32786885245901637,
229
+ "recall": 0.22727272727272727,
230
+ "f1": 0.2684563758389262,
231
  "gt_support": 880,
232
+ "predicted": 610
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
+ "accuracy": 0.311140849782536,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.06557377049180328,
243
+ "precision": 0.024096385542168676,
244
+ "recall": 0.06557377049180328,
245
+ "f1": 0.035242290748898675,
246
  "gt_support": 61,
247
+ "predicted": 166
248
  },
249
  "1": {
250
+ "accuracy": 0.40010245901639346,
251
+ "precision": 0.653010033444816,
252
+ "recall": 0.40010245901639346,
253
+ "f1": 0.4961880559085134,
254
  "gt_support": 1952,
255
+ "predicted": 1196
256
  },
257
  "2": {
258
+ "accuracy": 0.4262295081967213,
259
+ "precision": 0.02090032154340836,
260
+ "recall": 0.4262295081967213,
261
+ "f1": 0.03984674329501915,
262
  "gt_support": 61,
263
+ "predicted": 1244
264
  },
265
  "3": {
266
+ "accuracy": 0.13005464480874318,
267
+ "precision": 0.31070496083550914,
268
+ "recall": 0.13005464480874318,
269
+ "f1": 0.18335901386748846,
270
  "gt_support": 915,
271
+ "predicted": 383
272
  }
273
  },
274
  "total_samples": 2989
 
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 7
290
  }
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 286 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5793 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0035 0.0204 0.0035 0.0060 286 49
30
- # 0.9867 0.5873 0.9867 0.7363 8224 13818
31
- G 0.0105 0.0204 0.0105 0.0139 286 147
32
- _ 0.0000 0.0000 0.0000 0.0000 5218 0
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.4774 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 38 0
49
- # 0.9737 0.4898 0.9737 0.6517 912 1813
50
- G 0.0263 0.0204 0.0263 0.0230 38 49
51
- _ 0.0000 0.0000 0.0000 0.0000 874 0
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.5190 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 44 0
64
- # 0.9773 0.5306 0.9773 0.6878 1144 2107
65
- G 0.0227 0.0204 0.0227 0.0215 44 49
66
- _ 0.0000 0.0000 0.0000 0.0000 924 0
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5580 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0244 0.0204 0.0244 0.0222 41 49
79
- # 0.9756 0.5714 0.9756 0.7207 1148 1960
80
- G 0.0000 0.0000 0.0000 0.0000 41 0
81
- _ 0.0000 0.0000 0.0000 0.0000 779 0
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5918 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 47 0
94
- # 1.0000 0.5918 1.0000 0.7436 1363 2303
95
- G 0.0000 0.0000 0.0000 0.0000 47 0
96
- _ 0.0000 0.0000 0.0000 0.0000 846 0
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.6215 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 55 0
109
- # 0.9818 0.6327 0.9818 0.7695 1705 2646
110
- G 0.0182 0.0204 0.0182 0.0192 55 49
111
- _ 0.0000 0.0000 0.0000 0.0000 880 0
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.6531 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 61 0
124
- # 1.0000 0.6531 1.0000 0.7901 1952 2989
125
- G 0.0000 0.0000 0.0000 0.0000 61 0
126
- _ 0.0000 0.0000 0.0000 0.0000 915 0
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.2958 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.1014 0.0217 0.1014 0.0358 286 1334
30
+ # 0.2588 0.6211 0.2588 0.3653 8224 3426
31
+ G 0.3147 0.0199 0.3147 0.0374 286 4530
32
+ _ 0.3637 0.4018 0.3637 0.3818 5218 4724
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.3067 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0789 0.0248 0.0789 0.0377 38 121
49
+ # 0.0000 0.0000 0.0000 0.0000 912 0
50
+ G 0.2895 0.0203 0.2895 0.0379 38 543
51
+ _ 0.6373 0.4649 0.6373 0.5376 874 1198
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.2792 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.0909 0.0166 0.0909 0.0281 44 241
64
+ # 0.0332 0.5507 0.0332 0.0627 1144 69
65
+ G 0.2500 0.0194 0.2500 0.0360 44 567
66
+ _ 0.5942 0.4292 0.5942 0.4984 924 1279
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.2832 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.0488 0.0263 0.0488 0.0342 41 76
79
+ # 0.2134 0.5658 0.2134 0.3099 1148 433
80
+ G 0.3415 0.0201 0.3415 0.0379 41 697
81
+ _ 0.3954 0.3836 0.3954 0.3894 779 803
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.3009 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.1702 0.0200 0.1702 0.0358 47 400
94
+ # 0.3734 0.5946 0.3734 0.4588 1363 856
95
+ G 0.2340 0.0185 0.2340 0.0342 47 596
96
+ _ 0.1950 0.3659 0.1950 0.2544 846 451
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2894 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.1455 0.0242 0.1455 0.0416 55 330
109
+ # 0.3255 0.6365 0.3255 0.4307 1705 872
110
+ G 0.3091 0.0193 0.3091 0.0362 55 883
111
+ _ 0.2273 0.3279 0.2273 0.2685 880 610
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.3111 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.0656 0.0241 0.0656 0.0352 61 166
124
+ # 0.4001 0.6530 0.4001 0.4962 1952 1196
125
+ G 0.4262 0.0209 0.4262 0.0398 61 1244
126
+ _ 0.1301 0.3107 0.1301 0.1834 915 383
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size7.json
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.3426286208387376,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.1303501945525292,
8
- "precision": 0.012486023108460679,
9
- "recall": 0.1303501945525292,
10
- "f1": 0.022789115646258504,
11
  "gt_support": 514,
12
- "predicted": 5366
13
  },
14
  "1": {
15
- "accuracy": 0.4928539920599912,
16
- "precision": 0.5522712668676784,
17
- "recall": 0.4928539920599912,
18
- "f1": 0.520873639309107,
19
  "gt_support": 22670,
20
- "predicted": 20231
21
  },
22
  "2": {
23
- "accuracy": 0.2237354085603113,
24
- "precision": 0.01236160378372568,
25
- "recall": 0.2237354085603113,
26
- "f1": 0.023428746052765615,
27
  "gt_support": 514,
28
- "predicted": 9303
29
  },
30
  "3": {
31
- "accuracy": 0.1622435325602141,
32
- "precision": 0.43213543213543215,
33
- "recall": 0.1622435325602141,
34
- "f1": 0.2359140656668018,
35
  "gt_support": 17936,
36
- "predicted": 6734
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.2242798353909465,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.3148148148148148,
48
- "precision": 0.012345679012345678,
49
- "recall": 0.3148148148148148,
50
- "f1": 0.023759608665269043,
51
  "gt_support": 54,
52
- "predicted": 1377
53
  },
54
  "1": {
55
- "accuracy": 0.3888888888888889,
56
- "precision": 0.3950617283950617,
57
- "recall": 0.3888888888888889,
58
- "f1": 0.39195100612423445,
59
  "gt_support": 1728,
60
- "predicted": 1701
61
  },
62
  "2": {
63
- "accuracy": 0.18518518518518517,
64
- "precision": 0.012345679012345678,
65
- "recall": 0.18518518518518517,
66
- "f1": 0.023148148148148147,
67
  "gt_support": 54,
68
- "predicted": 810
69
  },
70
  "3": {
71
- "accuracy": 0.1111111111111111,
72
- "precision": 0.5802469135802469,
73
- "recall": 0.1111111111111111,
74
- "f1": 0.1865079365079365,
75
  "gt_support": 2538,
76
- "predicted": 486
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
- "accuracy": 0.35802469135802467,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.14893617021276595,
87
- "precision": 0.012345679012345678,
88
- "recall": 0.14893617021276595,
89
- "f1": 0.02280130293159609,
90
  "gt_support": 47,
91
- "predicted": 567
92
  },
93
  "1": {
94
- "accuracy": 0.46808510638297873,
95
- "precision": 0.4444444444444444,
96
- "recall": 0.46808510638297873,
97
- "f1": 0.45595854922279794,
98
  "gt_support": 1692,
99
- "predicted": 1782
100
  },
101
  "2": {
102
- "accuracy": 0.10638297872340426,
103
- "precision": 0.012345679012345678,
104
- "recall": 0.10638297872340426,
105
- "f1": 0.02212389380530973,
106
  "gt_support": 47,
107
- "predicted": 405
108
  },
109
  "3": {
110
- "accuracy": 0.2765957446808511,
111
- "precision": 0.5308641975308642,
112
- "recall": 0.2765957446808511,
113
- "f1": 0.36369551073519846,
114
  "gt_support": 2021,
115
- "predicted": 1053
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
- "accuracy": 0.28560899207665374,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.16417910447761194,
126
- "precision": 0.012485811577752554,
127
- "recall": 0.16417910447761194,
128
- "f1": 0.02320675105485232,
129
  "gt_support": 67,
130
- "predicted": 881
131
  },
132
  "1": {
133
- "accuracy": 0.45111940298507464,
134
- "precision": 0.4952888160589922,
135
- "recall": 0.45111940298507464,
136
- "f1": 0.47217340363210314,
137
  "gt_support": 2680,
138
- "predicted": 2441
139
  },
140
  "2": {
141
- "accuracy": 0.26865671641791045,
142
- "precision": 0.012345679012345678,
143
- "recall": 0.26865671641791045,
144
- "f1": 0.023606557377049177,
145
  "gt_support": 67,
146
- "predicted": 1458
147
  },
148
  "3": {
149
- "accuracy": 0.11940298507462686,
150
- "precision": 0.482225656877898,
151
- "recall": 0.11940298507462686,
152
- "f1": 0.19141104294478525,
153
  "gt_support": 2613,
154
- "predicted": 647
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
- "accuracy": 0.31127302165553533,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.13114754098360656,
165
- "precision": 0.012345679012345678,
166
- "recall": 0.13114754098360656,
167
- "f1": 0.022566995768688293,
168
  "gt_support": 61,
169
- "predicted": 648
170
  },
171
  "1": {
172
- "accuracy": 0.4262295081967213,
173
- "precision": 0.5308641975308642,
174
- "recall": 0.4262295081967213,
175
- "f1": 0.47282723620215694,
176
  "gt_support": 2623,
177
- "predicted": 2106
178
  },
179
  "2": {
180
- "accuracy": 0.26229508196721313,
181
  "precision": 0.012345679012345678,
182
- "recall": 0.26229508196721313,
183
- "f1": 0.023581429624170966,
184
  "gt_support": 61,
185
- "predicted": 1296
186
  },
187
  "3": {
188
- "accuracy": 0.18032786885245902,
189
- "precision": 0.4444444444444444,
190
- "recall": 0.18032786885245902,
191
- "f1": 0.2565597667638484,
192
  "gt_support": 2196,
193
- "predicted": 891
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
- "accuracy": 0.35631340911868964,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0891089108910891,
204
- "precision": 0.012345679012345678,
205
- "recall": 0.0891089108910891,
206
- "f1": 0.021686746987951807,
207
  "gt_support": 101,
208
- "predicted": 729
209
  },
210
  "1": {
211
- "accuracy": 0.48514851485148514,
212
- "precision": 0.5802469135802469,
213
- "recall": 0.48514851485148514,
214
- "f1": 0.5284534189995411,
215
  "gt_support": 4747,
216
- "predicted": 3969
217
  },
218
  "2": {
219
- "accuracy": 0.24752475247524752,
220
- "precision": 0.012419274714356682,
221
- "recall": 0.24752475247524752,
222
- "f1": 0.02365184484389783,
223
  "gt_support": 101,
224
- "predicted": 2013
225
  },
226
  "3": {
227
- "accuracy": 0.17883663366336633,
228
- "precision": 0.39319727891156464,
229
- "recall": 0.17883663366336633,
230
- "f1": 0.24585282858358146,
231
  "gt_support": 3232,
232
- "predicted": 1470
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
- "accuracy": 0.39707461084272677,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.08152173913043478,
243
- "precision": 0.01288659793814433,
244
- "recall": 0.08152173913043478,
245
- "f1": 0.02225519287833828,
246
  "gt_support": 184,
247
- "predicted": 1164
248
  },
249
  "1": {
250
- "accuracy": 0.5520652173913043,
251
- "precision": 0.6169825072886297,
252
- "recall": 0.5520652173913043,
253
- "f1": 0.5827214318494721,
254
  "gt_support": 9200,
255
- "predicted": 8232
256
  },
257
  "2": {
258
- "accuracy": 0.22282608695652173,
259
- "precision": 0.012345679012345678,
260
- "recall": 0.22282608695652173,
261
- "f1": 0.02339514978601997,
262
  "gt_support": 184,
263
- "predicted": 3321
264
  },
265
  "3": {
266
- "accuracy": 0.14673913043478262,
267
- "precision": 0.35802469135802467,
268
- "recall": 0.14673913043478262,
269
- "f1": 0.2081616376445567,
270
  "gt_support": 5336,
271
- "predicted": 2187
272
  }
273
  },
274
  "total_samples": 14904
@@ -278,13 +278,13 @@
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 9
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.31524715376855456,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.2529182879377432,
8
+ "precision": 0.012632397240307064,
9
+ "recall": 0.2529182879377432,
10
+ "f1": 0.024062933826931976,
11
  "gt_support": 514,
12
+ "predicted": 10291
13
  },
14
  "1": {
15
+ "accuracy": 0.4254080282311425,
16
+ "precision": 0.5909675837980268,
17
+ "recall": 0.4254080282311425,
18
+ "f1": 0.4947036343584088,
19
  "gt_support": 22670,
20
+ "predicted": 16319
21
  },
22
  "2": {
23
+ "accuracy": 0.21011673151750973,
24
+ "precision": 0.012098129270751653,
25
+ "recall": 0.21011673151750973,
26
+ "f1": 0.022878932316491896,
27
  "gt_support": 514,
28
+ "predicted": 8927
29
  },
30
  "3": {
31
+ "accuracy": 0.18080954504906332,
32
+ "precision": 0.5319009348860095,
33
+ "recall": 0.18080954504906332,
34
+ "f1": 0.2698789164898265,
35
  "gt_support": 17936,
36
+ "predicted": 6097
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.38065843621399176,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.05555555555555555,
48
+ "precision": 0.015625,
49
+ "recall": 0.05555555555555555,
50
+ "f1": 0.024390243902439025,
51
  "gt_support": 54,
52
+ "predicted": 192
53
  },
54
  "1": {
55
+ "accuracy": 0.0,
56
+ "precision": 0.0,
57
+ "recall": 0.0,
58
+ "f1": 0.0,
59
  "gt_support": 1728,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.2962962962962963,
64
+ "precision": 0.011922503725782414,
65
+ "recall": 0.2962962962962963,
66
+ "f1": 0.022922636103151862,
67
  "gt_support": 54,
68
+ "predicted": 1342
69
  },
70
  "3": {
71
+ "accuracy": 0.6485421591804571,
72
+ "precision": 0.5795774647887324,
73
+ "recall": 0.6485421591804571,
74
+ "f1": 0.6121234659724805,
75
  "gt_support": 2538,
76
+ "predicted": 2840
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
+ "accuracy": 0.23036511688993958,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.2553191489361702,
87
+ "precision": 0.012307692307692308,
88
+ "recall": 0.2553191489361702,
89
+ "f1": 0.023483365949119376,
90
  "gt_support": 47,
91
+ "predicted": 975
92
  },
93
  "1": {
94
+ "accuracy": 0.03723404255319149,
95
+ "precision": 0.4117647058823529,
96
+ "recall": 0.03723404255319149,
97
+ "f1": 0.06829268292682926,
98
  "gt_support": 1692,
99
+ "predicted": 153
100
  },
101
  "2": {
102
+ "accuracy": 0.3191489361702128,
103
+ "precision": 0.012520868113522538,
104
+ "recall": 0.3191489361702128,
105
+ "f1": 0.024096385542168676,
106
  "gt_support": 47,
107
+ "predicted": 1198
108
  },
109
  "3": {
110
+ "accuracy": 0.38941118258287977,
111
+ "precision": 0.5313977042538826,
112
+ "recall": 0.38941118258287977,
113
+ "f1": 0.44945745288406624,
114
  "gt_support": 2021,
115
+ "predicted": 1481
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
+ "accuracy": 0.1369080523309379,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.3880597014925373,
126
+ "precision": 0.012357414448669201,
127
+ "recall": 0.3880597014925373,
128
+ "f1": 0.02395209580838323,
129
  "gt_support": 67,
130
+ "predicted": 2104
131
  },
132
  "1": {
133
+ "accuracy": 0.11119402985074626,
134
+ "precision": 0.48932676518883417,
135
+ "recall": 0.11119402985074626,
136
+ "f1": 0.1812100942535725,
137
  "gt_support": 2680,
138
+ "predicted": 609
139
  },
140
  "2": {
141
+ "accuracy": 0.34328358208955223,
142
+ "precision": 0.012188659247482777,
143
+ "recall": 0.34328358208955223,
144
+ "f1": 0.02354145342886387,
145
  "gt_support": 67,
146
+ "predicted": 1887
147
  },
148
  "3": {
149
+ "accuracy": 0.1515499425947187,
150
+ "precision": 0.4788391777509069,
151
+ "recall": 0.1515499425947187,
152
+ "f1": 0.2302325581395349,
153
  "gt_support": 2613,
154
+ "predicted": 827
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
+ "accuracy": 0.30459421169803685,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.26229508196721313,
165
+ "precision": 0.01228878648233487,
166
+ "recall": 0.26229508196721313,
167
+ "f1": 0.023477622890682316,
168
  "gt_support": 61,
169
+ "predicted": 1302
170
  },
171
  "1": {
172
+ "accuracy": 0.43080442241707967,
173
+ "precision": 0.5322656617993405,
174
+ "recall": 0.43080442241707967,
175
+ "f1": 0.47619047619047616,
176
  "gt_support": 2623,
177
+ "predicted": 2123
178
  },
179
  "2": {
180
+ "accuracy": 0.14754098360655737,
181
  "precision": 0.012345679012345678,
182
+ "recall": 0.14754098360655737,
183
+ "f1": 0.022784810126582275,
184
  "gt_support": 61,
185
+ "predicted": 729
186
  },
187
  "3": {
188
+ "accuracy": 0.15938069216757741,
189
+ "precision": 0.44472681067344344,
190
+ "recall": 0.15938069216757741,
191
+ "f1": 0.23466309084813944,
192
  "gt_support": 2196,
193
+ "predicted": 787
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
+ "accuracy": 0.28260603838161596,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.33663366336633666,
204
+ "precision": 0.012849584278155708,
205
+ "recall": 0.33663366336633666,
206
+ "f1": 0.024754277393520204,
207
  "gt_support": 101,
208
+ "predicted": 2646
209
  },
210
  "1": {
211
+ "accuracy": 0.46218664419633454,
212
+ "precision": 0.5813460519342872,
213
+ "recall": 0.46218664419633454,
214
+ "f1": 0.5149630325079216,
215
  "gt_support": 4747,
216
+ "predicted": 3774
217
  },
218
  "2": {
219
+ "accuracy": 0.19801980198019803,
220
+ "precision": 0.012507817385866166,
221
+ "recall": 0.19801980198019803,
222
+ "f1": 0.023529411764705882,
223
  "gt_support": 101,
224
+ "predicted": 1599
225
  },
226
  "3": {
227
+ "accuracy": 0.019801980198019802,
228
+ "precision": 0.3950617283950617,
229
+ "recall": 0.019801980198019802,
230
+ "f1": 0.037713612256923976,
231
  "gt_support": 3232,
232
+ "predicted": 162
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
+ "accuracy": 0.4041196994095545,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.21195652173913043,
243
+ "precision": 0.0126953125,
244
+ "recall": 0.21195652173913043,
245
+ "f1": 0.023955773955773956,
246
  "gt_support": 184,
247
+ "predicted": 3072
248
  },
249
  "1": {
250
+ "accuracy": 0.6477173913043478,
251
+ "precision": 0.6168737060041408,
252
+ "recall": 0.6477173913043478,
253
+ "f1": 0.6319194061505833,
254
  "gt_support": 9200,
255
+ "predicted": 9660
256
  },
257
  "2": {
258
+ "accuracy": 0.1358695652173913,
259
+ "precision": 0.01151012891344383,
260
+ "recall": 0.1358695652173913,
261
+ "f1": 0.021222410865874366,
262
  "gt_support": 184,
263
+ "predicted": 2172
264
  },
265
  "3": {
266
+ "accuracy": 0.0,
267
+ "precision": 0.0,
268
+ "recall": 0.0,
269
+ "f1": 0.0,
270
  "gt_support": 5336,
271
+ "predicted": 0
272
  }
273
  },
274
  "total_samples": 14904
 
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 9
290
  }
layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 514 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.3426 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.1304 0.0125 0.1304 0.0228 514 5366
30
- # 0.4929 0.5523 0.4929 0.5209 22670 20231
31
- G 0.2237 0.0124 0.2237 0.0234 514 9303
32
- _ 0.1622 0.4321 0.1622 0.2359 17936 6734
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.2243 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.3148 0.0123 0.3148 0.0238 54 1377
49
- # 0.3889 0.3951 0.3889 0.3920 1728 1701
50
- G 0.1852 0.0123 0.1852 0.0231 54 810
51
- _ 0.1111 0.5802 0.1111 0.1865 2538 486
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.3580 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.1489 0.0123 0.1489 0.0228 47 567
64
- # 0.4681 0.4444 0.4681 0.4560 1692 1782
65
- G 0.1064 0.0123 0.1064 0.0221 47 405
66
- _ 0.2766 0.5309 0.2766 0.3637 2021 1053
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.2856 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.1642 0.0125 0.1642 0.0232 67 881
79
- # 0.4511 0.4953 0.4511 0.4722 2680 2441
80
- G 0.2687 0.0123 0.2687 0.0236 67 1458
81
- _ 0.1194 0.4822 0.1194 0.1914 2613 647
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.3113 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.1311 0.0123 0.1311 0.0226 61 648
94
- # 0.4262 0.5309 0.4262 0.4728 2623 2106
95
- G 0.2623 0.0123 0.2623 0.0236 61 1296
96
- _ 0.1803 0.4444 0.1803 0.2566 2196 891
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.3563 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0891 0.0123 0.0891 0.0217 101 729
109
- # 0.4851 0.5802 0.4851 0.5285 4747 3969
110
- G 0.2475 0.0124 0.2475 0.0237 101 2013
111
- _ 0.1788 0.3932 0.1788 0.2459 3232 1470
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.3971 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0815 0.0129 0.0815 0.0223 184 1164
124
- # 0.5521 0.6170 0.5521 0.5827 9200 8232
125
- G 0.2228 0.0123 0.2228 0.0234 184 3321
126
- _ 0.1467 0.3580 0.1467 0.2082 5336 2187
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.3152 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.2529 0.0126 0.2529 0.0241 514 10291
30
+ # 0.4254 0.5910 0.4254 0.4947 22670 16319
31
+ G 0.2101 0.0121 0.2101 0.0229 514 8927
32
+ _ 0.1808 0.5319 0.1808 0.2699 17936 6097
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.3807 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0556 0.0156 0.0556 0.0244 54 192
49
+ # 0.0000 0.0000 0.0000 0.0000 1728 0
50
+ G 0.2963 0.0119 0.2963 0.0229 54 1342
51
+ _ 0.6485 0.5796 0.6485 0.6121 2538 2840
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.2304 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.2553 0.0123 0.2553 0.0235 47 975
64
+ # 0.0372 0.4118 0.0372 0.0683 1692 153
65
+ G 0.3191 0.0125 0.3191 0.0241 47 1198
66
+ _ 0.3894 0.5314 0.3894 0.4495 2021 1481
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.1369 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.3881 0.0124 0.3881 0.0240 67 2104
79
+ # 0.1112 0.4893 0.1112 0.1812 2680 609
80
+ G 0.3433 0.0122 0.3433 0.0235 67 1887
81
+ _ 0.1515 0.4788 0.1515 0.2302 2613 827
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.3046 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.2623 0.0123 0.2623 0.0235 61 1302
94
+ # 0.4308 0.5323 0.4308 0.4762 2623 2123
95
+ G 0.1475 0.0123 0.1475 0.0228 61 729
96
+ _ 0.1594 0.4447 0.1594 0.2347 2196 787
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.2826 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.3366 0.0128 0.3366 0.0248 101 2646
109
+ # 0.4622 0.5813 0.4622 0.5150 4747 3774
110
+ G 0.1980 0.0125 0.1980 0.0235 101 1599
111
+ _ 0.0198 0.3951 0.0198 0.0377 3232 162
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.4041 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2120 0.0127 0.2120 0.0240 184 3072
124
+ # 0.6477 0.6169 0.6477 0.6319 9200 9660
125
+ G 0.1359 0.0115 0.1359 0.0212 184 2172
126
+ _ 0.0000 0.0000 0.0000 0.0000 5336 0
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_size9.json
layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.json CHANGED
@@ -1,600 +1,614 @@
1
  {
2
  "global": {
3
- "accuracy": 0.07962078119074706,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.700279242941359,
8
- "precision": 0.004599279846921621,
9
- "recall": 0.700279242941359,
10
- "f1": 0.009138539777144338,
11
  "gt_support": 3223,
12
- "predicted": 490729
13
  },
14
  "1": {
15
- "accuracy": 0.08599139288338406,
16
- "precision": 0.2977538707567057,
17
- "recall": 0.08599139288338406,
18
- "f1": 0.13344409695074275,
19
  "gt_support": 238175,
20
- "predicted": 68785
21
  },
22
  "2": {
23
- "accuracy": 0.15141172820353707,
24
- "precision": 0.0045629651793394925,
25
- "recall": 0.15141172820353707,
26
- "f1": 0.008858955623530694,
27
  "gt_support": 3223,
28
- "predicted": 106948
29
  },
30
  "3": {
31
- "accuracy": 0.05422320481947515,
32
- "precision": 0.41664035364698454,
33
- "recall": 0.05422320481947515,
34
- "f1": 0.0959580533496233,
35
  "gt_support": 243346,
36
- "predicted": 31670
37
  },
38
  "4": {
39
- "accuracy": 0.08987049340663046,
40
- "precision": 0.7883001146322524,
41
- "recall": 0.08987049340663046,
42
- "f1": 0.1613465984991542,
43
  "gt_support": 237208,
44
- "predicted": 27043
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
- "by_size": {
50
- "7": {
51
- "accuracy": 0.11361305361305361,
52
- "baseline_accuracy": 0.7822222222222223,
53
  "per_class": {
54
  "0": {
55
- "accuracy": 0.8146853146853147,
56
- "precision": 0.004848510071583153,
57
- "recall": 0.8146853146853147,
58
- "f1": 0.009639650821232054,
59
- "gt_support": 286,
60
- "predicted": 48056
61
  },
62
  "1": {
63
- "accuracy": 0.16427529182879377,
64
- "precision": 0.14203111858704795,
65
- "recall": 0.16427529182879377,
66
- "f1": 0.15234551195308976,
67
- "gt_support": 8224,
68
- "predicted": 9512
69
  },
70
  "2": {
71
- "accuracy": 0.017482517482517484,
72
- "precision": 0.005861664712778429,
73
- "recall": 0.017482517482517484,
74
- "f1": 0.00877963125548727,
75
- "gt_support": 286,
76
- "predicted": 853
77
  },
78
  "3": {
79
- "accuracy": 0.0034495975469528554,
80
- "precision": 0.08,
81
- "recall": 0.0034495975469528554,
82
- "f1": 0.006613999632555575,
83
- "gt_support": 5218,
84
- "predicted": 225
85
  },
86
  "4": {
87
- "accuracy": 0.11331849968213605,
88
- "precision": 1.0,
89
- "recall": 0.11331849968213605,
90
- "f1": 0.2035688793718772,
91
- "gt_support": 50336,
92
- "predicted": 5704
93
  }
94
  },
95
- "total_samples": 64350
96
  },
97
- "9": {
98
- "accuracy": 0.08400345871162992,
99
- "baseline_accuracy": 0.64,
100
  "per_class": {
101
  "0": {
102
- "accuracy": 0.8132295719844358,
103
- "precision": 0.004670443245175924,
104
- "recall": 0.8132295719844358,
105
- "f1": 0.00928754735427105,
106
- "gt_support": 514,
107
- "predicted": 89499
108
  },
109
  "1": {
110
- "accuracy": 0.14261138067931187,
111
- "precision": 0.19663058022138427,
112
- "recall": 0.14261138067931187,
113
- "f1": 0.16532010636121908,
114
- "gt_support": 22670,
115
- "predicted": 16442
116
  },
117
  "2": {
118
- "accuracy": 0.017509727626459144,
119
- "precision": 0.005064715813168261,
120
- "recall": 0.017509727626459144,
121
- "f1": 0.00785683107813182,
122
- "gt_support": 514,
123
- "predicted": 1777
124
  },
125
  "3": {
126
- "accuracy": 0.020851917930419267,
127
- "precision": 0.1671134941912422,
128
- "recall": 0.020851917930419267,
129
- "f1": 0.03707742639040349,
130
- "gt_support": 17936,
131
- "predicted": 2238
132
  },
133
  "4": {
134
- "accuracy": 0.07675367488110679,
135
- "precision": 0.997716894977169,
136
- "recall": 0.07675367488110679,
137
- "f1": 0.14254171371220678,
138
- "gt_support": 74016,
139
- "predicted": 5694
140
  }
141
  },
142
- "total_samples": 115650
143
  },
144
- "11": {
145
- "accuracy": 0.0723095360383496,
146
- "baseline_accuracy": 0.4622222222222222,
147
  "per_class": {
148
  "0": {
149
- "accuracy": 0.7442218798151001,
150
- "precision": 0.0046079432164016065,
151
- "recall": 0.7442218798151001,
152
- "f1": 0.009159176243031062,
153
- "gt_support": 649,
154
- "predicted": 104819
155
  },
156
  "1": {
157
- "accuracy": 0.08017839947291065,
158
- "precision": 0.2636446962753104,
159
- "recall": 0.08017839947291065,
160
- "f1": 0.12296212813089015,
161
- "gt_support": 39462,
162
- "predicted": 12001
163
  },
164
  "2": {
165
- "accuracy": 0.14329738058551617,
166
- "precision": 0.0047369225283960675,
167
- "recall": 0.14329738058551617,
168
- "f1": 0.009170693225520164,
169
- "gt_support": 649,
170
- "predicted": 19633
171
  },
172
  "3": {
173
- "accuracy": 0.022002171092695068,
174
- "precision": 0.25514276941971137,
175
- "recall": 0.022002171092695068,
176
- "f1": 0.040510895529664116,
177
- "gt_support": 37769,
178
- "predicted": 3257
179
  },
180
  "4": {
181
- "accuracy": 0.08871636837738532,
182
- "precision": 0.9482185273159145,
183
- "recall": 0.08871636837738532,
184
- "f1": 0.16225223882619122,
185
- "gt_support": 67496,
186
- "predicted": 6315
187
  }
188
  },
189
- "total_samples": 146025
190
  },
191
- "13": {
192
- "accuracy": 0.08479012345679013,
193
- "baseline_accuracy": 0.37721810699588476,
194
  "per_class": {
195
  "0": {
196
- "accuracy": 0.645679012345679,
197
- "precision": 0.004540955423967215,
198
- "recall": 0.645679012345679,
199
- "f1": 0.009018485308318388,
200
- "gt_support": 810,
201
- "predicted": 115174
202
  },
203
  "1": {
204
- "accuracy": 0.07970295541324675,
205
- "precision": 0.36275314723590585,
206
- "recall": 0.07970295541324675,
207
- "f1": 0.13069092163967563,
208
- "gt_support": 66522,
209
- "predicted": 14616
210
  },
211
  "2": {
212
- "accuracy": 0.19135802469135801,
213
- "precision": 0.004562983897082634,
214
- "recall": 0.19135802469135801,
215
- "f1": 0.008913424767819662,
216
- "gt_support": 810,
217
- "predicted": 33969
218
  },
219
  "3": {
220
- "accuracy": 0.080409611915983,
221
- "precision": 0.4025047327799621,
222
- "recall": 0.080409611915983,
223
- "f1": 0.13404136660119784,
224
- "gt_support": 68748,
225
- "predicted": 13734
226
  },
227
  "4": {
228
- "accuracy": 0.08697089947089948,
229
- "precision": 0.8293041833088081,
230
- "recall": 0.08697089947089948,
231
- "f1": 0.15743161003252393,
232
- "gt_support": 45360,
233
- "predicted": 4757
234
  }
235
  },
236
- "total_samples": 182250
237
  },
238
- "15": {
239
- "accuracy": 0.06777777777777778,
240
- "baseline_accuracy": 0.5240894421392347,
241
  "per_class": {
242
  "0": {
243
- "accuracy": 0.6224066390041494,
244
- "precision": 0.004505147130596706,
245
- "recall": 0.6224066390041494,
246
- "f1": 0.008945544000894555,
247
- "gt_support": 964,
248
- "predicted": 133181
249
  },
250
  "1": {
251
- "accuracy": 0.07335853973957768,
252
- "precision": 0.45830763537683483,
253
- "recall": 0.07335853973957768,
254
- "f1": 0.12647326633251357,
255
- "gt_support": 101297,
256
- "predicted": 16214
257
  },
258
  "2": {
259
- "accuracy": 0.23443983402489627,
260
- "precision": 0.0044561873964823724,
261
- "recall": 0.23443983402489627,
262
- "f1": 0.008746130030959753,
263
- "gt_support": 964,
264
- "predicted": 50716
265
  },
266
  "3": {
267
- "accuracy": 0.056687926105124256,
268
- "precision": 0.5275049115913556,
269
- "recall": 0.056687926105124256,
270
- "f1": 0.10237427615953483,
271
- "gt_support": 113675,
272
- "predicted": 12216
273
  },
274
  "4": {
275
- "accuracy": 0.0,
276
- "precision": 0.0,
277
- "recall": 0.0,
278
- "f1": 0.0,
279
- "gt_support": 0,
280
- "predicted": 4573
281
  }
282
  },
283
- "total_samples": 216900
284
- }
285
- },
286
- "by_complexity": {
287
- "0.0": {
288
- "accuracy": 0.11257347670250896,
289
- "baseline_accuracy": 0.4410322580645161,
290
  "per_class": {
291
  "0": {
292
- "accuracy": 0.4838709677419355,
293
- "precision": 0.004835745833199007,
294
- "recall": 0.4838709677419355,
295
- "f1": 0.009575792396820836,
296
- "gt_support": 310,
297
- "predicted": 31019
298
  },
299
  "1": {
300
- "accuracy": 0.11058272946859904,
301
- "precision": 0.17933651609744156,
302
- "recall": 0.11058272946859904,
303
- "f1": 0.13680720922631553,
304
- "gt_support": 13248,
305
- "predicted": 8169
306
  },
307
  "2": {
308
- "accuracy": 0.32903225806451614,
309
- "precision": 0.004619983694175197,
310
- "recall": 0.32903225806451614,
311
- "f1": 0.009112024298731461,
312
- "gt_support": 310,
313
- "predicted": 22078
314
  },
315
  "3": {
316
- "accuracy": 0.07801833430856252,
317
- "precision": 0.547945205479452,
318
- "recall": 0.07801833430856252,
319
- "f1": 0.13658869728529965,
320
- "gt_support": 30762,
321
- "predicted": 4380
322
  },
323
  "4": {
324
- "accuracy": 0.14868630573248406,
325
- "precision": 0.9100877192982456,
326
- "recall": 0.14868630573248406,
327
- "f1": 0.25561182589652337,
328
- "gt_support": 25120,
329
- "predicted": 4104
330
  }
331
  },
332
- "total_samples": 69750
 
 
 
 
 
 
 
 
 
 
 
 
 
333
  },
334
- "0.2": {
335
- "accuracy": 0.09407179487179487,
336
- "baseline_accuracy": 0.3972239316239316,
 
 
 
337
  "per_class": {
338
  "0": {
339
- "accuracy": 0.6738461538461539,
340
- "precision": 0.0046609628399948925,
341
- "recall": 0.6738461538461539,
342
- "f1": 0.009257889285789776,
343
- "gt_support": 325,
344
- "predicted": 46986
345
  },
346
  "1": {
347
- "accuracy": 0.08564201397253673,
348
- "precision": 0.20563991323210412,
349
- "recall": 0.08564201397253673,
350
- "f1": 0.12092350865257877,
351
- "gt_support": 16604,
352
- "predicted": 6915
353
  },
354
  "2": {
355
- "accuracy": 0.16,
356
- "precision": 0.004732004732004732,
357
- "recall": 0.16,
358
- "f1": 0.009192151316952448,
359
- "gt_support": 325,
360
- "predicted": 10989
361
  },
362
  "3": {
363
- "accuracy": 0.08627396977312632,
364
- "precision": 0.5101791530944625,
365
- "recall": 0.08627396977312632,
366
- "f1": 0.1475897405695103,
367
- "gt_support": 29047,
368
- "predicted": 4912
369
  },
370
  "4": {
371
- "accuracy": 0.09991052788547569,
372
- "precision": 0.806500150466446,
373
- "recall": 0.09991052788547569,
374
- "f1": 0.17779546886920752,
375
- "gt_support": 26824,
376
- "predicted": 3323
377
  }
378
  },
379
- "total_samples": 73125
380
  },
381
- "0.4": {
382
- "accuracy": 0.07845911949685534,
383
- "baseline_accuracy": 0.4000524109014675,
384
  "per_class": {
385
  "0": {
386
- "accuracy": 0.6863207547169812,
387
- "precision": 0.004570728489303553,
388
- "recall": 0.6863207547169812,
389
- "f1": 0.009080979872054923,
390
- "gt_support": 424,
391
- "predicted": 63666
392
  },
393
  "1": {
394
- "accuracy": 0.14222288549789203,
395
- "precision": 0.26072088092469736,
396
- "recall": 0.14222288549789203,
397
- "f1": 0.18404789494013135,
398
- "gt_support": 26803,
399
- "predicted": 14621
400
  },
401
  "2": {
402
- "accuracy": 0.13443396226415094,
403
- "precision": 0.004588633070359041,
404
- "recall": 0.13443396226415094,
405
- "f1": 0.008874357776739842,
406
- "gt_support": 424,
407
- "predicted": 12422
408
  },
409
  "3": {
410
- "accuracy": 0.01915367483296214,
411
- "precision": 0.5236389684813754,
412
- "recall": 0.01915367483296214,
413
- "f1": 0.03695558757362048,
414
- "gt_support": 38165,
415
- "predicted": 1396
416
  },
417
  "4": {
418
- "accuracy": 0.08768253109789075,
419
- "precision": 0.7872534142640364,
420
- "recall": 0.08768253109789075,
421
- "f1": 0.15779068706469176,
422
- "gt_support": 29584,
423
- "predicted": 3295
424
  }
425
  },
426
- "total_samples": 95400
427
  },
428
- "0.6": {
429
- "accuracy": 0.07742798353909465,
430
- "baseline_accuracy": 0.3553909465020576,
431
  "per_class": {
432
  "0": {
433
- "accuracy": 0.7361111111111112,
434
- "precision": 0.004585634562417985,
435
- "recall": 0.7361111111111112,
436
- "f1": 0.009114490032817897,
437
- "gt_support": 432,
438
- "predicted": 69347
439
  },
440
  "1": {
441
- "accuracy": 0.08617078773138764,
442
- "precision": 0.30075453902381516,
443
- "recall": 0.08617078773138764,
444
- "f1": 0.13395998529643438,
445
- "gt_support": 29604,
446
- "predicted": 8482
447
  },
448
  "2": {
449
- "accuracy": 0.12268518518518519,
450
- "precision": 0.004508719693747341,
451
- "recall": 0.12268518518518519,
452
- "f1": 0.00869779272995815,
453
- "gt_support": 432,
454
- "predicted": 11755
455
  },
456
  "3": {
457
- "accuracy": 0.04908661613023487,
458
- "precision": 0.4066924066924067,
459
- "recall": 0.04908661613023487,
460
- "f1": 0.08760014415213596,
461
- "gt_support": 32188,
462
- "predicted": 3885
463
  },
464
  "4": {
465
- "accuracy": 0.08754052802223251,
466
- "precision": 0.8105065666041276,
467
- "recall": 0.08754052802223251,
468
- "f1": 0.1580143696930111,
469
- "gt_support": 34544,
470
- "predicted": 3731
471
  }
472
  },
473
- "total_samples": 97200
474
  },
475
- "0.8": {
476
- "accuracy": 0.07645909645909646,
477
- "baseline_accuracy": 0.3470899470899471,
478
  "per_class": {
479
  "0": {
480
- "accuracy": 0.7326007326007326,
481
- "precision": 0.0045910520395748685,
482
- "recall": 0.7326007326007326,
483
- "f1": 0.009124920156948626,
484
- "gt_support": 546,
485
- "predicted": 87126
486
  },
487
  "1": {
488
- "accuracy": 0.07938555347091933,
489
- "precision": 0.2968256752016836,
490
- "recall": 0.07938555347091933,
491
- "f1": 0.1252682999037821,
492
- "gt_support": 42640,
493
- "predicted": 11404
494
  },
495
  "2": {
496
- "accuracy": 0.13003663003663005,
497
- "precision": 0.00464143296071125,
498
- "recall": 0.13003663003663005,
499
- "f1": 0.008962948936438803,
500
- "gt_support": 546,
501
- "predicted": 15297
502
  },
503
  "3": {
504
- "accuracy": 0.04387205570363923,
505
- "precision": 0.3726033726033726,
506
- "recall": 0.04387205570363923,
507
- "f1": 0.0785010341890741,
508
- "gt_support": 36766,
509
- "predicted": 4329
510
  },
511
  "4": {
512
- "accuracy": 0.092652058934643,
513
- "precision": 0.835960801022582,
514
- "recall": 0.092652058934643,
515
- "f1": 0.16681545721209032,
516
- "gt_support": 42352,
517
- "predicted": 4694
518
  }
519
  },
520
- "total_samples": 122850
521
  },
522
- "1.0": {
523
- "accuracy": 0.06971706951470864,
524
- "baseline_accuracy": 0.4095034663668728,
525
  "per_class": {
526
  "0": {
527
- "accuracy": 0.7411467116357504,
528
- "precision": 0.004564218397071423,
529
- "recall": 0.7411467116357504,
530
- "f1": 0.009072565038111997,
531
- "gt_support": 1186,
532
- "predicted": 192585
533
  },
534
  "1": {
535
- "accuracy": 0.0717998462608441,
536
- "precision": 0.40877357507554446,
537
- "recall": 0.0717998462608441,
538
- "f1": 0.1221452479178018,
539
- "gt_support": 109276,
540
- "predicted": 19194
541
  },
542
  "2": {
543
- "accuracy": 0.12900505902192244,
544
- "precision": 0.004446769552707298,
545
- "recall": 0.12900505902192244,
546
- "f1": 0.008597196077880483,
547
- "gt_support": 1186,
548
- "predicted": 34407
549
  },
550
  "3": {
551
- "accuracy": 0.057120050249941115,
552
- "precision": 0.3418703007518797,
553
- "recall": 0.057120050249941115,
554
- "f1": 0.09788531832350371,
555
- "gt_support": 76418,
556
- "predicted": 12768
557
  },
558
  "4": {
559
- "accuracy": 0.06804681153533712,
560
- "precision": 0.6789513677811551,
561
- "recall": 0.06804681153533712,
562
- "f1": 0.12369635440701429,
563
- "gt_support": 78784,
564
- "predicted": 7896
565
  }
566
  },
567
- "total_samples": 266850
568
  }
569
  },
570
  "by_size_complexity": {
571
  "7_0.0": {
572
- "accuracy": 0.14818713450292398,
573
  "baseline_accuracy": 0.7822222222222223,
574
  "per_class": {
575
  "0": {
576
- "accuracy": 0.7894736842105263,
577
- "precision": 0.005098572399728076,
578
- "recall": 0.7894736842105263,
579
- "f1": 0.010131712259371832,
580
  "gt_support": 38,
581
- "predicted": 5884
582
  },
583
  "1": {
584
- "accuracy": 0.13157894736842105,
585
- "precision": 0.11374407582938388,
586
- "recall": 0.13157894736842105,
587
- "f1": 0.12201321809862735,
588
  "gt_support": 912,
589
- "predicted": 1055
590
  },
591
  "2": {
592
- "accuracy": 0.07894736842105263,
593
- "precision": 0.006036217303822937,
594
- "recall": 0.07894736842105263,
595
- "f1": 0.011214953271028038,
596
  "gt_support": 38,
597
- "predicted": 497
598
  },
599
  "3": {
600
  "accuracy": 0.0,
@@ -605,43 +619,43 @@
605
  "predicted": 0
606
  },
607
  "4": {
608
- "accuracy": 0.166566985645933,
609
- "precision": 1.0,
610
- "recall": 0.166566985645933,
611
- "f1": 0.2855678031274032,
612
  "gt_support": 6688,
613
- "predicted": 1114
614
  }
615
  },
616
  "total_samples": 8550
617
  },
618
  "7_0.2": {
619
- "accuracy": 0.0994949494949495,
620
  "baseline_accuracy": 0.7822222222222223,
621
  "per_class": {
622
  "0": {
623
- "accuracy": 0.7727272727272727,
624
- "precision": 0.004771929824561404,
625
- "recall": 0.7727272727272727,
626
- "f1": 0.00948528386106849,
627
  "gt_support": 44,
628
- "predicted": 7125
629
  },
630
  "1": {
631
- "accuracy": 0.18181818181818182,
632
- "precision": 0.12395709177592372,
633
- "recall": 0.18181818181818182,
634
- "f1": 0.147413182140326,
635
  "gt_support": 1144,
636
- "predicted": 1678
637
  },
638
  "2": {
639
- "accuracy": 0.045454545454545456,
640
- "precision": 0.0056179775280898875,
641
- "recall": 0.045454545454545456,
642
- "f1": 0.01,
643
  "gt_support": 44,
644
- "predicted": 356
645
  },
646
  "3": {
647
  "accuracy": 0.0,
@@ -652,44 +666,44 @@
652
  "predicted": 0
653
  },
654
  "4": {
655
- "accuracy": 0.09568698347107438,
656
- "precision": 1.0,
657
- "recall": 0.09568698347107438,
658
- "f1": 0.1746611667648792,
659
  "gt_support": 7744,
660
- "predicted": 741
661
  }
662
  },
663
  "total_samples": 9900
664
  },
665
  "7_0.4": {
666
- "accuracy": 0.11349593495934959,
667
  "baseline_accuracy": 0.7822222222222223,
668
  "per_class": {
669
  "0": {
670
- "accuracy": 0.7560975609756098,
671
- "precision": 0.004810676598386096,
672
- "recall": 0.7560975609756098,
673
- "f1": 0.009560524286815728,
674
  "gt_support": 41,
675
- "predicted": 6444
676
  },
677
  "1": {
678
- "accuracy": 0.24390243902439024,
679
- "precision": 0.13691931540342298,
680
- "recall": 0.24390243902439024,
681
- "f1": 0.17538365173817724,
682
- "gt_support": 1148,
683
- "predicted": 2045
684
- },
685
- "2": {
686
  "accuracy": 0.0,
687
  "precision": 0.0,
688
  "recall": 0.0,
689
  "f1": 0.0,
690
- "gt_support": 41,
691
  "predicted": 0
692
  },
 
 
 
 
 
 
 
 
693
  "3": {
694
  "accuracy": 0.0,
695
  "precision": 0.0,
@@ -699,91 +713,91 @@
699
  "predicted": 0
700
  },
701
  "4": {
702
- "accuracy": 0.10199556541019955,
703
- "precision": 1.0,
704
- "recall": 0.10199556541019955,
705
- "f1": 0.1851106639839034,
706
  "gt_support": 7216,
707
- "predicted": 736
708
  }
709
  },
710
  "total_samples": 9225
711
  },
712
  "7_0.6": {
713
- "accuracy": 0.09342789598108747,
714
  "baseline_accuracy": 0.7822222222222223,
715
  "per_class": {
716
  "0": {
717
- "accuracy": 0.8723404255319149,
718
- "precision": 0.004802623872554762,
719
- "recall": 0.8723404255319149,
720
- "f1": 0.009552656104380242,
721
  "gt_support": 47,
722
- "predicted": 8537
723
  },
724
  "1": {
725
- "accuracy": 0.10638297872340426,
726
- "precision": 0.14091350826044705,
727
- "recall": 0.10638297872340426,
728
- "f1": 0.12123745819397994,
729
- "gt_support": 1363,
730
- "predicted": 1029
731
- },
732
- "2": {
733
  "accuracy": 0.0,
734
  "precision": 0.0,
735
  "recall": 0.0,
736
  "f1": 0.0,
737
- "gt_support": 47,
738
  "predicted": 0
739
  },
 
 
 
 
 
 
 
 
740
  "3": {
741
- "accuracy": 0.02127659574468085,
742
- "precision": 0.08,
743
- "recall": 0.02127659574468085,
744
- "f1": 0.03361344537815126,
745
  "gt_support": 846,
746
- "predicted": 225
747
  },
748
  "4": {
749
- "accuracy": 0.09477756286266925,
750
- "precision": 1.0,
751
- "recall": 0.09477756286266925,
752
- "f1": 0.17314487632508835,
753
  "gt_support": 8272,
754
- "predicted": 784
755
  }
756
  },
757
  "total_samples": 10575
758
  },
759
  "7_0.8": {
760
- "accuracy": 0.12654545454545454,
761
  "baseline_accuracy": 0.7822222222222223,
762
  "per_class": {
763
  "0": {
764
- "accuracy": 0.8181818181818182,
765
- "precision": 0.004866969500324465,
766
- "recall": 0.8181818181818182,
767
- "f1": 0.009676378883990968,
768
  "gt_support": 55,
769
- "predicted": 9246
770
  },
771
  "1": {
772
- "accuracy": 0.18181818181818182,
773
- "precision": 0.1616266944734098,
774
- "recall": 0.18181818181818182,
775
- "f1": 0.17112889870273257,
776
- "gt_support": 1705,
777
- "predicted": 1918
778
- },
779
- "2": {
780
  "accuracy": 0.0,
781
  "precision": 0.0,
782
  "recall": 0.0,
783
  "f1": 0.0,
784
- "gt_support": 55,
785
  "predicted": 0
786
  },
 
 
 
 
 
 
 
 
787
  "3": {
788
  "accuracy": 0.0,
789
  "precision": 0.0,
@@ -793,44 +807,44 @@
793
  "predicted": 0
794
  },
795
  "4": {
796
- "accuracy": 0.12510330578512396,
797
- "precision": 1.0,
798
- "recall": 0.12510330578512396,
799
- "f1": 0.22238545588100264,
800
  "gt_support": 9680,
801
- "predicted": 1211
802
  }
803
  },
804
  "total_samples": 12375
805
  },
806
  "7_1.0": {
807
- "accuracy": 0.1062295081967213,
808
  "baseline_accuracy": 0.7822222222222223,
809
  "per_class": {
810
  "0": {
811
- "accuracy": 0.8524590163934426,
812
- "precision": 0.004805914972273568,
813
- "recall": 0.8524590163934426,
814
- "f1": 0.00955794504181601,
815
  "gt_support": 61,
816
- "predicted": 10820
817
  },
818
  "1": {
819
- "accuracy": 0.14754098360655737,
820
- "precision": 0.16116396194739788,
821
- "recall": 0.14754098360655737,
822
- "f1": 0.1540518855308906,
823
- "gt_support": 1952,
824
- "predicted": 1787
825
- },
826
- "2": {
827
  "accuracy": 0.0,
828
  "precision": 0.0,
829
  "recall": 0.0,
830
  "f1": 0.0,
831
- "gt_support": 61,
832
  "predicted": 0
833
  },
 
 
 
 
 
 
 
 
834
  "3": {
835
  "accuracy": 0.0,
836
  "precision": 0.0,
@@ -840,897 +854,897 @@
840
  "predicted": 0
841
  },
842
  "4": {
843
- "accuracy": 0.10413561847988077,
844
- "precision": 1.0,
845
- "recall": 0.10413561847988077,
846
- "f1": 0.18862831111860975,
847
  "gt_support": 10736,
848
- "predicted": 1118
849
  }
850
  },
851
  "total_samples": 13725
852
  },
853
  "9_0.0": {
854
- "accuracy": 0.16238683127572018,
855
  "baseline_accuracy": 0.64,
856
  "per_class": {
857
  "0": {
858
- "accuracy": 0.7037037037037037,
859
- "precision": 0.005288796102992345,
860
- "recall": 0.7037037037037037,
861
- "f1": 0.010498687664041995,
862
  "gt_support": 54,
863
- "predicted": 7185
864
  },
865
  "1": {
866
- "accuracy": 0.19444444444444445,
867
- "precision": 0.1415929203539823,
868
- "recall": 0.19444444444444445,
869
- "f1": 0.16386247256766645,
870
  "gt_support": 1728,
871
- "predicted": 2373
872
  },
873
  "2": {
874
- "accuracy": 0.05555555555555555,
875
- "precision": 0.0047169811320754715,
876
- "recall": 0.05555555555555555,
877
- "f1": 0.008695652173913044,
878
  "gt_support": 54,
879
- "predicted": 636
880
  },
881
  "3": {
882
- "accuracy": 0.037037037037037035,
883
- "precision": 0.2088888888888889,
884
- "recall": 0.037037037037037035,
885
- "f1": 0.06291834002677375,
886
  "gt_support": 2538,
887
- "predicted": 450
888
  },
889
  "4": {
890
- "accuracy": 0.19315843621399176,
891
- "precision": 0.99734395750332,
892
- "recall": 0.19315843621399176,
893
- "f1": 0.3236371471665589,
894
  "gt_support": 7776,
895
- "predicted": 1506
896
  }
897
  },
898
  "total_samples": 12150
899
  },
900
  "9_0.2": {
901
- "accuracy": 0.1008983451536643,
902
  "baseline_accuracy": 0.64,
903
  "per_class": {
904
  "0": {
905
- "accuracy": 0.7872340425531915,
906
- "precision": 0.004744806360605283,
907
- "recall": 0.7872340425531915,
908
- "f1": 0.009432759719566603,
909
  "gt_support": 47,
910
- "predicted": 7798
911
  },
912
  "1": {
913
- "accuracy": 0.1702127659574468,
914
- "precision": 0.1722488038277512,
915
- "recall": 0.1702127659574468,
916
- "f1": 0.1712247324613555,
917
- "gt_support": 1692,
918
- "predicted": 1672
919
- },
920
- "2": {
921
  "accuracy": 0.0,
922
  "precision": 0.0,
923
  "recall": 0.0,
924
  "f1": 0.0,
925
- "gt_support": 47,
926
  "predicted": 0
927
  },
 
 
 
 
 
 
 
 
928
  "3": {
929
- "accuracy": 0.0425531914893617,
930
- "precision": 0.1915367483296214,
931
- "recall": 0.0425531914893617,
932
- "f1": 0.06963562753036437,
933
  "gt_support": 2021,
934
- "predicted": 449
935
  },
936
  "4": {
937
- "accuracy": 0.09692671394799054,
938
- "precision": 1.0,
939
- "recall": 0.09692671394799054,
940
- "f1": 0.17672413793103448,
941
  "gt_support": 6768,
942
- "predicted": 656
943
  }
944
  },
945
  "total_samples": 10575
946
  },
947
  "9_0.4": {
948
- "accuracy": 0.08398009950248757,
949
  "baseline_accuracy": 0.64,
950
  "per_class": {
951
  "0": {
952
- "accuracy": 0.7611940298507462,
953
- "precision": 0.004619983694175197,
954
- "recall": 0.7611940298507462,
955
- "f1": 0.009184224743381956,
956
  "gt_support": 67,
957
- "predicted": 11039
958
  },
959
  "1": {
960
- "accuracy": 0.23880597014925373,
961
- "precision": 0.1849176538572667,
962
- "recall": 0.23880597014925373,
963
- "f1": 0.20843510828855236,
964
- "gt_support": 2680,
965
- "predicted": 3461
966
- },
967
- "2": {
968
  "accuracy": 0.0,
969
  "precision": 0.0,
970
  "recall": 0.0,
971
  "f1": 0.0,
972
- "gt_support": 67,
973
  "predicted": 0
974
  },
 
 
 
 
 
 
 
 
975
  "3": {
976
- "accuracy": 0.0,
977
- "precision": 0.0,
978
- "recall": 0.0,
979
- "f1": 0.0,
980
  "gt_support": 2613,
981
- "predicted": 0
982
  },
983
  "4": {
984
- "accuracy": 0.059597844112769485,
985
- "precision": 1.0,
986
- "recall": 0.059597844112769485,
987
- "f1": 0.11249144086862956,
988
  "gt_support": 9648,
989
- "predicted": 575
990
  }
991
  },
992
  "total_samples": 15075
993
  },
994
  "9_0.6": {
995
- "accuracy": 0.07919854280510018,
996
  "baseline_accuracy": 0.64,
997
  "per_class": {
998
  "0": {
999
- "accuracy": 0.819672131147541,
1000
- "precision": 0.004605747973470892,
1001
- "recall": 0.819672131147541,
1002
- "f1": 0.009160025648071815,
1003
  "gt_support": 61,
1004
- "predicted": 10856
1005
  },
1006
  "1": {
1007
- "accuracy": 0.14754098360655737,
1008
- "precision": 0.2102118413905486,
1009
- "recall": 0.14754098360655737,
1010
- "f1": 0.17338709677419353,
1011
- "gt_support": 2623,
1012
- "predicted": 1841
1013
- },
1014
- "2": {
1015
  "accuracy": 0.0,
1016
  "precision": 0.0,
1017
  "recall": 0.0,
1018
  "f1": 0.0,
1019
- "gt_support": 61,
1020
  "predicted": 0
1021
  },
 
 
 
 
 
 
 
 
1022
  "3": {
1023
- "accuracy": 0.03278688524590164,
1024
- "precision": 0.16,
1025
- "recall": 0.03278688524590164,
1026
- "f1": 0.05442176870748299,
1027
  "gt_support": 2196,
1028
- "predicted": 450
1029
  },
1030
  "4": {
1031
- "accuracy": 0.06580145719489981,
1032
- "precision": 1.0,
1033
- "recall": 0.06580145719489981,
1034
- "f1": 0.12347788933988464,
1035
  "gt_support": 8784,
1036
- "predicted": 578
1037
  }
1038
  },
1039
  "total_samples": 13725
1040
  },
1041
  "9_0.8": {
1042
- "accuracy": 0.09826182618261826,
1043
  "baseline_accuracy": 0.64,
1044
  "per_class": {
1045
  "0": {
1046
- "accuracy": 0.7425742574257426,
1047
- "precision": 0.0046210720887245845,
1048
- "recall": 0.7425742574257426,
1049
- "f1": 0.009184985610189211,
1050
  "gt_support": 101,
1051
- "predicted": 16230
1052
  },
1053
  "1": {
1054
- "accuracy": 0.20686749526016432,
1055
- "precision": 0.22292849035187287,
1056
- "recall": 0.20686749526016432,
1057
- "f1": 0.21459790209790208,
1058
  "gt_support": 4747,
1059
- "predicted": 4405
1060
  },
1061
  "2": {
1062
- "accuracy": 0.0297029702970297,
1063
- "precision": 0.005639097744360902,
1064
- "recall": 0.0297029702970297,
1065
- "f1": 0.00947867298578199,
1066
  "gt_support": 101,
1067
- "predicted": 532
1068
  },
1069
  "3": {
1070
- "accuracy": 0.019801980198019802,
1071
- "precision": 0.14545454545454545,
1072
- "recall": 0.019801980198019802,
1073
- "f1": 0.034858387799564274,
1074
  "gt_support": 3232,
1075
- "predicted": 440
1076
  },
1077
  "4": {
1078
- "accuracy": 0.07625137513751375,
1079
- "precision": 0.9919499105545617,
1080
- "recall": 0.07625137513751375,
1081
- "f1": 0.14161665176861193,
1082
  "gt_support": 14544,
1083
- "predicted": 1118
1084
  }
1085
  },
1086
  "total_samples": 22725
1087
  },
1088
  "9_1.0": {
1089
- "accuracy": 0.050458937198067634,
1090
  "baseline_accuracy": 0.64,
1091
  "per_class": {
1092
  "0": {
1093
- "accuracy": 0.907608695652174,
1094
- "precision": 0.004589046742326399,
1095
- "recall": 0.907608695652174,
1096
- "f1": 0.00913192071086808,
1097
  "gt_support": 184,
1098
- "predicted": 36391
1099
  },
1100
  "1": {
1101
- "accuracy": 0.06521739130434782,
1102
- "precision": 0.22304832713754646,
1103
- "recall": 0.06521739130434782,
1104
- "f1": 0.1009251471825063,
1105
  "gt_support": 9200,
1106
- "predicted": 2690
1107
  },
1108
  "2": {
1109
- "accuracy": 0.016304347826086956,
1110
- "precision": 0.0049261083743842365,
1111
- "recall": 0.016304347826086956,
1112
- "f1": 0.007566204287515763,
1113
  "gt_support": 184,
1114
- "predicted": 609
1115
  },
1116
  "3": {
1117
- "accuracy": 0.010869565217391304,
1118
- "precision": 0.1291759465478842,
1119
- "recall": 0.010869565217391304,
1120
- "f1": 0.020051858254105444,
1121
  "gt_support": 5336,
1122
- "predicted": 449
1123
  },
1124
  "4": {
1125
- "accuracy": 0.04759208937198068,
1126
- "precision": 1.0,
1127
- "recall": 0.04759208937198068,
1128
- "f1": 0.09085996325251287,
1129
  "gt_support": 26496,
1130
- "predicted": 1261
1131
  }
1132
  },
1133
  "total_samples": 41400
1134
  },
1135
  "11_0.0": {
1136
- "accuracy": 0.09581699346405229,
1137
  "baseline_accuracy": 0.4622222222222222,
1138
  "per_class": {
1139
  "0": {
1140
- "accuracy": 0.4411764705882353,
1141
- "precision": 0.004691164972634871,
1142
- "recall": 0.4411764705882353,
1143
- "f1": 0.009283614420547733,
1144
  "gt_support": 68,
1145
- "predicted": 6395
1146
  },
1147
  "1": {
1148
- "accuracy": 0.15,
1149
- "precision": 0.18681318681318682,
1150
- "recall": 0.15,
1151
- "f1": 0.16639477977161501,
1152
  "gt_support": 2720,
1153
- "predicted": 2184
1154
  },
1155
  "2": {
1156
- "accuracy": 0.36764705882352944,
1157
- "precision": 0.0046702783485895755,
1158
- "recall": 0.36764705882352944,
1159
- "f1": 0.009223390518354546,
1160
  "gt_support": 68,
1161
- "predicted": 5353
1162
  },
1163
  "3": {
1164
- "accuracy": 0.026247207743857037,
1165
- "precision": 0.3472906403940887,
1166
- "recall": 0.026247207743857037,
1167
- "f1": 0.04880581516095535,
1168
  "gt_support": 5372,
1169
- "predicted": 406
1170
  },
1171
  "4": {
1172
- "accuracy": 0.12188914027149321,
1173
- "precision": 0.896049896049896,
1174
- "recall": 0.12188914027149321,
1175
- "f1": 0.214588000995768,
1176
  "gt_support": 7072,
1177
- "predicted": 962
1178
  }
1179
  },
1180
  "total_samples": 15300
1181
  },
1182
  "11_0.2": {
1183
- "accuracy": 0.0764967320261438,
1184
  "baseline_accuracy": 0.4622222222222222,
1185
  "per_class": {
1186
  "0": {
1187
- "accuracy": 0.788235294117647,
1188
- "precision": 0.004684659488183471,
1189
- "recall": 0.788235294117647,
1190
- "f1": 0.009313963995273511,
1191
  "gt_support": 85,
1192
- "predicted": 14302
1193
  },
1194
  "1": {
1195
- "accuracy": 0.058823529411764705,
1196
- "precision": 0.21133093525179855,
1197
- "recall": 0.058823529411764705,
1198
- "f1": 0.09203054630898766,
1199
  "gt_support": 3995,
1200
- "predicted": 1112
1201
  },
1202
  "2": {
1203
- "accuracy": 0.12941176470588237,
1204
- "precision": 0.004963898916967509,
1205
- "recall": 0.12941176470588237,
1206
- "f1": 0.009561060408518037,
1207
  "gt_support": 85,
1208
- "predicted": 2216
1209
  },
1210
  "3": {
1211
- "accuracy": 0.023529411764705882,
1212
- "precision": 0.32,
1213
- "recall": 0.023529411764705882,
1214
- "f1": 0.04383561643835617,
1215
  "gt_support": 6120,
1216
- "predicted": 450
1217
  },
1218
  "4": {
1219
- "accuracy": 0.11380090497737556,
1220
- "precision": 0.9626794258373206,
1221
- "recall": 0.11380090497737556,
1222
- "f1": 0.20354071825998987,
1223
  "gt_support": 8840,
1224
- "predicted": 1045
1225
  }
1226
  },
1227
  "total_samples": 19125
1228
  },
1229
  "11_0.4": {
1230
- "accuracy": 0.10036363636363636,
1231
  "baseline_accuracy": 0.4622222222222222,
1232
  "per_class": {
1233
  "0": {
1234
- "accuracy": 0.7090909090909091,
1235
- "precision": 0.00467065868263473,
1236
- "recall": 0.7090909090909091,
1237
- "f1": 0.00928019036287924,
1238
  "gt_support": 55,
1239
- "predicted": 8350
1240
  },
1241
  "1": {
1242
- "accuracy": 0.21783876500857632,
1243
- "precision": 0.24310872894333843,
1244
- "recall": 0.21783876500857632,
1245
- "f1": 0.22978107472408177,
1246
  "gt_support": 2915,
1247
- "predicted": 2612
1248
  },
1249
  "2": {
1250
- "accuracy": 0.07272727272727272,
1251
- "precision": 0.004722550177095631,
1252
- "recall": 0.07272727272727272,
1253
- "f1": 0.008869179600886918,
1254
  "gt_support": 55,
1255
- "predicted": 847
1256
  },
1257
  "3": {
1258
- "accuracy": 0.0,
1259
- "precision": 0.0,
1260
- "recall": 0.0,
1261
- "f1": 0.0,
1262
  "gt_support": 3630,
1263
- "predicted": 0
1264
  },
1265
  "4": {
1266
- "accuracy": 0.0986013986013986,
1267
- "precision": 0.9964664310954063,
1268
- "recall": 0.0986013986013986,
1269
- "f1": 0.1794463888005091,
1270
  "gt_support": 5720,
1271
- "predicted": 566
1272
  }
1273
  },
1274
  "total_samples": 12375
1275
  },
1276
  "11_0.6": {
1277
- "accuracy": 0.07529665587918015,
1278
  "baseline_accuracy": 0.4622222222222222,
1279
  "per_class": {
1280
  "0": {
1281
- "accuracy": 0.8252427184466019,
1282
- "precision": 0.00463291001253611,
1283
- "recall": 0.8252427184466019,
1284
- "f1": 0.00921409214092141,
1285
  "gt_support": 103,
1286
- "predicted": 18347
1287
  },
1288
  "1": {
1289
- "accuracy": 0.07766990291262135,
1290
- "precision": 0.2803738317757009,
1291
- "recall": 0.07766990291262135,
1292
- "f1": 0.12164216928535222,
1293
  "gt_support": 6180,
1294
- "predicted": 1712
1295
  },
1296
  "2": {
1297
- "accuracy": 0.07766990291262135,
1298
- "precision": 0.0049079754601227,
1299
- "recall": 0.07766990291262135,
1300
- "f1": 0.00923254472013849,
1301
  "gt_support": 103,
1302
- "predicted": 1630
1303
  },
1304
  "3": {
1305
- "accuracy": 0.01053151225933849,
1306
- "precision": 0.26666666666666666,
1307
- "recall": 0.01053151225933849,
1308
- "f1": 0.020262782966598065,
1309
  "gt_support": 6077,
1310
- "predicted": 240
1311
  },
1312
  "4": {
1313
- "accuracy": 0.10343539955190441,
1314
- "precision": 0.8892455858747994,
1315
- "recall": 0.10343539955190441,
1316
- "f1": 0.18531527011205887,
1317
  "gt_support": 10712,
1318
- "predicted": 1246
1319
  }
1320
  },
1321
  "total_samples": 23175
1322
  },
1323
  "11_0.8": {
1324
- "accuracy": 0.05769980506822612,
1325
  "baseline_accuracy": 0.4622222222222222,
1326
  "per_class": {
1327
  "0": {
1328
- "accuracy": 0.8070175438596491,
1329
- "precision": 0.004566436690326103,
1330
- "recall": 0.8070175438596491,
1331
- "f1": 0.009081486599871675,
1332
  "gt_support": 114,
1333
- "predicted": 20147
1334
  },
1335
  "1": {
1336
- "accuracy": 0.043859649122807015,
1337
- "precision": 0.30358785648574055,
1338
- "recall": 0.043859649122807015,
1339
- "f1": 0.07664615027290675,
1340
  "gt_support": 7524,
1341
- "predicted": 1087
1342
  },
1343
  "2": {
1344
- "accuracy": 0.13157894736842105,
1345
- "precision": 0.004932587964485367,
1346
- "recall": 0.13157894736842105,
1347
- "f1": 0.009508716323296355,
1348
  "gt_support": 114,
1349
- "predicted": 3041
1350
  },
1351
  "3": {
1352
- "accuracy": 0.017543859649122806,
1353
- "precision": 0.25118483412322273,
1354
- "recall": 0.017543859649122806,
1355
- "f1": 0.03279702970297029,
1356
  "gt_support": 6042,
1357
- "predicted": 422
1358
  },
1359
  "4": {
1360
- "accuracy": 0.07903171390013496,
1361
- "precision": 0.9832109129066107,
1362
- "recall": 0.07903171390013496,
1363
- "f1": 0.14630338043563124,
1364
  "gt_support": 11856,
1365
- "predicted": 953
1366
  }
1367
  },
1368
  "total_samples": 25650
1369
  },
1370
  "11_1.0": {
1371
- "accuracy": 0.06275793650793651,
1372
  "baseline_accuracy": 0.4622222222222222,
1373
  "per_class": {
1374
  "0": {
1375
- "accuracy": 0.7589285714285714,
1376
- "precision": 0.004560330489833146,
1377
- "recall": 0.7589285714285714,
1378
- "f1": 0.009066183136899365,
1379
  "gt_support": 224,
1380
- "predicted": 37278
1381
  },
1382
  "1": {
1383
- "accuracy": 0.06671626984126984,
1384
- "precision": 0.32665452337583484,
1385
- "recall": 0.06671626984126984,
1386
- "f1": 0.11080218309133973,
1387
  "gt_support": 16128,
1388
- "predicted": 3294
1389
  },
1390
  "2": {
1391
- "accuracy": 0.13392857142857142,
1392
- "precision": 0.00458295142071494,
1393
- "recall": 0.13392857142857142,
1394
- "f1": 0.008862629246676515,
1395
  "gt_support": 224,
1396
- "predicted": 6546
1397
  },
1398
  "3": {
1399
- "accuracy": 0.03571428571428571,
1400
- "precision": 0.21621621621621623,
1401
- "recall": 0.03571428571428571,
1402
- "f1": 0.06130268199233717,
1403
  "gt_support": 10528,
1404
- "predicted": 1739
1405
  },
1406
  "4": {
1407
- "accuracy": 0.06486092032967034,
1408
- "precision": 0.9792611795204148,
1409
- "recall": 0.06486092032967034,
1410
- "f1": 0.12166351302387376,
1411
  "gt_support": 23296,
1412
- "predicted": 1543
1413
  }
1414
  },
1415
  "total_samples": 50400
1416
  },
1417
  "13_0.0": {
1418
- "accuracy": 0.12055555555555555,
1419
  "baseline_accuracy": 0.5288888888888889,
1420
  "per_class": {
1421
  "0": {
1422
- "accuracy": 0.359375,
1423
- "precision": 0.004502740798747063,
1424
- "recall": 0.359375,
1425
- "f1": 0.008894044856921886,
1426
  "gt_support": 64,
1427
- "predicted": 5108
1428
  },
1429
  "1": {
1430
- "accuracy": 0.078125,
1431
- "precision": 0.21505376344086022,
1432
- "recall": 0.078125,
1433
- "f1": 0.11461318051575929,
1434
  "gt_support": 3072,
1435
- "predicted": 1116
1436
  },
1437
  "2": {
1438
- "accuracy": 0.40625,
1439
- "precision": 0.004573438874230431,
1440
- "recall": 0.40625,
1441
- "f1": 0.009045051313271874,
1442
  "gt_support": 64,
1443
- "predicted": 5685
1444
  },
1445
  "3": {
1446
- "accuracy": 0.15625,
1447
- "precision": 0.536036036036036,
1448
- "recall": 0.15625,
1449
- "f1": 0.24196827978853194,
1450
  "gt_support": 7616,
1451
- "predicted": 2220
1452
  },
1453
  "4": {
1454
- "accuracy": 0.07170758928571429,
1455
- "precision": 0.948339483394834,
1456
- "recall": 0.07170758928571429,
1457
- "f1": 0.13333333333333336,
1458
  "gt_support": 3584,
1459
- "predicted": 271
1460
  }
1461
  },
1462
  "total_samples": 14400
1463
  },
1464
  "13_0.2": {
1465
- "accuracy": 0.12903225806451613,
1466
  "baseline_accuracy": 0.48444444444444446,
1467
  "per_class": {
1468
  "0": {
1469
- "accuracy": 0.5645161290322581,
1470
- "precision": 0.004562043795620438,
1471
- "recall": 0.5645161290322581,
1472
- "f1": 0.009050943884147919,
1473
  "gt_support": 62,
1474
- "predicted": 7672
1475
  },
1476
  "1": {
1477
- "accuracy": 0.0967741935483871,
1478
- "precision": 0.25892857142857145,
1479
- "recall": 0.0967741935483871,
1480
- "f1": 0.14089068825910933,
1481
  "gt_support": 3596,
1482
- "predicted": 1344
1483
  },
1484
  "2": {
1485
- "accuracy": 0.16129032258064516,
1486
- "precision": 0.004351610095735422,
1487
- "recall": 0.16129032258064516,
1488
- "f1": 0.008474576271186439,
1489
  "gt_support": 62,
1490
- "predicted": 2298
1491
  },
1492
  "3": {
1493
- "accuracy": 0.1672092335010358,
1494
- "precision": 0.48812095032397407,
1495
- "recall": 0.1672092335010358,
1496
- "f1": 0.2490907086961314,
1497
  "gt_support": 6758,
1498
- "predicted": 2315
1499
  },
1500
  "4": {
1501
- "accuracy": 0.0797811059907834,
1502
- "precision": 0.8629283489096573,
1503
- "recall": 0.0797811059907834,
1504
- "f1": 0.14605852886896917,
1505
  "gt_support": 3472,
1506
- "predicted": 321
1507
  }
1508
  },
1509
  "total_samples": 13950
1510
  },
1511
  "13_0.4": {
1512
- "accuracy": 0.07409777777777778,
1513
  "baseline_accuracy": 0.44,
1514
  "per_class": {
1515
  "0": {
1516
- "accuracy": 0.688,
1517
- "precision": 0.00454281337488775,
1518
- "recall": 0.688,
1519
- "f1": 0.009026028547439127,
1520
  "gt_support": 125,
1521
- "predicted": 18931
1522
  },
1523
  "1": {
1524
- "accuracy": 0.10141176470588235,
1525
- "precision": 0.30918220946915353,
1526
- "recall": 0.10141176470588235,
1527
- "f1": 0.1527285613040397,
1528
  "gt_support": 8500,
1529
- "predicted": 2788
1530
  },
1531
  "2": {
1532
- "accuracy": 0.176,
1533
- "precision": 0.004674883127921802,
1534
- "recall": 0.176,
1535
- "f1": 0.009107845166632166,
1536
  "gt_support": 125,
1537
- "predicted": 4706
1538
  },
1539
  "3": {
1540
- "accuracy": 0.03191919191919192,
1541
- "precision": 0.45454545454545453,
1542
- "recall": 0.03191919191919192,
1543
- "f1": 0.05964965267290848,
1544
  "gt_support": 12375,
1545
- "predicted": 869
1546
  },
1547
  "4": {
1548
- "accuracy": 0.10271428571428572,
1549
- "precision": 0.865222623345367,
1550
- "recall": 0.10271428571428572,
1551
- "f1": 0.18362916613459326,
1552
  "gt_support": 7000,
1553
- "predicted": 831
1554
  }
1555
  },
1556
  "total_samples": 28125
1557
  },
1558
  "13_0.6": {
1559
- "accuracy": 0.0725068870523416,
1560
  "baseline_accuracy": 0.39555555555555555,
1561
  "per_class": {
1562
  "0": {
1563
- "accuracy": 0.6776859504132231,
1564
- "precision": 0.004549489569462938,
1565
- "recall": 0.6776859504132231,
1566
- "f1": 0.009038302562689447,
1567
  "gt_support": 121,
1568
- "predicted": 18024
1569
  },
1570
  "1": {
1571
- "accuracy": 0.07840644204280568,
1572
- "precision": 0.35221323179438363,
1573
- "recall": 0.07840644204280568,
1574
- "f1": 0.12826068116821215,
1575
  "gt_support": 9438,
1576
- "predicted": 2101
1577
  },
1578
  "2": {
1579
- "accuracy": 0.18181818181818182,
1580
- "precision": 0.004348685510970548,
1581
- "recall": 0.18181818181818182,
1582
- "f1": 0.008494208494208495,
1583
  "gt_support": 121,
1584
- "predicted": 5059
1585
  },
1586
  "3": {
1587
- "accuracy": 0.053486860432723556,
1588
- "precision": 0.40336134453781514,
1589
- "recall": 0.053486860432723556,
1590
- "f1": 0.09444945478396327,
1591
  "gt_support": 10769,
1592
- "predicted": 1428
1593
  },
1594
  "4": {
1595
- "accuracy": 0.08175914994096813,
1596
- "precision": 0.9037520391517129,
1597
- "recall": 0.08175914994096813,
1598
- "f1": 0.14995263229124375,
1599
  "gt_support": 6776,
1600
- "predicted": 613
1601
  }
1602
  },
1603
  "total_samples": 27225
1604
  },
1605
  "13_0.8": {
1606
- "accuracy": 0.09583333333333334,
1607
  "baseline_accuracy": 0.39555555555555555,
1608
  "per_class": {
1609
  "0": {
1610
- "accuracy": 0.6160714285714286,
1611
- "precision": 0.004555056773171376,
1612
- "recall": 0.6160714285714286,
1613
- "f1": 0.009043250327653999,
1614
  "gt_support": 112,
1615
- "predicted": 15148
1616
  },
1617
  "1": {
1618
- "accuracy": 0.09741171749598716,
1619
- "precision": 0.4004123711340206,
1620
- "recall": 0.09741171749598716,
1621
- "f1": 0.15670136367304124,
1622
  "gt_support": 9968,
1623
- "predicted": 2425
1624
  },
1625
  "2": {
1626
- "accuracy": 0.20535714285714285,
1627
- "precision": 0.004686226568867156,
1628
- "recall": 0.20535714285714285,
1629
- "f1": 0.009163346613545816,
1630
  "gt_support": 112,
1631
- "predicted": 4908
1632
  },
1633
  "3": {
1634
- "accuracy": 0.07841117216117216,
1635
- "precision": 0.3590146750524109,
1636
- "recall": 0.07841117216117216,
1637
- "f1": 0.1287110108981586,
1638
  "gt_support": 8736,
1639
- "predicted": 1908
1640
  },
1641
  "4": {
1642
- "accuracy": 0.10634566326530612,
1643
- "precision": 0.8224414303329223,
1644
- "recall": 0.10634566326530612,
1645
- "f1": 0.1883382747423408,
1646
  "gt_support": 6272,
1647
- "predicted": 811
1648
  }
1649
  },
1650
  "total_samples": 25200
1651
  },
1652
  "13_1.0": {
1653
- "accuracy": 0.07421949556918882,
1654
  "baseline_accuracy": 0.43555555555555553,
1655
  "per_class": {
1656
  "0": {
1657
- "accuracy": 0.6993865030674846,
1658
- "precision": 0.004533614364399197,
1659
- "recall": 0.6993865030674846,
1660
- "f1": 0.009008831025149653,
1661
  "gt_support": 326,
1662
- "predicted": 50291
1663
  },
1664
  "1": {
1665
- "accuracy": 0.06701514961812946,
1666
- "precision": 0.44217265592730276,
1667
- "recall": 0.06701514961812946,
1668
- "f1": 0.11639032345746127,
1669
  "gt_support": 31948,
1670
- "predicted": 4842
1671
  },
1672
  "2": {
1673
- "accuracy": 0.15950920245398773,
1674
- "precision": 0.004596481923450897,
1675
- "recall": 0.15950920245398773,
1676
- "f1": 0.008935475556319272,
1677
  "gt_support": 326,
1678
- "predicted": 11313
1679
  },
1680
  "3": {
1681
- "accuracy": 0.06899617675824664,
1682
- "precision": 0.3107729275130156,
1683
- "recall": 0.06899617675824664,
1684
- "f1": 0.1129220023282887,
1685
  "gt_support": 22494,
1686
- "predicted": 4994
1687
  },
1688
  "4": {
1689
- "accuracy": 0.08057624890446977,
1690
- "precision": 0.7701570680628272,
1691
- "recall": 0.08057624890446977,
1692
- "f1": 0.1458891203014976,
1693
  "gt_support": 18256,
1694
- "predicted": 1910
1695
  }
1696
  },
1697
  "total_samples": 73350
1698
  },
1699
  "15_0.0": {
1700
- "accuracy": 0.07286821705426356,
1701
  "baseline_accuracy": 0.7422222222222222,
1702
  "per_class": {
1703
  "0": {
1704
- "accuracy": 0.3372093023255814,
1705
- "precision": 0.004498216224600589,
1706
- "recall": 0.3372093023255814,
1707
- "f1": 0.008878003979794886,
1708
  "gt_support": 86,
1709
- "predicted": 6447
1710
  },
1711
  "1": {
1712
- "accuracy": 0.07495847176079734,
1713
- "precision": 0.2505204718945177,
1714
- "recall": 0.07495847176079734,
1715
- "f1": 0.11539076234617228,
1716
  "gt_support": 4816,
1717
- "predicted": 1441
1718
  },
1719
  "2": {
1720
- "accuracy": 0.5232558139534884,
1721
- "precision": 0.004542242858584839,
1722
- "recall": 0.5232558139534884,
1723
- "f1": 0.009006304413089162,
1724
  "gt_support": 86,
1725
- "predicted": 9907
1726
  },
1727
  "3": {
1728
- "accuracy": 0.067887480852249,
1729
- "precision": 0.7476993865030674,
1730
- "recall": 0.067887480852249,
1731
- "f1": 0.12447338184603601,
1732
  "gt_support": 14362,
1733
- "predicted": 1304
1734
  },
1735
  "4": {
1736
  "accuracy": 0.0,
@@ -1738,46 +1752,46 @@
1738
  "recall": 0.0,
1739
  "f1": 0.0,
1740
  "gt_support": 0,
1741
- "predicted": 251
1742
  }
1743
  },
1744
  "total_samples": 19350
1745
  },
1746
  "15_0.2": {
1747
- "accuracy": 0.07989782886334611,
1748
  "baseline_accuracy": 0.6755555555555556,
1749
  "per_class": {
1750
  "0": {
1751
- "accuracy": 0.5287356321839081,
1752
- "precision": 0.00455942115174943,
1753
- "recall": 0.5287356321839081,
1754
- "f1": 0.009040880503144654,
1755
  "gt_support": 87,
1756
- "predicted": 10089
1757
  },
1758
  "1": {
1759
- "accuracy": 0.05552857374129837,
1760
- "precision": 0.30928764652840396,
1761
- "recall": 0.05552857374129837,
1762
- "f1": 0.09415317046390338,
1763
  "gt_support": 6177,
1764
- "predicted": 1109
1765
  },
1766
  "2": {
1767
- "accuracy": 0.3333333333333333,
1768
- "precision": 0.004739336492890996,
1769
- "recall": 0.3333333333333333,
1770
- "f1": 0.009345794392523366,
1771
  "gt_support": 87,
1772
- "predicted": 6119
1773
  },
1774
  "3": {
1775
- "accuracy": 0.0866606170598911,
1776
- "precision": 0.6749116607773852,
1777
- "recall": 0.0866606170598911,
1778
- "f1": 0.15359871330920785,
1779
  "gt_support": 13224,
1780
- "predicted": 1698
1781
  },
1782
  "4": {
1783
  "accuracy": 0.0,
@@ -1785,46 +1799,46 @@
1785
  "recall": 0.0,
1786
  "f1": 0.0,
1787
  "gt_support": 0,
1788
- "predicted": 560
1789
  }
1790
  },
1791
  "total_samples": 19575
1792
  },
1793
  "15_0.4": {
1794
- "accuracy": 0.06032679738562091,
1795
  "baseline_accuracy": 0.6133333333333333,
1796
  "per_class": {
1797
  "0": {
1798
- "accuracy": 0.6176470588235294,
1799
- "precision": 0.004443974182626177,
1800
- "recall": 0.6176470588235294,
1801
- "f1": 0.00882445635045698,
1802
  "gt_support": 136,
1803
- "predicted": 18902
1804
  },
1805
  "1": {
1806
- "accuracy": 0.12067474048442907,
1807
- "precision": 0.3755047106325707,
1808
- "recall": 0.12067474048442907,
1809
- "f1": 0.18265139116202947,
1810
  "gt_support": 11560,
1811
- "predicted": 3715
1812
  },
1813
  "2": {
1814
- "accuracy": 0.22794117647058823,
1815
- "precision": 0.004513029553064493,
1816
- "recall": 0.22794117647058823,
1817
- "f1": 0.008850820842255532,
1818
  "gt_support": 136,
1819
- "predicted": 6869
1820
  },
1821
  "3": {
1822
- "accuracy": 0.017902813299232736,
1823
- "precision": 0.6375711574952562,
1824
- "recall": 0.017902813299232736,
1825
- "f1": 0.034827675563617516,
1826
  "gt_support": 18768,
1827
- "predicted": 527
1828
  },
1829
  "4": {
1830
  "accuracy": 0.0,
@@ -1832,46 +1846,46 @@
1832
  "recall": 0.0,
1833
  "f1": 0.0,
1834
  "gt_support": 0,
1835
- "predicted": 587
1836
  }
1837
  },
1838
  "total_samples": 30600
1839
  },
1840
  "15_0.6": {
1841
- "accuracy": 0.07697777777777778,
1842
  "baseline_accuracy": 0.5466666666666666,
1843
  "per_class": {
1844
  "0": {
1845
- "accuracy": 0.6,
1846
- "precision": 0.004417286313774571,
1847
- "recall": 0.6,
1848
- "f1": 0.008770006577504933,
1849
  "gt_support": 100,
1850
- "predicted": 13583
1851
  },
1852
  "1": {
1853
- "accuracy": 0.0799,
1854
- "precision": 0.4441356309060589,
1855
- "recall": 0.0799,
1856
- "f1": 0.1354352063734215,
1857
  "gt_support": 10000,
1858
- "predicted": 1799
1859
  },
1860
  "2": {
1861
- "accuracy": 0.23,
1862
- "precision": 0.0045400710619818395,
1863
- "recall": 0.23,
1864
- "f1": 0.008904374758033295,
1865
  "gt_support": 100,
1866
- "predicted": 5066
1867
  },
1868
  "3": {
1869
- "accuracy": 0.06910569105691057,
1870
- "precision": 0.5512321660181583,
1871
- "recall": 0.06910569105691057,
1872
- "f1": 0.12281462216442711,
1873
  "gt_support": 12300,
1874
- "predicted": 1542
1875
  },
1876
  "4": {
1877
  "accuracy": 0.0,
@@ -1879,46 +1893,46 @@
1879
  "recall": 0.0,
1880
  "f1": 0.0,
1881
  "gt_support": 0,
1882
- "predicted": 510
1883
  }
1884
  },
1885
  "total_samples": 22500
1886
  },
1887
  "15_0.8": {
1888
- "accuracy": 0.04604336043360434,
1889
  "baseline_accuracy": 0.5066666666666667,
1890
  "per_class": {
1891
  "0": {
1892
- "accuracy": 0.725609756097561,
1893
- "precision": 0.0045152722443559095,
1894
- "recall": 0.725609756097561,
1895
- "f1": 0.00897469738677929,
1896
  "gt_support": 164,
1897
- "predicted": 26355
1898
  },
1899
  "1": {
1900
- "accuracy": 0.04236200256739409,
1901
- "precision": 0.5047801147227533,
1902
- "recall": 0.04236200256739409,
1903
- "f1": 0.07816432272390822,
1904
  "gt_support": 18696,
1905
- "predicted": 1569
1906
  },
1907
  "2": {
1908
- "accuracy": 0.18292682926829268,
1909
- "precision": 0.0044014084507042256,
1910
- "recall": 0.18292682926829268,
1911
- "f1": 0.008595988538681949,
1912
  "gt_support": 164,
1913
- "predicted": 6816
1914
  },
1915
  "3": {
1916
- "accuracy": 0.04240322219735959,
1917
- "precision": 0.48620910840282233,
1918
- "recall": 0.04240322219735959,
1919
- "f1": 0.07800360174942114,
1920
  "gt_support": 17876,
1921
- "predicted": 1559
1922
  },
1923
  "4": {
1924
  "accuracy": 0.0,
@@ -1926,46 +1940,46 @@
1926
  "recall": 0.0,
1927
  "f1": 0.0,
1928
  "gt_support": 0,
1929
- "predicted": 601
1930
  }
1931
  },
1932
  "total_samples": 36900
1933
  },
1934
  "15_1.0": {
1935
- "accuracy": 0.0733162830349531,
1936
  "baseline_accuracy": 0.5688888888888889,
1937
  "per_class": {
1938
  "0": {
1939
- "accuracy": 0.670076726342711,
1940
- "precision": 0.004532479889282934,
1941
- "recall": 0.670076726342711,
1942
- "f1": 0.009004055261530001,
1943
  "gt_support": 391,
1944
- "predicted": 57805
1945
  },
1946
  "1": {
1947
- "accuracy": 0.07474824168797954,
1948
- "precision": 0.5684546421516486,
1949
- "recall": 0.07474824168797954,
1950
- "f1": 0.13212311713079872,
1951
  "gt_support": 50048,
1952
- "predicted": 6581
1953
  },
1954
  "2": {
1955
- "accuracy": 0.17391304347826086,
1956
- "precision": 0.004266265135830353,
1957
- "recall": 0.17391304347826086,
1958
- "f1": 0.008328230251071647,
1959
  "gt_support": 391,
1960
- "predicted": 15939
1961
  },
1962
  "3": {
1963
- "accuracy": 0.06404630502086418,
1964
- "precision": 0.42588614393125673,
1965
- "recall": 0.06404630502086418,
1966
- "f1": 0.11134773349558869,
1967
  "gt_support": 37145,
1968
- "predicted": 5586
1969
  },
1970
  "4": {
1971
  "accuracy": 0.0,
@@ -1973,23 +1987,10 @@
1973
  "recall": 0.0,
1974
  "f1": 0.0,
1975
  "gt_support": 0,
1976
- "predicted": 2064
1977
  }
1978
  },
1979
  "total_samples": 87975
1980
  }
1981
- },
1982
- "total_trajectories": 300,
1983
- "total_steps": 3223,
1984
- "config": {
1985
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_general.pt",
1986
- "trajectories_dir": "reveng/trajectories_test_full",
1987
- "activations_dir": "interp/activations_test_full",
1988
- "layers": "23",
1989
- "steps": "all",
1990
- "token_categories": {
1991
- "output": "all"
1992
- },
1993
- "pad_to_size": 15
1994
  }
1995
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.5710856000275796,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.1721998138380391,
8
+ "precision": 0.012563667232597622,
9
+ "recall": 0.1721998138380391,
10
+ "f1": 0.02341870965019621,
11
  "gt_support": 3223,
12
+ "predicted": 44175
13
  },
14
  "1": {
15
+ "accuracy": 0.6250194184948042,
16
+ "precision": 0.48428223337703447,
17
+ "recall": 0.6250194184948042,
18
+ "f1": 0.5457231572348717,
19
  "gt_support": 238175,
20
+ "predicted": 307391
21
  },
22
  "2": {
23
+ "accuracy": 0.1017685386286069,
24
+ "precision": 0.012620238553289727,
25
+ "recall": 0.1017685386286069,
26
+ "f1": 0.022455755999041524,
27
  "gt_support": 3223,
28
+ "predicted": 25990
29
  },
30
  "3": {
31
+ "accuracy": 0.2945641185801287,
32
+ "precision": 0.5986737156840637,
33
+ "recall": 0.2945641185801287,
34
+ "f1": 0.39485070742180073,
35
  "gt_support": 243346,
36
+ "predicted": 119733
37
  },
38
  "4": {
39
+ "accuracy": 0.8124051465380594,
40
+ "precision": 0.8456377311462749,
41
+ "recall": 0.8124051465380594,
42
+ "f1": 0.8286883941740809,
43
  "gt_support": 237208,
44
+ "predicted": 227886
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
+ "by_complexity": {
50
+ "0.0": {
51
+ "accuracy": 0.6723870967741935,
52
+ "baseline_accuracy": 0.4410322580645161,
53
  "per_class": {
54
  "0": {
55
+ "accuracy": 0.13225806451612904,
56
+ "precision": 0.012713178294573644,
57
+ "recall": 0.13225806451612904,
58
+ "f1": 0.0231966053748232,
59
+ "gt_support": 310,
60
+ "predicted": 3225
61
  },
62
  "1": {
63
+ "accuracy": 0.011020531400966184,
64
+ "precision": 0.22741433021806853,
65
+ "recall": 0.011020531400966184,
66
+ "f1": 0.021022318214542837,
67
+ "gt_support": 13248,
68
+ "predicted": 642
69
  },
70
  "2": {
71
+ "accuracy": 0.13225806451612904,
72
+ "precision": 0.013509060955518945,
73
+ "recall": 0.13225806451612904,
74
+ "f1": 0.024514200298953657,
75
+ "gt_support": 310,
76
+ "predicted": 3035
77
  },
78
  "3": {
79
+ "accuracy": 0.8410376438463039,
80
+ "precision": 0.6683716965046889,
81
+ "recall": 0.8410376438463039,
82
+ "f1": 0.7448287774754934,
83
+ "gt_support": 30762,
84
+ "predicted": 38709
85
  },
86
  "4": {
87
+ "accuracy": 0.827985668789809,
88
+ "precision": 0.861634699034757,
89
+ "recall": 0.827985668789809,
90
+ "f1": 0.8444751212976308,
91
+ "gt_support": 25120,
92
+ "predicted": 24139
93
  }
94
  },
95
+ "total_samples": 69750
96
  },
97
+ "0.2": {
98
+ "accuracy": 0.5927247863247863,
99
+ "baseline_accuracy": 0.3972239316239316,
100
  "per_class": {
101
  "0": {
102
+ "accuracy": 0.24307692307692308,
103
+ "precision": 0.01158357771260997,
104
+ "recall": 0.24307692307692308,
105
+ "f1": 0.022113365990202936,
106
+ "gt_support": 325,
107
+ "predicted": 6820
108
  },
109
  "1": {
110
+ "accuracy": 0.08016140688990604,
111
+ "precision": 0.3042981252857796,
112
+ "recall": 0.08016140688990604,
113
+ "f1": 0.12689484221565447,
114
+ "gt_support": 16604,
115
+ "predicted": 4374
116
  },
117
  "2": {
118
+ "accuracy": 0.14153846153846153,
119
+ "precision": 0.011121856866537718,
120
+ "recall": 0.14153846153846153,
121
+ "f1": 0.020623178659493386,
122
+ "gt_support": 325,
123
+ "predicted": 4136
124
  },
125
  "3": {
126
+ "accuracy": 0.671050366647158,
127
+ "precision": 0.6168549637646761,
128
+ "recall": 0.671050366647158,
129
+ "f1": 0.642812386637206,
130
+ "gt_support": 29047,
131
+ "predicted": 31599
132
  },
133
  "4": {
134
+ "accuracy": 0.8348866686549359,
135
+ "precision": 0.8549015116811727,
136
+ "recall": 0.8348866686549359,
137
+ "f1": 0.8447755563938136,
138
+ "gt_support": 26824,
139
+ "predicted": 26196
140
  }
141
  },
142
+ "total_samples": 73125
143
  },
144
+ "0.4": {
145
+ "accuracy": 0.5446331236897275,
146
+ "baseline_accuracy": 0.4000524109014675,
147
  "per_class": {
148
  "0": {
149
+ "accuracy": 0.1957547169811321,
150
+ "precision": 0.011751380433243664,
151
+ "recall": 0.1957547169811321,
152
+ "f1": 0.022171764391612125,
153
+ "gt_support": 424,
154
+ "predicted": 7063
155
  },
156
  "1": {
157
+ "accuracy": 0.25601611759877624,
158
+ "precision": 0.361881658052948,
159
+ "recall": 0.25601611759877624,
160
+ "f1": 0.2998798208237736,
161
+ "gt_support": 26803,
162
+ "predicted": 18962
163
  },
164
  "2": {
165
+ "accuracy": 0.09905660377358491,
166
+ "precision": 0.012180974477958236,
167
+ "recall": 0.09905660377358491,
168
+ "f1": 0.021694214876033055,
169
+ "gt_support": 424,
170
+ "predicted": 3448
171
  },
172
  "3": {
173
+ "accuracy": 0.540783440324905,
174
+ "precision": 0.5618806490253729,
175
+ "recall": 0.540783440324905,
176
+ "f1": 0.5511302188338651,
177
+ "gt_support": 38165,
178
+ "predicted": 36732
179
  },
180
  "4": {
181
+ "accuracy": 0.8224716062736614,
182
+ "precision": 0.8334303819147114,
183
+ "recall": 0.8224716062736614,
184
+ "f1": 0.8279147314517089,
185
+ "gt_support": 29584,
186
+ "predicted": 29195
187
  }
188
  },
189
+ "total_samples": 95400
190
  },
191
+ "0.6": {
192
+ "accuracy": 0.5108641975308642,
193
+ "baseline_accuracy": 0.3553909465020576,
194
  "per_class": {
195
  "0": {
196
+ "accuracy": 0.25462962962962965,
197
+ "precision": 0.011439267886855242,
198
+ "recall": 0.25462962962962965,
199
+ "f1": 0.021894904458598728,
200
+ "gt_support": 432,
201
+ "predicted": 9616
202
  },
203
  "1": {
204
+ "accuracy": 0.5597892176732874,
205
+ "precision": 0.41580730146782086,
206
+ "recall": 0.5597892176732874,
207
+ "f1": 0.47717358441670626,
208
+ "gt_support": 29604,
209
+ "predicted": 39855
210
  },
211
  "2": {
212
+ "accuracy": 0.10648148148148148,
213
+ "precision": 0.01209889531825355,
214
+ "recall": 0.10648148148148148,
215
+ "f1": 0.021728861596598963,
216
+ "gt_support": 432,
217
+ "predicted": 3802
218
  },
219
  "3": {
220
+ "accuracy": 0.14844041257611532,
221
+ "precision": 0.46365841824357107,
222
+ "recall": 0.14844041257611532,
223
+ "f1": 0.22488409855740943,
224
+ "gt_support": 32188,
225
+ "predicted": 10305
226
  },
227
  "4": {
228
+ "accuracy": 0.8149027327466419,
229
+ "precision": 0.8372494200226043,
230
+ "recall": 0.8149027327466419,
231
+ "f1": 0.825924947921251,
232
+ "gt_support": 34544,
233
+ "predicted": 33622
234
  }
235
  },
236
+ "total_samples": 97200
237
  },
238
+ "0.8": {
239
+ "accuracy": 0.5538380138380138,
240
+ "baseline_accuracy": 0.3470899470899471,
241
  "per_class": {
242
  "0": {
243
+ "accuracy": 0.2032967032967033,
244
+ "precision": 0.012821993762273306,
245
+ "recall": 0.2032967032967033,
246
+ "f1": 0.02412256872758883,
247
+ "gt_support": 546,
248
+ "predicted": 8657
249
  },
250
  "1": {
251
+ "accuracy": 0.738015947467167,
252
+ "precision": 0.4737380884278983,
253
+ "recall": 0.738015947467167,
254
+ "f1": 0.577058138575371,
255
+ "gt_support": 42640,
256
+ "predicted": 66427
257
  },
258
  "2": {
259
+ "accuracy": 0.10622710622710622,
260
+ "precision": 0.015072765072765074,
261
+ "recall": 0.10622710622710622,
262
+ "f1": 0.02639963586709149,
263
+ "gt_support": 546,
264
+ "predicted": 3848
265
  },
266
  "3": {
267
+ "accuracy": 0.020426480987869228,
268
+ "precision": 0.3849308047155305,
269
+ "recall": 0.020426480987869228,
270
+ "f1": 0.03879432807293954,
271
+ "gt_support": 36766,
272
+ "predicted": 1951
273
  },
274
  "4": {
275
+ "accuracy": 0.8417548167737061,
276
+ "precision": 0.8494769700002383,
277
+ "recall": 0.8417548167737061,
278
+ "f1": 0.8455982637365244,
279
+ "gt_support": 42352,
280
+ "predicted": 41967
281
  }
282
  },
283
+ "total_samples": 122850
284
+ },
285
+ "1.0": {
286
+ "accuracy": 0.5780101180438448,
287
+ "baseline_accuracy": 0.4095034663668728,
 
 
288
  "per_class": {
289
  "0": {
290
+ "accuracy": 0.11045531197301856,
291
+ "precision": 0.014896520354787356,
292
+ "recall": 0.11045531197301856,
293
+ "f1": 0.02625250501002004,
294
+ "gt_support": 1186,
295
+ "predicted": 8794
296
  },
297
  "1": {
298
+ "accuracy": 0.8463340532230316,
299
+ "precision": 0.5221220452659331,
300
+ "recall": 0.8463340532230316,
301
+ "f1": 0.6458222040662414,
302
+ "gt_support": 109276,
303
+ "predicted": 177131
304
  },
305
  "2": {
306
+ "accuracy": 0.08010118043844856,
307
+ "precision": 0.012304105685791995,
308
+ "recall": 0.08010118043844856,
309
+ "f1": 0.021331536993375994,
310
+ "gt_support": 1186,
311
+ "predicted": 7721
312
  },
313
  "3": {
314
+ "accuracy": 0.0019498024025753095,
315
+ "precision": 0.34096109839816935,
316
+ "recall": 0.0019498024025753095,
317
+ "f1": 0.003877431526901308,
318
+ "gt_support": 76418,
319
+ "predicted": 437
320
  },
321
  "4": {
322
+ "accuracy": 0.7791302802599512,
323
+ "precision": 0.8435554578311597,
324
+ "recall": 0.7791302802599512,
325
+ "f1": 0.8100639388720627,
326
+ "gt_support": 78784,
327
+ "predicted": 72767
328
  }
329
  },
330
+ "total_samples": 266850
331
+ }
332
+ },
333
+ "total_trajectories": 300,
334
+ "total_steps": 3223,
335
+ "single_size_mode": false,
336
+ "config": {
337
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_general.pt",
338
+ "trajectories_dir": "reveng/trajectories_test_full",
339
+ "activations_dir": "interp/activations_test_full",
340
+ "layers": "23",
341
+ "steps": "all",
342
+ "token_categories": {
343
+ "prompt_suffix": "all"
344
  },
345
+ "pad_to_size": 15
346
+ },
347
+ "by_size": {
348
+ "7": {
349
+ "accuracy": 0.7522144522144523,
350
+ "baseline_accuracy": 0.7822222222222223,
351
  "per_class": {
352
  "0": {
353
+ "accuracy": 0.6818181818181818,
354
+ "precision": 0.020224019912881145,
355
+ "recall": 0.6818181818181818,
356
+ "f1": 0.039282836422240126,
357
+ "gt_support": 286,
358
+ "predicted": 9642
359
  },
360
  "1": {
361
+ "accuracy": 0.0,
362
+ "precision": 0.0,
363
+ "recall": 0.0,
364
+ "f1": 0.0,
365
+ "gt_support": 8224,
366
+ "predicted": 0
367
  },
368
  "2": {
369
+ "accuracy": 0.28321678321678323,
370
+ "precision": 0.02050113895216401,
371
+ "recall": 0.28321678321678323,
372
+ "f1": 0.038234599952796794,
373
+ "gt_support": 286,
374
+ "predicted": 3951
375
  },
376
  "3": {
377
+ "accuracy": 0.0,
378
+ "precision": 0.0,
379
+ "recall": 0.0,
380
+ "f1": 0.0,
381
+ "gt_support": 5218,
382
+ "predicted": 0
383
  },
384
  "4": {
385
+ "accuracy": 0.9561546408137317,
386
+ "precision": 0.9482238903008452,
387
+ "recall": 0.9561546408137317,
388
+ "f1": 0.9521727518225791,
389
+ "gt_support": 50336,
390
+ "predicted": 50757
391
  }
392
  },
393
+ "total_samples": 64350
394
  },
395
+ "9": {
396
+ "accuracy": 0.6163164721141375,
397
+ "baseline_accuracy": 0.64,
398
  "per_class": {
399
  "0": {
400
+ "accuracy": 0.377431906614786,
401
+ "precision": 0.013621682347984833,
402
+ "recall": 0.377431906614786,
403
+ "f1": 0.026294388723231227,
404
+ "gt_support": 514,
405
+ "predicted": 14242
406
  },
407
  "1": {
408
+ "accuracy": 0.006837229819144243,
409
+ "precision": 0.4754601226993865,
410
+ "recall": 0.006837229819144243,
411
+ "f1": 0.01348060532266481,
412
+ "gt_support": 22670,
413
+ "predicted": 326
414
  },
415
  "2": {
416
+ "accuracy": 0.4455252918287938,
417
+ "precision": 0.01214660796690182,
418
+ "recall": 0.4455252918287938,
419
+ "f1": 0.023648474208705528,
420
+ "gt_support": 514,
421
+ "predicted": 18853
422
  },
423
  "3": {
424
+ "accuracy": 0.12578055307760927,
425
+ "precision": 0.44744149147163825,
426
+ "recall": 0.12578055307760927,
427
+ "f1": 0.19636173731395246,
428
+ "gt_support": 17936,
429
+ "predicted": 5042
430
  },
431
  "4": {
432
+ "accuracy": 0.9247054690877649,
433
+ "precision": 0.8867166750877739,
434
+ "recall": 0.9247054690877649,
435
+ "f1": 0.9053127252766149,
436
+ "gt_support": 74016,
437
+ "predicted": 77187
438
  }
439
  },
440
+ "total_samples": 115650
441
  },
442
+ "11": {
443
+ "accuracy": 0.5468652627974662,
444
+ "baseline_accuracy": 0.4622222222222222,
445
  "per_class": {
446
  "0": {
447
+ "accuracy": 0.2526964560862866,
448
+ "precision": 0.008220139341386396,
449
+ "recall": 0.2526964560862866,
450
+ "f1": 0.015922330097087378,
451
+ "gt_support": 649,
452
+ "predicted": 19951
453
  },
454
  "1": {
455
+ "accuracy": 0.48147585018498806,
456
+ "precision": 0.44593611378411996,
457
+ "recall": 0.48147585018498806,
458
+ "f1": 0.4630250155357077,
459
+ "gt_support": 39462,
460
+ "predicted": 42607
461
  },
462
  "2": {
463
+ "accuracy": 0.026194144838212634,
464
+ "precision": 0.00555011426705844,
465
+ "recall": 0.026194144838212634,
466
+ "f1": 0.00915948275862069,
467
+ "gt_support": 649,
468
+ "predicted": 3063
469
  },
470
  "3": {
471
+ "accuracy": 0.20866318938812253,
472
+ "precision": 0.5225087847245243,
473
+ "recall": 0.20866318938812253,
474
+ "f1": 0.2982290168773178,
475
+ "gt_support": 37769,
476
+ "predicted": 15083
477
  },
478
  "4": {
479
+ "accuracy": 0.7821796847220576,
480
+ "precision": 0.8082240014696652,
481
+ "recall": 0.7821796847220576,
482
+ "f1": 0.7949885933276614,
483
+ "gt_support": 67496,
484
+ "predicted": 65321
485
  }
486
  },
487
+ "total_samples": 146025
488
  },
489
+ "13": {
490
+ "accuracy": 0.5145075445816186,
491
+ "baseline_accuracy": 0.37721810699588476,
492
  "per_class": {
493
  "0": {
494
+ "accuracy": 0.0024691358024691358,
495
+ "precision": 0.0058823529411764705,
496
+ "recall": 0.0024691358024691358,
497
+ "f1": 0.0034782608695652175,
498
+ "gt_support": 810,
499
+ "predicted": 340
500
  },
501
  "1": {
502
+ "accuracy": 0.6790535461952437,
503
+ "precision": 0.4468360816278081,
504
+ "recall": 0.6790535461952437,
505
+ "f1": 0.5389971064642186,
506
+ "gt_support": 66522,
507
+ "predicted": 101093
508
  },
509
  "2": {
510
+ "accuracy": 0.0012345679012345679,
511
+ "precision": 0.008130081300813009,
512
+ "recall": 0.0012345679012345679,
513
+ "f1": 0.0021436227224008574,
514
+ "gt_support": 810,
515
+ "predicted": 123
516
  },
517
  "3": {
518
+ "accuracy": 0.36729795775877117,
519
+ "precision": 0.548065027239381,
520
+ "recall": 0.36729795775877117,
521
+ "f1": 0.43983243483334933,
522
+ "gt_support": 68748,
523
+ "predicted": 46073
524
  },
525
  "4": {
526
+ "accuracy": 0.5146164021164021,
527
+ "precision": 0.6742439559804743,
528
+ "recall": 0.5146164021164021,
529
+ "f1": 0.5837136319875971,
530
+ "gt_support": 45360,
531
+ "predicted": 34621
532
  }
533
  },
534
+ "total_samples": 182250
535
  },
536
+ "15": {
537
+ "accuracy": 0.5570769940064546,
538
+ "baseline_accuracy": 0.5240894421392347,
539
  "per_class": {
540
  "0": {
541
+ "accuracy": 0.0,
542
+ "precision": 0.0,
543
+ "recall": 0.0,
544
+ "f1": 0.0,
545
+ "gt_support": 964,
546
+ "predicted": 0
547
  },
548
  "1": {
549
+ "accuracy": 0.8345459391689783,
550
+ "precision": 0.5174731429620788,
551
+ "recall": 0.8345459391689783,
552
+ "f1": 0.6388299038018302,
553
+ "gt_support": 101297,
554
+ "predicted": 163365
555
  },
556
  "2": {
557
+ "accuracy": 0.0,
558
+ "precision": 0.0,
559
+ "recall": 0.0,
560
+ "f1": 0.0,
561
+ "gt_support": 964,
562
+ "predicted": 0
563
  },
564
  "3": {
565
+ "accuracy": 0.31926984825159443,
566
+ "precision": 0.6779303259549827,
567
+ "recall": 0.31926984825159443,
568
+ "f1": 0.4341008312899946,
569
+ "gt_support": 113675,
570
+ "predicted": 53535
571
  },
572
  "4": {
573
+ "accuracy": 0.0,
574
+ "precision": 0.0,
575
+ "recall": 0.0,
576
+ "f1": 0.0,
577
+ "gt_support": 0,
578
+ "predicted": 0
579
  }
580
  },
581
+ "total_samples": 216900
582
  }
583
  },
584
  "by_size_complexity": {
585
  "7_0.0": {
586
+ "accuracy": 0.7471345029239767,
587
  "baseline_accuracy": 0.7822222222222223,
588
  "per_class": {
589
  "0": {
590
+ "accuracy": 0.6052631578947368,
591
+ "precision": 0.02060931899641577,
592
+ "recall": 0.6052631578947368,
593
+ "f1": 0.03986135181975737,
594
  "gt_support": 38,
595
+ "predicted": 1116
596
  },
597
  "1": {
598
+ "accuracy": 0.0,
599
+ "precision": 0.0,
600
+ "recall": 0.0,
601
+ "f1": 0.0,
602
  "gt_support": 912,
603
+ "predicted": 0
604
  },
605
  "2": {
606
+ "accuracy": 0.39473684210526316,
607
+ "precision": 0.01953125,
608
+ "recall": 0.39473684210526316,
609
+ "f1": 0.03722084367245658,
610
  "gt_support": 38,
611
+ "predicted": 768
612
  },
613
  "3": {
614
  "accuracy": 0.0,
 
619
  "predicted": 0
620
  },
621
  "4": {
622
+ "accuracy": 0.9494617224880383,
623
+ "precision": 0.9525952595259526,
624
+ "recall": 0.9494617224880383,
625
+ "f1": 0.9510259098397483,
626
  "gt_support": 6688,
627
+ "predicted": 6666
628
  }
629
  },
630
  "total_samples": 8550
631
  },
632
  "7_0.2": {
633
+ "accuracy": 0.7506060606060606,
634
  "baseline_accuracy": 0.7822222222222223,
635
  "per_class": {
636
  "0": {
637
+ "accuracy": 0.7045454545454546,
638
+ "precision": 0.020889487870619946,
639
+ "recall": 0.7045454545454546,
640
+ "f1": 0.04057591623036649,
641
  "gt_support": 44,
642
+ "predicted": 1484
643
  },
644
  "1": {
645
+ "accuracy": 0.0,
646
+ "precision": 0.0,
647
+ "recall": 0.0,
648
+ "f1": 0.0,
649
  "gt_support": 1144,
650
+ "predicted": 0
651
  },
652
  "2": {
653
+ "accuracy": 0.2727272727272727,
654
+ "precision": 0.018957345971563982,
655
+ "recall": 0.2727272727272727,
656
+ "f1": 0.03545051698670606,
657
  "gt_support": 44,
658
+ "predicted": 633
659
  },
660
  "3": {
661
  "accuracy": 0.0,
 
666
  "predicted": 0
667
  },
668
  "4": {
669
+ "accuracy": 0.9540289256198347,
670
+ "precision": 0.9492483618142105,
671
+ "recall": 0.9540289256198347,
672
+ "f1": 0.951632639917563,
673
  "gt_support": 7744,
674
+ "predicted": 7783
675
  }
676
  },
677
  "total_samples": 9900
678
  },
679
  "7_0.4": {
680
+ "accuracy": 0.7516531165311653,
681
  "baseline_accuracy": 0.7822222222222223,
682
  "per_class": {
683
  "0": {
684
+ "accuracy": 0.6585365853658537,
685
+ "precision": 0.01998519615099926,
686
+ "recall": 0.6585365853658537,
687
+ "f1": 0.038793103448275856,
688
  "gt_support": 41,
689
+ "predicted": 1351
690
  },
691
  "1": {
 
 
 
 
 
 
 
 
692
  "accuracy": 0.0,
693
  "precision": 0.0,
694
  "recall": 0.0,
695
  "f1": 0.0,
696
+ "gt_support": 1148,
697
  "predicted": 0
698
  },
699
+ "2": {
700
+ "accuracy": 0.2926829268292683,
701
+ "precision": 0.020134228187919462,
702
+ "recall": 0.2926829268292683,
703
+ "f1": 0.03767660910518053,
704
+ "gt_support": 41,
705
+ "predicted": 596
706
+ },
707
  "3": {
708
  "accuracy": 0.0,
709
  "precision": 0.0,
 
713
  "predicted": 0
714
  },
715
  "4": {
716
+ "accuracy": 0.9555155210643016,
717
+ "precision": 0.9473756526518274,
718
+ "recall": 0.9555155210643016,
719
+ "f1": 0.9514281771767629,
720
  "gt_support": 7216,
721
+ "predicted": 7278
722
  }
723
  },
724
  "total_samples": 9225
725
  },
726
  "7_0.6": {
727
+ "accuracy": 0.7512056737588653,
728
  "baseline_accuracy": 0.7822222222222223,
729
  "per_class": {
730
  "0": {
731
+ "accuracy": 0.6808510638297872,
732
+ "precision": 0.020138451856513532,
733
+ "recall": 0.6808510638297872,
734
+ "f1": 0.039119804400978,
735
  "gt_support": 47,
736
+ "predicted": 1589
737
  },
738
  "1": {
 
 
 
 
 
 
 
 
739
  "accuracy": 0.0,
740
  "precision": 0.0,
741
  "recall": 0.0,
742
  "f1": 0.0,
743
+ "gt_support": 1363,
744
  "predicted": 0
745
  },
746
+ "2": {
747
+ "accuracy": 0.2978723404255319,
748
+ "precision": 0.021116138763197588,
749
+ "recall": 0.2978723404255319,
750
+ "f1": 0.03943661971830986,
751
+ "gt_support": 47,
752
+ "predicted": 663
753
+ },
754
  "3": {
755
+ "accuracy": 0.0,
756
+ "precision": 0.0,
757
+ "recall": 0.0,
758
+ "f1": 0.0,
759
  "gt_support": 846,
760
+ "predicted": 0
761
  },
762
  "4": {
763
+ "accuracy": 0.9547872340425532,
764
+ "precision": 0.9489366814850414,
765
+ "recall": 0.9547872340425532,
766
+ "f1": 0.951852967761374,
767
  "gt_support": 8272,
768
+ "predicted": 8323
769
  }
770
  },
771
  "total_samples": 10575
772
  },
773
  "7_0.8": {
774
+ "accuracy": 0.7544242424242424,
775
  "baseline_accuracy": 0.7822222222222223,
776
  "per_class": {
777
  "0": {
778
+ "accuracy": 0.6181818181818182,
779
+ "precision": 0.019871420222092345,
780
+ "recall": 0.6181818181818182,
781
+ "f1": 0.03850509626274066,
782
  "gt_support": 55,
783
+ "predicted": 1711
784
  },
785
  "1": {
 
 
 
 
 
 
 
 
786
  "accuracy": 0.0,
787
  "precision": 0.0,
788
  "recall": 0.0,
789
  "f1": 0.0,
790
+ "gt_support": 1705,
791
  "predicted": 0
792
  },
793
+ "2": {
794
+ "accuracy": 0.32727272727272727,
795
+ "precision": 0.020761245674740483,
796
+ "recall": 0.32727272727272727,
797
+ "f1": 0.039045553145336226,
798
+ "gt_support": 55,
799
+ "predicted": 867
800
+ },
801
  "3": {
802
  "accuracy": 0.0,
803
  "precision": 0.0,
 
807
  "predicted": 0
808
  },
809
  "4": {
810
+ "accuracy": 0.9590909090909091,
811
+ "precision": 0.9476370317444116,
812
+ "recall": 0.9590909090909091,
813
+ "f1": 0.9533295682086563,
814
  "gt_support": 9680,
815
+ "predicted": 9797
816
  }
817
  },
818
  "total_samples": 12375
819
  },
820
  "7_1.0": {
821
+ "accuracy": 0.7557012750455373,
822
  "baseline_accuracy": 0.7822222222222223,
823
  "per_class": {
824
  "0": {
825
+ "accuracy": 0.7868852459016393,
826
+ "precision": 0.020075282308657464,
827
+ "recall": 0.7868852459016393,
828
+ "f1": 0.03915171288743882,
829
  "gt_support": 61,
830
+ "predicted": 2391
831
  },
832
  "1": {
 
 
 
 
 
 
 
 
833
  "accuracy": 0.0,
834
  "precision": 0.0,
835
  "recall": 0.0,
836
  "f1": 0.0,
837
+ "gt_support": 1952,
838
  "predicted": 0
839
  },
840
+ "2": {
841
+ "accuracy": 0.16393442622950818,
842
+ "precision": 0.02358490566037736,
843
+ "recall": 0.16393442622950818,
844
+ "f1": 0.04123711340206186,
845
+ "gt_support": 61,
846
+ "predicted": 424
847
+ },
848
  "3": {
849
  "accuracy": 0.0,
850
  "precision": 0.0,
 
854
  "predicted": 0
855
  },
856
  "4": {
857
+ "accuracy": 0.9606929955290611,
858
+ "precision": 0.9453712190650779,
859
+ "recall": 0.9606929955290611,
860
+ "f1": 0.9529705257322368,
861
  "gt_support": 10736,
862
+ "predicted": 10910
863
  }
864
  },
865
  "total_samples": 13725
866
  },
867
  "9_0.0": {
868
+ "accuracy": 0.6484773662551441,
869
  "baseline_accuracy": 0.64,
870
  "per_class": {
871
  "0": {
872
+ "accuracy": 0.16666666666666666,
873
+ "precision": 0.01079136690647482,
874
+ "recall": 0.16666666666666666,
875
+ "f1": 0.020270270270270275,
876
  "gt_support": 54,
877
+ "predicted": 834
878
  },
879
  "1": {
880
+ "accuracy": 0.0,
881
+ "precision": 0.0,
882
+ "recall": 0.0,
883
+ "f1": 0.0,
884
  "gt_support": 1728,
885
+ "predicted": 0
886
  },
887
  "2": {
888
+ "accuracy": 0.4444444444444444,
889
+ "precision": 0.012282497441146366,
890
+ "recall": 0.4444444444444444,
891
+ "f1": 0.02390438247011952,
892
  "gt_support": 54,
893
+ "predicted": 1954
894
  },
895
  "3": {
896
+ "accuracy": 0.31757289204097716,
897
+ "precision": 0.5230369889682025,
898
+ "recall": 0.31757289204097716,
899
+ "f1": 0.3951949007109586,
900
  "gt_support": 2538,
901
+ "predicted": 1541
902
  },
903
  "4": {
904
+ "accuracy": 0.9053497942386831,
905
+ "precision": 0.90014064697609,
906
+ "recall": 0.9053497942386831,
907
+ "f1": 0.9027377059690966,
908
  "gt_support": 7776,
909
+ "predicted": 7821
910
  }
911
  },
912
  "total_samples": 12150
913
  },
914
  "9_0.2": {
915
+ "accuracy": 0.6367848699763593,
916
  "baseline_accuracy": 0.64,
917
  "per_class": {
918
  "0": {
919
+ "accuracy": 0.23404255319148937,
920
+ "precision": 0.012925969447708578,
921
+ "recall": 0.23404255319148937,
922
+ "f1": 0.024498886414253896,
923
  "gt_support": 47,
924
+ "predicted": 851
925
  },
926
  "1": {
 
 
 
 
 
 
 
 
927
  "accuracy": 0.0,
928
  "precision": 0.0,
929
  "recall": 0.0,
930
  "f1": 0.0,
931
+ "gt_support": 1692,
932
  "predicted": 0
933
  },
934
+ "2": {
935
+ "accuracy": 0.46808510638297873,
936
+ "precision": 0.012650948821161587,
937
+ "recall": 0.46808510638297873,
938
+ "f1": 0.024636058230683093,
939
+ "gt_support": 47,
940
+ "predicted": 1739
941
+ },
942
  "3": {
943
+ "accuracy": 0.2528451261751608,
944
+ "precision": 0.47846441947565543,
945
+ "recall": 0.2528451261751608,
946
+ "f1": 0.33085140822272574,
947
  "gt_support": 2021,
948
+ "predicted": 1068
949
  },
950
  "4": {
951
+ "accuracy": 0.914598108747045,
952
+ "precision": 0.8948966314876392,
953
+ "recall": 0.914598108747045,
954
+ "f1": 0.9046401169163317,
955
  "gt_support": 6768,
956
+ "predicted": 6917
957
  }
958
  },
959
  "total_samples": 10575
960
  },
961
  "9_0.4": {
962
+ "accuracy": 0.6100829187396352,
963
  "baseline_accuracy": 0.64,
964
  "per_class": {
965
  "0": {
966
+ "accuracy": 0.3880597014925373,
967
+ "precision": 0.01278898180029513,
968
+ "recall": 0.3880597014925373,
969
+ "f1": 0.02476190476190476,
970
  "gt_support": 67,
971
+ "predicted": 2033
972
  },
973
  "1": {
 
 
 
 
 
 
 
 
974
  "accuracy": 0.0,
975
  "precision": 0.0,
976
  "recall": 0.0,
977
  "f1": 0.0,
978
+ "gt_support": 2680,
979
  "predicted": 0
980
  },
981
+ "2": {
982
+ "accuracy": 0.43283582089552236,
983
+ "precision": 0.011953833470733718,
984
+ "recall": 0.43283582089552236,
985
+ "f1": 0.023265142398716406,
986
+ "gt_support": 67,
987
+ "predicted": 2426
988
+ },
989
  "3": {
990
+ "accuracy": 0.12782242632988902,
991
+ "precision": 0.4400527009222661,
992
+ "recall": 0.12782242632988902,
993
+ "f1": 0.19810201660735466,
994
  "gt_support": 2613,
995
+ "predicted": 759
996
  },
997
  "4": {
998
+ "accuracy": 0.9129353233830846,
999
+ "precision": 0.8935781677995334,
1000
+ "recall": 0.9129353233830846,
1001
+ "f1": 0.9031530376826454,
1002
  "gt_support": 9648,
1003
+ "predicted": 9857
1004
  }
1005
  },
1006
  "total_samples": 15075
1007
  },
1008
  "9_0.6": {
1009
+ "accuracy": 0.5916939890710382,
1010
  "baseline_accuracy": 0.64,
1011
  "per_class": {
1012
  "0": {
1013
+ "accuracy": 0.45901639344262296,
1014
+ "precision": 0.013958125623130608,
1015
+ "recall": 0.45901639344262296,
1016
+ "f1": 0.02709240445089502,
1017
  "gt_support": 61,
1018
+ "predicted": 2006
1019
  },
1020
  "1": {
 
 
 
 
 
 
 
 
1021
  "accuracy": 0.0,
1022
  "precision": 0.0,
1023
  "recall": 0.0,
1024
  "f1": 0.0,
1025
+ "gt_support": 2623,
1026
  "predicted": 0
1027
  },
1028
+ "2": {
1029
+ "accuracy": 0.4918032786885246,
1030
+ "precision": 0.011641443538998836,
1031
+ "recall": 0.4918032786885246,
1032
+ "f1": 0.02274450341167551,
1033
+ "gt_support": 61,
1034
+ "predicted": 2577
1035
+ },
1036
  "3": {
1037
+ "accuracy": 0.04007285974499089,
1038
+ "precision": 0.3963963963963964,
1039
+ "recall": 0.04007285974499089,
1040
+ "f1": 0.0727874276261373,
1041
  "gt_support": 2196,
1042
+ "predicted": 222
1043
  },
1044
  "4": {
1045
+ "accuracy": 0.9079007285974499,
1046
+ "precision": 0.8940582959641256,
1047
+ "recall": 0.9079007285974499,
1048
+ "f1": 0.9009263443289652,
1049
  "gt_support": 8784,
1050
+ "predicted": 8920
1051
  }
1052
  },
1053
  "total_samples": 13725
1054
  },
1055
  "9_0.8": {
1056
+ "accuracy": 0.6227062706270627,
1057
  "baseline_accuracy": 0.64,
1058
  "per_class": {
1059
  "0": {
1060
+ "accuracy": 0.44554455445544555,
1061
+ "precision": 0.014023060143346837,
1062
+ "recall": 0.44554455445544555,
1063
+ "f1": 0.02719033232628399,
1064
  "gt_support": 101,
1065
+ "predicted": 3209
1066
  },
1067
  "1": {
1068
+ "accuracy": 0.007583737097113966,
1069
+ "precision": 0.46153846153846156,
1070
+ "recall": 0.007583737097113966,
1071
+ "f1": 0.014922279792746112,
1072
  "gt_support": 4747,
1073
+ "predicted": 78
1074
  },
1075
  "2": {
1076
+ "accuracy": 0.38613861386138615,
1077
+ "precision": 0.013565217391304348,
1078
+ "recall": 0.38613861386138615,
1079
+ "f1": 0.02620967741935484,
1080
  "gt_support": 101,
1081
+ "predicted": 2875
1082
  },
1083
  "3": {
1084
+ "accuracy": 0.11386138613861387,
1085
+ "precision": 0.3625615763546798,
1086
+ "recall": 0.11386138613861387,
1087
+ "f1": 0.17329879915234284,
1088
  "gt_support": 3232,
1089
+ "predicted": 1015
1090
  },
1091
  "4": {
1092
+ "accuracy": 0.9394251925192519,
1093
+ "precision": 0.8787625418060201,
1094
+ "recall": 0.9394251925192519,
1095
+ "f1": 0.9080818822278346,
1096
  "gt_support": 14544,
1097
+ "predicted": 15548
1098
  }
1099
  },
1100
  "total_samples": 22725
1101
  },
1102
  "9_1.0": {
1103
+ "accuracy": 0.6085748792270531,
1104
  "baseline_accuracy": 0.64,
1105
  "per_class": {
1106
  "0": {
1107
+ "accuracy": 0.4076086956521739,
1108
+ "precision": 0.014126954228668298,
1109
+ "recall": 0.4076086956521739,
1110
+ "f1": 0.027307482250136537,
1111
  "gt_support": 184,
1112
+ "predicted": 5309
1113
  },
1114
  "1": {
1115
+ "accuracy": 0.012934782608695652,
1116
+ "precision": 0.4798387096774194,
1117
+ "recall": 0.012934782608695652,
1118
+ "f1": 0.02519051651143099,
1119
  "gt_support": 9200,
1120
+ "predicted": 248
1121
  },
1122
  "2": {
1123
+ "accuracy": 0.46195652173913043,
1124
+ "precision": 0.011672617412798682,
1125
+ "recall": 0.46195652173913043,
1126
+ "f1": 0.02276989016876507,
1127
  "gt_support": 184,
1128
+ "predicted": 7282
1129
  },
1130
  "3": {
1131
+ "accuracy": 0.02792353823088456,
1132
+ "precision": 0.34096109839816935,
1133
+ "recall": 0.02792353823088456,
1134
+ "f1": 0.05161960852243201,
1135
  "gt_support": 5336,
1136
+ "predicted": 437
1137
  },
1138
  "4": {
1139
+ "accuracy": 0.9347448671497585,
1140
+ "precision": 0.8806357559379889,
1141
+ "recall": 0.9347448671497585,
1142
+ "f1": 0.9068839253020872,
1143
  "gt_support": 26496,
1144
+ "predicted": 28124
1145
  }
1146
  },
1147
  "total_samples": 41400
1148
  },
1149
  "11_0.0": {
1150
+ "accuracy": 0.619281045751634,
1151
  "baseline_accuracy": 0.4622222222222222,
1152
  "per_class": {
1153
  "0": {
1154
+ "accuracy": 0.1323529411764706,
1155
+ "precision": 0.007058823529411765,
1156
+ "recall": 0.1323529411764706,
1157
+ "f1": 0.013402829486224871,
1158
  "gt_support": 68,
1159
+ "predicted": 1275
1160
  },
1161
  "1": {
1162
+ "accuracy": 0.0,
1163
+ "precision": 0.0,
1164
+ "recall": 0.0,
1165
+ "f1": 0.0,
1166
  "gt_support": 2720,
1167
+ "predicted": 0
1168
  },
1169
  "2": {
1170
+ "accuracy": 0.014705882352941176,
1171
+ "precision": 0.0045871559633027525,
1172
+ "recall": 0.014705882352941176,
1173
+ "f1": 0.006993006993006993,
1174
  "gt_support": 68,
1175
+ "predicted": 218
1176
  },
1177
  "3": {
1178
+ "accuracy": 0.7352941176470589,
1179
+ "precision": 0.5635611356826937,
1180
+ "recall": 0.7352941176470589,
1181
+ "f1": 0.6380744689443503,
1182
  "gt_support": 5372,
1183
+ "predicted": 7009
1184
  },
1185
  "4": {
1186
+ "accuracy": 0.7798359728506787,
1187
+ "precision": 0.811268020005884,
1188
+ "recall": 0.7798359728506787,
1189
+ "f1": 0.7952415284787311,
1190
  "gt_support": 7072,
1191
+ "predicted": 6798
1192
  }
1193
  },
1194
  "total_samples": 15300
1195
  },
1196
  "11_0.2": {
1197
+ "accuracy": 0.4792156862745098,
1198
  "baseline_accuracy": 0.4622222222222222,
1199
  "per_class": {
1200
  "0": {
1201
+ "accuracy": 0.43529411764705883,
1202
+ "precision": 0.008249721293199555,
1203
+ "recall": 0.43529411764705883,
1204
+ "f1": 0.016192560175054705,
1205
  "gt_support": 85,
1206
+ "predicted": 4485
1207
  },
1208
  "1": {
1209
+ "accuracy": 0.009261576971214018,
1210
+ "precision": 0.27611940298507465,
1211
+ "recall": 0.009261576971214018,
1212
+ "f1": 0.017922015015742313,
1213
  "gt_support": 3995,
1214
+ "predicted": 134
1215
  },
1216
  "2": {
1217
+ "accuracy": 0.1411764705882353,
1218
+ "precision": 0.0069124423963133645,
1219
+ "recall": 0.1411764705882353,
1220
+ "f1": 0.013179571663920923,
1221
  "gt_support": 85,
1222
+ "predicted": 1736
1223
  },
1224
  "3": {
1225
+ "accuracy": 0.3633986928104575,
1226
+ "precision": 0.5139819736538017,
1227
+ "recall": 0.3633986928104575,
1228
+ "f1": 0.42576816310902654,
1229
  "gt_support": 6120,
1230
+ "predicted": 4327
1231
  },
1232
  "4": {
1233
+ "accuracy": 0.7754524886877828,
1234
+ "precision": 0.8119151960203719,
1235
+ "recall": 0.7754524886877828,
1236
+ "f1": 0.7932650581496269,
1237
  "gt_support": 8840,
1238
+ "predicted": 8443
1239
  }
1240
  },
1241
  "total_samples": 19125
1242
  },
1243
  "11_0.4": {
1244
+ "accuracy": 0.47814141414141414,
1245
  "baseline_accuracy": 0.4622222222222222,
1246
  "per_class": {
1247
  "0": {
1248
+ "accuracy": 0.509090909090909,
1249
+ "precision": 0.008385744234800839,
1250
+ "recall": 0.509090909090909,
1251
+ "f1": 0.016499705362404242,
1252
  "gt_support": 55,
1253
+ "predicted": 3339
1254
  },
1255
  "1": {
1256
+ "accuracy": 0.06415094339622641,
1257
+ "precision": 0.3070607553366174,
1258
+ "recall": 0.06415094339622641,
1259
+ "f1": 0.1061293984108967,
1260
  "gt_support": 2915,
1261
+ "predicted": 609
1262
  },
1263
  "2": {
1264
+ "accuracy": 0.01818181818181818,
1265
+ "precision": 0.002347417840375587,
1266
+ "recall": 0.01818181818181818,
1267
+ "f1": 0.004158004158004158,
1268
  "gt_support": 55,
1269
+ "predicted": 426
1270
  },
1271
  "3": {
1272
+ "accuracy": 0.2820936639118457,
1273
+ "precision": 0.4871550903901047,
1274
+ "recall": 0.2820936639118457,
1275
+ "f1": 0.3572923935799023,
1276
  "gt_support": 3630,
1277
+ "predicted": 2102
1278
  },
1279
  "4": {
1280
+ "accuracy": 0.8176573426573427,
1281
+ "precision": 0.7928462451262925,
1282
+ "recall": 0.8176573426573427,
1283
+ "f1": 0.8050606764781824,
1284
  "gt_support": 5720,
1285
+ "predicted": 5899
1286
  }
1287
  },
1288
  "total_samples": 12375
1289
  },
1290
  "11_0.6": {
1291
+ "accuracy": 0.4725782092772384,
1292
  "baseline_accuracy": 0.4622222222222222,
1293
  "per_class": {
1294
  "0": {
1295
+ "accuracy": 0.4854368932038835,
1296
+ "precision": 0.008304268393954492,
1297
+ "recall": 0.4854368932038835,
1298
+ "f1": 0.016329196603527107,
1299
  "gt_support": 103,
1300
+ "predicted": 6021
1301
  },
1302
  "1": {
1303
+ "accuracy": 0.27621359223300973,
1304
+ "precision": 0.3854143147437345,
1305
+ "recall": 0.27621359223300973,
1306
+ "f1": 0.32180224337826374,
1307
  "gt_support": 6180,
1308
+ "predicted": 4429
1309
  },
1310
  "2": {
1311
+ "accuracy": 0.019417475728155338,
1312
+ "precision": 0.0035587188612099642,
1313
+ "recall": 0.019417475728155338,
1314
+ "f1": 0.006015037593984962,
1315
  "gt_support": 103,
1316
+ "predicted": 562
1317
  },
1318
  "3": {
1319
+ "accuracy": 0.09577093960835939,
1320
+ "precision": 0.41930835734870314,
1321
+ "recall": 0.09577093960835939,
1322
+ "f1": 0.15592766242464837,
1323
  "gt_support": 6077,
1324
+ "predicted": 1388
1325
  },
1326
  "4": {
1327
+ "accuracy": 0.8038648244958925,
1328
+ "precision": 0.7991647331786543,
1329
+ "recall": 0.8038648244958925,
1330
+ "f1": 0.8015078884907154,
1331
  "gt_support": 10712,
1332
+ "predicted": 10775
1333
  }
1334
  },
1335
  "total_samples": 23175
1336
  },
1337
  "11_0.8": {
1338
+ "accuracy": 0.5363352826510721,
1339
  "baseline_accuracy": 0.4622222222222222,
1340
  "per_class": {
1341
  "0": {
1342
+ "accuracy": 0.2807017543859649,
1343
+ "precision": 0.008563018464008563,
1344
+ "recall": 0.2807017543859649,
1345
+ "f1": 0.016619059984419628,
1346
  "gt_support": 114,
1347
+ "predicted": 3737
1348
  },
1349
  "1": {
1350
+ "accuracy": 0.5477139819245083,
1351
+ "precision": 0.4266928970801408,
1352
+ "recall": 0.5477139819245083,
1353
+ "f1": 0.4796880456291468,
1354
  "gt_support": 7524,
1355
+ "predicted": 9658
1356
  },
1357
  "2": {
1358
+ "accuracy": 0.008771929824561403,
1359
+ "precision": 0.009433962264150943,
1360
+ "recall": 0.008771929824561403,
1361
+ "f1": 0.00909090909090909,
1362
  "gt_support": 114,
1363
+ "predicted": 106
1364
  },
1365
  "3": {
1366
+ "accuracy": 0.016716319099635883,
1367
+ "precision": 0.39299610894941633,
1368
+ "recall": 0.016716319099635883,
1369
+ "f1": 0.03206858231465312,
1370
  "gt_support": 6042,
1371
+ "predicted": 257
1372
  },
1373
  "4": {
1374
+ "accuracy": 0.801450742240216,
1375
+ "precision": 0.7990245543222334,
1376
+ "recall": 0.801450742240216,
1377
+ "f1": 0.8002358093313121,
1378
  "gt_support": 11856,
1379
+ "predicted": 11892
1380
  }
1381
  },
1382
  "total_samples": 25650
1383
  },
1384
  "11_1.0": {
1385
+ "accuracy": 0.6069444444444444,
1386
  "baseline_accuracy": 0.4622222222222222,
1387
  "per_class": {
1388
  "0": {
1389
+ "accuracy": 0.03571428571428571,
1390
+ "precision": 0.007312614259597806,
1391
+ "recall": 0.03571428571428571,
1392
+ "f1": 0.012139605462822457,
1393
  "gt_support": 224,
1394
+ "predicted": 1094
1395
  },
1396
  "1": {
1397
+ "accuracy": 0.8028273809523809,
1398
+ "precision": 0.4661410519494546,
1399
+ "recall": 0.8028273809523809,
1400
+ "f1": 0.5898189272292449,
1401
  "gt_support": 16128,
1402
+ "predicted": 27777
1403
  },
1404
  "2": {
1405
+ "accuracy": 0.0,
1406
+ "precision": 0.0,
1407
+ "recall": 0.0,
1408
+ "f1": 0.0,
1409
  "gt_support": 224,
1410
+ "predicted": 15
1411
  },
1412
  "3": {
1413
+ "accuracy": 0.0,
1414
+ "precision": 0.0,
1415
+ "recall": 0.0,
1416
+ "f1": 0.0,
1417
  "gt_support": 10528,
1418
+ "predicted": 0
1419
  },
1420
  "4": {
1421
+ "accuracy": 0.7569539835164835,
1422
+ "precision": 0.8196523194199126,
1423
+ "recall": 0.7569539835164835,
1424
+ "f1": 0.7870564606114707,
1425
  "gt_support": 23296,
1426
+ "predicted": 21514
1427
  }
1428
  },
1429
  "total_samples": 50400
1430
  },
1431
  "13_0.0": {
1432
+ "accuracy": 0.6127777777777778,
1433
  "baseline_accuracy": 0.5288888888888889,
1434
  "per_class": {
1435
  "0": {
1436
+ "accuracy": 0.0,
1437
+ "precision": 0.0,
1438
+ "recall": 0.0,
1439
+ "f1": 0.0,
1440
  "gt_support": 64,
1441
+ "predicted": 0
1442
  },
1443
  "1": {
1444
+ "accuracy": 0.0400390625,
1445
+ "precision": 0.21693121693121692,
1446
+ "recall": 0.0400390625,
1447
+ "f1": 0.06760098928276999,
1448
  "gt_support": 3072,
1449
+ "predicted": 567
1450
  },
1451
  "2": {
1452
+ "accuracy": 0.015625,
1453
+ "precision": 0.010526315789473684,
1454
+ "recall": 0.015625,
1455
+ "f1": 0.012578616352201259,
1456
  "gt_support": 64,
1457
+ "predicted": 95
1458
  },
1459
  "3": {
1460
+ "accuracy": 0.8936449579831933,
1461
+ "precision": 0.6253215729511209,
1462
+ "recall": 0.8936449579831933,
1463
+ "f1": 0.7357837837837838,
1464
  "gt_support": 7616,
1465
+ "predicted": 10884
1466
  },
1467
  "4": {
1468
+ "accuracy": 0.5284598214285714,
1469
+ "precision": 0.6636299929922915,
1470
+ "recall": 0.5284598214285714,
1471
+ "f1": 0.5883814849332091,
1472
  "gt_support": 3584,
1473
+ "predicted": 2854
1474
  }
1475
  },
1476
  "total_samples": 14400
1477
  },
1478
  "13_0.2": {
1479
+ "accuracy": 0.5681720430107526,
1480
  "baseline_accuracy": 0.48444444444444446,
1481
  "per_class": {
1482
  "0": {
1483
+ "accuracy": 0.0,
1484
+ "precision": 0.0,
1485
+ "recall": 0.0,
1486
+ "f1": 0.0,
1487
  "gt_support": 62,
1488
+ "predicted": 0
1489
  },
1490
  "1": {
1491
+ "accuracy": 0.0842602892102336,
1492
+ "precision": 0.27721866422689845,
1493
+ "recall": 0.0842602892102336,
1494
+ "f1": 0.12923864363403711,
1495
  "gt_support": 3596,
1496
+ "predicted": 1093
1497
  },
1498
  "2": {
1499
+ "accuracy": 0.0,
1500
+ "precision": 0.0,
1501
+ "recall": 0.0,
1502
+ "f1": 0.0,
1503
  "gt_support": 62,
1504
+ "predicted": 28
1505
  },
1506
  "3": {
1507
+ "accuracy": 0.8376738680082865,
1508
+ "precision": 0.5790711947626841,
1509
+ "recall": 0.8376738680082865,
1510
+ "f1": 0.6847707753719607,
1511
  "gt_support": 6758,
1512
+ "predicted": 9776
1513
  },
1514
  "4": {
1515
+ "accuracy": 0.5650921658986175,
1516
+ "precision": 0.642646577137242,
1517
+ "recall": 0.5650921658986175,
1518
+ "f1": 0.6013793103448276,
1519
  "gt_support": 3472,
1520
+ "predicted": 3053
1521
  }
1522
  },
1523
  "total_samples": 13950
1524
  },
1525
  "13_0.4": {
1526
+ "accuracy": 0.5088711111111112,
1527
  "baseline_accuracy": 0.44,
1528
  "per_class": {
1529
  "0": {
1530
+ "accuracy": 0.016,
1531
+ "precision": 0.0058823529411764705,
1532
+ "recall": 0.016,
1533
+ "f1": 0.008602150537634409,
1534
  "gt_support": 125,
1535
+ "predicted": 340
1536
  },
1537
  "1": {
1538
+ "accuracy": 0.18764705882352942,
1539
+ "precision": 0.3249134243226726,
1540
+ "recall": 0.18764705882352942,
1541
+ "f1": 0.23789991796554552,
1542
  "gt_support": 8500,
1543
+ "predicted": 4909
1544
  },
1545
  "2": {
1546
+ "accuracy": 0.0,
1547
+ "precision": 0.0,
1548
+ "recall": 0.0,
1549
+ "f1": 0.0,
1550
  "gt_support": 125,
1551
+ "predicted": 0
1552
  },
1553
  "3": {
1554
+ "accuracy": 0.7081212121212122,
1555
+ "precision": 0.5242596470236315,
1556
+ "recall": 0.7081212121212122,
1557
+ "f1": 0.6024750773461671,
1558
  "gt_support": 12375,
1559
+ "predicted": 16715
1560
  },
1561
  "4": {
1562
+ "accuracy": 0.5645714285714286,
1563
+ "precision": 0.6414543093653627,
1564
+ "recall": 0.5645714285714286,
1565
+ "f1": 0.600562267304916,
1566
  "gt_support": 7000,
1567
+ "predicted": 6161
1568
  }
1569
  },
1570
  "total_samples": 28125
1571
  },
1572
  "13_0.6": {
1573
+ "accuracy": 0.4617447199265381,
1574
  "baseline_accuracy": 0.39555555555555555,
1575
  "per_class": {
1576
  "0": {
1577
+ "accuracy": 0.0,
1578
+ "precision": 0.0,
1579
+ "recall": 0.0,
1580
+ "f1": 0.0,
1581
  "gt_support": 121,
1582
+ "predicted": 0
1583
  },
1584
  "1": {
1585
+ "accuracy": 0.547361729179911,
1586
+ "precision": 0.37979708866343187,
1587
+ "recall": 0.547361729179911,
1588
+ "f1": 0.4484375,
1589
  "gt_support": 9438,
1590
+ "predicted": 13602
1591
  },
1592
  "2": {
1593
+ "accuracy": 0.0,
1594
+ "precision": 0.0,
1595
+ "recall": 0.0,
1596
+ "f1": 0.0,
1597
  "gt_support": 121,
1598
+ "predicted": 0
1599
  },
1600
  "3": {
1601
+ "accuracy": 0.3472002971492246,
1602
+ "precision": 0.4662676144157626,
1603
+ "recall": 0.3472002971492246,
1604
+ "f1": 0.39802001277411114,
1605
  "gt_support": 10769,
1606
+ "predicted": 8019
1607
  },
1608
  "4": {
1609
+ "accuracy": 0.5410271546635183,
1610
+ "precision": 0.6541755888650964,
1611
+ "recall": 0.5410271546635183,
1612
+ "f1": 0.5922455573505654,
1613
  "gt_support": 6776,
1614
+ "predicted": 5604
1615
  }
1616
  },
1617
  "total_samples": 27225
1618
  },
1619
  "13_0.8": {
1620
+ "accuracy": 0.4801190476190476,
1621
  "baseline_accuracy": 0.39555555555555555,
1622
  "per_class": {
1623
  "0": {
1624
+ "accuracy": 0.0,
1625
+ "precision": 0.0,
1626
+ "recall": 0.0,
1627
+ "f1": 0.0,
1628
  "gt_support": 112,
1629
+ "predicted": 0
1630
  },
1631
  "1": {
1632
+ "accuracy": 0.8643659711075441,
1633
+ "precision": 0.43534940124298926,
1634
+ "recall": 0.8643659711075441,
1635
+ "f1": 0.5790517154474276,
1636
  "gt_support": 9968,
1637
+ "predicted": 19791
1638
  },
1639
  "2": {
1640
+ "accuracy": 0.0,
1641
+ "precision": 0.0,
1642
+ "recall": 0.0,
1643
+ "f1": 0.0,
1644
  "gt_support": 112,
1645
+ "predicted": 0
1646
  },
1647
  "3": {
1648
+ "accuracy": 0.032280219780219783,
1649
+ "precision": 0.41531664212076586,
1650
+ "recall": 0.032280219780219783,
1651
+ "f1": 0.0599044078597982,
1652
  "gt_support": 8736,
1653
+ "predicted": 679
1654
  },
1655
  "4": {
1656
+ "accuracy": 0.5103635204081632,
1657
+ "precision": 0.6767441860465117,
1658
+ "recall": 0.5103635204081632,
1659
+ "f1": 0.5818942010543537,
1660
  "gt_support": 6272,
1661
+ "predicted": 4730
1662
  }
1663
  },
1664
  "total_samples": 25200
1665
  },
1666
  "13_1.0": {
1667
+ "accuracy": 0.5185685071574642,
1668
  "baseline_accuracy": 0.43555555555555553,
1669
  "per_class": {
1670
  "0": {
1671
+ "accuracy": 0.0,
1672
+ "precision": 0.0,
1673
+ "recall": 0.0,
1674
+ "f1": 0.0,
1675
  "gt_support": 326,
1676
+ "predicted": 0
1677
  },
1678
  "1": {
1679
+ "accuracy": 0.919275071991987,
1680
+ "precision": 0.48042727912188576,
1681
+ "recall": 0.919275071991987,
1682
+ "f1": 0.6310553400874526,
1683
  "gt_support": 31948,
1684
+ "predicted": 61131
1685
  },
1686
  "2": {
1687
+ "accuracy": 0.0,
1688
+ "precision": 0.0,
1689
+ "recall": 0.0,
1690
+ "f1": 0.0,
1691
  "gt_support": 326,
1692
+ "predicted": 0
1693
  },
1694
  "3": {
1695
+ "accuracy": 0.0,
1696
+ "precision": 0.0,
1697
+ "recall": 0.0,
1698
+ "f1": 0.0,
1699
  "gt_support": 22494,
1700
+ "predicted": 0
1701
  },
1702
  "4": {
1703
+ "accuracy": 0.4748028045574058,
1704
+ "precision": 0.7093870202144201,
1705
+ "recall": 0.4748028045574058,
1706
+ "f1": 0.5688597210828549,
1707
  "gt_support": 18256,
1708
+ "predicted": 12219
1709
  }
1710
  },
1711
  "total_samples": 73350
1712
  },
1713
  "15_0.0": {
1714
+ "accuracy": 0.7407235142118863,
1715
  "baseline_accuracy": 0.7422222222222222,
1716
  "per_class": {
1717
  "0": {
1718
+ "accuracy": 0.0,
1719
+ "precision": 0.0,
1720
+ "recall": 0.0,
1721
+ "f1": 0.0,
1722
  "gt_support": 86,
1723
+ "predicted": 0
1724
  },
1725
  "1": {
1726
+ "accuracy": 0.004775747508305648,
1727
+ "precision": 0.30666666666666664,
1728
+ "recall": 0.004775747508305648,
1729
+ "f1": 0.009405029646289103,
1730
  "gt_support": 4816,
1731
+ "predicted": 75
1732
  },
1733
  "2": {
1734
+ "accuracy": 0.0,
1735
+ "precision": 0.0,
1736
+ "recall": 0.0,
1737
+ "f1": 0.0,
1738
  "gt_support": 86,
1739
+ "predicted": 0
1740
  },
1741
  "3": {
1742
+ "accuracy": 0.9963793343545467,
1743
+ "precision": 0.7424124513618677,
1744
+ "recall": 0.9963793343545467,
1745
+ "f1": 0.8508487677260159,
1746
  "gt_support": 14362,
1747
+ "predicted": 19275
1748
  },
1749
  "4": {
1750
  "accuracy": 0.0,
 
1752
  "recall": 0.0,
1753
  "f1": 0.0,
1754
  "gt_support": 0,
1755
+ "predicted": 0
1756
  }
1757
  },
1758
  "total_samples": 19350
1759
  },
1760
  "15_0.2": {
1761
+ "accuracy": 0.6174712643678161,
1762
  "baseline_accuracy": 0.6755555555555556,
1763
  "per_class": {
1764
  "0": {
1765
+ "accuracy": 0.0,
1766
+ "precision": 0.0,
1767
+ "recall": 0.0,
1768
+ "f1": 0.0,
1769
  "gt_support": 87,
1770
+ "predicted": 0
1771
  },
1772
  "1": {
1773
+ "accuracy": 0.16043386757325562,
1774
+ "precision": 0.31490308230060376,
1775
+ "recall": 0.16043386757325562,
1776
+ "f1": 0.21256971256971258,
1777
  "gt_support": 6177,
1778
+ "predicted": 3147
1779
  },
1780
  "2": {
1781
+ "accuracy": 0.0,
1782
+ "precision": 0.0,
1783
+ "recall": 0.0,
1784
+ "f1": 0.0,
1785
  "gt_support": 87,
1786
+ "predicted": 0
1787
  },
1788
  "3": {
1789
+ "accuracy": 0.8390804597701149,
1790
+ "precision": 0.6754321889457024,
1791
+ "recall": 0.8390804597701149,
1792
+ "f1": 0.74841494671523,
1793
  "gt_support": 13224,
1794
+ "predicted": 16428
1795
  },
1796
  "4": {
1797
  "accuracy": 0.0,
 
1799
  "recall": 0.0,
1800
  "f1": 0.0,
1801
  "gt_support": 0,
1802
+ "predicted": 0
1803
  }
1804
  },
1805
  "total_samples": 19575
1806
  },
1807
  "15_0.4": {
1808
+ "accuracy": 0.5097385620915033,
1809
  "baseline_accuracy": 0.6133333333333333,
1810
  "per_class": {
1811
  "0": {
1812
+ "accuracy": 0.0,
1813
+ "precision": 0.0,
1814
+ "recall": 0.0,
1815
+ "f1": 0.0,
1816
  "gt_support": 136,
1817
+ "predicted": 0
1818
  },
1819
  "1": {
1820
+ "accuracy": 0.43944636678200694,
1821
+ "precision": 0.3778637310324308,
1822
+ "recall": 0.43944636678200694,
1823
+ "f1": 0.40633498640217564,
1824
  "gt_support": 11560,
1825
+ "predicted": 13444
1826
  },
1827
  "2": {
1828
+ "accuracy": 0.0,
1829
+ "precision": 0.0,
1830
+ "recall": 0.0,
1831
+ "f1": 0.0,
1832
  "gt_support": 136,
1833
+ "predicted": 0
1834
  },
1835
  "3": {
1836
+ "accuracy": 0.5604219948849105,
1837
+ "precision": 0.6130799720214503,
1838
+ "recall": 0.5604219948849105,
1839
+ "f1": 0.5855695356864492,
1840
  "gt_support": 18768,
1841
+ "predicted": 17156
1842
  },
1843
  "4": {
1844
  "accuracy": 0.0,
 
1846
  "recall": 0.0,
1847
  "f1": 0.0,
1848
  "gt_support": 0,
1849
+ "predicted": 0
1850
  }
1851
  },
1852
  "total_samples": 30600
1853
  },
1854
  "15_0.6": {
1855
+ "accuracy": 0.4474666666666667,
1856
  "baseline_accuracy": 0.5466666666666666,
1857
  "per_class": {
1858
  "0": {
1859
+ "accuracy": 0.0,
1860
+ "precision": 0.0,
1861
+ "recall": 0.0,
1862
+ "f1": 0.0,
1863
  "gt_support": 100,
1864
+ "predicted": 0
1865
  },
1866
  "1": {
1867
+ "accuracy": 0.9699,
1868
+ "precision": 0.4444189882697947,
1869
+ "recall": 0.9699,
1870
+ "f1": 0.6095399698340874,
1871
  "gt_support": 10000,
1872
+ "predicted": 21824
1873
  },
1874
  "2": {
1875
+ "accuracy": 0.0,
1876
+ "precision": 0.0,
1877
+ "recall": 0.0,
1878
+ "f1": 0.0,
1879
  "gt_support": 100,
1880
+ "predicted": 0
1881
  },
1882
  "3": {
1883
+ "accuracy": 0.03,
1884
+ "precision": 0.5458579881656804,
1885
+ "recall": 0.03,
1886
+ "f1": 0.05687422934648581,
1887
  "gt_support": 12300,
1888
+ "predicted": 676
1889
  },
1890
  "4": {
1891
  "accuracy": 0.0,
 
1893
  "recall": 0.0,
1894
  "f1": 0.0,
1895
  "gt_support": 0,
1896
+ "predicted": 0
1897
  }
1898
  },
1899
  "total_samples": 22500
1900
  },
1901
  "15_0.8": {
1902
+ "accuracy": 0.5066666666666667,
1903
  "baseline_accuracy": 0.5066666666666667,
1904
  "per_class": {
1905
  "0": {
1906
+ "accuracy": 0.0,
1907
+ "precision": 0.0,
1908
+ "recall": 0.0,
1909
+ "f1": 0.0,
1910
  "gt_support": 164,
1911
+ "predicted": 0
1912
  },
1913
  "1": {
1914
+ "accuracy": 1.0,
1915
+ "precision": 0.5066666666666667,
1916
+ "recall": 1.0,
1917
+ "f1": 0.6725663716814159,
1918
  "gt_support": 18696,
1919
+ "predicted": 36900
1920
  },
1921
  "2": {
1922
+ "accuracy": 0.0,
1923
+ "precision": 0.0,
1924
+ "recall": 0.0,
1925
+ "f1": 0.0,
1926
  "gt_support": 164,
1927
+ "predicted": 0
1928
  },
1929
  "3": {
1930
+ "accuracy": 0.0,
1931
+ "precision": 0.0,
1932
+ "recall": 0.0,
1933
+ "f1": 0.0,
1934
  "gt_support": 17876,
1935
+ "predicted": 0
1936
  },
1937
  "4": {
1938
  "accuracy": 0.0,
 
1940
  "recall": 0.0,
1941
  "f1": 0.0,
1942
  "gt_support": 0,
1943
+ "predicted": 0
1944
  }
1945
  },
1946
  "total_samples": 36900
1947
  },
1948
  "15_1.0": {
1949
+ "accuracy": 0.5688888888888889,
1950
  "baseline_accuracy": 0.5688888888888889,
1951
  "per_class": {
1952
  "0": {
1953
+ "accuracy": 0.0,
1954
+ "precision": 0.0,
1955
+ "recall": 0.0,
1956
+ "f1": 0.0,
1957
  "gt_support": 391,
1958
+ "predicted": 0
1959
  },
1960
  "1": {
1961
+ "accuracy": 1.0,
1962
+ "precision": 0.5688888888888889,
1963
+ "recall": 1.0,
1964
+ "f1": 0.725212464589235,
1965
  "gt_support": 50048,
1966
+ "predicted": 87975
1967
  },
1968
  "2": {
1969
+ "accuracy": 0.0,
1970
+ "precision": 0.0,
1971
+ "recall": 0.0,
1972
+ "f1": 0.0,
1973
  "gt_support": 391,
1974
+ "predicted": 0
1975
  },
1976
  "3": {
1977
+ "accuracy": 0.0,
1978
+ "precision": 0.0,
1979
+ "recall": 0.0,
1980
+ "f1": 0.0,
1981
  "gt_support": 37145,
1982
+ "predicted": 0
1983
  },
1984
  "4": {
1985
  "accuracy": 0.0,
 
1987
  "recall": 0.0,
1988
  "f1": 0.0,
1989
  "gt_support": 0,
1990
+ "predicted": 0
1991
  }
1992
  },
1993
  "total_samples": 87975
1994
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
1995
  }
1996
  }
layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_lr_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
@@ -27,17 +27,17 @@ Processed 300 trajectories, 3223 steps
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
- Overall Accuracy: 0.0796 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
- A 0.7003 0.0046 0.7003 0.0091 3223 490729
37
- # 0.0860 0.2978 0.0860 0.1334 238175 68785
38
- G 0.1514 0.0046 0.1514 0.0089 3223 106948
39
- _ 0.0542 0.4166 0.0542 0.0960 243346 31670
40
- + 0.0899 0.7883 0.0899 0.1613 237208 27043
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
@@ -47,81 +47,81 @@ METRICS BY SIZE
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
- Overall Accuracy: 0.1136 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
- A 0.8147 0.0048 0.8147 0.0096 286 48056
57
- # 0.1643 0.1420 0.1643 0.1523 8224 9512
58
- G 0.0175 0.0059 0.0175 0.0088 286 853
59
- _ 0.0034 0.0800 0.0034 0.0066 5218 225
60
- + 0.1133 1.0000 0.1133 0.2036 50336 5704
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
- Overall Accuracy: 0.0840 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
- A 0.8132 0.0047 0.8132 0.0093 514 89499
73
- # 0.1426 0.1966 0.1426 0.1653 22670 16442
74
- G 0.0175 0.0051 0.0175 0.0079 514 1777
75
- _ 0.0209 0.1671 0.0209 0.0371 17936 2238
76
- + 0.0768 0.9977 0.0768 0.1425 74016 5694
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
- Overall Accuracy: 0.0723 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
- A 0.7442 0.0046 0.7442 0.0092 649 104819
89
- # 0.0802 0.2636 0.0802 0.1230 39462 12001
90
- G 0.1433 0.0047 0.1433 0.0092 649 19633
91
- _ 0.0220 0.2551 0.0220 0.0405 37769 3257
92
- + 0.0887 0.9482 0.0887 0.1623 67496 6315
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
- Overall Accuracy: 0.0848 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
- A 0.6457 0.0045 0.6457 0.0090 810 115174
105
- # 0.0797 0.3628 0.0797 0.1307 66522 14616
106
- G 0.1914 0.0046 0.1914 0.0089 810 33969
107
- _ 0.0804 0.4025 0.0804 0.1340 68748 13734
108
- + 0.0870 0.8293 0.0870 0.1574 45360 4757
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
- Overall Accuracy: 0.0678 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
- A 0.6224 0.0045 0.6224 0.0089 964 133181
121
- # 0.0734 0.4583 0.0734 0.1265 101297 16214
122
- G 0.2344 0.0045 0.2344 0.0087 964 50716
123
- _ 0.0567 0.5275 0.0567 0.1024 113675 12216
124
- + 0.0000 0.0000 0.0000 0.0000 0 4573
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
@@ -131,97 +131,97 @@ METRICS BY COMPLEXITY
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
- Overall Accuracy: 0.1126 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
- A 0.4839 0.0048 0.4839 0.0096 310 31019
141
- # 0.1106 0.1793 0.1106 0.1368 13248 8169
142
- G 0.3290 0.0046 0.3290 0.0091 310 22078
143
- _ 0.0780 0.5479 0.0780 0.1366 30762 4380
144
- + 0.1487 0.9101 0.1487 0.2556 25120 4104
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
- Overall Accuracy: 0.0941 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
- A 0.6738 0.0047 0.6738 0.0093 325 46986
157
- # 0.0856 0.2056 0.0856 0.1209 16604 6915
158
- G 0.1600 0.0047 0.1600 0.0092 325 10989
159
- _ 0.0863 0.5102 0.0863 0.1476 29047 4912
160
- + 0.0999 0.8065 0.0999 0.1778 26824 3323
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
- Overall Accuracy: 0.0785 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
- A 0.6863 0.0046 0.6863 0.0091 424 63666
173
- # 0.1422 0.2607 0.1422 0.1840 26803 14621
174
- G 0.1344 0.0046 0.1344 0.0089 424 12422
175
- _ 0.0192 0.5236 0.0192 0.0370 38165 1396
176
- + 0.0877 0.7873 0.0877 0.1578 29584 3295
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
- Overall Accuracy: 0.0774 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
- A 0.7361 0.0046 0.7361 0.0091 432 69347
189
- # 0.0862 0.3008 0.0862 0.1340 29604 8482
190
- G 0.1227 0.0045 0.1227 0.0087 432 11755
191
- _ 0.0491 0.4067 0.0491 0.0876 32188 3885
192
- + 0.0875 0.8105 0.0875 0.1580 34544 3731
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
- Overall Accuracy: 0.0765 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
- A 0.7326 0.0046 0.7326 0.0091 546 87126
205
- # 0.0794 0.2968 0.0794 0.1253 42640 11404
206
- G 0.1300 0.0046 0.1300 0.0090 546 15297
207
- _ 0.0439 0.3726 0.0439 0.0785 36766 4329
208
- + 0.0927 0.8360 0.0927 0.1668 42352 4694
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
- Overall Accuracy: 0.0697 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
- A 0.7411 0.0046 0.7411 0.0091 1186 192585
221
- # 0.0718 0.4088 0.0718 0.1221 109276 19194
222
- G 0.1290 0.0044 0.1290 0.0086 1186 34407
223
- _ 0.0571 0.3419 0.0571 0.0979 76418 12768
224
- + 0.0680 0.6790 0.0680 0.1237 78784 7896
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
@@ -231,481 +231,481 @@ METRICS BY SIZE-COMPLEXITY COMBINATION
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
- Overall Accuracy: 0.1482 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
- A 0.7895 0.0051 0.7895 0.0101 38 5884
241
- # 0.1316 0.1137 0.1316 0.1220 912 1055
242
- G 0.0789 0.0060 0.0789 0.0112 38 497
243
  _ 0.0000 0.0000 0.0000 0.0000 874 0
244
- + 0.1666 1.0000 0.1666 0.2856 6688 1114
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
- Overall Accuracy: 0.0995 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
- A 0.7727 0.0048 0.7727 0.0095 44 7125
257
- # 0.1818 0.1240 0.1818 0.1474 1144 1678
258
- G 0.0455 0.0056 0.0455 0.0100 44 356
259
  _ 0.0000 0.0000 0.0000 0.0000 924 0
260
- + 0.0957 1.0000 0.0957 0.1747 7744 741
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
- Overall Accuracy: 0.1135 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
- A 0.7561 0.0048 0.7561 0.0096 41 6444
273
- # 0.2439 0.1369 0.2439 0.1754 1148 2045
274
- G 0.0000 0.0000 0.0000 0.0000 41 0
275
  _ 0.0000 0.0000 0.0000 0.0000 779 0
276
- + 0.1020 1.0000 0.1020 0.1851 7216 736
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
- Overall Accuracy: 0.0934 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
- A 0.8723 0.0048 0.8723 0.0096 47 8537
289
- # 0.1064 0.1409 0.1064 0.1212 1363 1029
290
- G 0.0000 0.0000 0.0000 0.0000 47 0
291
- _ 0.0213 0.0800 0.0213 0.0336 846 225
292
- + 0.0948 1.0000 0.0948 0.1731 8272 784
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
- Overall Accuracy: 0.1265 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
- A 0.8182 0.0049 0.8182 0.0097 55 9246
305
- # 0.1818 0.1616 0.1818 0.1711 1705 1918
306
- G 0.0000 0.0000 0.0000 0.0000 55 0
307
  _ 0.0000 0.0000 0.0000 0.0000 880 0
308
- + 0.1251 1.0000 0.1251 0.2224 9680 1211
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
- Overall Accuracy: 0.1062 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
- A 0.8525 0.0048 0.8525 0.0096 61 10820
321
- # 0.1475 0.1612 0.1475 0.1541 1952 1787
322
- G 0.0000 0.0000 0.0000 0.0000 61 0
323
  _ 0.0000 0.0000 0.0000 0.0000 915 0
324
- + 0.1041 1.0000 0.1041 0.1886 10736 1118
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
- Overall Accuracy: 0.1624 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
- A 0.7037 0.0053 0.7037 0.0105 54 7185
337
- # 0.1944 0.1416 0.1944 0.1639 1728 2373
338
- G 0.0556 0.0047 0.0556 0.0087 54 636
339
- _ 0.0370 0.2089 0.0370 0.0629 2538 450
340
- + 0.1932 0.9973 0.1932 0.3236 7776 1506
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
- Overall Accuracy: 0.1009 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
- A 0.7872 0.0047 0.7872 0.0094 47 7798
353
- # 0.1702 0.1722 0.1702 0.1712 1692 1672
354
- G 0.0000 0.0000 0.0000 0.0000 47 0
355
- _ 0.0426 0.1915 0.0426 0.0696 2021 449
356
- + 0.0969 1.0000 0.0969 0.1767 6768 656
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
- Overall Accuracy: 0.0840 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
- A 0.7612 0.0046 0.7612 0.0092 67 11039
369
- # 0.2388 0.1849 0.2388 0.2084 2680 3461
370
- G 0.0000 0.0000 0.0000 0.0000 67 0
371
- _ 0.0000 0.0000 0.0000 0.0000 2613 0
372
- + 0.0596 1.0000 0.0596 0.1125 9648 575
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
- Overall Accuracy: 0.0792 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
- A 0.8197 0.0046 0.8197 0.0092 61 10856
385
- # 0.1475 0.2102 0.1475 0.1734 2623 1841
386
- G 0.0000 0.0000 0.0000 0.0000 61 0
387
- _ 0.0328 0.1600 0.0328 0.0544 2196 450
388
- + 0.0658 1.0000 0.0658 0.1235 8784 578
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
- Overall Accuracy: 0.0983 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
- A 0.7426 0.0046 0.7426 0.0092 101 16230
401
- # 0.2069 0.2229 0.2069 0.2146 4747 4405
402
- G 0.0297 0.0056 0.0297 0.0095 101 532
403
- _ 0.0198 0.1455 0.0198 0.0349 3232 440
404
- + 0.0763 0.9919 0.0763 0.1416 14544 1118
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
- Overall Accuracy: 0.0505 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
- A 0.9076 0.0046 0.9076 0.0091 184 36391
417
- # 0.0652 0.2230 0.0652 0.1009 9200 2690
418
- G 0.0163 0.0049 0.0163 0.0076 184 609
419
- _ 0.0109 0.1292 0.0109 0.0201 5336 449
420
- + 0.0476 1.0000 0.0476 0.0909 26496 1261
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
- Overall Accuracy: 0.0958 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
- A 0.4412 0.0047 0.4412 0.0093 68 6395
433
- # 0.1500 0.1868 0.1500 0.1664 2720 2184
434
- G 0.3676 0.0047 0.3676 0.0092 68 5353
435
- _ 0.0262 0.3473 0.0262 0.0488 5372 406
436
- + 0.1219 0.8960 0.1219 0.2146 7072 962
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
- Overall Accuracy: 0.0765 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
- A 0.7882 0.0047 0.7882 0.0093 85 14302
449
- # 0.0588 0.2113 0.0588 0.0920 3995 1112
450
- G 0.1294 0.0050 0.1294 0.0096 85 2216
451
- _ 0.0235 0.3200 0.0235 0.0438 6120 450
452
- + 0.1138 0.9627 0.1138 0.2035 8840 1045
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
- Overall Accuracy: 0.1004 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
- A 0.7091 0.0047 0.7091 0.0093 55 8350
465
- # 0.2178 0.2431 0.2178 0.2298 2915 2612
466
- G 0.0727 0.0047 0.0727 0.0089 55 847
467
- _ 0.0000 0.0000 0.0000 0.0000 3630 0
468
- + 0.0986 0.9965 0.0986 0.1794 5720 566
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
- Overall Accuracy: 0.0753 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
- A 0.8252 0.0046 0.8252 0.0092 103 18347
481
- # 0.0777 0.2804 0.0777 0.1216 6180 1712
482
- G 0.0777 0.0049 0.0777 0.0092 103 1630
483
- _ 0.0105 0.2667 0.0105 0.0203 6077 240
484
- + 0.1034 0.8892 0.1034 0.1853 10712 1246
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
- Overall Accuracy: 0.0577 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
- A 0.8070 0.0046 0.8070 0.0091 114 20147
497
- # 0.0439 0.3036 0.0439 0.0766 7524 1087
498
- G 0.1316 0.0049 0.1316 0.0095 114 3041
499
- _ 0.0175 0.2512 0.0175 0.0328 6042 422
500
- + 0.0790 0.9832 0.0790 0.1463 11856 953
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
- Overall Accuracy: 0.0628 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
- A 0.7589 0.0046 0.7589 0.0091 224 37278
513
- # 0.0667 0.3267 0.0667 0.1108 16128 3294
514
- G 0.1339 0.0046 0.1339 0.0089 224 6546
515
- _ 0.0357 0.2162 0.0357 0.0613 10528 1739
516
- + 0.0649 0.9793 0.0649 0.1217 23296 1543
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
- Overall Accuracy: 0.1206 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
- A 0.3594 0.0045 0.3594 0.0089 64 5108
529
- # 0.0781 0.2151 0.0781 0.1146 3072 1116
530
- G 0.4062 0.0046 0.4062 0.0090 64 5685
531
- _ 0.1562 0.5360 0.1562 0.2420 7616 2220
532
- + 0.0717 0.9483 0.0717 0.1333 3584 271
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
- Overall Accuracy: 0.1290 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
- A 0.5645 0.0046 0.5645 0.0091 62 7672
545
- # 0.0968 0.2589 0.0968 0.1409 3596 1344
546
- G 0.1613 0.0044 0.1613 0.0085 62 2298
547
- _ 0.1672 0.4881 0.1672 0.2491 6758 2315
548
- + 0.0798 0.8629 0.0798 0.1461 3472 321
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
- Overall Accuracy: 0.0741 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
- A 0.6880 0.0045 0.6880 0.0090 125 18931
561
- # 0.1014 0.3092 0.1014 0.1527 8500 2788
562
- G 0.1760 0.0047 0.1760 0.0091 125 4706
563
- _ 0.0319 0.4545 0.0319 0.0596 12375 869
564
- + 0.1027 0.8652 0.1027 0.1836 7000 831
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
- Overall Accuracy: 0.0725 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
- A 0.6777 0.0045 0.6777 0.0090 121 18024
577
- # 0.0784 0.3522 0.0784 0.1283 9438 2101
578
- G 0.1818 0.0043 0.1818 0.0085 121 5059
579
- _ 0.0535 0.4034 0.0535 0.0944 10769 1428
580
- + 0.0818 0.9038 0.0818 0.1500 6776 613
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
- Overall Accuracy: 0.0958 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
- A 0.6161 0.0046 0.6161 0.0090 112 15148
593
- # 0.0974 0.4004 0.0974 0.1567 9968 2425
594
- G 0.2054 0.0047 0.2054 0.0092 112 4908
595
- _ 0.0784 0.3590 0.0784 0.1287 8736 1908
596
- + 0.1063 0.8224 0.1063 0.1883 6272 811
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
- Overall Accuracy: 0.0742 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
- A 0.6994 0.0045 0.6994 0.0090 326 50291
609
- # 0.0670 0.4422 0.0670 0.1164 31948 4842
610
- G 0.1595 0.0046 0.1595 0.0089 326 11313
611
- _ 0.0690 0.3108 0.0690 0.1129 22494 4994
612
- + 0.0806 0.7702 0.0806 0.1459 18256 1910
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
- Overall Accuracy: 0.0729 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
- A 0.3372 0.0045 0.3372 0.0089 86 6447
625
- # 0.0750 0.2505 0.0750 0.1154 4816 1441
626
- G 0.5233 0.0045 0.5233 0.0090 86 9907
627
- _ 0.0679 0.7477 0.0679 0.1245 14362 1304
628
- + 0.0000 0.0000 0.0000 0.0000 0 251
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
- Overall Accuracy: 0.0799 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
- A 0.5287 0.0046 0.5287 0.0090 87 10089
641
- # 0.0555 0.3093 0.0555 0.0942 6177 1109
642
- G 0.3333 0.0047 0.3333 0.0093 87 6119
643
- _ 0.0867 0.6749 0.0867 0.1536 13224 1698
644
- + 0.0000 0.0000 0.0000 0.0000 0 560
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
- Overall Accuracy: 0.0603 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
- A 0.6176 0.0044 0.6176 0.0088 136 18902
657
- # 0.1207 0.3755 0.1207 0.1827 11560 3715
658
- G 0.2279 0.0045 0.2279 0.0089 136 6869
659
- _ 0.0179 0.6376 0.0179 0.0348 18768 527
660
- + 0.0000 0.0000 0.0000 0.0000 0 587
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
- Overall Accuracy: 0.0770 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
- A 0.6000 0.0044 0.6000 0.0088 100 13583
673
- # 0.0799 0.4441 0.0799 0.1354 10000 1799
674
- G 0.2300 0.0045 0.2300 0.0089 100 5066
675
- _ 0.0691 0.5512 0.0691 0.1228 12300 1542
676
- + 0.0000 0.0000 0.0000 0.0000 0 510
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
- Overall Accuracy: 0.0460 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
- A 0.7256 0.0045 0.7256 0.0090 164 26355
689
- # 0.0424 0.5048 0.0424 0.0782 18696 1569
690
- G 0.1829 0.0044 0.1829 0.0086 164 6816
691
- _ 0.0424 0.4862 0.0424 0.0780 17876 1559
692
- + 0.0000 0.0000 0.0000 0.0000 0 601
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
- Overall Accuracy: 0.0733 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
- A 0.6701 0.0045 0.6701 0.0090 391 57805
705
- # 0.0747 0.5685 0.0747 0.1321 50048 6581
706
- G 0.1739 0.0043 0.1739 0.0083 391 15939
707
- _ 0.0640 0.4259 0.0640 0.1113 37145 5586
708
- + 0.0000 0.0000 0.0000 0.0000 0 2064
709
  ---------------------------------------------------------------------------------------
710
 
711
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_lr_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer23_lr_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
 
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
+ Overall Accuracy: 0.5711 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
+ A 0.1722 0.0126 0.1722 0.0234 3223 44175
37
+ # 0.6250 0.4843 0.6250 0.5457 238175 307391
38
+ G 0.1018 0.0126 0.1018 0.0225 3223 25990
39
+ _ 0.2946 0.5987 0.2946 0.3949 243346 119733
40
+ + 0.8124 0.8456 0.8124 0.8287 237208 227886
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
 
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
+ Overall Accuracy: 0.7522 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
+ A 0.6818 0.0202 0.6818 0.0393 286 9642
57
+ # 0.0000 0.0000 0.0000 0.0000 8224 0
58
+ G 0.2832 0.0205 0.2832 0.0382 286 3951
59
+ _ 0.0000 0.0000 0.0000 0.0000 5218 0
60
+ + 0.9562 0.9482 0.9562 0.9522 50336 50757
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
+ Overall Accuracy: 0.6163 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
+ A 0.3774 0.0136 0.3774 0.0263 514 14242
73
+ # 0.0068 0.4755 0.0068 0.0135 22670 326
74
+ G 0.4455 0.0121 0.4455 0.0236 514 18853
75
+ _ 0.1258 0.4474 0.1258 0.1964 17936 5042
76
+ + 0.9247 0.8867 0.9247 0.9053 74016 77187
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
+ Overall Accuracy: 0.5469 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
+ A 0.2527 0.0082 0.2527 0.0159 649 19951
89
+ # 0.4815 0.4459 0.4815 0.4630 39462 42607
90
+ G 0.0262 0.0056 0.0262 0.0092 649 3063
91
+ _ 0.2087 0.5225 0.2087 0.2982 37769 15083
92
+ + 0.7822 0.8082 0.7822 0.7950 67496 65321
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
+ Overall Accuracy: 0.5145 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
+ A 0.0025 0.0059 0.0025 0.0035 810 340
105
+ # 0.6791 0.4468 0.6791 0.5390 66522 101093
106
+ G 0.0012 0.0081 0.0012 0.0021 810 123
107
+ _ 0.3673 0.5481 0.3673 0.4398 68748 46073
108
+ + 0.5146 0.6742 0.5146 0.5837 45360 34621
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
+ Overall Accuracy: 0.5571 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
+ A 0.0000 0.0000 0.0000 0.0000 964 0
121
+ # 0.8345 0.5175 0.8345 0.6388 101297 163365
122
+ G 0.0000 0.0000 0.0000 0.0000 964 0
123
+ _ 0.3193 0.6779 0.3193 0.4341 113675 53535
124
+ + 0.0000 0.0000 0.0000 0.0000 0 0
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
 
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
+ Overall Accuracy: 0.6724 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
+ A 0.1323 0.0127 0.1323 0.0232 310 3225
141
+ # 0.0110 0.2274 0.0110 0.0210 13248 642
142
+ G 0.1323 0.0135 0.1323 0.0245 310 3035
143
+ _ 0.8410 0.6684 0.8410 0.7448 30762 38709
144
+ + 0.8280 0.8616 0.8280 0.8445 25120 24139
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
+ Overall Accuracy: 0.5927 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
+ A 0.2431 0.0116 0.2431 0.0221 325 6820
157
+ # 0.0802 0.3043 0.0802 0.1269 16604 4374
158
+ G 0.1415 0.0111 0.1415 0.0206 325 4136
159
+ _ 0.6711 0.6169 0.6711 0.6428 29047 31599
160
+ + 0.8349 0.8549 0.8349 0.8448 26824 26196
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
+ Overall Accuracy: 0.5446 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
+ A 0.1958 0.0118 0.1958 0.0222 424 7063
173
+ # 0.2560 0.3619 0.2560 0.2999 26803 18962
174
+ G 0.0991 0.0122 0.0991 0.0217 424 3448
175
+ _ 0.5408 0.5619 0.5408 0.5511 38165 36732
176
+ + 0.8225 0.8334 0.8225 0.8279 29584 29195
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
+ Overall Accuracy: 0.5109 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
+ A 0.2546 0.0114 0.2546 0.0219 432 9616
189
+ # 0.5598 0.4158 0.5598 0.4772 29604 39855
190
+ G 0.1065 0.0121 0.1065 0.0217 432 3802
191
+ _ 0.1484 0.4637 0.1484 0.2249 32188 10305
192
+ + 0.8149 0.8372 0.8149 0.8259 34544 33622
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
+ Overall Accuracy: 0.5538 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
+ A 0.2033 0.0128 0.2033 0.0241 546 8657
205
+ # 0.7380 0.4737 0.7380 0.5771 42640 66427
206
+ G 0.1062 0.0151 0.1062 0.0264 546 3848
207
+ _ 0.0204 0.3849 0.0204 0.0388 36766 1951
208
+ + 0.8418 0.8495 0.8418 0.8456 42352 41967
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
+ Overall Accuracy: 0.5780 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
+ A 0.1105 0.0149 0.1105 0.0263 1186 8794
221
+ # 0.8463 0.5221 0.8463 0.6458 109276 177131
222
+ G 0.0801 0.0123 0.0801 0.0213 1186 7721
223
+ _ 0.0019 0.3410 0.0019 0.0039 76418 437
224
+ + 0.7791 0.8436 0.7791 0.8101 78784 72767
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
 
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
+ Overall Accuracy: 0.7471 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
+ A 0.6053 0.0206 0.6053 0.0399 38 1116
241
+ # 0.0000 0.0000 0.0000 0.0000 912 0
242
+ G 0.3947 0.0195 0.3947 0.0372 38 768
243
  _ 0.0000 0.0000 0.0000 0.0000 874 0
244
+ + 0.9495 0.9526 0.9495 0.9510 6688 6666
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
+ Overall Accuracy: 0.7506 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
+ A 0.7045 0.0209 0.7045 0.0406 44 1484
257
+ # 0.0000 0.0000 0.0000 0.0000 1144 0
258
+ G 0.2727 0.0190 0.2727 0.0355 44 633
259
  _ 0.0000 0.0000 0.0000 0.0000 924 0
260
+ + 0.9540 0.9492 0.9540 0.9516 7744 7783
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
+ Overall Accuracy: 0.7517 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
+ A 0.6585 0.0200 0.6585 0.0388 41 1351
273
+ # 0.0000 0.0000 0.0000 0.0000 1148 0
274
+ G 0.2927 0.0201 0.2927 0.0377 41 596
275
  _ 0.0000 0.0000 0.0000 0.0000 779 0
276
+ + 0.9555 0.9474 0.9555 0.9514 7216 7278
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
+ Overall Accuracy: 0.7512 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
+ A 0.6809 0.0201 0.6809 0.0391 47 1589
289
+ # 0.0000 0.0000 0.0000 0.0000 1363 0
290
+ G 0.2979 0.0211 0.2979 0.0394 47 663
291
+ _ 0.0000 0.0000 0.0000 0.0000 846 0
292
+ + 0.9548 0.9489 0.9548 0.9519 8272 8323
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
+ Overall Accuracy: 0.7544 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
+ A 0.6182 0.0199 0.6182 0.0385 55 1711
305
+ # 0.0000 0.0000 0.0000 0.0000 1705 0
306
+ G 0.3273 0.0208 0.3273 0.0390 55 867
307
  _ 0.0000 0.0000 0.0000 0.0000 880 0
308
+ + 0.9591 0.9476 0.9591 0.9533 9680 9797
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
+ Overall Accuracy: 0.7557 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
+ A 0.7869 0.0201 0.7869 0.0392 61 2391
321
+ # 0.0000 0.0000 0.0000 0.0000 1952 0
322
+ G 0.1639 0.0236 0.1639 0.0412 61 424
323
  _ 0.0000 0.0000 0.0000 0.0000 915 0
324
+ + 0.9607 0.9454 0.9607 0.9530 10736 10910
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
+ Overall Accuracy: 0.6485 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
+ A 0.1667 0.0108 0.1667 0.0203 54 834
337
+ # 0.0000 0.0000 0.0000 0.0000 1728 0
338
+ G 0.4444 0.0123 0.4444 0.0239 54 1954
339
+ _ 0.3176 0.5230 0.3176 0.3952 2538 1541
340
+ + 0.9053 0.9001 0.9053 0.9027 7776 7821
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
+ Overall Accuracy: 0.6368 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
+ A 0.2340 0.0129 0.2340 0.0245 47 851
353
+ # 0.0000 0.0000 0.0000 0.0000 1692 0
354
+ G 0.4681 0.0127 0.4681 0.0246 47 1739
355
+ _ 0.2528 0.4785 0.2528 0.3309 2021 1068
356
+ + 0.9146 0.8949 0.9146 0.9046 6768 6917
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
+ Overall Accuracy: 0.6101 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
+ A 0.3881 0.0128 0.3881 0.0248 67 2033
369
+ # 0.0000 0.0000 0.0000 0.0000 2680 0
370
+ G 0.4328 0.0120 0.4328 0.0233 67 2426
371
+ _ 0.1278 0.4401 0.1278 0.1981 2613 759
372
+ + 0.9129 0.8936 0.9129 0.9032 9648 9857
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
+ Overall Accuracy: 0.5917 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
+ A 0.4590 0.0140 0.4590 0.0271 61 2006
385
+ # 0.0000 0.0000 0.0000 0.0000 2623 0
386
+ G 0.4918 0.0116 0.4918 0.0227 61 2577
387
+ _ 0.0401 0.3964 0.0401 0.0728 2196 222
388
+ + 0.9079 0.8941 0.9079 0.9009 8784 8920
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
+ Overall Accuracy: 0.6227 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
+ A 0.4455 0.0140 0.4455 0.0272 101 3209
401
+ # 0.0076 0.4615 0.0076 0.0149 4747 78
402
+ G 0.3861 0.0136 0.3861 0.0262 101 2875
403
+ _ 0.1139 0.3626 0.1139 0.1733 3232 1015
404
+ + 0.9394 0.8788 0.9394 0.9081 14544 15548
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
+ Overall Accuracy: 0.6086 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
+ A 0.4076 0.0141 0.4076 0.0273 184 5309
417
+ # 0.0129 0.4798 0.0129 0.0252 9200 248
418
+ G 0.4620 0.0117 0.4620 0.0228 184 7282
419
+ _ 0.0279 0.3410 0.0279 0.0516 5336 437
420
+ + 0.9347 0.8806 0.9347 0.9069 26496 28124
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
+ Overall Accuracy: 0.6193 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
+ A 0.1324 0.0071 0.1324 0.0134 68 1275
433
+ # 0.0000 0.0000 0.0000 0.0000 2720 0
434
+ G 0.0147 0.0046 0.0147 0.0070 68 218
435
+ _ 0.7353 0.5636 0.7353 0.6381 5372 7009
436
+ + 0.7798 0.8113 0.7798 0.7952 7072 6798
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
+ Overall Accuracy: 0.4792 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
+ A 0.4353 0.0082 0.4353 0.0162 85 4485
449
+ # 0.0093 0.2761 0.0093 0.0179 3995 134
450
+ G 0.1412 0.0069 0.1412 0.0132 85 1736
451
+ _ 0.3634 0.5140 0.3634 0.4258 6120 4327
452
+ + 0.7755 0.8119 0.7755 0.7933 8840 8443
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
+ Overall Accuracy: 0.4781 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
+ A 0.5091 0.0084 0.5091 0.0165 55 3339
465
+ # 0.0642 0.3071 0.0642 0.1061 2915 609
466
+ G 0.0182 0.0023 0.0182 0.0042 55 426
467
+ _ 0.2821 0.4872 0.2821 0.3573 3630 2102
468
+ + 0.8177 0.7928 0.8177 0.8051 5720 5899
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
+ Overall Accuracy: 0.4726 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
+ A 0.4854 0.0083 0.4854 0.0163 103 6021
481
+ # 0.2762 0.3854 0.2762 0.3218 6180 4429
482
+ G 0.0194 0.0036 0.0194 0.0060 103 562
483
+ _ 0.0958 0.4193 0.0958 0.1559 6077 1388
484
+ + 0.8039 0.7992 0.8039 0.8015 10712 10775
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
+ Overall Accuracy: 0.5363 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
+ A 0.2807 0.0086 0.2807 0.0166 114 3737
497
+ # 0.5477 0.4267 0.5477 0.4797 7524 9658
498
+ G 0.0088 0.0094 0.0088 0.0091 114 106
499
+ _ 0.0167 0.3930 0.0167 0.0321 6042 257
500
+ + 0.8015 0.7990 0.8015 0.8002 11856 11892
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
+ Overall Accuracy: 0.6069 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
+ A 0.0357 0.0073 0.0357 0.0121 224 1094
513
+ # 0.8028 0.4661 0.8028 0.5898 16128 27777
514
+ G 0.0000 0.0000 0.0000 0.0000 224 15
515
+ _ 0.0000 0.0000 0.0000 0.0000 10528 0
516
+ + 0.7570 0.8197 0.7570 0.7871 23296 21514
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
+ Overall Accuracy: 0.6128 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
+ A 0.0000 0.0000 0.0000 0.0000 64 0
529
+ # 0.0400 0.2169 0.0400 0.0676 3072 567
530
+ G 0.0156 0.0105 0.0156 0.0126 64 95
531
+ _ 0.8936 0.6253 0.8936 0.7358 7616 10884
532
+ + 0.5285 0.6636 0.5285 0.5884 3584 2854
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
+ Overall Accuracy: 0.5682 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
+ A 0.0000 0.0000 0.0000 0.0000 62 0
545
+ # 0.0843 0.2772 0.0843 0.1292 3596 1093
546
+ G 0.0000 0.0000 0.0000 0.0000 62 28
547
+ _ 0.8377 0.5791 0.8377 0.6848 6758 9776
548
+ + 0.5651 0.6426 0.5651 0.6014 3472 3053
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
+ Overall Accuracy: 0.5089 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
+ A 0.0160 0.0059 0.0160 0.0086 125 340
561
+ # 0.1876 0.3249 0.1876 0.2379 8500 4909
562
+ G 0.0000 0.0000 0.0000 0.0000 125 0
563
+ _ 0.7081 0.5243 0.7081 0.6025 12375 16715
564
+ + 0.5646 0.6415 0.5646 0.6006 7000 6161
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
+ Overall Accuracy: 0.4617 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
+ A 0.0000 0.0000 0.0000 0.0000 121 0
577
+ # 0.5474 0.3798 0.5474 0.4484 9438 13602
578
+ G 0.0000 0.0000 0.0000 0.0000 121 0
579
+ _ 0.3472 0.4663 0.3472 0.3980 10769 8019
580
+ + 0.5410 0.6542 0.5410 0.5922 6776 5604
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
+ Overall Accuracy: 0.4801 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
+ A 0.0000 0.0000 0.0000 0.0000 112 0
593
+ # 0.8644 0.4353 0.8644 0.5791 9968 19791
594
+ G 0.0000 0.0000 0.0000 0.0000 112 0
595
+ _ 0.0323 0.4153 0.0323 0.0599 8736 679
596
+ + 0.5104 0.6767 0.5104 0.5819 6272 4730
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
+ Overall Accuracy: 0.5186 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
+ A 0.0000 0.0000 0.0000 0.0000 326 0
609
+ # 0.9193 0.4804 0.9193 0.6311 31948 61131
610
+ G 0.0000 0.0000 0.0000 0.0000 326 0
611
+ _ 0.0000 0.0000 0.0000 0.0000 22494 0
612
+ + 0.4748 0.7094 0.4748 0.5689 18256 12219
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
+ Overall Accuracy: 0.7407 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
+ A 0.0000 0.0000 0.0000 0.0000 86 0
625
+ # 0.0048 0.3067 0.0048 0.0094 4816 75
626
+ G 0.0000 0.0000 0.0000 0.0000 86 0
627
+ _ 0.9964 0.7424 0.9964 0.8508 14362 19275
628
+ + 0.0000 0.0000 0.0000 0.0000 0 0
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
+ Overall Accuracy: 0.6175 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
+ A 0.0000 0.0000 0.0000 0.0000 87 0
641
+ # 0.1604 0.3149 0.1604 0.2126 6177 3147
642
+ G 0.0000 0.0000 0.0000 0.0000 87 0
643
+ _ 0.8391 0.6754 0.8391 0.7484 13224 16428
644
+ + 0.0000 0.0000 0.0000 0.0000 0 0
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
+ Overall Accuracy: 0.5097 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
+ A 0.0000 0.0000 0.0000 0.0000 136 0
657
+ # 0.4394 0.3779 0.4394 0.4063 11560 13444
658
+ G 0.0000 0.0000 0.0000 0.0000 136 0
659
+ _ 0.5604 0.6131 0.5604 0.5856 18768 17156
660
+ + 0.0000 0.0000 0.0000 0.0000 0 0
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
+ Overall Accuracy: 0.4475 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
+ A 0.0000 0.0000 0.0000 0.0000 100 0
673
+ # 0.9699 0.4444 0.9699 0.6095 10000 21824
674
+ G 0.0000 0.0000 0.0000 0.0000 100 0
675
+ _ 0.0300 0.5459 0.0300 0.0569 12300 676
676
+ + 0.0000 0.0000 0.0000 0.0000 0 0
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
+ Overall Accuracy: 0.5067 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
+ A 0.0000 0.0000 0.0000 0.0000 164 0
689
+ # 1.0000 0.5067 1.0000 0.6726 18696 36900
690
+ G 0.0000 0.0000 0.0000 0.0000 164 0
691
+ _ 0.0000 0.0000 0.0000 0.0000 17876 0
692
+ + 0.0000 0.0000 0.0000 0.0000 0 0
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
+ Overall Accuracy: 0.5689 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
+ A 0.0000 0.0000 0.0000 0.0000 391 0
705
+ # 1.0000 0.5689 1.0000 0.7252 50048 87975
706
+ G 0.0000 0.0000 0.0000 0.0000 391 0
707
+ _ 0.0000 0.0000 0.0000 0.0000 37145 0
708
+ + 0.0000 0.0000 0.0000 0.0000 0 0
709
  ---------------------------------------------------------------------------------------
710
 
711
+ Results saved to: reveng/cognitive_map_probes_results/layer23/lr_general/pre_reasoning/eval_cognitive_map_probe_layer23_lr_pre_reasoning_all_general.json
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.4888894548510741,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 649,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.23169124727586032,
16
- "precision": 0.509104070382538,
17
- "recall": 0.23169124727586032,
18
- "f1": 0.31845492067362113,
19
  "gt_support": 39462,
20
- "predicted": 17959
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 649,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.7744181736344621,
32
- "precision": 0.48289582301469375,
33
- "recall": 0.7744181736344621,
34
- "f1": 0.5948606351498389,
35
  "gt_support": 37769,
36
- "predicted": 60570
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.5865337870685464,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 68,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.20588235294117646,
56
- "precision": 0.3305785123966942,
57
- "recall": 0.20588235294117646,
58
- "f1": 0.25373810602628005,
59
  "gt_support": 2720,
60
- "predicted": 1694
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 68,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.7941176470588235,
72
- "precision": 0.6528925619834711,
73
- "recall": 0.7941176470588235,
74
- "f1": 0.7166134721988913,
75
  "gt_support": 5372,
76
- "predicted": 6534
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
- "accuracy": 0.5496353913466213,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 85,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.2202753441802253,
95
- "precision": 0.38852097130242824,
96
- "recall": 0.2202753441802253,
97
- "f1": 0.28115015974440893,
98
  "gt_support": 3995,
99
- "predicted": 2265
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 85,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.7799019607843137,
111
- "precision": 0.5951371571072319,
112
- "recall": 0.7799019607843137,
113
- "f1": 0.6751060820367751,
114
  "gt_support": 6120,
115
- "predicted": 8020
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
- "accuracy": 0.5129977460555973,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 55,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.31457975986277875,
134
- "precision": 0.44001919385796545,
135
- "recall": 0.31457975986277875,
136
- "f1": 0.36687337467493497,
137
  "gt_support": 2915,
138
- "predicted": 2084
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 55,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.6878787878787879,
150
- "precision": 0.5462699628090133,
151
- "recall": 0.6878787878787879,
152
- "f1": 0.6089501280331667,
153
  "gt_support": 3630,
154
- "predicted": 4571
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
- "accuracy": 0.49009066837840004,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 103,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.18171521035598706,
173
- "precision": 0.49888938249666814,
174
- "recall": 0.18171521035598706,
175
- "f1": 0.26639781757798603,
176
  "gt_support": 6180,
177
- "predicted": 2251
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 103,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.820306072075037,
189
- "precision": 0.48815119467293383,
190
- "recall": 0.820306072075037,
191
- "f1": 0.6120694947510591,
192
  "gt_support": 6077,
193
- "predicted": 10212
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
- "accuracy": 0.4552704074235175,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 114,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.14965443912812335,
212
- "precision": 0.5495363591996095,
213
- "recall": 0.14965443912812335,
214
- "f1": 0.23524495978272222,
215
  "gt_support": 7524,
216
- "predicted": 2049
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 114,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.8530287984111221,
228
- "precision": 0.4388250319284802,
229
- "recall": 0.8530287984111221,
230
- "f1": 0.579524371732164,
231
  "gt_support": 6042,
232
- "predicted": 11745
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
- "accuracy": 0.4468344155844156,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 224,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.28131200396825395,
251
- "precision": 0.5957195378151261,
252
- "recall": 0.28131200396825395,
253
- "f1": 0.3821597035040431,
254
  "gt_support": 16128,
255
- "predicted": 7616
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 224,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.7194148936170213,
267
- "precision": 0.3886494252873563,
268
- "recall": 0.7194148936170213,
269
- "f1": 0.5046641791044776,
270
  "gt_support": 10528,
271
- "predicted": 19488
272
  }
273
  },
274
  "total_samples": 27104
@@ -278,13 +278,13 @@
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 11
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7146913878949178,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.386748844375963,
8
+ "precision": 0.13252375923970433,
9
+ "recall": 0.386748844375963,
10
+ "f1": 0.19740464018875342,
11
  "gt_support": 649,
12
+ "predicted": 1894
13
  },
14
  "1": {
15
+ "accuracy": 0.6175814707820182,
16
+ "precision": 0.8468621863923831,
17
+ "recall": 0.6175814707820182,
18
+ "f1": 0.7142731535756155,
19
  "gt_support": 39462,
20
+ "predicted": 28778
21
  },
22
  "2": {
23
+ "accuracy": 0.22342064714946072,
24
+ "precision": 0.0701839303000968,
25
+ "recall": 0.22342064714946072,
26
+ "f1": 0.10681399631675872,
27
  "gt_support": 649,
28
+ "predicted": 2066
29
  },
30
  "3": {
31
+ "accuracy": 0.8302311419418041,
32
+ "precision": 0.6847852198030181,
33
+ "recall": 0.8302311419418041,
34
+ "f1": 0.7505265677357587,
35
  "gt_support": 37769,
36
+ "predicted": 45791
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8737238697131745,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.38235294117647056,
48
+ "precision": 0.12322274881516587,
49
+ "recall": 0.38235294117647056,
50
+ "f1": 0.1863799283154122,
51
  "gt_support": 68,
52
+ "predicted": 211
53
  },
54
  "1": {
55
+ "accuracy": 0.7988970588235295,
56
+ "precision": 0.9342218400687876,
57
+ "recall": 0.7988970588235295,
58
+ "f1": 0.8612762584225129,
59
  "gt_support": 2720,
60
+ "predicted": 2326
61
  },
62
  "2": {
63
+ "accuracy": 0.08823529411764706,
64
+ "precision": 0.07142857142857142,
65
+ "recall": 0.08823529411764706,
66
+ "f1": 0.07894736842105264,
67
  "gt_support": 68,
68
+ "predicted": 84
69
  },
70
  "3": {
71
+ "accuracy": 0.9277736411020104,
72
+ "precision": 0.8888888888888888,
73
+ "recall": 0.9277736411020104,
74
+ "f1": 0.9079151106658166,
75
  "gt_support": 5372,
76
+ "predicted": 5607
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8005833738454059,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.47058823529411764,
87
+ "precision": 0.11494252873563218,
88
+ "recall": 0.47058823529411764,
89
+ "f1": 0.18475750577367206,
90
  "gt_support": 85,
91
+ "predicted": 348
92
  },
93
  "1": {
94
+ "accuracy": 0.6623279098873592,
95
+ "precision": 0.9280954051210102,
96
+ "recall": 0.6623279098873592,
97
+ "f1": 0.7730061349693251,
98
  "gt_support": 3995,
99
+ "predicted": 2851
100
  },
101
  "2": {
102
+ "accuracy": 0.10588235294117647,
103
+ "precision": 0.05232558139534884,
104
+ "recall": 0.10588235294117647,
105
+ "f1": 0.07003891050583658,
106
  "gt_support": 85,
107
+ "predicted": 172
108
  },
109
  "3": {
110
+ "accuracy": 0.9050653594771242,
111
+ "precision": 0.8011281457911484,
112
+ "recall": 0.9050653594771242,
113
+ "f1": 0.8499309498235385,
114
  "gt_support": 6120,
115
+ "predicted": 6914
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
+ "accuracy": 0.7546205860255447,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.2909090909090909,
126
+ "precision": 0.12403100775193798,
127
+ "recall": 0.2909090909090909,
128
+ "f1": 0.17391304347826086,
129
  "gt_support": 55,
130
+ "predicted": 129
131
  },
132
  "1": {
133
+ "accuracy": 0.6253859348198971,
134
+ "precision": 0.8574788334901223,
135
+ "recall": 0.6253859348198971,
136
+ "f1": 0.7232691926205118,
137
  "gt_support": 2915,
138
+ "predicted": 2126
139
  },
140
  "2": {
141
+ "accuracy": 0.21818181818181817,
142
+ "precision": 0.09836065573770492,
143
+ "recall": 0.21818181818181817,
144
+ "f1": 0.13559322033898308,
145
  "gt_support": 55,
146
+ "predicted": 122
147
  },
148
  "3": {
149
+ "accuracy": 0.8735537190082645,
150
+ "precision": 0.741234221598878,
151
+ "recall": 0.8735537190082645,
152
+ "f1": 0.8019726858877086,
153
  "gt_support": 3630,
154
+ "predicted": 4278
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
+ "accuracy": 0.7421166653293749,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.39805825242718446,
165
+ "precision": 0.15891472868217055,
166
+ "recall": 0.39805825242718446,
167
+ "f1": 0.2271468144044321,
168
  "gt_support": 103,
169
+ "predicted": 258
170
  },
171
  "1": {
172
+ "accuracy": 0.6237864077669902,
173
+ "precision": 0.8872266973532796,
174
+ "recall": 0.6237864077669902,
175
+ "f1": 0.7325415676959619,
176
  "gt_support": 6180,
177
+ "predicted": 4345
178
  },
179
  "2": {
180
+ "accuracy": 0.1650485436893204,
181
+ "precision": 0.08542713567839195,
182
+ "recall": 0.1650485436893204,
183
+ "f1": 0.11258278145695363,
184
  "gt_support": 103,
185
+ "predicted": 199
186
  },
187
  "3": {
188
+ "accuracy": 0.8780648346223465,
189
+ "precision": 0.696514815298264,
190
+ "recall": 0.8780648346223465,
191
+ "f1": 0.7768234095210366,
192
  "gt_support": 6077,
193
+ "predicted": 7661
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
+ "accuracy": 0.6923299985500942,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.24561403508771928,
204
+ "precision": 0.14,
205
+ "recall": 0.24561403508771928,
206
+ "f1": 0.17834394904458598,
207
  "gt_support": 114,
208
+ "predicted": 200
209
  },
210
  "1": {
211
+ "accuracy": 0.5804093567251462,
212
+ "precision": 0.8687089715536105,
213
+ "recall": 0.5804093567251462,
214
+ "f1": 0.6958808063102542,
215
  "gt_support": 7524,
216
+ "predicted": 5027
217
  },
218
  "2": {
219
+ "accuracy": 0.23684210526315788,
220
+ "precision": 0.07541899441340782,
221
+ "recall": 0.23684210526315788,
222
+ "f1": 0.11440677966101695,
223
  "gt_support": 114,
224
+ "predicted": 358
225
  },
226
  "3": {
227
+ "accuracy": 0.8487255875537901,
228
+ "precision": 0.6246802290169327,
229
+ "recall": 0.8487255875537901,
230
+ "f1": 0.7196687951722687,
231
  "gt_support": 6042,
232
+ "predicted": 8209
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
+ "accuracy": 0.6227863046044865,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.44642857142857145,
243
+ "precision": 0.13368983957219252,
244
+ "recall": 0.44642857142857145,
245
+ "f1": 0.205761316872428,
246
  "gt_support": 224,
247
+ "predicted": 748
248
  },
249
  "1": {
250
+ "accuracy": 0.5894717261904762,
251
+ "precision": 0.7855077253573494,
252
+ "recall": 0.5894717261904762,
253
+ "f1": 0.6735149303956643,
254
  "gt_support": 16128,
255
+ "predicted": 12103
256
  },
257
  "2": {
258
+ "accuracy": 0.33035714285714285,
259
+ "precision": 0.06542882404951371,
260
+ "recall": 0.33035714285714285,
261
+ "f1": 0.10922509225092251,
262
  "gt_support": 224,
263
+ "predicted": 1131
264
  },
265
  "3": {
266
+ "accuracy": 0.6837955927051672,
267
+ "precision": 0.5486206370980033,
268
+ "recall": 0.6837955927051672,
269
+ "f1": 0.6087949260042284,
270
  "gt_support": 10528,
271
+ "predicted": 13122
272
  }
273
  },
274
  "total_samples": 27104
 
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 11
290
  }
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 649 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.4889 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 649 0
30
- # 0.2317 0.5091 0.2317 0.3185 39462 17959
31
- G 0.0000 0.0000 0.0000 0.0000 649 0
32
- _ 0.7744 0.4829 0.7744 0.5949 37769 60570
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.5865 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 68 0
49
- # 0.2059 0.3306 0.2059 0.2537 2720 1694
50
- G 0.0000 0.0000 0.0000 0.0000 68 0
51
- _ 0.7941 0.6529 0.7941 0.7166 5372 6534
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.5496 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 85 0
64
- # 0.2203 0.3885 0.2203 0.2812 3995 2265
65
- G 0.0000 0.0000 0.0000 0.0000 85 0
66
- _ 0.7799 0.5951 0.7799 0.6751 6120 8020
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5130 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 55 0
79
- # 0.3146 0.4400 0.3146 0.3669 2915 2084
80
- G 0.0000 0.0000 0.0000 0.0000 55 0
81
- _ 0.6879 0.5463 0.6879 0.6090 3630 4571
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.4901 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 103 0
94
- # 0.1817 0.4989 0.1817 0.2664 6180 2251
95
- G 0.0000 0.0000 0.0000 0.0000 103 0
96
- _ 0.8203 0.4882 0.8203 0.6121 6077 10212
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.4553 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 114 0
109
- # 0.1497 0.5495 0.1497 0.2352 7524 2049
110
- G 0.0000 0.0000 0.0000 0.0000 114 0
111
- _ 0.8530 0.4388 0.8530 0.5795 6042 11745
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.4468 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 224 0
124
- # 0.2813 0.5957 0.2813 0.3822 16128 7616
125
- G 0.0000 0.0000 0.0000 0.0000 224 0
126
- _ 0.7194 0.3886 0.7194 0.5047 10528 19488
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.7147 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.3867 0.1325 0.3867 0.1974 649 1894
30
+ # 0.6176 0.8469 0.6176 0.7143 39462 28778
31
+ G 0.2234 0.0702 0.2234 0.1068 649 2066
32
+ _ 0.8302 0.6848 0.8302 0.7505 37769 45791
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8737 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.3824 0.1232 0.3824 0.1864 68 211
49
+ # 0.7989 0.9342 0.7989 0.8613 2720 2326
50
+ G 0.0882 0.0714 0.0882 0.0789 68 84
51
+ _ 0.9278 0.8889 0.9278 0.9079 5372 5607
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8006 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.4706 0.1149 0.4706 0.1848 85 348
64
+ # 0.6623 0.9281 0.6623 0.7730 3995 2851
65
+ G 0.1059 0.0523 0.1059 0.0700 85 172
66
+ _ 0.9051 0.8011 0.9051 0.8499 6120 6914
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7546 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.2909 0.1240 0.2909 0.1739 55 129
79
+ # 0.6254 0.8575 0.6254 0.7233 2915 2126
80
+ G 0.2182 0.0984 0.2182 0.1356 55 122
81
+ _ 0.8736 0.7412 0.8736 0.8020 3630 4278
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.7421 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.3981 0.1589 0.3981 0.2271 103 258
94
+ # 0.6238 0.8872 0.6238 0.7325 6180 4345
95
+ G 0.1650 0.0854 0.1650 0.1126 103 199
96
+ _ 0.8781 0.6965 0.8781 0.7768 6077 7661
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6923 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.2456 0.1400 0.2456 0.1783 114 200
109
+ # 0.5804 0.8687 0.5804 0.6959 7524 5027
110
+ G 0.2368 0.0754 0.2368 0.1144 114 358
111
+ _ 0.8487 0.6247 0.8487 0.7197 6042 8209
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.6228 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.4464 0.1337 0.4464 0.2058 224 748
124
+ # 0.5895 0.7855 0.5895 0.6735 16128 12103
125
+ G 0.3304 0.0654 0.3304 0.1092 224 1131
126
+ _ 0.6838 0.5486 0.6838 0.6088 10528 13122
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size11.json
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5422090729783038,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 810,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.5211659300682481,
16
- "precision": 0.5367964697685221,
17
- "recall": 0.5211659300682481,
18
- "f1": 0.5288657356205236,
19
  "gt_support": 66522,
20
- "predicted": 64585
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 810,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.5753476464769884,
32
- "precision": 0.547043772906438,
33
- "recall": 0.5753476464769884,
34
- "f1": 0.5608388336299122,
35
  "gt_support": 68748,
36
- "predicted": 72305
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.6916605029585798,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 64,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.15755208333333334,
56
- "precision": 0.43564356435643564,
57
- "recall": 0.15755208333333334,
58
- "f1": 0.23141286158259627,
59
  "gt_support": 3072,
60
- "predicted": 1111
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 64,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.9187237394957983,
72
- "precision": 0.7209685729005667,
73
- "recall": 0.9187237394957983,
74
- "f1": 0.807921020726286,
75
  "gt_support": 7616,
76
- "predicted": 9705
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
- "accuracy": 0.6025004771903035,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 62,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.3729143492769744,
95
- "precision": 0.4243670886075949,
96
- "recall": 0.3729143492769744,
97
- "f1": 0.39698046181172286,
98
  "gt_support": 3596,
99
- "predicted": 3160
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 62,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.7357206274045576,
111
- "precision": 0.6794206067231484,
112
- "recall": 0.7357206274045576,
113
- "f1": 0.7064506962205173,
114
  "gt_support": 6758,
115
- "predicted": 7318
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
- "accuracy": 0.5406390532544378,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 125,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.4510588235294118,
134
- "precision": 0.44003213588890167,
135
- "recall": 0.4510588235294118,
136
- "f1": 0.44547725556265616,
137
  "gt_support": 8500,
138
- "predicted": 8713
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 125,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.6130909090909091,
150
- "precision": 0.6112632935868514,
151
- "recall": 0.6130909090909091,
152
- "f1": 0.6121757372816395,
153
  "gt_support": 12375,
154
- "predicted": 12412
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
- "accuracy": 0.5158687466379774,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 121,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.45136681500317866,
173
- "precision": 0.48206404888536836,
174
- "recall": 0.45136681500317866,
175
- "f1": 0.46621067031463753,
176
  "gt_support": 9438,
177
- "predicted": 8837
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 121,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.5839910855232612,
189
- "precision": 0.5415949018256976,
190
- "recall": 0.5839910855232612,
191
- "f1": 0.5619945489477682,
192
  "gt_support": 10769,
193
- "predicted": 11612
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
- "accuracy": 0.5100380388841927,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 112,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.5027086677367576,
212
- "precision": 0.5436693067158511,
213
- "recall": 0.5027086677367576,
214
- "f1": 0.5223872817305186,
215
  "gt_support": 9968,
216
- "predicted": 9217
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 112,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.5314789377289377,
228
- "precision": 0.4781175985995263,
229
- "recall": 0.5314789377289377,
230
- "f1": 0.5033880847834336,
231
  "gt_support": 8736,
232
- "predicted": 9711
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
- "accuracy": 0.5228337023995353,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 326,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.6178477525979718,
251
- "precision": 0.5883983664709214,
252
- "recall": 0.6178477525979718,
253
- "f1": 0.6027635697381479,
254
  "gt_support": 31948,
255
- "predicted": 33547
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 326,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.4030408108829021,
267
- "precision": 0.4207546294147677,
268
- "recall": 0.4030408108829021,
269
- "f1": 0.41170727276855656,
270
  "gt_support": 22494,
271
- "predicted": 21547
272
  }
273
  },
274
  "total_samples": 55094
@@ -278,13 +278,13 @@
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 13
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.6934545985828037,
4
  "baseline_accuracy": 0.5022134560596099,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.4160493827160494,
8
+ "precision": 0.1067469116249604,
9
+ "recall": 0.4160493827160494,
10
+ "f1": 0.16990168893370305,
11
  "gt_support": 810,
12
+ "predicted": 3157
13
  },
14
  "1": {
15
+ "accuracy": 0.613135504043775,
16
+ "precision": 0.7808814519834584,
17
+ "recall": 0.613135504043775,
18
+ "f1": 0.6869158091516917,
19
  "gt_support": 66522,
20
+ "predicted": 52232
21
  },
22
  "2": {
23
+ "accuracy": 0.32839506172839505,
24
+ "precision": 0.06131857999077916,
25
+ "recall": 0.32839506172839505,
26
+ "f1": 0.10334110334110334,
27
  "gt_support": 810,
28
+ "predicted": 4338
29
  },
30
  "3": {
31
+ "accuracy": 0.7787426543317624,
32
+ "precision": 0.693816984824333,
33
+ "recall": 0.7787426543317624,
34
+ "f1": 0.7338308969166136,
35
  "gt_support": 68748,
36
+ "predicted": 77163
37
  }
38
  },
39
  "total_samples": 136890
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8410687869822485,
44
  "baseline_accuracy": 0.7041420118343196,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.375,
48
+ "precision": 0.09486166007905138,
49
+ "recall": 0.375,
50
+ "f1": 0.15141955835962143,
51
  "gt_support": 64,
52
+ "predicted": 253
53
  },
54
  "1": {
55
+ "accuracy": 0.7320963541666666,
56
+ "precision": 0.8464433571697403,
57
+ "recall": 0.7320963541666666,
58
+ "f1": 0.7851282946412987,
59
  "gt_support": 3072,
60
+ "predicted": 2657
61
  },
62
  "2": {
63
+ "accuracy": 0.234375,
64
+ "precision": 0.061224489795918366,
65
+ "recall": 0.234375,
66
+ "f1": 0.0970873786407767,
67
  "gt_support": 64,
68
+ "predicted": 245
69
  },
70
  "3": {
71
+ "accuracy": 0.8940388655462185,
72
+ "precision": 0.8887873645738155,
73
+ "recall": 0.8940388655462185,
74
+ "f1": 0.8914053806375598,
75
  "gt_support": 7616,
76
+ "predicted": 7661
77
  }
78
  },
79
  "total_samples": 10816
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8365146020232869,
83
  "baseline_accuracy": 0.6449704142011834,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.3548387096774194,
87
+ "precision": 0.10837438423645321,
88
+ "recall": 0.3548387096774194,
89
+ "f1": 0.16603773584905662,
90
  "gt_support": 62,
91
+ "predicted": 203
92
  },
93
  "1": {
94
+ "accuracy": 0.756952169076752,
95
+ "precision": 0.8769329896907216,
96
+ "recall": 0.756952169076752,
97
+ "f1": 0.8125373134328359,
98
  "gt_support": 3596,
99
+ "predicted": 3104
100
  },
101
  "2": {
102
+ "accuracy": 0.2903225806451613,
103
+ "precision": 0.06293706293706294,
104
+ "recall": 0.2903225806451613,
105
+ "f1": 0.10344827586206898,
106
  "gt_support": 62,
107
+ "predicted": 286
108
  },
109
  "3": {
110
+ "accuracy": 0.8882805563776265,
111
+ "precision": 0.8718954248366013,
112
+ "recall": 0.8882805563776265,
113
+ "f1": 0.8800117276258888,
114
  "gt_support": 6758,
115
+ "predicted": 6885
116
  }
117
  },
118
  "total_samples": 10478
119
  },
120
  "0.4": {
121
+ "accuracy": 0.7854674556213018,
122
  "baseline_accuracy": 0.5857988165680473,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.424,
126
+ "precision": 0.12589073634204276,
127
+ "recall": 0.424,
128
+ "f1": 0.19413919413919414,
129
  "gt_support": 125,
130
+ "predicted": 421
131
  },
132
  "1": {
133
+ "accuracy": 0.6316470588235295,
134
+ "precision": 0.8875847247478922,
135
+ "recall": 0.6316470588235295,
136
+ "f1": 0.7380575984603754,
137
  "gt_support": 8500,
138
+ "predicted": 6049
139
  },
140
  "2": {
141
+ "accuracy": 0.296,
142
+ "precision": 0.08314606741573034,
143
+ "recall": 0.296,
144
+ "f1": 0.1298245614035088,
145
  "gt_support": 125,
146
+ "predicted": 445
147
  },
148
  "3": {
149
+ "accuracy": 0.8997171717171717,
150
+ "precision": 0.7835327234342012,
151
+ "recall": 0.8997171717171717,
152
+ "f1": 0.8376151965394019,
153
  "gt_support": 12375,
154
+ "predicted": 14210
155
  }
156
  },
157
  "total_samples": 21125
158
  },
159
  "0.6": {
160
+ "accuracy": 0.7262457821898382,
161
  "baseline_accuracy": 0.5266272189349113,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.4049586776859504,
165
+ "precision": 0.12039312039312039,
166
+ "recall": 0.4049586776859504,
167
+ "f1": 0.18560606060606064,
168
  "gt_support": 121,
169
+ "predicted": 407
170
  },
171
  "1": {
172
+ "accuracy": 0.5770290315744862,
173
+ "precision": 0.8615725359911407,
174
+ "recall": 0.5770290315744862,
175
+ "f1": 0.6911606066374771,
176
  "gt_support": 9438,
177
+ "predicted": 6321
178
  },
179
  "2": {
180
+ "accuracy": 0.4049586776859504,
181
+ "precision": 0.07060518731988473,
182
+ "recall": 0.4049586776859504,
183
+ "f1": 0.12024539877300613,
184
  "gt_support": 121,
185
+ "predicted": 694
186
  },
187
  "3": {
188
+ "accuracy": 0.8642399479988857,
189
+ "precision": 0.71443924157519,
190
+ "recall": 0.8642399479988857,
191
+ "f1": 0.7822323079509161,
192
  "gt_support": 10769,
193
+ "predicted": 13027
194
  }
195
  },
196
  "total_samples": 20449
197
  },
198
  "0.8": {
199
+ "accuracy": 0.636094674556213,
200
  "baseline_accuracy": 0.5266272189349113,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.45535714285714285,
204
+ "precision": 0.10647181628392484,
205
+ "recall": 0.45535714285714285,
206
+ "f1": 0.17258883248730963,
207
  "gt_support": 112,
208
+ "predicted": 479
209
  },
210
  "1": {
211
+ "accuracy": 0.5965088282504013,
212
+ "precision": 0.7266283759012587,
213
+ "recall": 0.5965088282504013,
214
+ "f1": 0.655170514021266,
215
  "gt_support": 9968,
216
+ "predicted": 8183
217
  },
218
  "2": {
219
+ "accuracy": 0.36607142857142855,
220
+ "precision": 0.07130434782608695,
221
+ "recall": 0.36607142857142855,
222
+ "f1": 0.11935953420669576,
223
  "gt_support": 112,
224
+ "predicted": 575
225
  },
226
  "3": {
227
+ "accuracy": 0.6870421245421245,
228
+ "precision": 0.6193375296667011,
229
+ "recall": 0.6870421245421245,
230
+ "f1": 0.6514353937157433,
231
  "gt_support": 8736,
232
+ "predicted": 9691
233
  }
234
  },
235
  "total_samples": 18928
236
  },
237
  "1.0": {
238
+ "accuracy": 0.6095219080117618,
239
  "baseline_accuracy": 0.5798816568047337,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.4233128834355828,
243
+ "precision": 0.09899569583931134,
244
+ "recall": 0.4233128834355828,
245
+ "f1": 0.16046511627906979,
246
  "gt_support": 326,
247
+ "predicted": 1394
248
  },
249
  "1": {
250
+ "accuracy": 0.5964379616877425,
251
+ "precision": 0.7352033335905548,
252
+ "recall": 0.5964379616877425,
253
+ "f1": 0.6585905367573359,
254
  "gt_support": 31948,
255
+ "predicted": 25918
256
  },
257
  "2": {
258
+ "accuracy": 0.32515337423312884,
259
+ "precision": 0.050645007166746296,
260
+ "recall": 0.32515337423312884,
261
+ "f1": 0.08763952046300123,
262
  "gt_support": 326,
263
+ "predicted": 2093
264
  },
265
  "3": {
266
+ "accuracy": 0.6349248688539166,
267
+ "precision": 0.5559578029506793,
268
+ "recall": 0.6349248688539166,
269
+ "f1": 0.5928231949027666,
270
  "gt_support": 22494,
271
+ "predicted": 25689
272
  }
273
  },
274
  "total_samples": 55094
 
278
  "total_steps": 810,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size13",
283
  "activations_dir": "interp/activations_test_full/size13",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 13
290
  }
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 810 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.5422 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 810 0
30
- # 0.5212 0.5368 0.5212 0.5289 66522 64585
31
- G 0.0000 0.0000 0.0000 0.0000 810 0
32
- _ 0.5753 0.5470 0.5753 0.5608 68748 72305
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.6917 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 64 0
49
- # 0.1576 0.4356 0.1576 0.2314 3072 1111
50
- G 0.0000 0.0000 0.0000 0.0000 64 0
51
- _ 0.9187 0.7210 0.9187 0.8079 7616 9705
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.6025 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 62 0
64
- # 0.3729 0.4244 0.3729 0.3970 3596 3160
65
- G 0.0000 0.0000 0.0000 0.0000 62 0
66
- _ 0.7357 0.6794 0.7357 0.7065 6758 7318
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.5406 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 125 0
79
- # 0.4511 0.4400 0.4511 0.4455 8500 8713
80
- G 0.0000 0.0000 0.0000 0.0000 125 0
81
- _ 0.6131 0.6113 0.6131 0.6122 12375 12412
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.5159 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 121 0
94
- # 0.4514 0.4821 0.4514 0.4662 9438 8837
95
- G 0.0000 0.0000 0.0000 0.0000 121 0
96
- _ 0.5840 0.5416 0.5840 0.5620 10769 11612
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.5100 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 112 0
109
- # 0.5027 0.5437 0.5027 0.5224 9968 9217
110
- G 0.0000 0.0000 0.0000 0.0000 112 0
111
- _ 0.5315 0.4781 0.5315 0.5034 8736 9711
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.5228 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 326 0
124
- # 0.6178 0.5884 0.6178 0.6028 31948 33547
125
- G 0.0000 0.0000 0.0000 0.0000 326 0
126
- _ 0.4030 0.4208 0.4030 0.4117 22494 21547
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.6935 (Baseline: 0.5022, 136890 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.4160 0.1067 0.4160 0.1699 810 3157
30
+ # 0.6131 0.7809 0.6131 0.6869 66522 52232
31
+ G 0.3284 0.0613 0.3284 0.1033 810 4338
32
+ _ 0.7787 0.6938 0.7787 0.7338 68748 77163
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8411 (Baseline: 0.7041, 10816 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.3750 0.0949 0.3750 0.1514 64 253
49
+ # 0.7321 0.8464 0.7321 0.7851 3072 2657
50
+ G 0.2344 0.0612 0.2344 0.0971 64 245
51
+ _ 0.8940 0.8888 0.8940 0.8914 7616 7661
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8365 (Baseline: 0.6450, 10478 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.3548 0.1084 0.3548 0.1660 62 203
64
+ # 0.7570 0.8769 0.7570 0.8125 3596 3104
65
+ G 0.2903 0.0629 0.2903 0.1034 62 286
66
+ _ 0.8883 0.8719 0.8883 0.8800 6758 6885
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7855 (Baseline: 0.5858, 21125 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.4240 0.1259 0.4240 0.1941 125 421
79
+ # 0.6316 0.8876 0.6316 0.7381 8500 6049
80
+ G 0.2960 0.0831 0.2960 0.1298 125 445
81
+ _ 0.8997 0.7835 0.8997 0.8376 12375 14210
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.7262 (Baseline: 0.5266, 20449 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.4050 0.1204 0.4050 0.1856 121 407
94
+ # 0.5770 0.8616 0.5770 0.6912 9438 6321
95
+ G 0.4050 0.0706 0.4050 0.1202 121 694
96
+ _ 0.8642 0.7144 0.8642 0.7822 10769 13027
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6361 (Baseline: 0.5266, 18928 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.4554 0.1065 0.4554 0.1726 112 479
109
+ # 0.5965 0.7266 0.5965 0.6552 9968 8183
110
+ G 0.3661 0.0713 0.3661 0.1194 112 575
111
+ _ 0.6870 0.6193 0.6870 0.6514 8736 9691
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.6095 (Baseline: 0.5799, 55094 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.4233 0.0990 0.4233 0.1605 326 1394
124
+ # 0.5964 0.7352 0.5964 0.6586 31948 25918
125
+ G 0.3252 0.0506 0.3252 0.0876 326 2093
126
+ _ 0.6349 0.5560 0.6349 0.5928 22494 25689
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size13.json
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.46370216689718763,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 964,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.9369477872000158,
16
- "precision": 0.46356807236565756,
17
- "recall": 0.9369477872000158,
18
- "f1": 0.620255853088699,
19
  "gt_support": 101297,
20
- "predicted": 204738
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 964,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.049852650098966354,
32
- "precision": 0.4659595461272817,
33
- "recall": 0.049852650098966354,
34
- "f1": 0.09006889865460874,
35
  "gt_support": 113675,
36
- "predicted": 12162
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.26036175710594317,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 86,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.9767441860465116,
56
- "precision": 0.24888888888888888,
57
- "recall": 0.9767441860465116,
58
- "f1": 0.396694214876033,
59
  "gt_support": 4816,
60
- "predicted": 18900
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 86,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.023255813953488372,
72
- "precision": 0.7422222222222222,
73
- "recall": 0.023255813953488372,
74
- "f1": 0.04509856872805833,
75
  "gt_support": 14362,
76
- "predicted": 450
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
- "accuracy": 0.3215836526181354,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 87,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.9898008742107819,
95
- "precision": 0.3163450095721012,
96
- "recall": 0.9898008742107819,
97
- "f1": 0.4794542032622333,
98
  "gt_support": 6177,
99
- "predicted": 19327
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 87,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.013687235329703569,
111
- "precision": 0.7298387096774194,
112
- "recall": 0.013687235329703569,
113
- "f1": 0.026870546318289786,
114
  "gt_support": 13224,
115
- "predicted": 248
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
- "accuracy": 0.38222222222222224,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 136,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.9878027681660899,
134
- "precision": 0.37837569170615326,
135
- "recall": 0.9878027681660899,
136
- "f1": 0.5471621265483121,
137
  "gt_support": 11560,
138
- "predicted": 30179
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 136,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.01475916453537937,
150
- "precision": 0.6579572446555819,
151
- "recall": 0.01475916453537937,
152
- "f1": 0.028870707175986242,
153
  "gt_support": 18768,
154
- "predicted": 421
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
- "accuracy": 0.4457333333333333,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 100,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.9855,
173
- "precision": 0.44433923982145274,
174
- "recall": 0.9855,
175
- "f1": 0.6125112651107866,
176
  "gt_support": 10000,
177
- "predicted": 22179
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 100,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.014146341463414635,
189
- "precision": 0.5420560747663551,
190
- "recall": 0.014146341463414635,
191
- "f1": 0.027573092464939386,
192
  "gt_support": 12300,
193
- "predicted": 321
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
- "accuracy": 0.5063414634146342,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 164,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.9475823705605477,
212
- "precision": 0.5071132102476027,
213
- "recall": 0.9475823705605477,
214
- "f1": 0.6606626764371353,
215
  "gt_support": 18696,
216
- "predicted": 34935
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 164,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.05415081673752517,
228
- "precision": 0.4926208651399491,
229
- "recall": 0.05415081673752517,
230
- "f1": 0.09757572702988761,
231
  "gt_support": 17876,
232
- "predicted": 1965
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
- "accuracy": 0.555100880932083,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 391,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.9011748721227621,
251
- "precision": 0.5693403014466409,
252
- "recall": 0.9011748721227621,
253
- "f1": 0.6978169046771773,
254
  "gt_support": 50048,
255
- "predicted": 79218
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 391,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.10049804818952753,
267
- "precision": 0.42628754139545505,
268
- "recall": 0.10049804818952753,
269
- "f1": 0.16265086488606162,
270
  "gt_support": 37145,
271
- "predicted": 8757
272
  }
273
  },
274
  "total_samples": 87975
@@ -278,13 +278,13 @@
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 15
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.674817888427847,
4
  "baseline_accuracy": 0.5240894421392347,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.33402489626556015,
8
+ "precision": 0.0720196823976739,
9
+ "recall": 0.33402489626556015,
10
+ "f1": 0.11849126034958601,
11
  "gt_support": 964,
12
+ "predicted": 4471
13
  },
14
  "1": {
15
+ "accuracy": 0.7249474318094317,
16
+ "precision": 0.6655880940080304,
17
+ "recall": 0.7249474318094317,
18
+ "f1": 0.6940007938458049,
19
  "gt_support": 101297,
20
+ "predicted": 110331
21
  },
22
  "2": {
23
+ "accuracy": 0.22717842323651452,
24
+ "precision": 0.05547112462006079,
25
+ "recall": 0.22717842323651452,
26
+ "f1": 0.08916938110749185,
27
  "gt_support": 964,
28
+ "predicted": 3948
29
  },
30
  "3": {
31
+ "accuracy": 0.6368330767539037,
32
+ "precision": 0.7375649516046867,
33
+ "recall": 0.6368330767539037,
34
+ "f1": 0.6835076124159094,
35
  "gt_support": 113675,
36
+ "predicted": 98150
37
  }
38
  },
39
  "total_samples": 216900
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.8611369509043928,
44
  "baseline_accuracy": 0.7422222222222222,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.4069767441860465,
48
+ "precision": 0.08728179551122195,
49
+ "recall": 0.4069767441860465,
50
+ "f1": 0.1437371663244353,
51
  "gt_support": 86,
52
+ "predicted": 401
53
  },
54
  "1": {
55
+ "accuracy": 0.7230066445182725,
56
+ "precision": 0.8862305930262153,
57
+ "recall": 0.7230066445182725,
58
+ "f1": 0.7963407661520869,
59
  "gt_support": 4816,
60
+ "predicted": 3929
61
  },
62
  "2": {
63
+ "accuracy": 0.3953488372093023,
64
+ "precision": 0.06007067137809187,
65
+ "recall": 0.3953488372093023,
66
+ "f1": 0.10429447852760736,
67
  "gt_support": 86,
68
+ "predicted": 566
69
  },
70
  "3": {
71
+ "accuracy": 0.9129647681381423,
72
+ "precision": 0.9071537290715372,
73
+ "recall": 0.9129647681381423,
74
+ "f1": 0.9100499722376457,
75
  "gt_support": 14362,
76
+ "predicted": 14454
77
  }
78
  },
79
  "total_samples": 19350
80
  },
81
  "0.2": {
82
+ "accuracy": 0.8364240102171137,
83
  "baseline_accuracy": 0.6755555555555556,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.25287356321839083,
87
+ "precision": 0.06321839080459771,
88
+ "recall": 0.25287356321839083,
89
+ "f1": 0.10114942528735633,
90
  "gt_support": 87,
91
+ "predicted": 348
92
  },
93
  "1": {
94
+ "accuracy": 0.6841508823053262,
95
+ "precision": 0.8891226593730276,
96
+ "recall": 0.6841508823053262,
97
+ "f1": 0.773284537968893,
98
  "gt_support": 6177,
99
+ "predicted": 4753
100
  },
101
  "2": {
102
+ "accuracy": 0.25287356321839083,
103
+ "precision": 0.05405405405405406,
104
+ "recall": 0.25287356321839083,
105
+ "f1": 0.08906882591093118,
106
  "gt_support": 87,
107
+ "predicted": 407
108
  },
109
  "3": {
110
+ "accuracy": 0.9152298850574713,
111
+ "precision": 0.8603824553920523,
112
+ "recall": 0.9152298850574713,
113
+ "f1": 0.8869590707559268,
114
  "gt_support": 13224,
115
+ "predicted": 14067
116
  }
117
  },
118
  "total_samples": 19575
119
  },
120
  "0.4": {
121
+ "accuracy": 0.772124183006536,
122
  "baseline_accuracy": 0.6133333333333333,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.375,
126
+ "precision": 0.08056872037914692,
127
+ "recall": 0.375,
128
+ "f1": 0.13263979193758127,
129
  "gt_support": 136,
130
+ "predicted": 633
131
  },
132
  "1": {
133
+ "accuracy": 0.5525951557093426,
134
+ "precision": 0.8984528832630099,
135
+ "recall": 0.5525951557093426,
136
+ "f1": 0.6843063738618104,
137
  "gt_support": 11560,
138
+ "predicted": 7110
139
  },
140
  "2": {
141
+ "accuracy": 0.3161764705882353,
142
+ "precision": 0.06231884057971015,
143
+ "recall": 0.3161764705882353,
144
+ "f1": 0.10411622276029056,
145
  "gt_support": 136,
146
+ "predicted": 690
147
  },
148
  "3": {
149
+ "accuracy": 0.9135230179028133,
150
+ "precision": 0.7734470158343484,
151
+ "recall": 0.9135230179028133,
152
+ "f1": 0.8376694759985344,
153
  "gt_support": 18768,
154
+ "predicted": 22167
155
  }
156
  },
157
  "total_samples": 30600
158
  },
159
  "0.6": {
160
+ "accuracy": 0.6992888888888888,
161
  "baseline_accuracy": 0.5466666666666666,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.34,
165
+ "precision": 0.07188160676532769,
166
+ "recall": 0.34,
167
+ "f1": 0.11867364746945899,
168
  "gt_support": 100,
169
+ "predicted": 473
170
  },
171
  "1": {
172
+ "accuracy": 0.5682,
173
+ "precision": 0.7686688311688312,
174
+ "recall": 0.5682,
175
+ "f1": 0.6534038638454462,
176
  "gt_support": 10000,
177
+ "predicted": 7392
178
  },
179
  "2": {
180
+ "accuracy": 0.23,
181
+ "precision": 0.047520661157024795,
182
+ "recall": 0.23,
183
+ "f1": 0.07876712328767124,
184
  "gt_support": 100,
185
+ "predicted": 484
186
  },
187
  "3": {
188
+ "accuracy": 0.8126016260162602,
189
+ "precision": 0.7063105080913009,
190
+ "recall": 0.8126016260162602,
191
+ "f1": 0.7557370231749273,
192
  "gt_support": 12300,
193
+ "predicted": 14151
194
  }
195
  },
196
  "total_samples": 22500
197
  },
198
  "0.8": {
199
+ "accuracy": 0.6090514905149051,
200
  "baseline_accuracy": 0.5066666666666667,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.3475609756097561,
204
+ "precision": 0.06312292358803986,
205
+ "recall": 0.3475609756097561,
206
+ "f1": 0.10684161199625117,
207
  "gt_support": 164,
208
+ "predicted": 903
209
  },
210
  "1": {
211
+ "accuracy": 0.6822314933675653,
212
+ "precision": 0.6259815469179427,
213
+ "recall": 0.6822314933675653,
214
+ "f1": 0.6528972153972153,
215
  "gt_support": 18696,
216
+ "predicted": 20376
217
  },
218
  "2": {
219
+ "accuracy": 0.17073170731707318,
220
+ "precision": 0.05982905982905983,
221
+ "recall": 0.17073170731707318,
222
+ "f1": 0.08860759493670886,
223
  "gt_support": 164,
224
+ "predicted": 468
225
  },
226
  "3": {
227
+ "accuracy": 0.5389348847616916,
228
+ "precision": 0.6357816933940473,
229
+ "recall": 0.5389348847616916,
230
+ "f1": 0.5833661327923945,
231
  "gt_support": 17876,
232
+ "predicted": 15153
233
  }
234
  },
235
  "total_samples": 36900
236
  },
237
  "1.0": {
238
+ "accuracy": 0.585359477124183,
239
  "baseline_accuracy": 0.5688888888888889,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.3145780051150895,
243
+ "precision": 0.07180385288966724,
244
+ "recall": 0.3145780051150895,
245
+ "f1": 0.11692015209125474,
246
  "gt_support": 391,
247
+ "predicted": 1713
248
  },
249
  "1": {
250
+ "accuracy": 0.8172554347826086,
251
+ "precision": 0.6125713258750056,
252
+ "recall": 0.8172554347826086,
253
+ "f1": 0.7002627997157996,
254
  "gt_support": 50048,
255
+ "predicted": 66771
256
  },
257
  "2": {
258
+ "accuracy": 0.17647058823529413,
259
+ "precision": 0.051762940735183796,
260
+ "recall": 0.17647058823529413,
261
+ "f1": 0.08004640371229697,
262
  "gt_support": 391,
263
+ "predicted": 1333
264
  },
265
  "3": {
266
+ "accuracy": 0.2800646116570198,
267
+ "precision": 0.5729155193303227,
268
+ "recall": 0.2800646116570198,
269
+ "f1": 0.3762182883387881,
270
  "gt_support": 37145,
271
+ "predicted": 18158
272
  }
273
  },
274
  "total_samples": 87975
 
278
  "total_steps": 964,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size15",
283
  "activations_dir": "interp/activations_test_full/size15",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 15
290
  }
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 964 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.4637 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 964 0
30
- # 0.9369 0.4636 0.9369 0.6203 101297 204738
31
- G 0.0000 0.0000 0.0000 0.0000 964 0
32
- _ 0.0499 0.4660 0.0499 0.0901 113675 12162
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.2604 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 86 0
49
- # 0.9767 0.2489 0.9767 0.3967 4816 18900
50
- G 0.0000 0.0000 0.0000 0.0000 86 0
51
- _ 0.0233 0.7422 0.0233 0.0451 14362 450
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.3216 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 87 0
64
- # 0.9898 0.3163 0.9898 0.4795 6177 19327
65
- G 0.0000 0.0000 0.0000 0.0000 87 0
66
- _ 0.0137 0.7298 0.0137 0.0269 13224 248
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.3822 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 136 0
79
- # 0.9878 0.3784 0.9878 0.5472 11560 30179
80
- G 0.0000 0.0000 0.0000 0.0000 136 0
81
- _ 0.0148 0.6580 0.0148 0.0289 18768 421
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.4457 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 100 0
94
- # 0.9855 0.4443 0.9855 0.6125 10000 22179
95
- G 0.0000 0.0000 0.0000 0.0000 100 0
96
- _ 0.0141 0.5421 0.0141 0.0276 12300 321
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.5063 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 164 0
109
- # 0.9476 0.5071 0.9476 0.6607 18696 34935
110
- G 0.0000 0.0000 0.0000 0.0000 164 0
111
- _ 0.0542 0.4926 0.0542 0.0976 17876 1965
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.5551 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 391 0
124
- # 0.9012 0.5693 0.9012 0.6978 50048 79218
125
- G 0.0000 0.0000 0.0000 0.0000 391 0
126
- _ 0.1005 0.4263 0.1005 0.1627 37145 8757
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.6748 (Baseline: 0.5241, 216900 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.3340 0.0720 0.3340 0.1185 964 4471
30
+ # 0.7249 0.6656 0.7249 0.6940 101297 110331
31
+ G 0.2272 0.0555 0.2272 0.0892 964 3948
32
+ _ 0.6368 0.7376 0.6368 0.6835 113675 98150
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.8611 (Baseline: 0.7422, 19350 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.4070 0.0873 0.4070 0.1437 86 401
49
+ # 0.7230 0.8862 0.7230 0.7963 4816 3929
50
+ G 0.3953 0.0601 0.3953 0.1043 86 566
51
+ _ 0.9130 0.9072 0.9130 0.9100 14362 14454
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.8364 (Baseline: 0.6756, 19575 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.2529 0.0632 0.2529 0.1011 87 348
64
+ # 0.6842 0.8891 0.6842 0.7733 6177 4753
65
+ G 0.2529 0.0541 0.2529 0.0891 87 407
66
+ _ 0.9152 0.8604 0.9152 0.8870 13224 14067
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.7721 (Baseline: 0.6133, 30600 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.3750 0.0806 0.3750 0.1326 136 633
79
+ # 0.5526 0.8985 0.5526 0.6843 11560 7110
80
+ G 0.3162 0.0623 0.3162 0.1041 136 690
81
+ _ 0.9135 0.7734 0.9135 0.8377 18768 22167
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.6993 (Baseline: 0.5467, 22500 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.3400 0.0719 0.3400 0.1187 100 473
94
+ # 0.5682 0.7687 0.5682 0.6534 10000 7392
95
+ G 0.2300 0.0475 0.2300 0.0788 100 484
96
+ _ 0.8126 0.7063 0.8126 0.7557 12300 14151
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.6091 (Baseline: 0.5067, 36900 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.3476 0.0631 0.3476 0.1068 164 903
109
+ # 0.6822 0.6260 0.6822 0.6529 18696 20376
110
+ G 0.1707 0.0598 0.1707 0.0886 164 468
111
+ _ 0.5389 0.6358 0.5389 0.5834 17876 15153
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.5854 (Baseline: 0.5689, 87975 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.3146 0.0718 0.3146 0.1169 391 1713
124
+ # 0.8173 0.6126 0.8173 0.7003 50048 66771
125
+ G 0.1765 0.0518 0.1765 0.0800 391 1333
126
+ _ 0.2801 0.5729 0.2801 0.3762 37145 18158
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size15.json
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.39203653489367773,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 286,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.09800583657587549,
16
- "precision": 0.5790229885057471,
17
- "recall": 0.09800583657587549,
18
- "f1": 0.16763727121464225,
19
  "gt_support": 8224,
20
- "predicted": 1392
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 286,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.8984285166730548,
32
- "precision": 0.3714149897005229,
33
- "recall": 0.8984285166730548,
34
- "f1": 0.525560538116592,
35
  "gt_support": 5218,
36
- "predicted": 12622
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.47207303974221265,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 38,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.13157894736842105,
56
- "precision": 0.4897959183673469,
57
- "recall": 0.13157894736842105,
58
- "f1": 0.20743301642178044,
59
  "gt_support": 912,
60
- "predicted": 245
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 38,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.868421052631579,
72
- "precision": 0.46938775510204084,
73
- "recall": 0.868421052631579,
74
- "f1": 0.6093938177438779,
75
  "gt_support": 874,
76
- "predicted": 1617
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
- "accuracy": 0.4359925788497217,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 44,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.06905594405594405,
95
- "precision": 0.5337837837837838,
96
- "recall": 0.06905594405594405,
97
- "f1": 0.12229102167182662,
98
  "gt_support": 1144,
99
- "predicted": 148
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 44,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.9318181818181818,
111
- "precision": 0.42878486055776893,
112
- "recall": 0.9318181818181818,
113
- "f1": 0.58731241473397,
114
  "gt_support": 924,
115
- "predicted": 2008
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
- "accuracy": 0.4235938277750124,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 41,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.1951219512195122,
134
- "precision": 0.5714285714285714,
135
- "recall": 0.1951219512195122,
136
- "f1": 0.29090909090909095,
137
  "gt_support": 1148,
138
- "predicted": 392
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 41,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.8048780487804879,
150
- "precision": 0.3877551020408163,
151
- "recall": 0.8048780487804879,
152
- "f1": 0.5233722871452421,
153
  "gt_support": 779,
154
- "predicted": 1617
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
- "accuracy": 0.3816760746851932,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 47,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.06382978723404255,
173
- "precision": 0.5918367346938775,
174
- "recall": 0.06382978723404255,
175
- "f1": 0.1152317880794702,
176
  "gt_support": 1363,
177
- "predicted": 147
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 47,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.9361702127659575,
189
- "precision": 0.3673469387755102,
190
- "recall": 0.9361702127659575,
191
- "f1": 0.5276482345103265,
192
  "gt_support": 846,
193
- "predicted": 2156
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
- "accuracy": 0.34879406307977734,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 55,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.07272727272727272,
212
- "precision": 0.6326530612244898,
213
- "recall": 0.07272727272727272,
214
- "f1": 0.1304576538663861,
215
  "gt_support": 1705,
216
- "predicted": 196
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 55,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.9272727272727272,
228
- "precision": 0.32653061224489793,
229
- "recall": 0.9272727272727272,
230
- "f1": 0.4829831311038769,
231
  "gt_support": 880,
232
- "predicted": 2499
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
- "accuracy": 0.3362328537972566,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 61,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.08811475409836066,
251
- "precision": 0.6515151515151515,
252
- "recall": 0.08811475409836066,
253
- "f1": 0.1552346570397112,
254
  "gt_support": 1952,
255
- "predicted": 264
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 61,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.9103825136612022,
267
- "precision": 0.3056880733944954,
268
- "recall": 0.9103825136612022,
269
- "f1": 0.4576923076923076,
270
  "gt_support": 915,
271
- "predicted": 2725
272
  }
273
  },
274
  "total_samples": 2989
@@ -278,13 +278,13 @@
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 7
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.8741972313400885,
4
  "baseline_accuracy": 0.5868417296988726,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.9090909090909091,
8
+ "precision": 0.6951871657754011,
9
+ "recall": 0.9090909090909091,
10
+ "f1": 0.7878787878787878,
11
  "gt_support": 286,
12
+ "predicted": 374
13
  },
14
  "1": {
15
+ "accuracy": 0.8453307392996109,
16
+ "precision": 0.958235699517574,
17
+ "recall": 0.8453307392996109,
18
+ "f1": 0.8982492409070353,
19
  "gt_support": 8224,
20
+ "predicted": 7255
21
  },
22
  "2": {
23
+ "accuracy": 0.8706293706293706,
24
+ "precision": 0.6766304347826086,
25
+ "recall": 0.8706293706293706,
26
+ "f1": 0.7614678899082569,
27
  "gt_support": 286,
28
+ "predicted": 368
29
  },
30
  "3": {
31
+ "accuracy": 0.9179762361057877,
32
+ "precision": 0.7960777796243975,
33
+ "recall": 0.9179762361057877,
34
+ "f1": 0.8526924788607032,
35
  "gt_support": 5218,
36
+ "predicted": 6017
37
  }
38
  },
39
  "total_samples": 14014
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.9833512352309345,
44
  "baseline_accuracy": 0.4897959183673469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.8947368421052632,
48
+ "precision": 0.8095238095238095,
49
+ "recall": 0.8947368421052632,
50
+ "f1": 0.8500000000000001,
51
  "gt_support": 38,
52
+ "predicted": 42
53
  },
54
  "1": {
55
+ "accuracy": 0.9945175438596491,
56
+ "precision": 0.9967032967032967,
57
+ "recall": 0.9945175438596491,
58
+ "f1": 0.995609220636663,
59
  "gt_support": 912,
60
+ "predicted": 910
61
  },
62
  "2": {
63
+ "accuracy": 0.9210526315789473,
64
+ "precision": 0.7777777777777778,
65
+ "recall": 0.9210526315789473,
66
+ "f1": 0.8433734939759038,
67
  "gt_support": 38,
68
+ "predicted": 45
69
  },
70
  "3": {
71
+ "accuracy": 0.9782608695652174,
72
+ "precision": 0.9884393063583815,
73
+ "recall": 0.9782608695652174,
74
+ "f1": 0.9833237492811961,
75
  "gt_support": 874,
76
+ "predicted": 865
77
  }
78
  },
79
  "total_samples": 1862
80
  },
81
  "0.2": {
82
+ "accuracy": 0.9318181818181818,
83
  "baseline_accuracy": 0.5306122448979592,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.9772727272727273,
87
+ "precision": 0.6935483870967742,
88
+ "recall": 0.9772727272727273,
89
+ "f1": 0.8113207547169812,
90
  "gt_support": 44,
91
+ "predicted": 62
92
  },
93
  "1": {
94
+ "accuracy": 0.9230769230769231,
95
+ "precision": 0.9688073394495413,
96
+ "recall": 0.9230769230769231,
97
+ "f1": 0.945389435989257,
98
  "gt_support": 1144,
99
+ "predicted": 1090
100
  },
101
  "2": {
102
+ "accuracy": 0.9090909090909091,
103
+ "precision": 0.8,
104
+ "recall": 0.9090909090909091,
105
+ "f1": 0.8510638297872342,
106
  "gt_support": 44,
107
+ "predicted": 50
108
  },
109
  "3": {
110
+ "accuracy": 0.9415584415584416,
111
+ "precision": 0.9119496855345912,
112
+ "recall": 0.9415584415584416,
113
+ "f1": 0.9265175718849841,
114
  "gt_support": 924,
115
+ "predicted": 954
116
  }
117
  },
118
  "total_samples": 2156
119
  },
120
  "0.4": {
121
+ "accuracy": 0.8929815828770533,
122
  "baseline_accuracy": 0.5714285714285714,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.9024390243902439,
126
+ "precision": 0.6491228070175439,
127
+ "recall": 0.9024390243902439,
128
+ "f1": 0.7551020408163265,
129
  "gt_support": 41,
130
+ "predicted": 57
131
  },
132
  "1": {
133
+ "accuracy": 0.8571428571428571,
134
+ "precision": 0.983016983016983,
135
+ "recall": 0.8571428571428571,
136
+ "f1": 0.9157747789669612,
137
  "gt_support": 1148,
138
+ "predicted": 1001
139
  },
140
  "2": {
141
+ "accuracy": 0.9512195121951219,
142
+ "precision": 0.6842105263157895,
143
+ "recall": 0.9512195121951219,
144
+ "f1": 0.7959183673469389,
145
  "gt_support": 41,
146
+ "predicted": 57
147
  },
148
  "3": {
149
+ "accuracy": 0.9422336328626444,
150
+ "precision": 0.8210290827740492,
151
+ "recall": 0.9422336328626444,
152
+ "f1": 0.8774656306037059,
153
  "gt_support": 779,
154
+ "predicted": 894
155
  }
156
  },
157
  "total_samples": 2009
158
  },
159
  "0.6": {
160
+ "accuracy": 0.8597481545809813,
161
  "baseline_accuracy": 0.5918367346938775,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.9361702127659575,
165
+ "precision": 0.7096774193548387,
166
+ "recall": 0.9361702127659575,
167
+ "f1": 0.8073394495412844,
168
  "gt_support": 47,
169
+ "predicted": 62
170
  },
171
  "1": {
172
+ "accuracy": 0.8327219369038885,
173
+ "precision": 0.9442595673876872,
174
+ "recall": 0.8327219369038885,
175
+ "f1": 0.884990253411306,
176
  "gt_support": 1363,
177
+ "predicted": 1202
178
  },
179
  "2": {
180
+ "accuracy": 0.8936170212765957,
181
+ "precision": 0.6666666666666666,
182
+ "recall": 0.8936170212765957,
183
+ "f1": 0.7636363636363636,
184
  "gt_support": 47,
185
+ "predicted": 63
186
  },
187
  "3": {
188
+ "accuracy": 0.8971631205673759,
189
+ "precision": 0.7776639344262295,
190
+ "recall": 0.8971631205673759,
191
+ "f1": 0.8331503841931943,
192
  "gt_support": 846,
193
+ "predicted": 976
194
  }
195
  },
196
  "total_samples": 2303
197
  },
198
  "0.8": {
199
+ "accuracy": 0.8330241187384044,
200
  "baseline_accuracy": 0.6326530612244898,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.9272727272727272,
204
+ "precision": 0.6891891891891891,
205
+ "recall": 0.9272727272727272,
206
+ "f1": 0.7906976744186047,
207
  "gt_support": 55,
208
+ "predicted": 74
209
  },
210
  "1": {
211
+ "accuracy": 0.8023460410557185,
212
+ "precision": 0.9440993788819876,
213
+ "recall": 0.8023460410557185,
214
+ "f1": 0.8674698795180723,
215
  "gt_support": 1705,
216
+ "predicted": 1449
217
  },
218
  "2": {
219
+ "accuracy": 0.8727272727272727,
220
+ "precision": 0.6,
221
+ "recall": 0.8727272727272727,
222
+ "f1": 0.711111111111111,
223
  "gt_support": 55,
224
+ "predicted": 80
225
  },
226
  "3": {
227
+ "accuracy": 0.884090909090909,
228
+ "precision": 0.7124542124542125,
229
+ "recall": 0.884090909090909,
230
+ "f1": 0.7890466531440163,
231
  "gt_support": 880,
232
+ "predicted": 1092
233
  }
234
  },
235
  "total_samples": 2695
236
  },
237
  "1.0": {
238
+ "accuracy": 0.8002676480428237,
239
  "baseline_accuracy": 0.6530612244897959,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.8360655737704918,
243
+ "precision": 0.6623376623376623,
244
+ "recall": 0.8360655737704918,
245
+ "f1": 0.7391304347826089,
246
  "gt_support": 61,
247
+ "predicted": 77
248
  },
249
  "1": {
250
+ "accuracy": 0.7694672131147541,
251
+ "precision": 0.9369931378665003,
252
+ "recall": 0.7694672131147541,
253
+ "f1": 0.8450070323488045,
254
  "gt_support": 1952,
255
+ "predicted": 1603
256
  },
257
  "2": {
258
+ "accuracy": 0.7377049180327869,
259
+ "precision": 0.6164383561643836,
260
+ "recall": 0.7377049180327869,
261
+ "f1": 0.6716417910447762,
262
  "gt_support": 61,
263
+ "predicted": 73
264
  },
265
  "3": {
266
+ "accuracy": 0.8677595628415301,
267
+ "precision": 0.6423948220064725,
268
+ "recall": 0.8677595628415301,
269
+ "f1": 0.7382612738261274,
270
  "gt_support": 915,
271
+ "predicted": 1236
272
  }
273
  },
274
  "total_samples": 2989
 
278
  "total_steps": 286,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size7",
283
  "activations_dir": "interp/activations_test_full/size7",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 7
290
  }
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 286 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.3920 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 286 0
30
- # 0.0980 0.5790 0.0980 0.1676 8224 1392
31
- G 0.0000 0.0000 0.0000 0.0000 286 0
32
- _ 0.8984 0.3714 0.8984 0.5256 5218 12622
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.4721 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 38 0
49
- # 0.1316 0.4898 0.1316 0.2074 912 245
50
- G 0.0000 0.0000 0.0000 0.0000 38 0
51
- _ 0.8684 0.4694 0.8684 0.6094 874 1617
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.4360 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 44 0
64
- # 0.0691 0.5338 0.0691 0.1223 1144 148
65
- G 0.0000 0.0000 0.0000 0.0000 44 0
66
- _ 0.9318 0.4288 0.9318 0.5873 924 2008
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.4236 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 41 0
79
- # 0.1951 0.5714 0.1951 0.2909 1148 392
80
- G 0.0000 0.0000 0.0000 0.0000 41 0
81
- _ 0.8049 0.3878 0.8049 0.5234 779 1617
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.3817 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 47 0
94
- # 0.0638 0.5918 0.0638 0.1152 1363 147
95
- G 0.0000 0.0000 0.0000 0.0000 47 0
96
- _ 0.9362 0.3673 0.9362 0.5276 846 2156
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.3488 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 55 0
109
- # 0.0727 0.6327 0.0727 0.1305 1705 196
110
- G 0.0000 0.0000 0.0000 0.0000 55 0
111
- _ 0.9273 0.3265 0.9273 0.4830 880 2499
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.3362 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 61 0
124
- # 0.0881 0.6515 0.0881 0.1552 1952 264
125
- G 0.0000 0.0000 0.0000 0.0000 61 0
126
- _ 0.9104 0.3057 0.9104 0.4577 915 2725
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.8742 (Baseline: 0.5868, 14014 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.9091 0.6952 0.9091 0.7879 286 374
30
+ # 0.8453 0.9582 0.8453 0.8982 8224 7255
31
+ G 0.8706 0.6766 0.8706 0.7615 286 368
32
+ _ 0.9180 0.7961 0.9180 0.8527 5218 6017
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.9834 (Baseline: 0.4898, 1862 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.8947 0.8095 0.8947 0.8500 38 42
49
+ # 0.9945 0.9967 0.9945 0.9956 912 910
50
+ G 0.9211 0.7778 0.9211 0.8434 38 45
51
+ _ 0.9783 0.9884 0.9783 0.9833 874 865
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.9318 (Baseline: 0.5306, 2156 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.9773 0.6935 0.9773 0.8113 44 62
64
+ # 0.9231 0.9688 0.9231 0.9454 1144 1090
65
+ G 0.9091 0.8000 0.9091 0.8511 44 50
66
+ _ 0.9416 0.9119 0.9416 0.9265 924 954
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.8930 (Baseline: 0.5714, 2009 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.9024 0.6491 0.9024 0.7551 41 57
79
+ # 0.8571 0.9830 0.8571 0.9158 1148 1001
80
+ G 0.9512 0.6842 0.9512 0.7959 41 57
81
+ _ 0.9422 0.8210 0.9422 0.8775 779 894
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.8597 (Baseline: 0.5918, 2303 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.9362 0.7097 0.9362 0.8073 47 62
94
+ # 0.8327 0.9443 0.8327 0.8850 1363 1202
95
+ G 0.8936 0.6667 0.8936 0.7636 47 63
96
+ _ 0.8972 0.7777 0.8972 0.8332 846 976
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.8330 (Baseline: 0.6327, 2695 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.9273 0.6892 0.9273 0.7907 55 74
109
+ # 0.8023 0.9441 0.8023 0.8675 1705 1449
110
+ G 0.8727 0.6000 0.8727 0.7111 55 80
111
+ _ 0.8841 0.7125 0.8841 0.7890 880 1092
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.8003 (Baseline: 0.6531, 2989 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.8361 0.6623 0.8361 0.7391 61 77
124
+ # 0.7695 0.9370 0.7695 0.8450 1952 1603
125
+ G 0.7377 0.6164 0.7377 0.6716 61 73
126
+ _ 0.8678 0.6424 0.8678 0.7383 915 1236
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size7.json
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.45998462794831146,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.0,
8
- "precision": 0.0,
9
- "recall": 0.0,
10
- "f1": 0.0,
11
  "gt_support": 514,
12
- "predicted": 0
13
  },
14
  "1": {
15
- "accuracy": 0.19069254521393914,
16
- "precision": 0.5679936933385888,
17
- "recall": 0.19069254521393914,
18
- "f1": 0.28552557709454773,
19
  "gt_support": 22670,
20
- "predicted": 7611
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 514,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.8267172167707404,
32
- "precision": 0.4358228257355318,
33
- "recall": 0.8267172167707404,
34
- "f1": 0.5707577128120248,
35
  "gt_support": 17936,
36
- "predicted": 34023
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.5539551897576589,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.0,
48
- "precision": 0.0,
49
- "recall": 0.0,
50
- "f1": 0.0,
51
  "gt_support": 54,
52
- "predicted": 0
53
  },
54
  "1": {
55
- "accuracy": 0.23842592592592593,
56
- "precision": 0.4287200832466181,
57
- "recall": 0.23842592592592593,
58
- "f1": 0.30643361844551875,
59
  "gt_support": 1728,
60
- "predicted": 961
61
  },
62
  "2": {
63
- "accuracy": 0.0,
64
- "precision": 0.0,
65
- "recall": 0.0,
66
- "f1": 0.0,
67
  "gt_support": 54,
68
- "predicted": 0
69
  },
70
  "3": {
71
- "accuracy": 0.7923561859732072,
72
- "precision": 0.5892176970407267,
73
- "recall": 0.7923561859732072,
74
- "f1": 0.675852797849101,
75
  "gt_support": 2538,
76
- "predicted": 3413
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
- "accuracy": 0.527712109272393,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.0,
87
- "precision": 0.0,
88
- "recall": 0.0,
89
- "f1": 0.0,
90
  "gt_support": 47,
91
- "predicted": 0
92
  },
93
  "1": {
94
- "accuracy": 0.17671394799054374,
95
- "precision": 0.4784,
96
- "recall": 0.17671394799054374,
97
- "f1": 0.2580923608113941,
98
  "gt_support": 1692,
99
- "predicted": 625
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 47,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.8461157842652153,
111
- "precision": 0.5373978629792583,
112
- "recall": 0.8461157842652153,
113
- "f1": 0.6573130886027292,
114
  "gt_support": 2021,
115
- "predicted": 3182
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
- "accuracy": 0.4895890915791413,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.0,
126
- "precision": 0.0,
127
- "recall": 0.0,
128
- "f1": 0.0,
129
  "gt_support": 67,
130
- "predicted": 0
131
  },
132
  "1": {
133
- "accuracy": 0.18880597014925374,
134
- "precision": 0.5090543259557344,
135
- "recall": 0.18880597014925374,
136
- "f1": 0.2754491017964072,
137
  "gt_support": 2680,
138
- "predicted": 994
139
  },
140
  "2": {
141
- "accuracy": 0.0,
142
- "precision": 0.0,
143
- "recall": 0.0,
144
- "f1": 0.0,
145
  "gt_support": 67,
146
- "predicted": 0
147
  },
148
  "3": {
149
- "accuracy": 0.8231917336394948,
150
- "precision": 0.48522445296638844,
151
- "recall": 0.8231917336394948,
152
- "f1": 0.610559182514902,
153
  "gt_support": 2613,
154
- "predicted": 4433
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
- "accuracy": 0.46124266342845577,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.0,
165
- "precision": 0.0,
166
- "recall": 0.0,
167
- "f1": 0.0,
168
  "gt_support": 61,
169
- "predicted": 0
170
  },
171
  "1": {
172
- "accuracy": 0.14182234083110942,
173
- "precision": 0.5494830132939439,
174
- "recall": 0.14182234083110942,
175
- "f1": 0.22545454545454544,
176
  "gt_support": 2623,
177
- "predicted": 677
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 61,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.8683970856102003,
189
- "precision": 0.4472326454033771,
190
- "recall": 0.8683970856102003,
191
- "f1": 0.5904024767801858,
192
  "gt_support": 2196,
193
- "predicted": 4264
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
- "accuracy": 0.43979953550910644,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.0,
204
- "precision": 0.0,
205
- "recall": 0.0,
206
- "f1": 0.0,
207
  "gt_support": 101,
208
- "predicted": 0
209
  },
210
  "1": {
211
- "accuracy": 0.20834211080682535,
212
- "precision": 0.5986682808716707,
213
- "recall": 0.20834211080682535,
214
- "f1": 0.3091107985622753,
215
  "gt_support": 4747,
216
- "predicted": 1652
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 101,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.807240099009901,
228
- "precision": 0.39960177668862,
229
- "recall": 0.807240099009901,
230
- "f1": 0.5345763753713758,
231
  "gt_support": 3232,
232
- "predicted": 6529
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
- "accuracy": 0.4149892646269458,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.0,
243
- "precision": 0.0,
244
- "recall": 0.0,
245
- "f1": 0.0,
246
  "gt_support": 184,
247
- "predicted": 0
248
  },
249
  "1": {
250
- "accuracy": 0.18967391304347825,
251
- "precision": 0.6458179126572909,
252
- "recall": 0.18967391304347825,
253
- "f1": 0.2932280289027054,
254
  "gt_support": 9200,
255
- "predicted": 2702
256
  },
257
  "2": {
258
- "accuracy": 0.0,
259
- "precision": 0.0,
260
- "recall": 0.0,
261
- "f1": 0.0,
262
  "gt_support": 184,
263
- "predicted": 0
264
  },
265
  "3": {
266
- "accuracy": 0.8320839580209896,
267
- "precision": 0.3638747746271103,
268
- "recall": 0.8320839580209896,
269
- "f1": 0.5063291139240507,
270
  "gt_support": 5336,
271
- "predicted": 12202
272
  }
273
  },
274
  "total_samples": 14904
@@ -278,13 +278,13 @@
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 9
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7963683527885862,
4
  "baseline_accuracy": 0.544506893404429,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.6498054474708171,
8
+ "precision": 0.33669354838709675,
9
+ "recall": 0.6498054474708171,
10
+ "f1": 0.4435590969455511,
11
  "gt_support": 514,
12
+ "predicted": 992
13
  },
14
  "1": {
15
+ "accuracy": 0.7436259373621527,
16
+ "precision": 0.9181917211328976,
17
+ "recall": 0.7436259373621527,
18
+ "f1": 0.8217401901048015,
19
  "gt_support": 22670,
20
+ "predicted": 18360
21
  },
22
  "2": {
23
+ "accuracy": 0.5972762645914397,
24
+ "precision": 0.2712014134275618,
25
+ "recall": 0.5972762645914397,
26
+ "f1": 0.3730255164034021,
27
  "gt_support": 514,
28
+ "predicted": 1132
29
  },
30
  "3": {
31
+ "accuracy": 0.8729371097234612,
32
+ "precision": 0.7402836879432624,
33
+ "recall": 0.8729371097234612,
34
+ "f1": 0.8011564242951441,
35
  "gt_support": 17936,
36
+ "predicted": 21150
37
  }
38
  },
39
  "total_samples": 41634
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.9286694101508917,
44
  "baseline_accuracy": 0.5802469135802469,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.5555555555555556,
48
+ "precision": 0.2459016393442623,
49
+ "recall": 0.5555555555555556,
50
+ "f1": 0.3409090909090909,
51
  "gt_support": 54,
52
+ "predicted": 122
53
  },
54
  "1": {
55
+ "accuracy": 0.9577546296296297,
56
+ "precision": 0.981031416716064,
57
+ "recall": 0.9577546296296297,
58
+ "f1": 0.9692532942898975,
59
  "gt_support": 1728,
60
+ "predicted": 1687
61
  },
62
  "2": {
63
+ "accuracy": 0.5370370370370371,
64
+ "precision": 0.23015873015873015,
65
+ "recall": 0.5370370370370371,
66
+ "f1": 0.32222222222222224,
67
  "gt_support": 54,
68
+ "predicted": 126
69
  },
70
  "3": {
71
+ "accuracy": 0.9251379038613081,
72
+ "precision": 0.9626896268962689,
73
+ "recall": 0.9251379038613081,
74
+ "f1": 0.9435402853124372,
75
  "gt_support": 2538,
76
+ "predicted": 2439
77
  }
78
  },
79
  "total_samples": 4374
80
  },
81
  "0.2": {
82
+ "accuracy": 0.9014972419227738,
83
  "baseline_accuracy": 0.5308641975308642,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.7021276595744681,
87
+ "precision": 0.4125,
88
+ "recall": 0.7021276595744681,
89
+ "f1": 0.5196850393700788,
90
  "gt_support": 47,
91
+ "predicted": 80
92
  },
93
  "1": {
94
+ "accuracy": 0.8729314420803782,
95
+ "precision": 0.9666230366492147,
96
+ "recall": 0.8729314420803782,
97
+ "f1": 0.917391304347826,
98
  "gt_support": 1692,
99
+ "predicted": 1528
100
  },
101
  "2": {
102
+ "accuracy": 0.48936170212765956,
103
+ "precision": 0.32857142857142857,
104
+ "recall": 0.48936170212765956,
105
+ "f1": 0.39316239316239315,
106
  "gt_support": 47,
107
+ "predicted": 70
108
  },
109
  "3": {
110
+ "accuracy": 0.9396338446313706,
111
+ "precision": 0.891968060122123,
112
+ "recall": 0.9396338446313706,
113
+ "f1": 0.9151807228915663,
114
  "gt_support": 2021,
115
+ "predicted": 2129
116
  }
117
  },
118
  "total_samples": 3807
119
  },
120
  "0.4": {
121
+ "accuracy": 0.8586696148885203,
122
  "baseline_accuracy": 0.49382716049382713,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.6716417910447762,
126
+ "precision": 0.4368932038834951,
127
+ "recall": 0.6716417910447762,
128
+ "f1": 0.5294117647058824,
129
  "gt_support": 67,
130
+ "predicted": 103
131
  },
132
  "1": {
133
+ "accuracy": 0.7951492537313433,
134
+ "precision": 0.9492204899777282,
135
+ "recall": 0.7951492537313433,
136
+ "f1": 0.8653807106598985,
137
  "gt_support": 2680,
138
+ "predicted": 2245
139
  },
140
  "2": {
141
+ "accuracy": 0.6865671641791045,
142
+ "precision": 0.46938775510204084,
143
+ "recall": 0.6865671641791045,
144
+ "f1": 0.5575757575757576,
145
  "gt_support": 67,
146
+ "predicted": 98
147
  },
148
  "3": {
149
+ "accuracy": 0.933027171833142,
150
+ "precision": 0.8178463602817846,
151
+ "recall": 0.933027171833142,
152
+ "f1": 0.8716481944941007,
153
  "gt_support": 2613,
154
+ "predicted": 2981
155
  }
156
  },
157
  "total_samples": 5427
158
  },
159
  "0.6": {
160
+ "accuracy": 0.8008500303582271,
161
  "baseline_accuracy": 0.5308641975308642,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.7049180327868853,
165
+ "precision": 0.40186915887850466,
166
+ "recall": 0.7049180327868853,
167
+ "f1": 0.5119047619047619,
168
  "gt_support": 61,
169
+ "predicted": 107
170
  },
171
  "1": {
172
+ "accuracy": 0.7354174609226077,
173
+ "precision": 0.9242932438907523,
174
+ "recall": 0.7354174609226077,
175
+ "f1": 0.819108280254777,
176
  "gt_support": 2623,
177
+ "predicted": 2087
178
  },
179
  "2": {
180
+ "accuracy": 0.6557377049180327,
181
+ "precision": 0.32,
182
+ "recall": 0.6557377049180327,
183
+ "f1": 0.4301075268817205,
184
  "gt_support": 61,
185
+ "predicted": 125
186
  },
187
  "3": {
188
+ "accuracy": 0.8857012750455373,
189
+ "precision": 0.7418001525553013,
190
+ "recall": 0.8857012750455373,
191
+ "f1": 0.8073889580738896,
192
  "gt_support": 2196,
193
+ "predicted": 2622
194
  }
195
  },
196
  "total_samples": 4941
197
  },
198
  "0.8": {
199
+ "accuracy": 0.760298252047427,
200
  "baseline_accuracy": 0.5802469135802469,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.6534653465346535,
204
+ "precision": 0.3055555555555556,
205
+ "recall": 0.6534653465346535,
206
+ "f1": 0.416403785488959,
207
  "gt_support": 101,
208
+ "predicted": 216
209
  },
210
  "1": {
211
+ "accuracy": 0.6958078786602064,
212
+ "precision": 0.9213389121338912,
213
+ "recall": 0.6958078786602064,
214
+ "f1": 0.7928468554968794,
215
  "gt_support": 4747,
216
+ "predicted": 3585
217
  },
218
  "2": {
219
+ "accuracy": 0.5247524752475248,
220
+ "precision": 0.2054263565891473,
221
+ "recall": 0.5247524752475248,
222
+ "f1": 0.29526462395543174,
223
  "gt_support": 101,
224
+ "predicted": 258
225
  },
226
  "3": {
227
+ "accuracy": 0.8657178217821783,
228
+ "precision": 0.6787967006307618,
229
+ "recall": 0.8657178217821783,
230
+ "f1": 0.7609464237149852,
231
  "gt_support": 3232,
232
+ "predicted": 4122
233
  }
234
  },
235
  "total_samples": 8181
236
  },
237
  "1.0": {
238
+ "accuracy": 0.7263150831991412,
239
  "baseline_accuracy": 0.6172839506172839,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.6358695652173914,
243
+ "precision": 0.32142857142857145,
244
+ "recall": 0.6358695652173914,
245
+ "f1": 0.4270072992700731,
246
  "gt_support": 184,
247
+ "predicted": 364
248
  },
249
  "1": {
250
+ "accuracy": 0.6916304347826087,
251
+ "precision": 0.8803265080243498,
252
+ "recall": 0.6916304347826087,
253
+ "f1": 0.7746530314097881,
254
  "gt_support": 9200,
255
+ "predicted": 7228
256
  },
257
  "2": {
258
+ "accuracy": 0.6304347826086957,
259
+ "precision": 0.2549450549450549,
260
+ "recall": 0.6304347826086957,
261
+ "f1": 0.3630672926447574,
262
  "gt_support": 184,
263
+ "predicted": 455
264
  },
265
  "3": {
266
+ "accuracy": 0.7925412293853074,
267
+ "precision": 0.6167420154586554,
268
+ "recall": 0.7925412293853074,
269
+ "f1": 0.6936766997457557,
270
  "gt_support": 5336,
271
+ "predicted": 6857
272
  }
273
  },
274
  "total_samples": 14904
 
278
  "total_steps": 514,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size9",
283
  "activations_dir": "interp/activations_test_full/size9",
284
  "layers": "23",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 9
290
  }
layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 514 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.4600 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.0000 0.0000 0.0000 0.0000 514 0
30
- # 0.1907 0.5680 0.1907 0.2855 22670 7611
31
- G 0.0000 0.0000 0.0000 0.0000 514 0
32
- _ 0.8267 0.4358 0.8267 0.5708 17936 34023
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.5540 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.0000 0.0000 0.0000 0.0000 54 0
49
- # 0.2384 0.4287 0.2384 0.3064 1728 961
50
- G 0.0000 0.0000 0.0000 0.0000 54 0
51
- _ 0.7924 0.5892 0.7924 0.6759 2538 3413
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.5277 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.0000 0.0000 0.0000 0.0000 47 0
64
- # 0.1767 0.4784 0.1767 0.2581 1692 625
65
- G 0.0000 0.0000 0.0000 0.0000 47 0
66
- _ 0.8461 0.5374 0.8461 0.6573 2021 3182
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.4896 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.0000 0.0000 0.0000 0.0000 67 0
79
- # 0.1888 0.5091 0.1888 0.2754 2680 994
80
- G 0.0000 0.0000 0.0000 0.0000 67 0
81
- _ 0.8232 0.4852 0.8232 0.6106 2613 4433
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.4612 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.0000 0.0000 0.0000 0.0000 61 0
94
- # 0.1418 0.5495 0.1418 0.2255 2623 677
95
- G 0.0000 0.0000 0.0000 0.0000 61 0
96
- _ 0.8684 0.4472 0.8684 0.5904 2196 4264
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.4398 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.0000 0.0000 0.0000 0.0000 101 0
109
- # 0.2083 0.5987 0.2083 0.3091 4747 1652
110
- G 0.0000 0.0000 0.0000 0.0000 101 0
111
- _ 0.8072 0.3996 0.8072 0.5346 3232 6529
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.4150 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.0000 0.0000 0.0000 0.0000 184 0
124
- # 0.1897 0.6458 0.1897 0.2932 9200 2702
125
- G 0.0000 0.0000 0.0000 0.0000 184 0
126
- _ 0.8321 0.3639 0.8321 0.5063 5336 12202
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.7964 (Baseline: 0.5445, 41634 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.6498 0.3367 0.6498 0.4436 514 992
30
+ # 0.7436 0.9182 0.7436 0.8217 22670 18360
31
+ G 0.5973 0.2712 0.5973 0.3730 514 1132
32
+ _ 0.8729 0.7403 0.8729 0.8012 17936 21150
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.9287 (Baseline: 0.5802, 4374 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.5556 0.2459 0.5556 0.3409 54 122
49
+ # 0.9578 0.9810 0.9578 0.9693 1728 1687
50
+ G 0.5370 0.2302 0.5370 0.3222 54 126
51
+ _ 0.9251 0.9627 0.9251 0.9435 2538 2439
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.9015 (Baseline: 0.5309, 3807 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.7021 0.4125 0.7021 0.5197 47 80
64
+ # 0.8729 0.9666 0.8729 0.9174 1692 1528
65
+ G 0.4894 0.3286 0.4894 0.3932 47 70
66
+ _ 0.9396 0.8920 0.9396 0.9152 2021 2129
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.8587 (Baseline: 0.4938, 5427 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.6716 0.4369 0.6716 0.5294 67 103
79
+ # 0.7951 0.9492 0.7951 0.8654 2680 2245
80
+ G 0.6866 0.4694 0.6866 0.5576 67 98
81
+ _ 0.9330 0.8178 0.9330 0.8716 2613 2981
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.8009 (Baseline: 0.5309, 4941 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.7049 0.4019 0.7049 0.5119 61 107
94
+ # 0.7354 0.9243 0.7354 0.8191 2623 2087
95
+ G 0.6557 0.3200 0.6557 0.4301 61 125
96
+ _ 0.8857 0.7418 0.8857 0.8074 2196 2622
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.7603 (Baseline: 0.5802, 8181 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.6535 0.3056 0.6535 0.4164 101 216
109
+ # 0.6958 0.9213 0.6958 0.7928 4747 3585
110
+ G 0.5248 0.2054 0.5248 0.2953 101 258
111
+ _ 0.8657 0.6788 0.8657 0.7609 3232 4122
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.7263 (Baseline: 0.6173, 14904 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.6359 0.3214 0.6359 0.4270 184 364
124
+ # 0.6916 0.8803 0.6916 0.7747 9200 7228
125
+ G 0.6304 0.2549 0.6304 0.3631 184 455
126
+ _ 0.7925 0.6167 0.7925 0.6937 5336 6857
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_size9.json
layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.json CHANGED
@@ -1,1178 +1,1192 @@
1
  {
2
  "global": {
3
- "accuracy": 0.5123122005033268,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.000930809804529941,
8
- "precision": 0.015873015873015872,
9
- "recall": 0.000930809804529941,
10
- "f1": 0.0017584994138335288,
11
  "gt_support": 3223,
12
- "predicted": 189
13
  },
14
  "1": {
15
- "accuracy": 0.16149469927574261,
16
- "precision": 0.45177354944796805,
17
- "recall": 0.16149469927574261,
18
- "f1": 0.23793514065230503,
19
  "gt_support": 238175,
20
- "predicted": 85140
21
  },
22
  "2": {
23
- "accuracy": 0.0,
24
- "precision": 0.0,
25
- "recall": 0.0,
26
- "f1": 0.0,
27
  "gt_support": 3223,
28
- "predicted": 0
29
  },
30
  "3": {
31
- "accuracy": 0.5685690333927823,
32
- "precision": 0.48665864708199674,
33
- "recall": 0.5685690333927823,
34
- "f1": 0.524434757888752,
35
  "gt_support": 243346,
36
- "predicted": 284304
37
  },
38
  "4": {
39
- "accuracy": 0.8207564668982497,
40
- "precision": 0.5475865017353787,
41
- "recall": 0.8207564668982497,
42
- "f1": 0.6569042598059891,
43
  "gt_support": 237208,
44
- "predicted": 355542
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
- "by_size": {
50
- "7": {
51
- "accuracy": 0.6499922299922299,
52
- "baseline_accuracy": 0.7822222222222223,
53
  "per_class": {
54
  "0": {
55
- "accuracy": 0.0,
56
- "precision": 0.0,
57
- "recall": 0.0,
58
- "f1": 0.0,
59
- "gt_support": 286,
60
- "predicted": 0
61
  },
62
  "1": {
63
- "accuracy": 0.22957198443579765,
64
- "precision": 0.23778337531486146,
65
- "recall": 0.22957198443579765,
66
- "f1": 0.23360554318238058,
67
- "gt_support": 8224,
68
- "predicted": 7940
69
  },
70
  "2": {
71
- "accuracy": 0.0,
72
- "precision": 0.0,
73
- "recall": 0.0,
74
- "f1": 0.0,
75
- "gt_support": 286,
76
- "predicted": 0
77
  },
78
  "3": {
79
- "accuracy": 0.7157914909927175,
80
- "precision": 0.19512067704524083,
81
- "recall": 0.7157914909927175,
82
- "f1": 0.3066502463054187,
83
- "gt_support": 5218,
84
- "predicted": 19142
85
  },
86
  "4": {
87
- "accuracy": 0.7192466624284806,
88
- "precision": 0.97145003756574,
89
- "recall": 0.7192466624284806,
90
- "f1": 0.8265376010227843,
91
- "gt_support": 50336,
92
- "predicted": 37268
93
  }
94
  },
95
- "total_samples": 64350
96
  },
97
- "9": {
98
- "accuracy": 0.6679377431906615,
99
- "baseline_accuracy": 0.64,
100
  "per_class": {
101
  "0": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
- "gt_support": 514,
107
- "predicted": 0
108
  },
109
  "1": {
110
- "accuracy": 0.23281870313189237,
111
- "precision": 0.35439468206539987,
112
- "recall": 0.23281870313189237,
113
- "f1": 0.2810212176876181,
114
- "gt_support": 22670,
115
- "predicted": 14893
116
  },
117
  "2": {
118
- "accuracy": 0.0,
119
- "precision": 0.0,
120
- "recall": 0.0,
121
- "f1": 0.0,
122
- "gt_support": 514,
123
- "predicted": 0
124
  },
125
  "3": {
126
- "accuracy": 0.7106935771632471,
127
- "precision": 0.3527702440914374,
128
- "recall": 0.7106935771632471,
129
- "f1": 0.47149990752727944,
130
- "gt_support": 17936,
131
- "predicted": 36134
132
  },
133
  "4": {
134
- "accuracy": 0.8001242974492002,
135
- "precision": 0.9164229453909599,
136
- "recall": 0.8001242974492002,
137
- "f1": 0.854333917584518,
138
- "gt_support": 74016,
139
- "predicted": 64623
140
  }
141
  },
142
- "total_samples": 115650
143
  },
144
- "11": {
145
- "accuracy": 0.6246738572162301,
146
- "baseline_accuracy": 0.4622222222222222,
147
  "per_class": {
148
  "0": {
149
- "accuracy": 0.0015408320493066256,
150
- "precision": 0.09090909090909091,
151
- "recall": 0.0015408320493066256,
152
- "f1": 0.0030303030303030307,
153
- "gt_support": 649,
154
- "predicted": 11
155
  },
156
  "1": {
157
- "accuracy": 0.19981247782677006,
158
- "precision": 0.4639872896316347,
159
- "recall": 0.19981247782677006,
160
- "f1": 0.2793325775825422,
161
- "gt_support": 39462,
162
- "predicted": 16994
163
  },
164
  "2": {
165
- "accuracy": 0.0,
166
- "precision": 0.0,
167
- "recall": 0.0,
168
- "f1": 0.0,
169
- "gt_support": 649,
170
- "predicted": 0
171
  },
172
  "3": {
173
- "accuracy": 0.698932987370595,
174
- "precision": 0.456034274263207,
175
- "recall": 0.698932987370595,
176
- "f1": 0.5519418744446187,
177
- "gt_support": 37769,
178
- "predicted": 57886
179
  },
180
  "4": {
181
- "accuracy": 0.8435166528386867,
182
- "precision": 0.8003767537323924,
183
- "recall": 0.8435166528386867,
184
- "f1": 0.8213806535381951,
185
- "gt_support": 67496,
186
- "predicted": 71134
187
  }
188
  },
189
- "total_samples": 146025
190
  },
191
- "13": {
192
- "accuracy": 0.526633744855967,
193
- "baseline_accuracy": 0.37721810699588476,
194
  "per_class": {
195
  "0": {
196
- "accuracy": 0.0024691358024691358,
197
- "precision": 0.01652892561983471,
198
- "recall": 0.0024691358024691358,
199
- "f1": 0.004296455424274973,
200
- "gt_support": 810,
201
- "predicted": 121
202
  },
203
  "1": {
204
- "accuracy": 0.15354318871952136,
205
- "precision": 0.513498567191192,
206
- "recall": 0.15354318871952136,
207
- "f1": 0.23639961579854882,
208
- "gt_support": 66522,
209
- "predicted": 19891
210
  },
211
  "2": {
212
- "accuracy": 0.0,
213
- "precision": 0.0,
214
- "recall": 0.0,
215
- "f1": 0.0,
216
- "gt_support": 810,
217
- "predicted": 0
218
  },
219
  "3": {
220
- "accuracy": 0.6317711060685401,
221
- "precision": 0.5376369375502879,
222
- "recall": 0.6317711060685401,
223
- "f1": 0.5809152494767043,
224
- "gt_support": 68748,
225
- "predicted": 80785
226
  },
227
  "4": {
228
- "accuracy": 0.9332010582010583,
229
- "precision": 0.5196861994033369,
230
- "recall": 0.9332010582010583,
231
- "f1": 0.6675971706370799,
232
- "gt_support": 45360,
233
- "predicted": 81453
234
  }
235
  },
236
- "total_samples": 182250
237
  },
238
- "15": {
239
- "accuracy": 0.3008068234209313,
240
- "baseline_accuracy": 0.5240894421392347,
241
  "per_class": {
242
  "0": {
243
- "accuracy": 0.0,
244
- "precision": 0.0,
245
- "recall": 0.0,
246
- "f1": 0.0,
247
- "gt_support": 964,
248
- "predicted": 57
249
  },
250
  "1": {
251
- "accuracy": 0.1303000088847646,
252
- "precision": 0.5191959719927621,
253
- "recall": 0.1303000088847646,
254
- "f1": 0.20831919443808739,
255
- "gt_support": 101297,
256
- "predicted": 25422
257
  },
258
  "2": {
259
- "accuracy": 0.0,
260
- "precision": 0.0,
261
- "recall": 0.0,
262
- "f1": 0.0,
263
- "gt_support": 964,
264
- "predicted": 0
265
  },
266
  "3": {
267
- "accuracy": 0.4578491312953596,
268
- "precision": 0.5760040727337118,
269
- "recall": 0.4578491312953596,
270
- "f1": 0.5101748745294856,
271
- "gt_support": 113675,
272
- "predicted": 90357
273
  },
274
  "4": {
275
- "accuracy": 0.0,
276
- "precision": 0.0,
277
- "recall": 0.0,
278
- "f1": 0.0,
279
- "gt_support": 0,
280
- "predicted": 101064
281
  }
282
  },
283
- "total_samples": 216900
284
- }
285
- },
286
- "by_complexity": {
287
- "0.0": {
288
- "accuracy": 0.5855770609318997,
289
- "baseline_accuracy": 0.4410322580645161,
290
  "per_class": {
291
  "0": {
292
- "accuracy": 0.0,
293
- "precision": 0.0,
294
- "recall": 0.0,
295
- "f1": 0.0,
296
- "gt_support": 310,
297
- "predicted": 0
298
  },
299
  "1": {
300
- "accuracy": 0.20969202898550723,
301
- "precision": 0.29845294370434033,
302
- "recall": 0.20969202898550723,
303
- "f1": 0.24632026955133887,
304
- "gt_support": 13248,
305
- "predicted": 9308
306
  },
307
  "2": {
308
- "accuracy": 0.0,
309
- "precision": 0.0,
310
- "recall": 0.0,
311
- "f1": 0.0,
312
- "gt_support": 310,
313
- "predicted": 0
314
  },
315
  "3": {
316
- "accuracy": 0.6021715103049217,
317
- "precision": 0.6702366307258123,
318
- "recall": 0.6021715103049217,
319
- "f1": 0.6343835616438356,
320
- "gt_support": 30762,
321
- "predicted": 27638
322
  },
323
  "4": {
324
- "accuracy": 0.7779458598726114,
325
- "precision": 0.5957200341421778,
326
- "recall": 0.7779458598726114,
327
- "f1": 0.674746219183758,
328
- "gt_support": 25120,
329
- "predicted": 32804
330
  }
331
  },
332
- "total_samples": 69750
 
 
 
 
 
 
 
 
 
 
 
 
 
333
  },
334
- "0.2": {
335
- "accuracy": 0.5804854700854701,
336
- "baseline_accuracy": 0.3972239316239316,
 
 
 
337
  "per_class": {
338
  "0": {
339
- "accuracy": 0.0,
340
- "precision": 0.0,
341
- "recall": 0.0,
342
- "f1": 0.0,
343
- "gt_support": 325,
344
- "predicted": 0
345
  },
346
  "1": {
347
- "accuracy": 0.14376053962900506,
348
- "precision": 0.32252398324550735,
349
- "recall": 0.14376053962900506,
350
- "f1": 0.19887523432618204,
351
- "gt_support": 16604,
352
- "predicted": 7401
353
  },
354
  "2": {
355
- "accuracy": 0.0,
356
- "precision": 0.0,
357
- "recall": 0.0,
358
- "f1": 0.0,
359
- "gt_support": 325,
360
- "predicted": 0
361
  },
362
  "3": {
363
- "accuracy": 0.6406169311805006,
364
- "precision": 0.5994072928746296,
365
- "recall": 0.6406169311805006,
366
- "f1": 0.6193273535138374,
367
- "gt_support": 29047,
368
- "predicted": 31044
369
  },
370
  "4": {
371
- "accuracy": 0.7997688637041456,
372
- "precision": 0.618598615916955,
373
- "recall": 0.7997688637041456,
374
- "f1": 0.697613163371488,
375
- "gt_support": 26824,
376
- "predicted": 34680
377
  }
378
  },
379
- "total_samples": 73125
380
  },
381
- "0.4": {
382
- "accuracy": 0.5197589098532495,
383
- "baseline_accuracy": 0.4000524109014675,
384
  "per_class": {
385
  "0": {
386
- "accuracy": 0.0,
387
- "precision": 0.0,
388
- "recall": 0.0,
389
- "f1": 0.0,
390
- "gt_support": 424,
391
- "predicted": 0
392
  },
393
  "1": {
394
- "accuracy": 0.13879043390665224,
395
- "precision": 0.379746835443038,
396
- "recall": 0.13879043390665224,
397
- "f1": 0.20328424273887266,
398
- "gt_support": 26803,
399
- "predicted": 9796
400
  },
401
  "2": {
402
- "accuracy": 0.0,
403
- "precision": 0.0,
404
- "recall": 0.0,
405
- "f1": 0.0,
406
- "gt_support": 424,
407
- "predicted": 0
408
  },
409
  "3": {
410
- "accuracy": 0.5594130748067602,
411
- "precision": 0.5653532464781273,
412
- "recall": 0.5594130748067602,
413
- "f1": 0.5623674748778464,
414
- "gt_support": 38165,
415
- "predicted": 37764
416
  },
417
  "4": {
418
- "accuracy": 0.828657382368848,
419
- "precision": 0.5124372909698997,
420
- "recall": 0.828657382368848,
421
- "f1": 0.6332661706964249,
422
- "gt_support": 29584,
423
- "predicted": 47840
424
  }
425
  },
426
- "total_samples": 95400
427
  },
428
- "0.6": {
429
- "accuracy": 0.536738683127572,
430
- "baseline_accuracy": 0.3553909465020576,
431
  "per_class": {
432
  "0": {
433
- "accuracy": 0.0,
434
- "precision": 0.0,
435
- "recall": 0.0,
436
- "f1": 0.0,
437
- "gt_support": 432,
438
- "predicted": 48
439
  },
440
  "1": {
441
- "accuracy": 0.1539994595324956,
442
- "precision": 0.42871920255783336,
443
- "recall": 0.1539994595324956,
444
- "f1": 0.22660171976738405,
445
- "gt_support": 29604,
446
- "predicted": 10634
447
  },
448
  "2": {
449
- "accuracy": 0.0,
450
- "precision": 0.0,
451
- "recall": 0.0,
452
- "f1": 0.0,
453
- "gt_support": 432,
454
- "predicted": 0
455
  },
456
  "3": {
457
- "accuracy": 0.6097303342860694,
458
- "precision": 0.4859364167574527,
459
- "recall": 0.6097303342860694,
460
- "f1": 0.5408399470899471,
461
- "gt_support": 32188,
462
- "predicted": 40388
463
  },
464
  "4": {
465
- "accuracy": 0.8101551644279759,
466
- "precision": 0.6066767830045523,
467
- "recall": 0.8101551644279759,
468
- "f1": 0.6938046954409103,
469
- "gt_support": 34544,
470
- "predicted": 46130
471
  }
472
  },
473
- "total_samples": 97200
474
  },
475
- "0.8": {
476
- "accuracy": 0.5011640211640211,
477
- "baseline_accuracy": 0.3470899470899471,
478
  "per_class": {
479
  "0": {
480
- "accuracy": 0.0,
481
- "precision": 0.0,
482
- "recall": 0.0,
483
- "f1": 0.0,
484
- "gt_support": 546,
485
- "predicted": 0
486
  },
487
  "1": {
488
- "accuracy": 0.1501641651031895,
489
- "precision": 0.4623772385904102,
490
- "recall": 0.1501641651031895,
491
- "f1": 0.22670301656989097,
492
- "gt_support": 42640,
493
- "predicted": 13848
494
  },
495
  "2": {
496
- "accuracy": 0.0,
497
- "precision": 0.0,
498
- "recall": 0.0,
499
- "f1": 0.0,
500
- "gt_support": 546,
501
- "predicted": 0
502
  },
503
  "3": {
504
- "accuracy": 0.5758037317086439,
505
- "precision": 0.42842976544634004,
506
- "recall": 0.5758037317086439,
507
- "f1": 0.49130298564615504,
508
- "gt_support": 36766,
509
- "predicted": 49413
510
  },
511
  "4": {
512
- "accuracy": 0.8026775595013222,
513
- "precision": 0.5704911980399067,
514
- "recall": 0.8026775595013222,
515
- "f1": 0.6669544148085657,
516
- "gt_support": 42352,
517
- "predicted": 59589
518
  }
519
  },
520
- "total_samples": 122850
521
  },
522
- "1.0": {
523
- "accuracy": 0.46805321341577666,
524
- "baseline_accuracy": 0.4095034663668728,
525
  "per_class": {
526
  "0": {
527
- "accuracy": 0.002529510961214165,
528
- "precision": 0.02127659574468085,
529
- "recall": 0.002529510961214165,
530
- "f1": 0.004521477015825169,
531
- "gt_support": 1186,
532
- "predicted": 141
533
  },
534
  "1": {
535
- "accuracy": 0.1703667777005015,
536
- "precision": 0.5451058472169356,
537
- "recall": 0.1703667777005015,
538
- "f1": 0.2595988258999226,
539
- "gt_support": 109276,
540
- "predicted": 34153
541
  },
542
  "2": {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
543
  "accuracy": 0.0,
544
  "precision": 0.0,
545
  "recall": 0.0,
546
  "f1": 0.0,
547
- "gt_support": 1186,
548
  "predicted": 0
549
- },
550
- "3": {
551
- "accuracy": 0.5114109241278233,
552
- "precision": 0.3985539023221188,
553
- "recall": 0.5114109241278233,
554
- "f1": 0.4479839518555667,
555
- "gt_support": 76418,
556
- "predicted": 98057
557
- },
558
- "4": {
559
- "accuracy": 0.852952376116978,
560
- "precision": 0.49962453252440536,
561
- "recall": 0.852952376116978,
562
- "f1": 0.6301392984907376,
563
- "gt_support": 78784,
564
- "predicted": 134499
565
  }
566
  },
567
- "total_samples": 266850
568
  }
569
  },
570
  "by_size_complexity": {
571
  "7_0.0": {
572
- "accuracy": 0.6457309941520468,
573
  "baseline_accuracy": 0.7822222222222223,
574
  "per_class": {
575
  "0": {
576
- "accuracy": 0.0,
577
- "precision": 0.0,
578
- "recall": 0.0,
579
- "f1": 0.0,
580
  "gt_support": 38,
581
- "predicted": 0
582
  },
583
  "1": {
584
- "accuracy": 0.37280701754385964,
585
- "precision": 0.21505376344086022,
586
- "recall": 0.37280701754385964,
587
- "f1": 0.27276373846770957,
588
  "gt_support": 912,
589
- "predicted": 1581
590
  },
591
  "2": {
592
- "accuracy": 0.0,
593
- "precision": 0.0,
594
- "recall": 0.0,
595
- "f1": 0.0,
596
  "gt_support": 38,
597
- "predicted": 0
598
  },
599
  "3": {
600
- "accuracy": 0.631578947368421,
601
- "precision": 0.2497737556561086,
602
- "recall": 0.631578947368421,
603
- "f1": 0.35797665369649806,
604
  "gt_support": 874,
605
- "predicted": 2210
606
  },
607
  "4": {
608
- "accuracy": 0.6921351674641149,
609
- "precision": 0.9726833368354696,
610
- "recall": 0.6921351674641149,
611
- "f1": 0.8087708569930986,
612
  "gt_support": 6688,
613
- "predicted": 4759
614
  }
615
  },
616
  "total_samples": 8550
617
  },
618
  "7_0.2": {
619
- "accuracy": 0.6617171717171717,
620
  "baseline_accuracy": 0.7822222222222223,
621
  "per_class": {
622
  "0": {
623
- "accuracy": 0.0,
624
- "precision": 0.0,
625
- "recall": 0.0,
626
- "f1": 0.0,
627
  "gt_support": 44,
628
- "predicted": 0
629
  },
630
  "1": {
631
- "accuracy": 0.17832167832167833,
632
- "precision": 0.1883656509695291,
633
- "recall": 0.17832167832167833,
634
- "f1": 0.18320610687022904,
635
  "gt_support": 1144,
636
- "predicted": 1083
637
  },
638
  "2": {
639
- "accuracy": 0.0,
640
- "precision": 0.0,
641
- "recall": 0.0,
642
- "f1": 0.0,
643
  "gt_support": 44,
644
- "predicted": 0
645
  },
646
  "3": {
647
- "accuracy": 0.7748917748917749,
648
- "precision": 0.2378737541528239,
649
- "recall": 0.7748917748917749,
650
- "f1": 0.36400610066090494,
651
  "gt_support": 924,
652
- "predicted": 3010
653
  },
654
  "4": {
655
- "accuracy": 0.7271435950413223,
656
- "precision": 0.9696917513345962,
657
- "recall": 0.7271435950413223,
658
- "f1": 0.8310825769315917,
659
  "gt_support": 7744,
660
- "predicted": 5807
661
  }
662
  },
663
  "total_samples": 9900
664
  },
665
  "7_0.4": {
666
- "accuracy": 0.6750135501355014,
667
  "baseline_accuracy": 0.7822222222222223,
668
  "per_class": {
669
  "0": {
670
- "accuracy": 0.0,
671
- "precision": 0.0,
672
- "recall": 0.0,
673
- "f1": 0.0,
674
  "gt_support": 41,
675
- "predicted": 0
676
  },
677
  "1": {
678
- "accuracy": 0.25522648083623695,
679
- "precision": 0.25107112253641817,
680
- "recall": 0.25522648083623695,
681
- "f1": 0.2531317494600432,
682
  "gt_support": 1148,
683
- "predicted": 1167
684
  },
685
  "2": {
686
- "accuracy": 0.0,
687
- "precision": 0.0,
688
- "recall": 0.0,
689
- "f1": 0.0,
690
  "gt_support": 41,
691
- "predicted": 0
692
  },
693
  "3": {
694
- "accuracy": 0.5943517329910141,
695
- "precision": 0.205595026642984,
696
- "recall": 0.5943517329910141,
697
- "f1": 0.30550973276146487,
698
  "gt_support": 779,
699
- "predicted": 2252
700
  },
701
  "4": {
702
- "accuracy": 0.7581762749445676,
703
- "precision": 0.9423010678608337,
704
- "recall": 0.7581762749445676,
705
- "f1": 0.8402703117800646,
706
  "gt_support": 7216,
707
- "predicted": 5806
708
  }
709
  },
710
  "total_samples": 9225
711
  },
712
  "7_0.6": {
713
- "accuracy": 0.6356501182033097,
714
  "baseline_accuracy": 0.7822222222222223,
715
  "per_class": {
716
  "0": {
717
- "accuracy": 0.0,
718
- "precision": 0.0,
719
- "recall": 0.0,
720
- "f1": 0.0,
721
  "gt_support": 47,
722
- "predicted": 0
723
  },
724
  "1": {
725
- "accuracy": 0.17461482024944974,
726
- "precision": 0.21854912764003673,
727
- "recall": 0.17461482024944974,
728
- "f1": 0.19412724306688414,
729
  "gt_support": 1363,
730
- "predicted": 1089
731
  },
732
  "2": {
733
- "accuracy": 0.0,
734
- "precision": 0.0,
735
- "recall": 0.0,
736
- "f1": 0.0,
737
  "gt_support": 47,
738
- "predicted": 0
739
  },
740
  "3": {
741
- "accuracy": 0.8132387706855791,
742
- "precision": 0.19244755244755246,
743
- "recall": 0.8132387706855791,
744
- "f1": 0.31124180049762495,
745
  "gt_support": 846,
746
- "predicted": 3575
747
  },
748
  "4": {
749
- "accuracy": 0.7006769825918762,
750
- "precision": 0.980544747081712,
751
- "recall": 0.7006769825918762,
752
- "f1": 0.8173165056758092,
753
  "gt_support": 8272,
754
- "predicted": 5911
755
  }
756
  },
757
  "total_samples": 10575
758
  },
759
  "7_0.8": {
760
- "accuracy": 0.6393535353535353,
761
  "baseline_accuracy": 0.7822222222222223,
762
  "per_class": {
763
  "0": {
764
- "accuracy": 0.0,
765
- "precision": 0.0,
766
- "recall": 0.0,
767
- "f1": 0.0,
768
  "gt_support": 55,
769
- "predicted": 0
770
  },
771
  "1": {
772
- "accuracy": 0.2469208211143695,
773
- "precision": 0.26214196762141967,
774
- "recall": 0.2469208211143695,
775
- "f1": 0.2543038356991846,
776
  "gt_support": 1705,
777
- "predicted": 1606
778
  },
779
  "2": {
780
- "accuracy": 0.0,
781
- "precision": 0.0,
782
- "recall": 0.0,
783
- "f1": 0.0,
784
  "gt_support": 55,
785
- "predicted": 0
786
  },
787
  "3": {
788
- "accuracy": 0.7170454545454545,
789
- "precision": 0.1660089450144699,
790
- "recall": 0.7170454545454545,
791
- "f1": 0.2696005127109592,
792
  "gt_support": 880,
793
- "predicted": 3801
794
  },
795
  "4": {
796
- "accuracy": 0.7086776859504132,
797
- "precision": 0.9845005740528129,
798
- "recall": 0.7086776859504132,
799
- "f1": 0.8241230177799136,
800
  "gt_support": 9680,
801
- "predicted": 6968
802
  }
803
  },
804
  "total_samples": 12375
805
  },
806
  "7_1.0": {
807
- "accuracy": 0.6480145719489981,
808
  "baseline_accuracy": 0.7822222222222223,
809
  "per_class": {
810
  "0": {
811
- "accuracy": 0.0,
812
- "precision": 0.0,
813
- "recall": 0.0,
814
- "f1": 0.0,
815
  "gt_support": 61,
816
- "predicted": 0
817
  },
818
  "1": {
819
- "accuracy": 0.20081967213114754,
820
- "precision": 0.27722772277227725,
821
- "recall": 0.20081967213114754,
822
- "f1": 0.23291740938799765,
823
  "gt_support": 1952,
824
- "predicted": 1414
825
  },
826
  "2": {
827
- "accuracy": 0.0,
828
- "precision": 0.0,
829
- "recall": 0.0,
830
- "f1": 0.0,
831
  "gt_support": 61,
832
- "predicted": 0
833
  },
834
  "3": {
835
- "accuracy": 0.7486338797814208,
836
- "precision": 0.15952491849091757,
837
- "recall": 0.7486338797814208,
838
- "f1": 0.26300633518909583,
839
  "gt_support": 915,
840
- "predicted": 4294
841
  },
842
  "4": {
843
- "accuracy": 0.7281110283159463,
844
- "precision": 0.9750530123487589,
845
- "recall": 0.7281110283159463,
846
- "f1": 0.8336799445422065,
847
  "gt_support": 10736,
848
- "predicted": 8017
849
  }
850
  },
851
  "total_samples": 13725
852
  },
853
  "9_0.0": {
854
- "accuracy": 0.6721810699588477,
855
  "baseline_accuracy": 0.64,
856
  "per_class": {
857
  "0": {
858
- "accuracy": 0.0,
859
- "precision": 0.0,
860
- "recall": 0.0,
861
- "f1": 0.0,
862
  "gt_support": 54,
863
- "predicted": 0
864
  },
865
  "1": {
866
- "accuracy": 0.25289351851851855,
867
- "precision": 0.23775843307943417,
868
- "recall": 0.25289351851851855,
869
- "f1": 0.24509254066180597,
870
  "gt_support": 1728,
871
- "predicted": 1838
872
  },
873
  "2": {
874
- "accuracy": 0.0,
875
- "precision": 0.0,
876
- "recall": 0.0,
877
- "f1": 0.0,
878
  "gt_support": 54,
879
- "predicted": 0
880
  },
881
  "3": {
882
- "accuracy": 0.6958234830575256,
883
- "precision": 0.4667019027484144,
884
- "recall": 0.6958234830575256,
885
- "f1": 0.5586839607719075,
886
  "gt_support": 2538,
887
- "predicted": 3784
888
  },
889
  "4": {
890
- "accuracy": 0.7669753086419753,
891
- "precision": 0.9136029411764706,
892
- "recall": 0.7669753086419753,
893
- "f1": 0.8338926174496645,
894
  "gt_support": 7776,
895
- "predicted": 6528
896
  }
897
  },
898
  "total_samples": 12150
899
  },
900
  "9_0.2": {
901
- "accuracy": 0.6950354609929078,
902
  "baseline_accuracy": 0.64,
903
  "per_class": {
904
  "0": {
905
- "accuracy": 0.0,
906
- "precision": 0.0,
907
- "recall": 0.0,
908
- "f1": 0.0,
909
  "gt_support": 47,
910
- "predicted": 0
911
  },
912
  "1": {
913
- "accuracy": 0.2565011820330969,
914
- "precision": 0.30370888733379986,
915
- "recall": 0.2565011820330969,
916
- "f1": 0.2781159884652355,
917
  "gt_support": 1692,
918
- "predicted": 1429
919
  },
920
  "2": {
921
- "accuracy": 0.0,
922
- "precision": 0.0,
923
- "recall": 0.0,
924
- "f1": 0.0,
925
  "gt_support": 47,
926
- "predicted": 0
927
  },
928
  "3": {
929
- "accuracy": 0.7318159327065809,
930
- "precision": 0.45132743362831856,
931
- "recall": 0.7318159327065809,
932
- "f1": 0.5583238958097395,
933
  "gt_support": 2021,
934
- "predicted": 3277
935
  },
936
  "4": {
937
- "accuracy": 0.8033392434988179,
938
- "precision": 0.9263929119100358,
939
- "recall": 0.8033392434988179,
940
- "f1": 0.8604890401202817,
941
  "gt_support": 6768,
942
- "predicted": 5869
943
  }
944
  },
945
  "total_samples": 10575
946
  },
947
  "9_0.4": {
948
- "accuracy": 0.6655389718076286,
949
  "baseline_accuracy": 0.64,
950
  "per_class": {
951
  "0": {
952
- "accuracy": 0.0,
953
- "precision": 0.0,
954
- "recall": 0.0,
955
- "f1": 0.0,
956
  "gt_support": 67,
957
- "predicted": 0
958
  },
959
  "1": {
960
- "accuracy": 0.18582089552238806,
961
- "precision": 0.30533415082771304,
962
- "recall": 0.18582089552238806,
963
- "f1": 0.23103688239387613,
964
  "gt_support": 2680,
965
- "predicted": 1631
966
  },
967
  "2": {
968
- "accuracy": 0.0,
969
- "precision": 0.0,
970
- "recall": 0.0,
971
- "f1": 0.0,
972
  "gt_support": 67,
973
- "predicted": 0
974
  },
975
  "3": {
976
- "accuracy": 0.7290470723306545,
977
- "precision": 0.38001196888090966,
978
- "recall": 0.7290470723306545,
979
- "f1": 0.49960660896931547,
980
  "gt_support": 2613,
981
- "predicted": 5013
982
  },
983
  "4": {
984
- "accuracy": 0.7908374792703151,
985
- "precision": 0.9049934764559364,
986
- "recall": 0.7908374792703151,
987
- "f1": 0.8440732341390564,
988
  "gt_support": 9648,
989
- "predicted": 8431
990
  }
991
  },
992
  "total_samples": 15075
993
  },
994
  "9_0.6": {
995
- "accuracy": 0.6485974499089253,
996
  "baseline_accuracy": 0.64,
997
  "per_class": {
998
  "0": {
999
- "accuracy": 0.0,
1000
- "precision": 0.0,
1001
- "recall": 0.0,
1002
- "f1": 0.0,
1003
  "gt_support": 61,
1004
- "predicted": 0
1005
  },
1006
  "1": {
1007
- "accuracy": 0.14868471216164697,
1008
- "precision": 0.28867505551443373,
1009
- "recall": 0.14868471216164697,
1010
- "f1": 0.19627579265223954,
1011
  "gt_support": 2623,
1012
- "predicted": 1351
1013
  },
1014
  "2": {
1015
- "accuracy": 0.0,
1016
- "precision": 0.0,
1017
- "recall": 0.0,
1018
- "f1": 0.0,
1019
  "gt_support": 61,
1020
- "predicted": 0
1021
  },
1022
  "3": {
1023
- "accuracy": 0.8110200364298725,
1024
- "precision": 0.34515503875968995,
1025
- "recall": 0.8110200364298725,
1026
- "f1": 0.4842305600870038,
1027
  "gt_support": 2196,
1028
- "predicted": 5160
1029
  },
1030
  "4": {
1031
- "accuracy": 0.7662795992714025,
1032
- "precision": 0.9330468533407263,
1033
- "recall": 0.7662795992714025,
1034
- "f1": 0.8414801850231278,
1035
  "gt_support": 8784,
1036
- "predicted": 7214
1037
  }
1038
  },
1039
  "total_samples": 13725
1040
  },
1041
  "9_0.8": {
1042
- "accuracy": 0.656017601760176,
1043
  "baseline_accuracy": 0.64,
1044
  "per_class": {
1045
  "0": {
1046
- "accuracy": 0.0,
1047
- "precision": 0.0,
1048
- "recall": 0.0,
1049
- "f1": 0.0,
1050
  "gt_support": 101,
1051
- "predicted": 0
1052
  },
1053
  "1": {
1054
- "accuracy": 0.20623551716873814,
1055
- "precision": 0.3570386579139314,
1056
- "recall": 0.20623551716873814,
1057
- "f1": 0.2614501268527173,
1058
  "gt_support": 4747,
1059
- "predicted": 2742
1060
  },
1061
  "2": {
1062
- "accuracy": 0.0,
1063
- "precision": 0.0,
1064
- "recall": 0.0,
1065
- "f1": 0.0,
1066
  "gt_support": 101,
1067
- "predicted": 0
1068
  },
1069
  "3": {
1070
- "accuracy": 0.6992574257425742,
1071
- "precision": 0.3182202196564348,
1072
- "recall": 0.6992574257425742,
1073
- "f1": 0.43739113605573837,
1074
  "gt_support": 3232,
1075
- "predicted": 7102
1076
  },
1077
  "4": {
1078
- "accuracy": 0.8023239823982399,
1079
- "precision": 0.9059079264032296,
1080
- "recall": 0.8023239823982399,
1081
- "f1": 0.850975387420237,
1082
  "gt_support": 14544,
1083
- "predicted": 12881
1084
  }
1085
  },
1086
  "total_samples": 22725
1087
  },
1088
  "9_1.0": {
1089
- "accuracy": 0.6735990338164252,
1090
  "baseline_accuracy": 0.64,
1091
  "per_class": {
1092
  "0": {
1093
- "accuracy": 0.0,
1094
- "precision": 0.0,
1095
- "recall": 0.0,
1096
- "f1": 0.0,
1097
  "gt_support": 184,
1098
- "predicted": 0
1099
  },
1100
  "1": {
1101
- "accuracy": 0.27608695652173915,
1102
- "precision": 0.4303625889528973,
1103
- "recall": 0.27608695652173915,
1104
- "f1": 0.33637928751158785,
1105
  "gt_support": 9200,
1106
- "predicted": 5902
1107
  },
1108
  "2": {
1109
- "accuracy": 0.0,
1110
- "precision": 0.0,
1111
- "recall": 0.0,
1112
- "f1": 0.0,
1113
  "gt_support": 184,
1114
- "predicted": 0
1115
  },
1116
  "3": {
1117
- "accuracy": 0.6664167916041979,
1118
- "precision": 0.30140701813866755,
1119
- "recall": 0.6664167916041979,
1120
- "f1": 0.4150811252480448,
1121
  "gt_support": 5336,
1122
- "predicted": 11798
1123
  },
1124
  "4": {
1125
- "accuracy": 0.8224260265700483,
1126
- "precision": 0.9194514767932489,
1127
- "recall": 0.8224260265700483,
1128
- "f1": 0.8682365128695513,
1129
  "gt_support": 26496,
1130
- "predicted": 23700
1131
  }
1132
  },
1133
  "total_samples": 41400
1134
  },
1135
  "11_0.0": {
1136
- "accuracy": 0.6680392156862746,
1137
  "baseline_accuracy": 0.4622222222222222,
1138
  "per_class": {
1139
  "0": {
1140
- "accuracy": 0.0,
1141
- "precision": 0.0,
1142
- "recall": 0.0,
1143
- "f1": 0.0,
1144
  "gt_support": 68,
1145
- "predicted": 0
1146
  },
1147
  "1": {
1148
- "accuracy": 0.22573529411764706,
1149
- "precision": 0.32834224598930484,
1150
- "recall": 0.22573529411764706,
1151
- "f1": 0.26753812636165575,
1152
  "gt_support": 2720,
1153
- "predicted": 1870
1154
  },
1155
  "2": {
1156
- "accuracy": 0.0,
1157
- "precision": 0.0,
1158
- "recall": 0.0,
1159
- "f1": 0.0,
1160
  "gt_support": 68,
1161
- "predicted": 0
1162
  },
1163
  "3": {
1164
- "accuracy": 0.7241250930752048,
1165
- "precision": 0.6129845572013867,
1166
- "recall": 0.7241250930752048,
1167
- "f1": 0.6639358252261477,
1168
  "gt_support": 5372,
1169
- "predicted": 6346
1170
  },
1171
  "4": {
1172
- "accuracy": 0.8083993212669683,
1173
- "precision": 0.8070299265951439,
1174
- "recall": 0.8083993212669683,
1175
- "f1": 0.8077140435151172,
1176
  "gt_support": 7072,
1177
  "predicted": 7084
1178
  }
@@ -1180,557 +1194,557 @@
1180
  "total_samples": 15300
1181
  },
1182
  "11_0.2": {
1183
- "accuracy": 0.6509803921568628,
1184
  "baseline_accuracy": 0.4622222222222222,
1185
  "per_class": {
1186
  "0": {
1187
- "accuracy": 0.0,
1188
- "precision": 0.0,
1189
- "recall": 0.0,
1190
- "f1": 0.0,
1191
  "gt_support": 85,
1192
- "predicted": 0
1193
  },
1194
  "1": {
1195
- "accuracy": 0.16170212765957448,
1196
- "precision": 0.34361702127659577,
1197
- "recall": 0.16170212765957448,
1198
- "f1": 0.2199148936170213,
1199
  "gt_support": 3995,
1200
- "predicted": 1880
1201
  },
1202
  "2": {
1203
- "accuracy": 0.0,
1204
- "precision": 0.0,
1205
- "recall": 0.0,
1206
- "f1": 0.0,
1207
  "gt_support": 85,
1208
- "predicted": 0
1209
  },
1210
  "3": {
1211
- "accuracy": 0.7490196078431373,
1212
- "precision": 0.553690059185892,
1213
- "recall": 0.7490196078431373,
1214
- "f1": 0.6367108827001875,
1215
  "gt_support": 6120,
1216
- "predicted": 8279
1217
  },
1218
  "4": {
1219
- "accuracy": 0.8167420814479638,
1220
- "precision": 0.8052643319205889,
1221
- "recall": 0.8167420814479638,
1222
- "f1": 0.810962596877457,
1223
  "gt_support": 8840,
1224
- "predicted": 8966
1225
  }
1226
  },
1227
  "total_samples": 19125
1228
  },
1229
  "11_0.4": {
1230
- "accuracy": 0.6288484848484849,
1231
  "baseline_accuracy": 0.4622222222222222,
1232
  "per_class": {
1233
  "0": {
1234
- "accuracy": 0.0,
1235
- "precision": 0.0,
1236
- "recall": 0.0,
1237
- "f1": 0.0,
1238
  "gt_support": 55,
1239
- "predicted": 0
1240
  },
1241
  "1": {
1242
- "accuracy": 0.18799313893653516,
1243
- "precision": 0.3832167832167832,
1244
- "recall": 0.18799313893653516,
1245
- "f1": 0.2522439585730725,
1246
  "gt_support": 2915,
1247
- "predicted": 1430
1248
  },
1249
  "2": {
1250
- "accuracy": 0.0,
1251
- "precision": 0.0,
1252
- "recall": 0.0,
1253
- "f1": 0.0,
1254
  "gt_support": 55,
1255
- "predicted": 0
1256
  },
1257
  "3": {
1258
- "accuracy": 0.6484848484848484,
1259
- "precision": 0.5166812993854258,
1260
- "recall": 0.6484848484848484,
1261
- "f1": 0.5751282677742487,
1262
  "gt_support": 3630,
1263
- "predicted": 4556
1264
  },
1265
  "4": {
1266
- "accuracy": 0.8531468531468531,
1267
- "precision": 0.7638128032555955,
1268
- "recall": 0.8531468531468531,
1269
- "f1": 0.8060120571475762,
1270
  "gt_support": 5720,
1271
- "predicted": 6389
1272
  }
1273
  },
1274
  "total_samples": 12375
1275
  },
1276
  "11_0.6": {
1277
- "accuracy": 0.6424163969795038,
1278
  "baseline_accuracy": 0.4622222222222222,
1279
  "per_class": {
1280
  "0": {
1281
- "accuracy": 0.0,
1282
- "precision": 0.0,
1283
- "recall": 0.0,
1284
- "f1": 0.0,
1285
  "gt_support": 103,
1286
- "predicted": 0
1287
  },
1288
  "1": {
1289
- "accuracy": 0.21844660194174756,
1290
- "precision": 0.4716981132075472,
1291
- "recall": 0.21844660194174756,
1292
- "f1": 0.298606502986065,
1293
  "gt_support": 6180,
1294
- "predicted": 2862
1295
  },
1296
  "2": {
1297
- "accuracy": 0.0,
1298
- "precision": 0.0,
1299
- "recall": 0.0,
1300
- "f1": 0.0,
1301
  "gt_support": 103,
1302
- "predicted": 0
1303
  },
1304
  "3": {
1305
- "accuracy": 0.7166364982721738,
1306
- "precision": 0.4707599178467193,
1307
- "recall": 0.7166364982721738,
1308
- "f1": 0.5682411273486431,
1309
  "gt_support": 6077,
1310
- "predicted": 9251
1311
  },
1312
  "4": {
1313
- "accuracy": 0.8572628827483196,
1314
- "precision": 0.8301392153317664,
1315
- "recall": 0.8572628827483196,
1316
- "f1": 0.8434830531826949,
1317
  "gt_support": 10712,
1318
- "predicted": 11062
1319
  }
1320
  },
1321
  "total_samples": 23175
1322
  },
1323
  "11_0.8": {
1324
- "accuracy": 0.5999220272904483,
1325
  "baseline_accuracy": 0.4622222222222222,
1326
  "per_class": {
1327
  "0": {
1328
- "accuracy": 0.0,
1329
- "precision": 0.0,
1330
- "recall": 0.0,
1331
- "f1": 0.0,
1332
  "gt_support": 114,
1333
- "predicted": 0
1334
  },
1335
  "1": {
1336
- "accuracy": 0.14380648591174908,
1337
- "precision": 0.4606215410813112,
1338
- "recall": 0.14380648591174908,
1339
- "f1": 0.21918363212802594,
1340
  "gt_support": 7524,
1341
- "predicted": 2349
1342
  },
1343
  "2": {
1344
- "accuracy": 0.0,
1345
- "precision": 0.0,
1346
- "recall": 0.0,
1347
- "f1": 0.0,
1348
  "gt_support": 114,
1349
- "predicted": 0
1350
  },
1351
  "3": {
1352
- "accuracy": 0.7725918570009931,
1353
- "precision": 0.4009620340147741,
1354
- "recall": 0.7725918570009931,
1355
- "f1": 0.5279348563673377,
1356
  "gt_support": 6042,
1357
- "predicted": 11642
1358
  },
1359
  "4": {
1360
- "accuracy": 0.8129217273954116,
1361
- "precision": 0.8266575177974097,
1362
- "recall": 0.8129217273954116,
1363
- "f1": 0.8197320859026154,
1364
  "gt_support": 11856,
1365
- "predicted": 11659
1366
  }
1367
  },
1368
  "total_samples": 25650
1369
  },
1370
  "11_1.0": {
1371
- "accuracy": 0.6049404761904762,
1372
  "baseline_accuracy": 0.4622222222222222,
1373
  "per_class": {
1374
  "0": {
1375
- "accuracy": 0.004464285714285714,
1376
- "precision": 0.09090909090909091,
1377
- "recall": 0.004464285714285714,
1378
- "f1": 0.00851063829787234,
1379
  "gt_support": 224,
1380
- "predicted": 11
1381
  },
1382
  "1": {
1383
- "accuracy": 0.22600446428571427,
1384
- "precision": 0.5520218082689686,
1385
- "recall": 0.22600446428571427,
1386
- "f1": 0.3207074039857463,
1387
  "gt_support": 16128,
1388
- "predicted": 6603
1389
  },
1390
  "2": {
1391
- "accuracy": 0.0,
1392
- "precision": 0.0,
1393
- "recall": 0.0,
1394
- "f1": 0.0,
1395
  "gt_support": 224,
1396
- "predicted": 0
1397
  },
1398
  "3": {
1399
- "accuracy": 0.6218655015197568,
1400
- "precision": 0.3675611947002021,
1401
- "recall": 0.6218655015197568,
1402
- "f1": 0.4620324629498941,
1403
  "gt_support": 10528,
1404
- "predicted": 17812
1405
  },
1406
  "4": {
1407
- "accuracy": 0.8712225274725275,
1408
- "precision": 0.7813967813967814,
1409
- "recall": 0.8712225274725275,
1410
- "f1": 0.8238684798051553,
1411
  "gt_support": 23296,
1412
- "predicted": 25974
1413
  }
1414
  },
1415
  "total_samples": 50400
1416
  },
1417
  "13_0.0": {
1418
- "accuracy": 0.6322222222222222,
1419
  "baseline_accuracy": 0.5288888888888889,
1420
  "per_class": {
1421
  "0": {
1422
- "accuracy": 0.0,
1423
- "precision": 0.0,
1424
- "recall": 0.0,
1425
- "f1": 0.0,
1426
  "gt_support": 64,
1427
- "predicted": 0
1428
  },
1429
  "1": {
1430
- "accuracy": 0.2509765625,
1431
- "precision": 0.3712084737602311,
1432
- "recall": 0.2509765625,
1433
- "f1": 0.29947562633521074,
1434
  "gt_support": 3072,
1435
- "predicted": 2077
1436
  },
1437
  "2": {
1438
- "accuracy": 0.0,
1439
- "precision": 0.0,
1440
- "recall": 0.0,
1441
- "f1": 0.0,
1442
  "gt_support": 64,
1443
- "predicted": 0
1444
  },
1445
  "3": {
1446
- "accuracy": 0.6697741596638656,
1447
- "precision": 0.7778286062824031,
1448
- "recall": 0.6697741596638656,
1449
- "f1": 0.7197685903767461,
1450
  "gt_support": 7616,
1451
- "predicted": 6558
1452
  },
1453
  "4": {
1454
- "accuracy": 0.9017857142857143,
1455
- "precision": 0.5606244579358196,
1456
- "recall": 0.9017857142857143,
1457
- "f1": 0.6914108460797947,
1458
  "gt_support": 3584,
1459
- "predicted": 5765
1460
  }
1461
  },
1462
  "total_samples": 14400
1463
  },
1464
  "13_0.2": {
1465
- "accuracy": 0.6028673835125448,
1466
  "baseline_accuracy": 0.48444444444444446,
1467
  "per_class": {
1468
  "0": {
1469
- "accuracy": 0.0,
1470
- "precision": 0.0,
1471
- "recall": 0.0,
1472
- "f1": 0.0,
1473
  "gt_support": 62,
1474
- "predicted": 0
1475
  },
1476
  "1": {
1477
- "accuracy": 0.15016685205784205,
1478
- "precision": 0.3837953091684435,
1479
- "recall": 0.15016685205784205,
1480
- "f1": 0.21587047771337198,
1481
  "gt_support": 3596,
1482
- "predicted": 1407
1483
  },
1484
  "2": {
1485
- "accuracy": 0.0,
1486
- "precision": 0.0,
1487
- "recall": 0.0,
1488
- "f1": 0.0,
1489
  "gt_support": 62,
1490
- "predicted": 0
1491
  },
1492
  "3": {
1493
- "accuracy": 0.69621189701095,
1494
- "precision": 0.6800115623645036,
1495
- "recall": 0.69621189701095,
1496
- "f1": 0.688016377860642,
1497
  "gt_support": 6758,
1498
- "predicted": 6919
1499
  },
1500
  "4": {
1501
- "accuracy": 0.9115783410138248,
1502
- "precision": 0.5627667140825036,
1503
- "recall": 0.9115783410138248,
1504
- "f1": 0.695910290237467,
1505
  "gt_support": 3472,
1506
- "predicted": 5624
1507
  }
1508
  },
1509
  "total_samples": 13950
1510
  },
1511
  "13_0.4": {
1512
- "accuracy": 0.5656177777777778,
1513
  "baseline_accuracy": 0.44,
1514
  "per_class": {
1515
  "0": {
1516
- "accuracy": 0.0,
1517
- "precision": 0.0,
1518
- "recall": 0.0,
1519
- "f1": 0.0,
1520
  "gt_support": 125,
1521
- "predicted": 0
1522
  },
1523
  "1": {
1524
- "accuracy": 0.12976470588235295,
1525
- "precision": 0.4433279742765273,
1526
- "recall": 0.12976470588235295,
1527
- "f1": 0.20076447033127048,
1528
  "gt_support": 8500,
1529
- "predicted": 2488
1530
  },
1531
  "2": {
1532
- "accuracy": 0.0,
1533
- "precision": 0.0,
1534
- "recall": 0.0,
1535
- "f1": 0.0,
1536
  "gt_support": 125,
1537
- "predicted": 0
1538
  },
1539
  "3": {
1540
- "accuracy": 0.6683636363636364,
1541
- "precision": 0.6207129455909943,
1542
- "recall": 0.6683636363636364,
1543
- "f1": 0.643657587548638,
1544
  "gt_support": 12375,
1545
- "predicted": 13325
1546
  },
1547
  "4": {
1548
- "accuracy": 0.9334285714285714,
1549
- "precision": 0.5307017543859649,
1550
- "recall": 0.9334285714285714,
1551
- "f1": 0.6766777133388567,
1552
  "gt_support": 7000,
1553
- "predicted": 12312
1554
  }
1555
  },
1556
  "total_samples": 28125
1557
  },
1558
  "13_0.6": {
1559
- "accuracy": 0.5388797061524334,
1560
  "baseline_accuracy": 0.39555555555555555,
1561
  "per_class": {
1562
  "0": {
1563
- "accuracy": 0.0,
1564
- "precision": 0.0,
1565
- "recall": 0.0,
1566
- "f1": 0.0,
1567
  "gt_support": 121,
1568
- "predicted": 48
1569
  },
1570
  "1": {
1571
- "accuracy": 0.15246874337783428,
1572
- "precision": 0.4863129435620142,
1573
- "recall": 0.15246874337783428,
1574
- "f1": 0.23215294022747437,
1575
  "gt_support": 9438,
1576
- "predicted": 2959
1577
  },
1578
  "2": {
1579
- "accuracy": 0.0,
1580
- "precision": 0.0,
1581
- "recall": 0.0,
1582
- "f1": 0.0,
1583
  "gt_support": 121,
1584
- "predicted": 0
1585
  },
1586
  "3": {
1587
- "accuracy": 0.6459281270312935,
1588
- "precision": 0.5577293136626043,
1589
- "recall": 0.6459281270312935,
1590
- "f1": 0.5985973064842305,
1591
  "gt_support": 10769,
1592
- "predicted": 12472
1593
  },
1594
  "4": {
1595
- "accuracy": 0.9262101534828807,
1596
- "precision": 0.5343095521879789,
1597
- "recall": 0.9262101534828807,
1598
- "f1": 0.6776805960479431,
1599
  "gt_support": 6776,
1600
- "predicted": 11746
1601
  }
1602
  },
1603
  "total_samples": 27225
1604
  },
1605
  "13_0.8": {
1606
- "accuracy": 0.5112301587301588,
1607
  "baseline_accuracy": 0.39555555555555555,
1608
  "per_class": {
1609
  "0": {
1610
- "accuracy": 0.0,
1611
- "precision": 0.0,
1612
- "recall": 0.0,
1613
- "f1": 0.0,
1614
  "gt_support": 112,
1615
- "predicted": 0
1616
  },
1617
  "1": {
1618
- "accuracy": 0.17686597110754415,
1619
- "precision": 0.5308641975308642,
1620
- "recall": 0.17686597110754415,
1621
- "f1": 0.2653322296636316,
1622
  "gt_support": 9968,
1623
- "predicted": 3321
1624
  },
1625
  "2": {
1626
- "accuracy": 0.0,
1627
- "precision": 0.0,
1628
- "recall": 0.0,
1629
- "f1": 0.0,
1630
  "gt_support": 112,
1631
- "predicted": 0
1632
  },
1633
  "3": {
1634
- "accuracy": 0.6057692307692307,
1635
- "precision": 0.486441768544903,
1636
- "recall": 0.6057692307692307,
1637
- "f1": 0.5395870507264847,
1638
  "gt_support": 8736,
1639
- "predicted": 10879
1640
  },
1641
  "4": {
1642
- "accuracy": 0.9292091836734694,
1643
- "precision": 0.5298181818181819,
1644
- "recall": 0.9292091836734694,
1645
- "f1": 0.6748494673459935,
1646
  "gt_support": 6272,
1647
- "predicted": 11000
1648
  }
1649
  },
1650
  "total_samples": 25200
1651
  },
1652
  "13_1.0": {
1653
- "accuracy": 0.47720518064076345,
1654
  "baseline_accuracy": 0.43555555555555553,
1655
  "per_class": {
1656
  "0": {
1657
- "accuracy": 0.006134969325153374,
1658
- "precision": 0.0273972602739726,
1659
- "recall": 0.006134969325153374,
1660
- "f1": 0.010025062656641603,
1661
  "gt_support": 326,
1662
- "predicted": 73
1663
  },
1664
  "1": {
1665
- "accuracy": 0.14392137222987356,
1666
- "precision": 0.6019112449273465,
1667
- "recall": 0.14392137222987356,
1668
- "f1": 0.2322984818248415,
1669
  "gt_support": 31948,
1670
- "predicted": 7639
1671
  },
1672
  "2": {
1673
- "accuracy": 0.0,
1674
- "precision": 0.0,
1675
- "recall": 0.0,
1676
- "f1": 0.0,
1677
  "gt_support": 326,
1678
- "predicted": 0
1679
  },
1680
  "3": {
1681
- "accuracy": 0.5827331732906553,
1682
- "precision": 0.42791851658396446,
1683
- "recall": 0.5827331732906553,
1684
- "f1": 0.4934683582426683,
1685
  "gt_support": 22494,
1686
- "predicted": 30632
1687
  },
1688
  "4": {
1689
- "accuracy": 0.9473597721297108,
1690
- "precision": 0.49405816145803577,
1691
- "recall": 0.9473597721297108,
1692
- "f1": 0.649431114115129,
1693
  "gt_support": 18256,
1694
- "predicted": 35006
1695
  }
1696
  },
1697
  "total_samples": 73350
1698
  },
1699
  "15_0.0": {
1700
- "accuracy": 0.404702842377261,
1701
  "baseline_accuracy": 0.7422222222222222,
1702
  "per_class": {
1703
  "0": {
1704
- "accuracy": 0.0,
1705
- "precision": 0.0,
1706
- "recall": 0.0,
1707
- "f1": 0.0,
1708
  "gt_support": 86,
1709
- "predicted": 0
1710
  },
1711
  "1": {
1712
- "accuracy": 0.12790697674418605,
1713
- "precision": 0.3171987641606591,
1714
- "recall": 0.12790697674418605,
1715
- "f1": 0.182302456348032,
1716
  "gt_support": 4816,
1717
- "predicted": 1942
1718
  },
1719
  "2": {
1720
- "accuracy": 0.0,
1721
- "precision": 0.0,
1722
- "recall": 0.0,
1723
- "f1": 0.0,
1724
  "gt_support": 86,
1725
- "predicted": 0
1726
  },
1727
  "3": {
1728
- "accuracy": 0.5023673583066425,
1729
- "precision": 0.8255148741418764,
1730
- "recall": 0.5023673583066425,
1731
- "f1": 0.6246212449138602,
1732
  "gt_support": 14362,
1733
- "predicted": 8740
1734
  },
1735
  "4": {
1736
  "accuracy": 0.0,
@@ -1738,46 +1752,46 @@
1738
  "recall": 0.0,
1739
  "f1": 0.0,
1740
  "gt_support": 0,
1741
- "predicted": 8668
1742
  }
1743
  },
1744
  "total_samples": 19350
1745
  },
1746
  "15_0.2": {
1747
- "accuracy": 0.3926947637292465,
1748
  "baseline_accuracy": 0.6755555555555556,
1749
  "per_class": {
1750
  "0": {
1751
- "accuracy": 0.0,
1752
- "precision": 0.0,
1753
- "recall": 0.0,
1754
- "f1": 0.0,
1755
  "gt_support": 87,
1756
- "predicted": 0
1757
  },
1758
  "1": {
1759
- "accuracy": 0.09114456856079002,
1760
- "precision": 0.35143570536828966,
1761
- "recall": 0.09114456856079002,
1762
- "f1": 0.14474868234991645,
1763
  "gt_support": 6177,
1764
- "predicted": 1602
1765
  },
1766
  "2": {
1767
- "accuracy": 0.0,
1768
- "precision": 0.0,
1769
- "recall": 0.0,
1770
- "f1": 0.0,
1771
  "gt_support": 87,
1772
- "predicted": 0
1773
  },
1774
  "3": {
1775
- "accuracy": 0.5387174833635814,
1776
- "precision": 0.7452662412386233,
1777
- "recall": 0.5387174833635814,
1778
- "f1": 0.6253785717420884,
1779
  "gt_support": 13224,
1780
- "predicted": 9559
1781
  },
1782
  "4": {
1783
  "accuracy": 0.0,
@@ -1785,46 +1799,46 @@
1785
  "recall": 0.0,
1786
  "f1": 0.0,
1787
  "gt_support": 0,
1788
- "predicted": 8414
1789
  }
1790
  },
1791
  "total_samples": 19575
1792
  },
1793
  "15_0.4": {
1794
- "accuracy": 0.3148692810457516,
1795
  "baseline_accuracy": 0.6133333333333333,
1796
  "per_class": {
1797
  "0": {
1798
- "accuracy": 0.0,
1799
- "precision": 0.0,
1800
- "recall": 0.0,
1801
- "f1": 0.0,
1802
  "gt_support": 136,
1803
- "predicted": 0
1804
  },
1805
  "1": {
1806
- "accuracy": 0.11055363321799308,
1807
- "precision": 0.41493506493506493,
1808
- "recall": 0.11055363321799308,
1809
- "f1": 0.17459016393442622,
1810
  "gt_support": 11560,
1811
- "predicted": 3080
1812
  },
1813
  "2": {
1814
- "accuracy": 0.0,
1815
- "precision": 0.0,
1816
- "recall": 0.0,
1817
- "f1": 0.0,
1818
  "gt_support": 136,
1819
- "predicted": 0
1820
  },
1821
  "3": {
1822
- "accuracy": 0.44527919863597615,
1823
- "precision": 0.6623078142336345,
1824
- "recall": 0.44527919863597615,
1825
- "f1": 0.5325304275791755,
1826
  "gt_support": 18768,
1827
- "predicted": 12618
1828
  },
1829
  "4": {
1830
  "accuracy": 0.0,
@@ -1832,46 +1846,46 @@
1832
  "recall": 0.0,
1833
  "f1": 0.0,
1834
  "gt_support": 0,
1835
- "predicted": 14902
1836
  }
1837
  },
1838
  "total_samples": 30600
1839
  },
1840
  "15_0.6": {
1841
- "accuracy": 0.3105777777777778,
1842
  "baseline_accuracy": 0.5466666666666666,
1843
  "per_class": {
1844
  "0": {
1845
- "accuracy": 0.0,
1846
- "precision": 0.0,
1847
- "recall": 0.0,
1848
- "f1": 0.0,
1849
  "gt_support": 100,
1850
- "predicted": 0
1851
  },
1852
  "1": {
1853
- "accuracy": 0.1142,
1854
- "precision": 0.4812473662031184,
1855
- "recall": 0.1142,
1856
- "f1": 0.18459549018023114,
1857
  "gt_support": 10000,
1858
- "predicted": 2373
1859
  },
1860
  "2": {
1861
- "accuracy": 0.0,
1862
- "precision": 0.0,
1863
- "recall": 0.0,
1864
- "f1": 0.0,
1865
  "gt_support": 100,
1866
- "predicted": 0
1867
  },
1868
  "3": {
1869
- "accuracy": 0.47528455284552845,
1870
- "precision": 0.5887210473313192,
1871
- "recall": 0.47528455284552845,
1872
- "f1": 0.5259559154295996,
1873
  "gt_support": 12300,
1874
- "predicted": 9930
1875
  },
1876
  "4": {
1877
  "accuracy": 0.0,
@@ -1879,46 +1893,46 @@
1879
  "recall": 0.0,
1880
  "f1": 0.0,
1881
  "gt_support": 0,
1882
- "predicted": 10197
1883
  }
1884
  },
1885
  "total_samples": 22500
1886
  },
1887
  "15_0.8": {
1888
- "accuracy": 0.28392953929539294,
1889
  "baseline_accuracy": 0.5066666666666667,
1890
  "per_class": {
1891
  "0": {
1892
- "accuracy": 0.0,
1893
- "precision": 0.0,
1894
- "recall": 0.0,
1895
- "f1": 0.0,
1896
  "gt_support": 164,
1897
- "predicted": 0
1898
  },
1899
  "1": {
1900
- "accuracy": 0.1154257595207531,
1901
- "precision": 0.5634464751958225,
1902
- "recall": 0.1154257595207531,
1903
- "f1": 0.19160081683388086,
1904
  "gt_support": 18696,
1905
- "predicted": 3830
1906
  },
1907
  "2": {
1908
- "accuracy": 0.0,
1909
- "precision": 0.0,
1910
- "recall": 0.0,
1911
- "f1": 0.0,
1912
  "gt_support": 164,
1913
- "predicted": 0
1914
  },
1915
  "3": {
1916
- "accuracy": 0.4653725665697024,
1917
- "precision": 0.5202952029520295,
1918
- "recall": 0.4653725665697024,
1919
- "f1": 0.49130370589103795,
1920
  "gt_support": 17876,
1921
- "predicted": 15989
1922
  },
1923
  "4": {
1924
  "accuracy": 0.0,
@@ -1926,46 +1940,46 @@
1926
  "recall": 0.0,
1927
  "f1": 0.0,
1928
  "gt_support": 0,
1929
- "predicted": 17081
1930
  }
1931
  },
1932
  "total_samples": 36900
1933
  },
1934
  "15_1.0": {
1935
- "accuracy": 0.25719806763285025,
1936
  "baseline_accuracy": 0.5688888888888889,
1937
  "per_class": {
1938
  "0": {
1939
- "accuracy": 0.0,
1940
- "precision": 0.0,
1941
- "recall": 0.0,
1942
- "f1": 0.0,
1943
  "gt_support": 391,
1944
- "predicted": 57
1945
  },
1946
  "1": {
1947
- "accuracy": 0.14869725063938619,
1948
- "precision": 0.5908693926161175,
1949
- "recall": 0.14869725063938619,
1950
- "f1": 0.23760037035263315,
1951
  "gt_support": 50048,
1952
- "predicted": 12595
1953
  },
1954
  "2": {
1955
- "accuracy": 0.0,
1956
- "precision": 0.0,
1957
- "recall": 0.0,
1958
- "f1": 0.0,
1959
  "gt_support": 391,
1960
- "predicted": 0
1961
  },
1962
  "3": {
1963
- "accuracy": 0.40880333826894605,
1964
- "precision": 0.4529996121834074,
1965
- "recall": 0.40880333826894605,
1966
- "f1": 0.4297682053604279,
1967
  "gt_support": 37145,
1968
- "predicted": 33521
1969
  },
1970
  "4": {
1971
  "accuracy": 0.0,
@@ -1973,23 +1987,10 @@
1973
  "recall": 0.0,
1974
  "f1": 0.0,
1975
  "gt_support": 0,
1976
- "predicted": 41802
1977
  }
1978
  },
1979
  "total_samples": 87975
1980
  }
1981
- },
1982
- "total_trajectories": 300,
1983
- "total_steps": 3223,
1984
- "config": {
1985
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.pt",
1986
- "trajectories_dir": "reveng/trajectories_test_full",
1987
- "activations_dir": "interp/activations_test_full",
1988
- "layers": "23",
1989
- "steps": "all",
1990
- "token_categories": {
1991
- "output": "all"
1992
- },
1993
- "pad_to_size": 15
1994
  }
1995
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.7944372048126315,
4
  "baseline_accuracy": 0.3355686558416934,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.7533354017995656,
8
+ "precision": 0.0965561123041438,
9
+ "recall": 0.7533354017995656,
10
+ "f1": 0.17117275899749726,
11
  "gt_support": 3223,
12
+ "predicted": 25146
13
  },
14
  "1": {
15
+ "accuracy": 0.714839928623911,
16
+ "precision": 0.7784493834385702,
17
+ "recall": 0.714839928623911,
18
+ "f1": 0.7452898741048135,
19
  "gt_support": 238175,
20
+ "predicted": 218713
21
  },
22
  "2": {
23
+ "accuracy": 0.7139311200744648,
24
+ "precision": 0.08174938714605465,
25
+ "recall": 0.7139311200744648,
26
+ "f1": 0.14670066942939114,
27
  "gt_support": 3223,
28
+ "predicted": 28147
29
  },
30
  "3": {
31
+ "accuracy": 0.6747224117100754,
32
+ "precision": 0.7625228722959605,
33
+ "recall": 0.6747224117100754,
34
+ "f1": 0.7159408030139185,
35
  "gt_support": 243346,
36
+ "predicted": 215326
37
  },
38
  "4": {
39
+ "accuracy": 0.9988238170719369,
40
+ "precision": 0.9961571288623168,
41
+ "recall": 0.9988238170719369,
42
+ "f1": 0.9974886906879472,
43
  "gt_support": 237208,
44
+ "predicted": 237843
45
  }
46
  },
47
  "total_samples": 725175
48
  },
49
+ "by_complexity": {
50
+ "0.0": {
51
+ "accuracy": 0.926594982078853,
52
+ "baseline_accuracy": 0.4410322580645161,
53
  "per_class": {
54
  "0": {
55
+ "accuracy": 0.8419354838709677,
56
+ "precision": 0.09,
57
+ "recall": 0.8419354838709677,
58
+ "f1": 0.16261682242990652,
59
+ "gt_support": 310,
60
+ "predicted": 2900
61
  },
62
  "1": {
63
+ "accuracy": 0.9621074879227053,
64
+ "precision": 0.9692038628241199,
65
+ "recall": 0.9621074879227053,
66
+ "f1": 0.9656426379787113,
67
+ "gt_support": 13248,
68
+ "predicted": 13151
69
  },
70
  "2": {
71
+ "accuracy": 0.7258064516129032,
72
+ "precision": 0.11633919338159256,
73
+ "recall": 0.7258064516129032,
74
+ "f1": 0.20053475935828877,
75
+ "gt_support": 310,
76
+ "predicted": 1934
77
  },
78
  "3": {
79
+ "accuracy": 0.8546908523503023,
80
+ "precision": 0.990096027113538,
81
+ "recall": 0.8546908523503023,
82
+ "f1": 0.91742414990317,
83
+ "gt_support": 30762,
84
+ "predicted": 26555
85
  },
86
  "4": {
87
+ "accuracy": 0.9994426751592357,
88
+ "precision": 0.9958746529155097,
89
+ "recall": 0.9994426751592357,
90
+ "f1": 0.9976554738724419,
91
+ "gt_support": 25120,
92
+ "predicted": 25210
93
  }
94
  },
95
+ "total_samples": 69750
96
  },
97
+ "0.2": {
98
+ "accuracy": 0.8868239316239316,
99
+ "baseline_accuracy": 0.3972239316239316,
100
  "per_class": {
101
  "0": {
102
+ "accuracy": 0.8153846153846154,
103
+ "precision": 0.08783559827643354,
104
+ "recall": 0.8153846153846154,
105
+ "f1": 0.15858767205266308,
106
+ "gt_support": 325,
107
+ "predicted": 3017
108
  },
109
  "1": {
110
+ "accuracy": 0.8082992050108407,
111
+ "precision": 0.9674884659746251,
112
+ "recall": 0.8082992050108407,
113
+ "f1": 0.8807586297414359,
114
+ "gt_support": 16604,
115
+ "predicted": 13872
116
  },
117
  "2": {
118
+ "accuracy": 0.7446153846153846,
119
+ "precision": 0.08645944980350125,
120
+ "recall": 0.7446153846153846,
121
+ "f1": 0.15492957746478872,
122
+ "gt_support": 325,
123
+ "predicted": 2799
124
  },
125
  "3": {
126
+ "accuracy": 0.8298963748407753,
127
+ "precision": 0.9090429142469266,
128
+ "recall": 0.8298963748407753,
129
+ "f1": 0.8676684963556195,
130
+ "gt_support": 29047,
131
+ "predicted": 26518
132
  },
133
  "4": {
134
+ "accuracy": 0.9996644795705338,
135
+ "precision": 0.996136557821613,
136
+ "recall": 0.9996644795705338,
137
+ "f1": 0.997897400591705,
138
+ "gt_support": 26824,
139
+ "predicted": 26919
140
  }
141
  },
142
+ "total_samples": 73125
143
  },
144
+ "0.4": {
145
+ "accuracy": 0.8479245283018868,
146
+ "baseline_accuracy": 0.4000524109014675,
147
  "per_class": {
148
  "0": {
149
+ "accuracy": 0.8514150943396226,
150
+ "precision": 0.09583222723652775,
151
+ "recall": 0.8514150943396226,
152
+ "f1": 0.17227392030541638,
153
+ "gt_support": 424,
154
+ "predicted": 3767
155
  },
156
  "1": {
157
+ "accuracy": 0.6873111218893407,
158
+ "precision": 0.9571859087602619,
159
+ "recall": 0.6873111218893407,
160
+ "f1": 0.800104236791244,
161
+ "gt_support": 26803,
162
+ "predicted": 19246
163
  },
164
  "2": {
165
+ "accuracy": 0.75,
166
+ "precision": 0.08397148138368102,
167
+ "recall": 0.75,
168
+ "f1": 0.1510330087865115,
169
+ "gt_support": 424,
170
+ "predicted": 3787
171
  },
172
  "3": {
173
+ "accuracy": 0.8446482379143194,
174
+ "precision": 0.8279227450174645,
175
+ "recall": 0.8446482379143194,
176
+ "f1": 0.8362018650860561,
177
+ "gt_support": 38165,
178
+ "predicted": 38936
179
  },
180
  "4": {
181
+ "accuracy": 0.9990197404002163,
182
+ "precision": 0.9963255124056095,
183
+ "recall": 0.9990197404002163,
184
+ "f1": 0.9976708074534161,
185
+ "gt_support": 29584,
186
+ "predicted": 29664
187
  }
188
  },
189
+ "total_samples": 95400
190
  },
191
+ "0.6": {
192
+ "accuracy": 0.8105555555555556,
193
+ "baseline_accuracy": 0.3553909465020576,
194
  "per_class": {
195
  "0": {
196
+ "accuracy": 0.7384259259259259,
197
+ "precision": 0.0915877117427505,
198
+ "recall": 0.7384259259259259,
199
+ "f1": 0.16296296296296295,
200
+ "gt_support": 432,
201
+ "predicted": 3483
202
  },
203
  "1": {
204
+ "accuracy": 0.6471084988515066,
205
+ "precision": 0.8582500784015054,
206
+ "recall": 0.6471084988515066,
207
+ "f1": 0.7378719306692345,
208
+ "gt_support": 29604,
209
+ "predicted": 22321
210
  },
211
  "2": {
212
+ "accuracy": 0.7453703703703703,
213
+ "precision": 0.08068153345026309,
214
+ "recall": 0.7453703703703703,
215
+ "f1": 0.1456025322179516,
216
+ "gt_support": 432,
217
+ "predicted": 3991
218
  },
219
  "3": {
220
+ "accuracy": 0.7601901329688082,
221
+ "precision": 0.7468941729495436,
222
+ "recall": 0.7601901329688082,
223
+ "f1": 0.7534835024403763,
224
+ "gt_support": 32188,
225
+ "predicted": 32761
226
  },
227
  "4": {
228
+ "accuracy": 0.9992762853172765,
229
+ "precision": 0.9963918716083593,
230
+ "recall": 0.9992762853172765,
231
+ "f1": 0.9978319939873966,
232
+ "gt_support": 34544,
233
+ "predicted": 34644
234
  }
235
  },
236
+ "total_samples": 97200
237
  },
238
+ "0.8": {
239
+ "accuracy": 0.7576638176638176,
240
+ "baseline_accuracy": 0.3470899470899471,
241
  "per_class": {
242
  "0": {
243
+ "accuracy": 0.7655677655677655,
244
+ "precision": 0.09865470852017937,
245
+ "recall": 0.7655677655677655,
246
+ "f1": 0.17478569935187122,
247
+ "gt_support": 546,
248
+ "predicted": 4237
249
  },
250
  "1": {
251
+ "accuracy": 0.6668621013133208,
252
+ "precision": 0.7276658904214756,
253
+ "recall": 0.6668621013133208,
254
+ "f1": 0.6959384216258551,
255
+ "gt_support": 42640,
256
+ "predicted": 39077
257
  },
258
  "2": {
259
+ "accuracy": 0.7344322344322345,
260
+ "precision": 0.08133874239350913,
261
+ "recall": 0.7344322344322345,
262
+ "f1": 0.1464572680788897,
263
+ "gt_support": 546,
264
+ "predicted": 4930
265
  },
266
  "3": {
267
+ "accuracy": 0.5844802262960344,
268
+ "precision": 0.6685228969636635,
269
+ "recall": 0.5844802262960344,
270
+ "f1": 0.6236830648672181,
271
+ "gt_support": 36766,
272
+ "predicted": 32144
273
  },
274
  "4": {
275
+ "accuracy": 0.9996222138269739,
276
+ "precision": 0.9970326409495549,
277
+ "recall": 0.9996222138269739,
278
+ "f1": 0.9983257481076236,
279
+ "gt_support": 42352,
280
+ "predicted": 42462
281
  }
282
  },
283
+ "total_samples": 122850
284
+ },
285
+ "1.0": {
286
+ "accuracy": 0.7265130222971707,
287
+ "baseline_accuracy": 0.4095034663668728,
 
 
288
  "per_class": {
289
  "0": {
290
+ "accuracy": 0.6779089376053963,
291
+ "precision": 0.10384913459054508,
292
+ "recall": 0.6779089376053963,
293
+ "f1": 0.18010752688172044,
294
+ "gt_support": 1186,
295
+ "predicted": 7742
296
  },
297
  "1": {
298
+ "accuracy": 0.714484424759325,
299
+ "precision": 0.7030960142643589,
300
+ "recall": 0.714484424759325,
301
+ "f1": 0.7087444739971496,
302
+ "gt_support": 109276,
303
+ "predicted": 111046
304
  },
305
  "2": {
306
+ "accuracy": 0.6686340640809444,
307
+ "precision": 0.07407061460863068,
308
+ "recall": 0.6686340640809444,
309
+ "f1": 0.13336696939118736,
310
+ "gt_support": 1186,
311
+ "predicted": 10706
312
  },
313
  "3": {
314
+ "accuracy": 0.46584574314951976,
315
+ "precision": 0.6094466890364993,
316
+ "recall": 0.46584574314951976,
317
+ "f1": 0.5280575539568345,
318
+ "gt_support": 76418,
319
+ "predicted": 58412
320
  },
321
  "4": {
322
+ "accuracy": 0.9976391145410236,
323
+ "precision": 0.9956171463315768,
324
+ "recall": 0.9976391145410236,
325
+ "f1": 0.9966271048894298,
326
+ "gt_support": 78784,
327
+ "predicted": 78944
328
  }
329
  },
330
+ "total_samples": 266850
331
+ }
332
+ },
333
+ "total_trajectories": 300,
334
+ "total_steps": 3223,
335
+ "single_size_mode": false,
336
+ "config": {
337
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.pt",
338
+ "trajectories_dir": "reveng/trajectories_test_full",
339
+ "activations_dir": "interp/activations_test_full",
340
+ "layers": "23",
341
+ "steps": "all",
342
+ "token_categories": {
343
+ "prompt_suffix": "all"
344
  },
345
+ "pad_to_size": 15
346
+ },
347
+ "by_size": {
348
+ "7": {
349
+ "accuracy": 0.9465889665889666,
350
+ "baseline_accuracy": 0.7822222222222223,
351
  "per_class": {
352
  "0": {
353
+ "accuracy": 0.965034965034965,
354
+ "precision": 0.23173803526448364,
355
+ "recall": 0.965034965034965,
356
+ "f1": 0.37373053486797564,
357
+ "gt_support": 286,
358
+ "predicted": 1191
359
  },
360
  "1": {
361
+ "accuracy": 0.8083657587548638,
362
+ "precision": 0.9191206968063045,
363
+ "recall": 0.8083657587548638,
364
+ "f1": 0.8601927929093616,
365
+ "gt_support": 8224,
366
+ "predicted": 7233
367
  },
368
  "2": {
369
+ "accuracy": 0.9790209790209791,
370
+ "precision": 0.1959412176347096,
371
+ "recall": 0.9790209790209791,
372
+ "f1": 0.32653061224489793,
373
+ "gt_support": 286,
374
+ "predicted": 1429
375
  },
376
  "3": {
377
+ "accuracy": 0.6500574932924492,
378
+ "precision": 0.8116774347930127,
379
+ "recall": 0.6500574932924492,
380
+ "f1": 0.7219325316590401,
381
+ "gt_support": 5218,
382
+ "predicted": 4179
383
  },
384
  "4": {
385
+ "accuracy": 0.9996225365543547,
386
+ "precision": 0.9999801263961207,
387
+ "recall": 0.9996225365543547,
388
+ "f1": 0.9998012995012618,
389
+ "gt_support": 50336,
390
+ "predicted": 50318
391
  }
392
  },
393
+ "total_samples": 64350
394
  },
395
+ "9": {
396
+ "accuracy": 0.8837354085603113,
397
+ "baseline_accuracy": 0.64,
398
  "per_class": {
399
  "0": {
400
+ "accuracy": 0.9085603112840467,
401
+ "precision": 0.11063728974176736,
402
+ "recall": 0.9085603112840467,
403
+ "f1": 0.1972544878563886,
404
+ "gt_support": 514,
405
+ "predicted": 4221
406
  },
407
  "1": {
408
+ "accuracy": 0.6361711513012792,
409
+ "precision": 0.9361287809944178,
410
+ "recall": 0.6361711513012792,
411
+ "f1": 0.7575375564660153,
412
+ "gt_support": 22670,
413
+ "predicted": 15406
414
  },
415
  "2": {
416
+ "accuracy": 0.8871595330739299,
417
+ "precision": 0.11776859504132231,
418
+ "recall": 0.8871595330739299,
419
+ "f1": 0.2079343365253078,
420
+ "gt_support": 514,
421
+ "predicted": 3872
422
  },
423
  "3": {
424
+ "accuracy": 0.7171052631578947,
425
+ "precision": 0.7400886126934806,
426
+ "recall": 0.7171052631578947,
427
+ "f1": 0.7284156873849639,
428
+ "gt_support": 17936,
429
+ "predicted": 17379
430
  },
431
  "4": {
432
+ "accuracy": 0.999743298746217,
433
+ "precision": 0.9896351575456053,
434
+ "recall": 0.999743298746217,
435
+ "f1": 0.9946635481356023,
436
+ "gt_support": 74016,
437
+ "predicted": 74772
438
  }
439
  },
440
+ "total_samples": 115650
441
  },
442
+ "11": {
443
+ "accuracy": 0.8392398561890088,
444
+ "baseline_accuracy": 0.4622222222222222,
445
  "per_class": {
446
  "0": {
447
+ "accuracy": 0.8258859784283513,
448
+ "precision": 0.10952186350633429,
449
+ "recall": 0.8258859784283513,
450
+ "f1": 0.19339707739491252,
451
+ "gt_support": 649,
452
+ "predicted": 4894
453
  },
454
  "1": {
455
+ "accuracy": 0.7662561451522985,
456
+ "precision": 0.7914463696801549,
457
+ "recall": 0.7662561451522985,
458
+ "f1": 0.7786475768656331,
459
+ "gt_support": 39462,
460
+ "predicted": 38206
461
  },
462
  "2": {
463
+ "accuracy": 0.7211093990755008,
464
+ "precision": 0.08861957962507101,
465
+ "recall": 0.7211093990755008,
466
+ "f1": 0.1578414839797639,
467
+ "gt_support": 649,
468
+ "predicted": 5281
469
  },
470
  "3": {
471
+ "accuracy": 0.6310201487992798,
472
+ "precision": 0.7921624675929003,
473
+ "recall": 0.6310201487992798,
474
+ "f1": 0.7024684990052317,
475
+ "gt_support": 37769,
476
+ "predicted": 30086
477
  },
478
  "4": {
479
+ "accuracy": 0.9996888704515823,
480
+ "precision": 0.9987714260339264,
481
+ "recall": 0.9996888704515823,
482
+ "f1": 0.9992299376545678,
483
+ "gt_support": 67496,
484
+ "predicted": 67558
485
  }
486
  },
487
+ "total_samples": 146025
488
  },
489
+ "13": {
490
+ "accuracy": 0.7878244170096022,
491
+ "baseline_accuracy": 0.37721810699588476,
492
  "per_class": {
493
  "0": {
494
+ "accuracy": 0.6728395061728395,
495
+ "precision": 0.09747808978715793,
496
+ "recall": 0.6728395061728395,
497
+ "f1": 0.17028589282924542,
498
+ "gt_support": 810,
499
+ "predicted": 5591
500
  },
501
  "1": {
502
+ "accuracy": 0.6752202278945312,
503
+ "precision": 0.8312113697768237,
504
+ "recall": 0.6752202278945312,
505
+ "f1": 0.7451393497013935,
506
+ "gt_support": 66522,
507
+ "predicted": 54038
508
  },
509
  "2": {
510
+ "accuracy": 0.6037037037037037,
511
+ "precision": 0.06906779661016949,
512
+ "recall": 0.6037037037037037,
513
+ "f1": 0.12395437262357414,
514
+ "gt_support": 810,
515
+ "predicted": 7080
516
  },
517
  "3": {
518
+ "accuracy": 0.7635131203816838,
519
+ "precision": 0.7461689364000796,
520
+ "recall": 0.7635131203816838,
521
+ "f1": 0.754741397903576,
522
+ "gt_support": 68748,
523
+ "predicted": 70346
524
  },
525
  "4": {
526
+ "accuracy": 0.9951499118165785,
527
+ "precision": 0.9987830512224803,
528
+ "recall": 0.9951499118165785,
529
+ "f1": 0.9969631715531997,
530
+ "gt_support": 45360,
531
+ "predicted": 45195
532
  }
533
  },
534
+ "total_samples": 182250
535
  },
536
+ "15": {
537
+ "accuracy": 0.6770769940064546,
538
+ "baseline_accuracy": 0.5240894421392347,
539
  "per_class": {
540
  "0": {
541
+ "accuracy": 0.6265560165975104,
542
+ "precision": 0.06530435722780842,
543
+ "recall": 0.6265560165975104,
544
+ "f1": 0.11828062273572898,
545
+ "gt_support": 964,
546
+ "predicted": 9249
547
  },
548
  "1": {
549
+ "accuracy": 0.7308409923294865,
550
+ "precision": 0.7130116536646441,
551
+ "recall": 0.7308409923294865,
552
+ "f1": 0.7218162406704139,
553
+ "gt_support": 101297,
554
+ "predicted": 103830
555
  },
556
  "2": {
557
+ "accuracy": 0.6307053941908713,
558
+ "precision": 0.05798760133524082,
559
+ "recall": 0.6307053941908713,
560
+ "f1": 0.10621014935802255,
561
+ "gt_support": 964,
562
+ "predicted": 10485
563
+ },
564
+ "3": {
565
+ "accuracy": 0.6299890037387288,
566
+ "precision": 0.7672709351161395,
567
+ "recall": 0.6299890037387288,
568
+ "f1": 0.6918859384283927,
569
+ "gt_support": 113675,
570
+ "predicted": 93336
571
+ },
572
+ "4": {
573
  "accuracy": 0.0,
574
  "precision": 0.0,
575
  "recall": 0.0,
576
  "f1": 0.0,
577
+ "gt_support": 0,
578
  "predicted": 0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
579
  }
580
  },
581
+ "total_samples": 216900
582
  }
583
  },
584
  "by_size_complexity": {
585
  "7_0.0": {
586
+ "accuracy": 0.9671345029239766,
587
  "baseline_accuracy": 0.7822222222222223,
588
  "per_class": {
589
  "0": {
590
+ "accuracy": 1.0,
591
+ "precision": 0.25675675675675674,
592
+ "recall": 1.0,
593
+ "f1": 0.4086021505376344,
594
  "gt_support": 38,
595
+ "predicted": 148
596
  },
597
  "1": {
598
+ "accuracy": 0.9758771929824561,
599
+ "precision": 0.9632034632034632,
600
+ "recall": 0.9758771929824561,
601
+ "f1": 0.9694989106753812,
602
  "gt_support": 912,
603
+ "predicted": 924
604
  },
605
  "2": {
606
+ "accuracy": 0.9736842105263158,
607
+ "precision": 0.2138728323699422,
608
+ "recall": 0.9736842105263158,
609
+ "f1": 0.35071090047393366,
610
  "gt_support": 38,
611
+ "predicted": 173
612
  },
613
  "3": {
614
+ "accuracy": 0.7059496567505721,
615
+ "precision": 0.9983818770226537,
616
+ "recall": 0.7059496567505721,
617
+ "f1": 0.8270777479892761,
618
  "gt_support": 874,
619
+ "predicted": 618
620
  },
621
  "4": {
622
+ "accuracy": 0.9998504784688995,
623
+ "precision": 1.0,
624
+ "recall": 0.9998504784688995,
625
+ "f1": 0.9999252336448599,
626
  "gt_support": 6688,
627
+ "predicted": 6687
628
  }
629
  },
630
  "total_samples": 8550
631
  },
632
  "7_0.2": {
633
+ "accuracy": 0.9588888888888889,
634
  "baseline_accuracy": 0.7822222222222223,
635
  "per_class": {
636
  "0": {
637
+ "accuracy": 0.9318181818181818,
638
+ "precision": 0.23295454545454544,
639
+ "recall": 0.9318181818181818,
640
+ "f1": 0.3727272727272727,
641
  "gt_support": 44,
642
+ "predicted": 176
643
  },
644
  "1": {
645
+ "accuracy": 0.8767482517482518,
646
+ "precision": 0.9672131147540983,
647
+ "recall": 0.8767482517482518,
648
+ "f1": 0.9197615772581386,
649
  "gt_support": 1144,
650
+ "predicted": 1037
651
  },
652
  "2": {
653
+ "accuracy": 1.0,
654
+ "precision": 0.19555555555555557,
655
+ "recall": 1.0,
656
+ "f1": 0.32713754646840154,
657
  "gt_support": 44,
658
+ "predicted": 225
659
  },
660
  "3": {
661
+ "accuracy": 0.7207792207792207,
662
+ "precision": 0.921161825726141,
663
+ "recall": 0.7207792207792207,
664
+ "f1": 0.808743169398907,
665
  "gt_support": 924,
666
+ "predicted": 723
667
  },
668
  "4": {
669
+ "accuracy": 0.9993543388429752,
670
+ "precision": 1.0,
671
+ "recall": 0.9993543388429752,
672
+ "f1": 0.999677065168249,
673
  "gt_support": 7744,
674
+ "predicted": 7739
675
  }
676
  },
677
  "total_samples": 9900
678
  },
679
  "7_0.4": {
680
+ "accuracy": 0.94579945799458,
681
  "baseline_accuracy": 0.7822222222222223,
682
  "per_class": {
683
  "0": {
684
+ "accuracy": 0.9512195121951219,
685
+ "precision": 0.21195652173913043,
686
+ "recall": 0.9512195121951219,
687
+ "f1": 0.3466666666666667,
688
  "gt_support": 41,
689
+ "predicted": 184
690
  },
691
  "1": {
692
+ "accuracy": 0.7883275261324042,
693
+ "precision": 0.9427083333333334,
694
+ "recall": 0.7883275261324042,
695
+ "f1": 0.8586337760910816,
696
  "gt_support": 1148,
697
+ "predicted": 960
698
  },
699
  "2": {
700
+ "accuracy": 0.975609756097561,
701
+ "precision": 0.1834862385321101,
702
+ "recall": 0.975609756097561,
703
+ "f1": 0.3088803088803089,
704
  "gt_support": 41,
705
+ "predicted": 218
706
  },
707
  "3": {
708
+ "accuracy": 0.6777920410783055,
709
+ "precision": 0.8123076923076923,
710
+ "recall": 0.6777920410783055,
711
+ "f1": 0.7389783065080475,
712
  "gt_support": 779,
713
+ "predicted": 650
714
  },
715
  "4": {
716
+ "accuracy": 0.9995842572062085,
717
+ "precision": 1.0,
718
+ "recall": 0.9995842572062085,
719
+ "f1": 0.9997920853836025,
720
  "gt_support": 7216,
721
+ "predicted": 7213
722
  }
723
  },
724
  "total_samples": 9225
725
  },
726
  "7_0.6": {
727
+ "accuracy": 0.948274231678487,
728
  "baseline_accuracy": 0.7822222222222223,
729
  "per_class": {
730
  "0": {
731
+ "accuracy": 0.9148936170212766,
732
+ "precision": 0.2529411764705882,
733
+ "recall": 0.9148936170212766,
734
+ "f1": 0.3963133640552996,
735
  "gt_support": 47,
736
+ "predicted": 170
737
  },
738
  "1": {
739
+ "accuracy": 0.8114453411592076,
740
+ "precision": 0.9278523489932886,
741
+ "recall": 0.8114453411592076,
742
+ "f1": 0.8657534246575342,
743
  "gt_support": 1363,
744
+ "predicted": 1192
745
  },
746
  "2": {
747
+ "accuracy": 0.9574468085106383,
748
+ "precision": 0.1836734693877551,
749
+ "recall": 0.9574468085106383,
750
+ "f1": 0.3082191780821918,
751
  "gt_support": 47,
752
+ "predicted": 245
753
  },
754
  "3": {
755
+ "accuracy": 0.6690307328605201,
756
+ "precision": 0.8085714285714286,
757
+ "recall": 0.6690307328605201,
758
+ "f1": 0.7322121604139714,
759
  "gt_support": 846,
760
+ "predicted": 700
761
  },
762
  "4": {
763
+ "accuracy": 0.9995164410058027,
764
+ "precision": 1.0,
765
+ "recall": 0.9995164410058027,
766
+ "f1": 0.9997581620314389,
767
  "gt_support": 8272,
768
+ "predicted": 8268
769
  }
770
  },
771
  "total_samples": 10575
772
  },
773
  "7_0.8": {
774
+ "accuracy": 0.9381010101010101,
775
  "baseline_accuracy": 0.7822222222222223,
776
  "per_class": {
777
  "0": {
778
+ "accuracy": 1.0,
779
+ "precision": 0.24017467248908297,
780
+ "recall": 1.0,
781
+ "f1": 0.3873239436619718,
782
  "gt_support": 55,
783
+ "predicted": 229
784
  },
785
  "1": {
786
+ "accuracy": 0.7601173020527859,
787
+ "precision": 0.8822328114363512,
788
+ "recall": 0.7601173020527859,
789
+ "f1": 0.8166351606805293,
790
  "gt_support": 1705,
791
+ "predicted": 1469
792
  },
793
  "2": {
794
+ "accuracy": 0.9636363636363636,
795
+ "precision": 0.20075757575757575,
796
+ "recall": 0.9636363636363636,
797
+ "f1": 0.3322884012539185,
798
  "gt_support": 55,
799
+ "predicted": 264
800
  },
801
  "3": {
802
+ "accuracy": 0.5977272727272728,
803
+ "precision": 0.7166212534059946,
804
+ "recall": 0.5977272727272728,
805
+ "f1": 0.6517967781908303,
806
  "gt_support": 880,
807
+ "predicted": 734
808
  },
809
  "4": {
810
+ "accuracy": 0.999896694214876,
811
+ "precision": 1.0,
812
+ "recall": 0.999896694214876,
813
+ "f1": 0.9999483444392789,
814
  "gt_support": 9680,
815
+ "predicted": 9679
816
  }
817
  },
818
  "total_samples": 12375
819
  },
820
  "7_1.0": {
821
+ "accuracy": 0.9318032786885246,
822
  "baseline_accuracy": 0.7822222222222223,
823
  "per_class": {
824
  "0": {
825
+ "accuracy": 0.9836065573770492,
826
+ "precision": 0.2112676056338028,
827
+ "recall": 0.9836065573770492,
828
+ "f1": 0.34782608695652173,
829
  "gt_support": 61,
830
+ "predicted": 284
831
  },
832
  "1": {
833
+ "accuracy": 0.7418032786885246,
834
+ "precision": 0.8770442156268928,
835
+ "recall": 0.7418032786885246,
836
+ "f1": 0.803774632250902,
837
  "gt_support": 1952,
838
+ "predicted": 1651
839
  },
840
  "2": {
841
+ "accuracy": 1.0,
842
+ "precision": 0.20065789473684212,
843
+ "recall": 1.0,
844
+ "f1": 0.33424657534246577,
845
  "gt_support": 61,
846
+ "predicted": 304
847
  },
848
  "3": {
849
+ "accuracy": 0.5344262295081967,
850
+ "precision": 0.6485411140583555,
851
+ "recall": 0.5344262295081967,
852
+ "f1": 0.5859796285200719,
853
  "gt_support": 915,
854
+ "predicted": 754
855
  },
856
  "4": {
857
+ "accuracy": 0.9995342771982116,
858
+ "precision": 0.9999068207230712,
859
+ "recall": 0.9995342771982116,
860
+ "f1": 0.9997205142537731,
861
  "gt_support": 10736,
862
+ "predicted": 10732
863
  }
864
  },
865
  "total_samples": 13725
866
  },
867
  "9_0.0": {
868
+ "accuracy": 0.9208230452674897,
869
  "baseline_accuracy": 0.64,
870
  "per_class": {
871
  "0": {
872
+ "accuracy": 0.7777777777777778,
873
+ "precision": 0.1065989847715736,
874
+ "recall": 0.7777777777777778,
875
+ "f1": 0.1875,
876
  "gt_support": 54,
877
+ "predicted": 394
878
  },
879
  "1": {
880
+ "accuracy": 0.8020833333333334,
881
+ "precision": 0.9914163090128756,
882
+ "recall": 0.8020833333333334,
883
+ "f1": 0.886756238003839,
884
  "gt_support": 1728,
885
+ "predicted": 1398
886
  },
887
  "2": {
888
+ "accuracy": 0.8518518518518519,
889
+ "precision": 0.1220159151193634,
890
+ "recall": 0.8518518518518519,
891
+ "f1": 0.21345707656612528,
892
  "gt_support": 54,
893
+ "predicted": 377
894
  },
895
  "3": {
896
+ "accuracy": 0.7635933806146572,
897
+ "precision": 0.9137199434229137,
898
+ "recall": 0.7635933806146572,
899
+ "f1": 0.831938184159691,
900
  "gt_support": 2538,
901
+ "predicted": 2121
902
  },
903
  "4": {
904
+ "accuracy": 1.0,
905
+ "precision": 0.9893129770992366,
906
+ "recall": 1.0,
907
+ "f1": 0.9946277820414428,
908
  "gt_support": 7776,
909
+ "predicted": 7860
910
  }
911
  },
912
  "total_samples": 12150
913
  },
914
  "9_0.2": {
915
+ "accuracy": 0.9122458628841608,
916
  "baseline_accuracy": 0.64,
917
  "per_class": {
918
  "0": {
919
+ "accuracy": 0.8085106382978723,
920
+ "precision": 0.10410958904109589,
921
+ "recall": 0.8085106382978723,
922
+ "f1": 0.18446601941747573,
923
  "gt_support": 47,
924
+ "predicted": 365
925
  },
926
  "1": {
927
+ "accuracy": 0.7559101654846335,
928
+ "precision": 0.9914728682170543,
929
+ "recall": 0.7559101654846335,
930
+ "f1": 0.8578135479543929,
931
  "gt_support": 1692,
932
+ "predicted": 1290
933
  },
934
  "2": {
935
+ "accuracy": 0.723404255319149,
936
+ "precision": 0.1059190031152648,
937
+ "recall": 0.723404255319149,
938
+ "f1": 0.1847826086956522,
939
  "gt_support": 47,
940
+ "predicted": 321
941
  },
942
  "3": {
943
+ "accuracy": 0.7570509648688768,
944
+ "precision": 0.8747855917667239,
945
+ "recall": 0.7570509648688768,
946
+ "f1": 0.8116710875331565,
947
  "gt_support": 2021,
948
+ "predicted": 1749
949
  },
950
  "4": {
951
+ "accuracy": 0.9997044917257684,
952
+ "precision": 0.9877372262773723,
953
+ "recall": 0.9997044917257684,
954
+ "f1": 0.9936848289029226,
955
  "gt_support": 6768,
956
+ "predicted": 6850
957
  }
958
  },
959
  "total_samples": 10575
960
  },
961
  "9_0.4": {
962
+ "accuracy": 0.9020895522388059,
963
  "baseline_accuracy": 0.64,
964
  "per_class": {
965
  "0": {
966
+ "accuracy": 0.9402985074626866,
967
+ "precision": 0.12209302325581395,
968
+ "recall": 0.9402985074626866,
969
+ "f1": 0.21612349914236706,
970
  "gt_support": 67,
971
+ "predicted": 516
972
  },
973
  "1": {
974
+ "accuracy": 0.6798507462686567,
975
+ "precision": 0.9785177228786252,
976
+ "recall": 0.6798507462686567,
977
+ "f1": 0.802289740202554,
978
  "gt_support": 2680,
979
+ "predicted": 1862
980
  },
981
  "2": {
982
+ "accuracy": 0.9850746268656716,
983
+ "precision": 0.12571428571428572,
984
+ "recall": 0.9850746268656716,
985
+ "f1": 0.222972972972973,
986
  "gt_support": 67,
987
+ "predicted": 525
988
  },
989
  "3": {
990
+ "accuracy": 0.7680826636050516,
991
+ "precision": 0.8215309046254605,
992
+ "recall": 0.7680826636050516,
993
+ "f1": 0.7939082278481012,
994
  "gt_support": 2613,
995
+ "predicted": 2443
996
  },
997
  "4": {
998
+ "accuracy": 0.9992744610281924,
999
+ "precision": 0.9909548771713435,
1000
+ "recall": 0.9992744610281924,
1001
+ "f1": 0.9950972802807454,
1002
  "gt_support": 9648,
1003
+ "predicted": 9729
1004
  }
1005
  },
1006
  "total_samples": 15075
1007
  },
1008
  "9_0.6": {
1009
+ "accuracy": 0.8903460837887067,
1010
  "baseline_accuracy": 0.64,
1011
  "per_class": {
1012
  "0": {
1013
+ "accuracy": 0.9180327868852459,
1014
+ "precision": 0.11522633744855967,
1015
+ "recall": 0.9180327868852459,
1016
+ "f1": 0.20475319926873858,
1017
  "gt_support": 61,
1018
+ "predicted": 486
1019
  },
1020
  "1": {
1021
+ "accuracy": 0.6523065192527641,
1022
+ "precision": 0.9672131147540983,
1023
+ "recall": 0.6523065192527641,
1024
+ "f1": 0.779143897996357,
1025
  "gt_support": 2623,
1026
+ "predicted": 1769
1027
  },
1028
  "2": {
1029
+ "accuracy": 0.9016393442622951,
1030
+ "precision": 0.1206140350877193,
1031
+ "recall": 0.9016393442622951,
1032
+ "f1": 0.2127659574468085,
1033
  "gt_support": 61,
1034
+ "predicted": 456
1035
  },
1036
  "3": {
1037
+ "accuracy": 0.7363387978142076,
1038
+ "precision": 0.7570224719101124,
1039
+ "recall": 0.7363387978142076,
1040
+ "f1": 0.7465373961218837,
1041
  "gt_support": 2196,
1042
+ "predicted": 2136
1043
  },
1044
  "4": {
1045
+ "accuracy": 0.9996584699453552,
1046
+ "precision": 0.989074115791845,
1047
+ "recall": 0.9996584699453552,
1048
+ "f1": 0.994338127052429,
1049
  "gt_support": 8784,
1050
+ "predicted": 8878
1051
  }
1052
  },
1053
  "total_samples": 13725
1054
  },
1055
  "9_0.8": {
1056
+ "accuracy": 0.8773157315731573,
1057
  "baseline_accuracy": 0.64,
1058
  "per_class": {
1059
  "0": {
1060
+ "accuracy": 0.9405940594059405,
1061
+ "precision": 0.10957324106113034,
1062
+ "recall": 0.9405940594059405,
1063
+ "f1": 0.1962809917355372,
1064
  "gt_support": 101,
1065
+ "predicted": 867
1066
  },
1067
  "1": {
1068
+ "accuracy": 0.6180745734147883,
1069
+ "precision": 0.9349904397705545,
1070
+ "recall": 0.6180745734147883,
1071
+ "f1": 0.7441978440076095,
1072
  "gt_support": 4747,
1073
+ "predicted": 3138
1074
  },
1075
  "2": {
1076
+ "accuracy": 0.8712871287128713,
1077
+ "precision": 0.1116751269035533,
1078
+ "recall": 0.8712871287128713,
1079
+ "f1": 0.19797525309336333,
1080
  "gt_support": 101,
1081
+ "predicted": 788
1082
  },
1083
  "3": {
1084
+ "accuracy": 0.7051361386138614,
1085
+ "precision": 0.6958778625954198,
1086
+ "recall": 0.7051361386138614,
1087
+ "f1": 0.7004764100199785,
1088
  "gt_support": 3232,
1089
+ "predicted": 3275
1090
  },
1091
  "4": {
1092
+ "accuracy": 0.9997937293729373,
1093
+ "precision": 0.9920856928430102,
1094
+ "recall": 0.9997937293729373,
1095
+ "f1": 0.9959247970959898,
1096
  "gt_support": 14544,
1097
+ "predicted": 14657
1098
  }
1099
  },
1100
  "total_samples": 22725
1101
  },
1102
  "9_1.0": {
1103
+ "accuracy": 0.8602173913043478,
1104
  "baseline_accuracy": 0.64,
1105
  "per_class": {
1106
  "0": {
1107
+ "accuracy": 0.9402173913043478,
1108
+ "precision": 0.10860012554927809,
1109
+ "recall": 0.9402173913043478,
1110
+ "f1": 0.19471018570624649,
1111
  "gt_support": 184,
1112
+ "predicted": 1593
1113
  },
1114
  "1": {
1115
+ "accuracy": 0.575,
1116
+ "precision": 0.8892250798453522,
1117
+ "recall": 0.575,
1118
+ "f1": 0.6983959337249983,
1119
  "gt_support": 9200,
1120
+ "predicted": 5949
1121
  },
1122
  "2": {
1123
+ "accuracy": 0.907608695652174,
1124
+ "precision": 0.11886120996441281,
1125
+ "recall": 0.907608695652174,
1126
+ "f1": 0.21019509125236,
1127
  "gt_support": 184,
1128
+ "predicted": 1405
1129
  },
1130
  "3": {
1131
+ "accuracy": 0.6542353823088456,
1132
+ "precision": 0.6173297966401414,
1133
+ "recall": 0.6542353823088456,
1134
+ "f1": 0.6352470202893276,
1135
  "gt_support": 5336,
1136
+ "predicted": 5655
1137
  },
1138
  "4": {
1139
+ "accuracy": 0.9998490338164251,
1140
+ "precision": 0.9885812374057765,
1141
+ "recall": 0.9998490338164251,
1142
+ "f1": 0.9941832101174616,
1143
  "gt_support": 26496,
1144
+ "predicted": 26798
1145
  }
1146
  },
1147
  "total_samples": 41400
1148
  },
1149
  "11_0.0": {
1150
+ "accuracy": 0.9388888888888889,
1151
  "baseline_accuracy": 0.4622222222222222,
1152
  "per_class": {
1153
  "0": {
1154
+ "accuracy": 0.9264705882352942,
1155
+ "precision": 0.09828393135725429,
1156
+ "recall": 0.9264705882352942,
1157
+ "f1": 0.1777150916784203,
1158
  "gt_support": 68,
1159
+ "predicted": 641
1160
  },
1161
  "1": {
1162
+ "accuracy": 0.9838235294117647,
1163
+ "precision": 0.9863619609288611,
1164
+ "recall": 0.9838235294117647,
1165
+ "f1": 0.985091109884042,
1166
  "gt_support": 2720,
1167
+ "predicted": 2713
1168
  },
1169
  "2": {
1170
+ "accuracy": 0.75,
1171
+ "precision": 0.1457142857142857,
1172
+ "recall": 0.75,
1173
+ "f1": 0.24401913875598089,
1174
  "gt_support": 68,
1175
+ "predicted": 350
1176
  },
1177
  "3": {
1178
+ "accuracy": 0.8382352941176471,
1179
+ "precision": 0.9980053191489362,
1180
+ "recall": 0.8382352941176471,
1181
+ "f1": 0.9111695669769324,
1182
  "gt_support": 5372,
1183
+ "predicted": 4512
1184
  },
1185
  "4": {
1186
+ "accuracy": 1.0,
1187
+ "precision": 0.9983060417843026,
1188
+ "recall": 1.0,
1189
+ "f1": 0.9991523029104267,
1190
  "gt_support": 7072,
1191
  "predicted": 7084
1192
  }
 
1194
  "total_samples": 15300
1195
  },
1196
  "11_0.2": {
1197
+ "accuracy": 0.8983006535947713,
1198
  "baseline_accuracy": 0.4622222222222222,
1199
  "per_class": {
1200
  "0": {
1201
+ "accuracy": 0.8941176470588236,
1202
+ "precision": 0.09417596034696406,
1203
+ "recall": 0.8941176470588236,
1204
+ "f1": 0.17040358744394618,
1205
  "gt_support": 85,
1206
+ "predicted": 807
1207
  },
1208
  "1": {
1209
+ "accuracy": 0.8382978723404255,
1210
+ "precision": 0.973263586166812,
1211
+ "recall": 0.8382978723404255,
1212
+ "f1": 0.9007530930607853,
1213
  "gt_support": 3995,
1214
+ "predicted": 3441
1215
  },
1216
  "2": {
1217
+ "accuracy": 0.8705882352941177,
1218
+ "precision": 0.10378681626928471,
1219
+ "recall": 0.8705882352941177,
1220
+ "f1": 0.18546365914786966,
1221
  "gt_support": 85,
1222
+ "predicted": 713
1223
  },
1224
  "3": {
1225
+ "accuracy": 0.7910130718954248,
1226
+ "precision": 0.9123633622314361,
1227
+ "recall": 0.7910130718954248,
1228
+ "f1": 0.8473656572728864,
1229
  "gt_support": 6120,
1230
+ "predicted": 5306
1231
  },
1232
  "4": {
1233
+ "accuracy": 1.0,
1234
+ "precision": 0.9979679385865884,
1235
+ "recall": 1.0,
1236
+ "f1": 0.9989829359249632,
1237
  "gt_support": 8840,
1238
+ "predicted": 8858
1239
  }
1240
  },
1241
  "total_samples": 19125
1242
  },
1243
  "11_0.4": {
1244
+ "accuracy": 0.8774141414141414,
1245
  "baseline_accuracy": 0.4622222222222222,
1246
  "per_class": {
1247
  "0": {
1248
+ "accuracy": 0.8545454545454545,
1249
+ "precision": 0.10514541387024609,
1250
+ "recall": 0.8545454545454545,
1251
+ "f1": 0.18725099601593626,
1252
  "gt_support": 55,
1253
+ "predicted": 447
1254
  },
1255
  "1": {
1256
+ "accuracy": 0.7495711835334476,
1257
+ "precision": 0.9446606139213143,
1258
+ "recall": 0.7495711835334476,
1259
+ "f1": 0.8358837031369547,
1260
  "gt_support": 2915,
1261
+ "predicted": 2313
1262
  },
1263
  "2": {
1264
+ "accuracy": 0.6909090909090909,
1265
+ "precision": 0.084070796460177,
1266
+ "recall": 0.6909090909090909,
1267
+ "f1": 0.14990138067061146,
1268
  "gt_support": 55,
1269
+ "predicted": 452
1270
  },
1271
  "3": {
1272
+ "accuracy": 0.7900826446280992,
1273
+ "precision": 0.8349344978165939,
1274
+ "recall": 0.7900826446280992,
1275
+ "f1": 0.8118895966029723,
1276
  "gt_support": 3630,
1277
+ "predicted": 3435
1278
  },
1279
  "4": {
1280
+ "accuracy": 1.0,
1281
+ "precision": 0.9986033519553073,
1282
+ "recall": 1.0,
1283
+ "f1": 0.9993011879804332,
1284
  "gt_support": 5720,
1285
+ "predicted": 5728
1286
  }
1287
  },
1288
  "total_samples": 12375
1289
  },
1290
  "11_0.6": {
1291
+ "accuracy": 0.8253721682847897,
1292
  "baseline_accuracy": 0.4622222222222222,
1293
  "per_class": {
1294
  "0": {
1295
+ "accuracy": 0.7378640776699029,
1296
+ "precision": 0.08597285067873303,
1297
+ "recall": 0.7378640776699029,
1298
+ "f1": 0.15400202634245186,
1299
  "gt_support": 103,
1300
+ "predicted": 884
1301
  },
1302
  "1": {
1303
+ "accuracy": 0.7351132686084142,
1304
+ "precision": 0.7740671323905265,
1305
+ "recall": 0.7351132686084142,
1306
+ "f1": 0.7540874761390987,
1307
  "gt_support": 6180,
1308
+ "predicted": 5869
1309
  },
1310
  "2": {
1311
+ "accuracy": 0.8058252427184466,
1312
+ "precision": 0.09617612977983778,
1313
+ "recall": 0.8058252427184466,
1314
+ "f1": 0.17184265010351968,
1315
  "gt_support": 103,
1316
+ "predicted": 863
1317
  },
1318
  "3": {
1319
+ "accuracy": 0.6111568207997368,
1320
+ "precision": 0.7692626346313173,
1321
+ "recall": 0.6111568207997368,
1322
+ "f1": 0.6811554332874828,
1323
  "gt_support": 6077,
1324
+ "predicted": 4828
1325
  },
1326
  "4": {
1327
+ "accuracy": 1.0,
1328
+ "precision": 0.9982294287578045,
1329
+ "recall": 1.0,
1330
+ "f1": 0.9991139299538311,
1331
  "gt_support": 10712,
1332
+ "predicted": 10731
1333
  }
1334
  },
1335
  "total_samples": 23175
1336
  },
1337
  "11_0.8": {
1338
+ "accuracy": 0.8061598440545809,
1339
  "baseline_accuracy": 0.4622222222222222,
1340
  "per_class": {
1341
  "0": {
1342
+ "accuracy": 0.8157894736842105,
1343
+ "precision": 0.13517441860465115,
1344
+ "recall": 0.8157894736842105,
1345
+ "f1": 0.23192019950124684,
1346
  "gt_support": 114,
1347
+ "predicted": 688
1348
  },
1349
  "1": {
1350
+ "accuracy": 0.736709197235513,
1351
+ "precision": 0.7253336822821251,
1352
+ "recall": 0.736709197235513,
1353
+ "f1": 0.7309771858103653,
1354
  "gt_support": 7524,
1355
+ "predicted": 7642
1356
  },
1357
  "2": {
1358
+ "accuracy": 0.6403508771929824,
1359
+ "precision": 0.07549120992761117,
1360
+ "recall": 0.6403508771929824,
1361
+ "f1": 0.13506012950971322,
1362
  "gt_support": 114,
1363
+ "predicted": 967
1364
  },
1365
  "3": {
1366
+ "accuracy": 0.5153922542204568,
1367
+ "precision": 0.6935412026726058,
1368
+ "recall": 0.5153922542204568,
1369
+ "f1": 0.5913406760349411,
1370
  "gt_support": 6042,
1371
+ "predicted": 4490
1372
  },
1373
  "4": {
1374
+ "accuracy": 0.9999156545209177,
1375
+ "precision": 0.9993256343252128,
1376
+ "recall": 0.9999156545209177,
1377
+ "f1": 0.9996205573590792,
1378
  "gt_support": 11856,
1379
+ "predicted": 11863
1380
  }
1381
  },
1382
  "total_samples": 25650
1383
  },
1384
  "11_1.0": {
1385
+ "accuracy": 0.8004166666666667,
1386
  "baseline_accuracy": 0.4622222222222222,
1387
  "per_class": {
1388
  "0": {
1389
+ "accuracy": 0.8080357142857143,
1390
+ "precision": 0.12683952347582342,
1391
+ "recall": 0.8080357142857143,
1392
+ "f1": 0.21926105390672324,
1393
  "gt_support": 224,
1394
+ "predicted": 1427
1395
  },
1396
  "1": {
1397
+ "accuracy": 0.7404513888888888,
1398
+ "precision": 0.7358885876263249,
1399
+ "recall": 0.7404513888888888,
1400
+ "f1": 0.7381629373222894,
1401
  "gt_support": 16128,
1402
+ "predicted": 16228
1403
  },
1404
  "2": {
1405
+ "accuracy": 0.6651785714285714,
1406
+ "precision": 0.07696280991735537,
1407
+ "recall": 0.6651785714285714,
1408
+ "f1": 0.13796296296296295,
1409
  "gt_support": 224,
1410
+ "predicted": 1936
1411
  },
1412
  "3": {
1413
+ "accuracy": 0.45526215805471126,
1414
+ "precision": 0.6377910844976713,
1415
+ "recall": 0.45526215805471126,
1416
+ "f1": 0.5312863714459902,
1417
  "gt_support": 10528,
1418
+ "predicted": 7515
1419
  },
1420
  "4": {
1421
+ "accuracy": 0.9991414835164835,
1422
+ "precision": 0.99922726882459,
1423
+ "recall": 0.9991414835164835,
1424
+ "f1": 0.9991843743292551,
1425
  "gt_support": 23296,
1426
+ "predicted": 23294
1427
  }
1428
  },
1429
  "total_samples": 50400
1430
  },
1431
  "13_0.0": {
1432
+ "accuracy": 0.9322916666666666,
1433
  "baseline_accuracy": 0.5288888888888889,
1434
  "per_class": {
1435
  "0": {
1436
+ "accuracy": 0.8125,
1437
+ "precision": 0.08538587848932677,
1438
+ "recall": 0.8125,
1439
+ "f1": 0.15453194650817237,
1440
  "gt_support": 64,
1441
+ "predicted": 609
1442
  },
1443
  "1": {
1444
+ "accuracy": 0.9827473958333334,
1445
+ "precision": 0.9792409990269219,
1446
+ "recall": 0.9827473958333334,
1447
+ "f1": 0.9809910641754672,
1448
  "gt_support": 3072,
1449
+ "predicted": 3083
1450
  },
1451
  "2": {
1452
+ "accuracy": 0.5,
1453
+ "precision": 0.09523809523809523,
1454
+ "recall": 0.5,
1455
+ "f1": 0.16,
1456
  "gt_support": 64,
1457
+ "predicted": 336
1458
  },
1459
  "3": {
1460
+ "accuracy": 0.8864233193277311,
1461
+ "precision": 0.9938171647283969,
1462
+ "recall": 0.8864233193277311,
1463
+ "f1": 0.9370532306197515,
1464
  "gt_support": 7616,
1465
+ "predicted": 6793
1466
  },
1467
  "4": {
1468
+ "accuracy": 0.9963727678571429,
1469
+ "precision": 0.9977647387538419,
1470
+ "recall": 0.9963727678571429,
1471
+ "f1": 0.9970682674856904,
1472
  "gt_support": 3584,
1473
+ "predicted": 3579
1474
  }
1475
  },
1476
  "total_samples": 14400
1477
  },
1478
  "13_0.2": {
1479
+ "accuracy": 0.886594982078853,
1480
  "baseline_accuracy": 0.48444444444444446,
1481
  "per_class": {
1482
  "0": {
1483
+ "accuracy": 0.8709677419354839,
1484
+ "precision": 0.09574468085106383,
1485
+ "recall": 0.8709677419354839,
1486
+ "f1": 0.17252396166134185,
1487
  "gt_support": 62,
1488
+ "predicted": 564
1489
  },
1490
  "1": {
1491
+ "accuracy": 0.82119021134594,
1492
+ "precision": 0.9989851150202977,
1493
+ "recall": 0.82119021134594,
1494
+ "f1": 0.9014041514041514,
1495
  "gt_support": 3596,
1496
+ "predicted": 2956
1497
  },
1498
  "2": {
1499
+ "accuracy": 0.5967741935483871,
1500
+ "precision": 0.06764168190127971,
1501
+ "recall": 0.5967741935483871,
1502
+ "f1": 0.12151067323481117,
1503
  "gt_support": 62,
1504
+ "predicted": 547
1505
  },
1506
  "3": {
1507
+ "accuracy": 0.8662326131991713,
1508
+ "precision": 0.9131180783029169,
1509
+ "recall": 0.8662326131991713,
1510
+ "f1": 0.8890576353557598,
1511
  "gt_support": 6758,
1512
+ "predicted": 6411
1513
  },
1514
  "4": {
1515
+ "accuracy": 0.9994239631336406,
1516
+ "precision": 0.9994239631336406,
1517
+ "recall": 0.9994239631336406,
1518
+ "f1": 0.9994239631336406,
1519
  "gt_support": 3472,
1520
+ "predicted": 3472
1521
  }
1522
  },
1523
  "total_samples": 13950
1524
  },
1525
  "13_0.4": {
1526
+ "accuracy": 0.8528,
1527
  "baseline_accuracy": 0.44,
1528
  "per_class": {
1529
  "0": {
1530
+ "accuracy": 0.848,
1531
+ "precision": 0.10056925996204934,
1532
+ "recall": 0.848,
1533
+ "f1": 0.179813401187447,
1534
  "gt_support": 125,
1535
+ "predicted": 1054
1536
  },
1537
  "1": {
1538
+ "accuracy": 0.698,
1539
+ "precision": 0.9803370786516854,
1540
+ "recall": 0.698,
1541
+ "f1": 0.8154205607476634,
1542
  "gt_support": 8500,
1543
+ "predicted": 6052
1544
  },
1545
  "2": {
1546
+ "accuracy": 0.68,
1547
+ "precision": 0.09042553191489362,
1548
+ "recall": 0.68,
1549
+ "f1": 0.1596244131455399,
1550
  "gt_support": 125,
1551
+ "predicted": 940
1552
  },
1553
  "3": {
1554
+ "accuracy": 0.8791919191919192,
1555
+ "precision": 0.8314864348490638,
1556
+ "recall": 0.8791919191919192,
1557
+ "f1": 0.8546739984289081,
1558
  "gt_support": 12375,
1559
+ "predicted": 13085
1560
  },
1561
  "4": {
1562
+ "accuracy": 0.9972857142857143,
1563
+ "precision": 0.9981412639405205,
1564
+ "recall": 0.9972857142857143,
1565
+ "f1": 0.9977133057024439,
1566
  "gt_support": 7000,
1567
+ "predicted": 6994
1568
  }
1569
  },
1570
  "total_samples": 28125
1571
  },
1572
  "13_0.6": {
1573
+ "accuracy": 0.8123048668503214,
1574
  "baseline_accuracy": 0.39555555555555555,
1575
  "per_class": {
1576
  "0": {
1577
+ "accuracy": 0.6115702479338843,
1578
+ "precision": 0.08026030368763558,
1579
+ "recall": 0.6115702479338843,
1580
+ "f1": 0.14189837008628955,
1581
  "gt_support": 121,
1582
+ "predicted": 922
1583
  },
1584
  "1": {
1585
+ "accuracy": 0.6160203432930705,
1586
+ "precision": 0.9586150041220115,
1587
+ "recall": 0.6160203432930705,
1588
+ "f1": 0.7500483777333419,
1589
  "gt_support": 9438,
1590
+ "predicted": 6065
1591
  },
1592
  "2": {
1593
+ "accuracy": 0.6198347107438017,
1594
+ "precision": 0.08232711306256861,
1595
+ "recall": 0.6198347107438017,
1596
+ "f1": 0.14534883720930233,
1597
  "gt_support": 121,
1598
+ "predicted": 911
1599
  },
1600
  "3": {
1601
+ "accuracy": 0.8723186925434117,
1602
+ "precision": 0.7479299363057325,
1603
+ "recall": 0.8723186925434117,
1604
+ "f1": 0.8053495649191994,
1605
  "gt_support": 10769,
1606
+ "predicted": 12560
1607
  },
1608
  "4": {
1609
+ "accuracy": 0.9973435655253837,
1610
+ "precision": 0.9986700162553569,
1611
+ "recall": 0.9973435655253837,
1612
+ "f1": 0.9980063501439859,
1613
  "gt_support": 6776,
1614
+ "predicted": 6767
1615
  }
1616
  },
1617
  "total_samples": 27225
1618
  },
1619
  "13_0.8": {
1620
+ "accuracy": 0.746984126984127,
1621
  "baseline_accuracy": 0.39555555555555555,
1622
  "per_class": {
1623
  "0": {
1624
+ "accuracy": 0.6875,
1625
+ "precision": 0.0980891719745223,
1626
+ "recall": 0.6875,
1627
+ "f1": 0.17168338907469344,
1628
  "gt_support": 112,
1629
+ "predicted": 785
1630
  },
1631
  "1": {
1632
+ "accuracy": 0.615569823434992,
1633
+ "precision": 0.7994788273615635,
1634
+ "recall": 0.615569823434992,
1635
+ "f1": 0.6955733151958284,
1636
  "gt_support": 9968,
1637
+ "predicted": 7675
1638
  },
1639
  "2": {
1640
+ "accuracy": 0.7767857142857143,
1641
+ "precision": 0.08487804878048781,
1642
+ "recall": 0.7767857142857143,
1643
+ "f1": 0.15303430079155675,
1644
  "gt_support": 112,
1645
+ "predicted": 1025
1646
  },
1647
  "3": {
1648
+ "accuracy": 0.7169184981684982,
1649
+ "precision": 0.6626110876005078,
1650
+ "recall": 0.7169184981684982,
1651
+ "f1": 0.6886958434132394,
1652
  "gt_support": 8736,
1653
+ "predicted": 9452
1654
  },
1655
  "4": {
1656
+ "accuracy": 0.9982461734693877,
1657
+ "precision": 0.9996806642184257,
1658
+ "recall": 0.9982461734693877,
1659
+ "f1": 0.9989629038691664,
1660
  "gt_support": 6272,
1661
+ "predicted": 6263
1662
  }
1663
  },
1664
  "total_samples": 25200
1665
  },
1666
  "13_1.0": {
1667
+ "accuracy": 0.7207089297886844,
1668
  "baseline_accuracy": 0.43555555555555553,
1669
  "per_class": {
1670
  "0": {
1671
+ "accuracy": 0.558282208588957,
1672
+ "precision": 0.10983705491852747,
1673
+ "recall": 0.558282208588957,
1674
+ "f1": 0.18356026222894606,
1675
  "gt_support": 326,
1676
+ "predicted": 1657
1677
  },
1678
  "1": {
1679
+ "accuracy": 0.659258795542757,
1680
+ "precision": 0.7466940830290354,
1681
+ "recall": 0.659258795542757,
1682
+ "f1": 0.7002576676917963,
1683
  "gt_support": 31948,
1684
+ "predicted": 28207
1685
  },
1686
  "2": {
1687
+ "accuracy": 0.5306748466257669,
1688
+ "precision": 0.05209274314965372,
1689
+ "recall": 0.5306748466257669,
1690
+ "f1": 0.09487249794351521,
1691
  "gt_support": 326,
1692
+ "predicted": 3321
1693
  },
1694
  "3": {
1695
+ "accuracy": 0.5934026851604872,
1696
+ "precision": 0.605488772964391,
1697
+ "recall": 0.5934026851604872,
1698
+ "f1": 0.5993848088192371,
1699
  "gt_support": 22494,
1700
+ "predicted": 22045
1701
  },
1702
  "4": {
1703
+ "accuracy": 0.9914000876424189,
1704
+ "precision": 0.998841059602649,
1705
+ "recall": 0.9914000876424189,
1706
+ "f1": 0.9951066637343303,
1707
  "gt_support": 18256,
1708
+ "predicted": 18120
1709
  }
1710
  },
1711
  "total_samples": 73350
1712
  },
1713
  "15_0.0": {
1714
+ "accuracy": 0.8983462532299742,
1715
  "baseline_accuracy": 0.7422222222222222,
1716
  "per_class": {
1717
  "0": {
1718
+ "accuracy": 0.7674418604651163,
1719
+ "precision": 0.05956678700361011,
1720
+ "recall": 0.7674418604651163,
1721
+ "f1": 0.11055276381909548,
1722
  "gt_support": 86,
1723
+ "predicted": 1108
1724
  },
1725
  "1": {
1726
+ "accuracy": 0.9914867109634552,
1727
+ "precision": 0.9487383270415259,
1728
+ "recall": 0.9914867109634552,
1729
+ "f1": 0.9696415879784749,
1730
  "gt_support": 4816,
1731
+ "predicted": 5033
1732
  },
1733
  "2": {
1734
+ "accuracy": 0.686046511627907,
1735
+ "precision": 0.08452722063037249,
1736
+ "recall": 0.686046511627907,
1737
+ "f1": 0.15051020408163265,
1738
  "gt_support": 86,
1739
+ "predicted": 698
1740
  },
1741
  "3": {
1742
+ "accuracy": 0.8691686394652556,
1743
+ "precision": 0.9977619694668691,
1744
+ "recall": 0.8691686394652556,
1745
+ "f1": 0.929036579466379,
1746
  "gt_support": 14362,
1747
+ "predicted": 12511
1748
  },
1749
  "4": {
1750
  "accuracy": 0.0,
 
1752
  "recall": 0.0,
1753
  "f1": 0.0,
1754
  "gt_support": 0,
1755
+ "predicted": 0
1756
  }
1757
  },
1758
  "total_samples": 19350
1759
  },
1760
  "15_0.2": {
1761
+ "accuracy": 0.8255938697318007,
1762
  "baseline_accuracy": 0.6755555555555556,
1763
  "per_class": {
1764
  "0": {
1765
+ "accuracy": 0.6436781609195402,
1766
+ "precision": 0.05067873303167421,
1767
+ "recall": 0.6436781609195402,
1768
+ "f1": 0.09395973154362415,
1769
  "gt_support": 87,
1770
+ "predicted": 1105
1771
  },
1772
  "1": {
1773
+ "accuracy": 0.7830662133721872,
1774
+ "precision": 0.9395881895881896,
1775
+ "recall": 0.7830662133721872,
1776
+ "f1": 0.8542163355408389,
1777
  "gt_support": 6177,
1778
+ "predicted": 5148
1779
  },
1780
  "2": {
1781
+ "accuracy": 0.6091954022988506,
1782
+ "precision": 0.05337361530715005,
1783
+ "recall": 0.6091954022988506,
1784
+ "f1": 0.09814814814814814,
1785
  "gt_support": 87,
1786
+ "predicted": 993
1787
  },
1788
  "3": {
1789
+ "accuracy": 0.8480792498487598,
1790
+ "precision": 0.9096439289480087,
1791
+ "recall": 0.8480792498487598,
1792
+ "f1": 0.8777834305169648,
1793
  "gt_support": 13224,
1794
+ "predicted": 12329
1795
  },
1796
  "4": {
1797
  "accuracy": 0.0,
 
1799
  "recall": 0.0,
1800
  "f1": 0.0,
1801
  "gt_support": 0,
1802
+ "predicted": 0
1803
  }
1804
  },
1805
  "total_samples": 19575
1806
  },
1807
  "15_0.4": {
1808
+ "accuracy": 0.7753267973856209,
1809
  "baseline_accuracy": 0.6133333333333333,
1810
  "per_class": {
1811
  "0": {
1812
+ "accuracy": 0.7794117647058824,
1813
+ "precision": 0.06768837803320563,
1814
+ "recall": 0.7794117647058824,
1815
+ "f1": 0.12455934195064632,
1816
  "gt_support": 136,
1817
+ "predicted": 1566
1818
  },
1819
  "1": {
1820
+ "accuracy": 0.6554498269896194,
1821
+ "precision": 0.9401910907060429,
1822
+ "recall": 0.6554498269896194,
1823
+ "f1": 0.7724144961516897,
1824
  "gt_support": 11560,
1825
+ "predicted": 8059
1826
  },
1827
  "2": {
1828
+ "accuracy": 0.6544117647058824,
1829
+ "precision": 0.05387409200968523,
1830
+ "recall": 0.6544117647058824,
1831
+ "f1": 0.09955257270693513,
1832
  "gt_support": 136,
1833
+ "predicted": 1652
1834
  },
1835
  "3": {
1836
+ "accuracy": 0.8500106564364877,
1837
+ "precision": 0.8255964394762718,
1838
+ "recall": 0.8500106564364877,
1839
+ "f1": 0.8376256858575517,
1840
  "gt_support": 18768,
1841
+ "predicted": 19323
1842
  },
1843
  "4": {
1844
  "accuracy": 0.0,
 
1846
  "recall": 0.0,
1847
  "f1": 0.0,
1848
  "gt_support": 0,
1849
+ "predicted": 0
1850
  }
1851
  },
1852
  "total_samples": 30600
1853
  },
1854
  "15_0.6": {
1855
+ "accuracy": 0.6797777777777778,
1856
  "baseline_accuracy": 0.5466666666666666,
1857
  "per_class": {
1858
  "0": {
1859
+ "accuracy": 0.7,
1860
+ "precision": 0.06856023506366307,
1861
+ "recall": 0.7,
1862
+ "f1": 0.12488849241748438,
1863
  "gt_support": 100,
1864
+ "predicted": 1021
1865
  },
1866
  "1": {
1867
+ "accuracy": 0.5983,
1868
+ "precision": 0.8056827363318072,
1869
+ "recall": 0.5983,
1870
+ "f1": 0.6866750832089982,
1871
  "gt_support": 10000,
1872
+ "predicted": 7426
1873
  },
1874
  "2": {
1875
+ "accuracy": 0.64,
1876
+ "precision": 0.04221635883905013,
1877
+ "recall": 0.64,
1878
+ "f1": 0.07920792079207921,
1879
  "gt_support": 100,
1880
+ "predicted": 1516
1881
  },
1882
  "3": {
1883
+ "accuracy": 0.7461788617886179,
1884
+ "precision": 0.7320730637313552,
1885
+ "recall": 0.7461788617886179,
1886
+ "f1": 0.7390586624793655,
1887
  "gt_support": 12300,
1888
+ "predicted": 12537
1889
  },
1890
  "4": {
1891
  "accuracy": 0.0,
 
1893
  "recall": 0.0,
1894
  "f1": 0.0,
1895
  "gt_support": 0,
1896
+ "predicted": 0
1897
  }
1898
  },
1899
  "total_samples": 22500
1900
  },
1901
  "15_0.8": {
1902
+ "accuracy": 0.5970460704607046,
1903
  "baseline_accuracy": 0.5066666666666667,
1904
  "per_class": {
1905
  "0": {
1906
+ "accuracy": 0.5975609756097561,
1907
+ "precision": 0.05875299760191847,
1908
+ "recall": 0.5975609756097561,
1909
+ "f1": 0.10698689956331879,
1910
  "gt_support": 164,
1911
+ "predicted": 1668
1912
  },
1913
  "1": {
1914
+ "accuracy": 0.6699828840393667,
1915
+ "precision": 0.6539967629092048,
1916
+ "recall": 0.6699828840393667,
1917
+ "f1": 0.6618933129012654,
1918
  "gt_support": 18696,
1919
+ "predicted": 19153
1920
  },
1921
  "2": {
1922
+ "accuracy": 0.6097560975609756,
1923
+ "precision": 0.053022269353128315,
1924
+ "recall": 0.6097560975609756,
1925
+ "f1": 0.09756097560975611,
1926
  "gt_support": 164,
1927
+ "predicted": 1886
1928
  },
1929
  "3": {
1930
+ "accuracy": 0.5206422018348624,
1931
+ "precision": 0.6557457901782568,
1932
+ "recall": 0.5206422018348624,
1933
+ "f1": 0.5804359350151237,
1934
  "gt_support": 17876,
1935
+ "predicted": 14193
1936
  },
1937
  "4": {
1938
  "accuracy": 0.0,
 
1940
  "recall": 0.0,
1941
  "f1": 0.0,
1942
  "gt_support": 0,
1943
+ "predicted": 0
1944
  }
1945
  },
1946
  "total_samples": 36900
1947
  },
1948
  "15_1.0": {
1949
+ "accuracy": 0.5940664961636829,
1950
  "baseline_accuracy": 0.5688888888888889,
1951
  "per_class": {
1952
  "0": {
1953
+ "accuracy": 0.5319693094629157,
1954
+ "precision": 0.07479323984178353,
1955
+ "recall": 0.5319693094629157,
1956
+ "f1": 0.13114754098360656,
1957
  "gt_support": 391,
1958
+ "predicted": 2781
1959
  },
1960
  "1": {
1961
+ "accuracy": 0.7659446930946292,
1962
+ "precision": 0.6496077002592737,
1963
+ "recall": 0.7659446930946292,
1964
+ "f1": 0.7029956262206696,
1965
  "gt_support": 50048,
1966
+ "predicted": 59011
1967
  },
1968
  "2": {
1969
+ "accuracy": 0.6214833759590793,
1970
+ "precision": 0.06497326203208556,
1971
+ "recall": 0.6214833759590793,
1972
+ "f1": 0.11764705882352942,
1973
  "gt_support": 391,
1974
+ "predicted": 3740
1975
  },
1976
  "3": {
1977
+ "accuracy": 0.3628482972136223,
1978
+ "precision": 0.6005435993405516,
1979
+ "recall": 0.3628482972136223,
1980
+ "f1": 0.45237296099885876,
1981
  "gt_support": 37145,
1982
+ "predicted": 22443
1983
  },
1984
  "4": {
1985
  "accuracy": 0.0,
 
1987
  "recall": 0.0,
1988
  "f1": 0.0,
1989
  "gt_support": 0,
1990
+ "predicted": 0
1991
  }
1992
  },
1993
  "total_samples": 87975
1994
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
1995
  }
1996
  }
layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
@@ -27,17 +27,17 @@ Processed 300 trajectories, 3223 steps
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
- Overall Accuracy: 0.5123 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
- A 0.0009 0.0159 0.0009 0.0018 3223 189
37
- # 0.1615 0.4518 0.1615 0.2379 238175 85140
38
- G 0.0000 0.0000 0.0000 0.0000 3223 0
39
- _ 0.5686 0.4867 0.5686 0.5244 243346 284304
40
- + 0.8208 0.5476 0.8208 0.6569 237208 355542
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
@@ -47,81 +47,81 @@ METRICS BY SIZE
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
- Overall Accuracy: 0.6500 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
- A 0.0000 0.0000 0.0000 0.0000 286 0
57
- # 0.2296 0.2378 0.2296 0.2336 8224 7940
58
- G 0.0000 0.0000 0.0000 0.0000 286 0
59
- _ 0.7158 0.1951 0.7158 0.3067 5218 19142
60
- + 0.7192 0.9715 0.7192 0.8265 50336 37268
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
- Overall Accuracy: 0.6679 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
- A 0.0000 0.0000 0.0000 0.0000 514 0
73
- # 0.2328 0.3544 0.2328 0.2810 22670 14893
74
- G 0.0000 0.0000 0.0000 0.0000 514 0
75
- _ 0.7107 0.3528 0.7107 0.4715 17936 36134
76
- + 0.8001 0.9164 0.8001 0.8543 74016 64623
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
- Overall Accuracy: 0.6247 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
- A 0.0015 0.0909 0.0015 0.0030 649 11
89
- # 0.1998 0.4640 0.1998 0.2793 39462 16994
90
- G 0.0000 0.0000 0.0000 0.0000 649 0
91
- _ 0.6989 0.4560 0.6989 0.5519 37769 57886
92
- + 0.8435 0.8004 0.8435 0.8214 67496 71134
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
- Overall Accuracy: 0.5266 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
- A 0.0025 0.0165 0.0025 0.0043 810 121
105
- # 0.1535 0.5135 0.1535 0.2364 66522 19891
106
- G 0.0000 0.0000 0.0000 0.0000 810 0
107
- _ 0.6318 0.5376 0.6318 0.5809 68748 80785
108
- + 0.9332 0.5197 0.9332 0.6676 45360 81453
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
- Overall Accuracy: 0.3008 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
- A 0.0000 0.0000 0.0000 0.0000 964 57
121
- # 0.1303 0.5192 0.1303 0.2083 101297 25422
122
- G 0.0000 0.0000 0.0000 0.0000 964 0
123
- _ 0.4578 0.5760 0.4578 0.5102 113675 90357
124
- + 0.0000 0.0000 0.0000 0.0000 0 101064
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
@@ -131,97 +131,97 @@ METRICS BY COMPLEXITY
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
- Overall Accuracy: 0.5856 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
- A 0.0000 0.0000 0.0000 0.0000 310 0
141
- # 0.2097 0.2985 0.2097 0.2463 13248 9308
142
- G 0.0000 0.0000 0.0000 0.0000 310 0
143
- _ 0.6022 0.6702 0.6022 0.6344 30762 27638
144
- + 0.7779 0.5957 0.7779 0.6747 25120 32804
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
- Overall Accuracy: 0.5805 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
- A 0.0000 0.0000 0.0000 0.0000 325 0
157
- # 0.1438 0.3225 0.1438 0.1989 16604 7401
158
- G 0.0000 0.0000 0.0000 0.0000 325 0
159
- _ 0.6406 0.5994 0.6406 0.6193 29047 31044
160
- + 0.7998 0.6186 0.7998 0.6976 26824 34680
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
- Overall Accuracy: 0.5198 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
- A 0.0000 0.0000 0.0000 0.0000 424 0
173
- # 0.1388 0.3797 0.1388 0.2033 26803 9796
174
- G 0.0000 0.0000 0.0000 0.0000 424 0
175
- _ 0.5594 0.5654 0.5594 0.5624 38165 37764
176
- + 0.8287 0.5124 0.8287 0.6333 29584 47840
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
- Overall Accuracy: 0.5367 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
- A 0.0000 0.0000 0.0000 0.0000 432 48
189
- # 0.1540 0.4287 0.1540 0.2266 29604 10634
190
- G 0.0000 0.0000 0.0000 0.0000 432 0
191
- _ 0.6097 0.4859 0.6097 0.5408 32188 40388
192
- + 0.8102 0.6067 0.8102 0.6938 34544 46130
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
- Overall Accuracy: 0.5012 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
- A 0.0000 0.0000 0.0000 0.0000 546 0
205
- # 0.1502 0.4624 0.1502 0.2267 42640 13848
206
- G 0.0000 0.0000 0.0000 0.0000 546 0
207
- _ 0.5758 0.4284 0.5758 0.4913 36766 49413
208
- + 0.8027 0.5705 0.8027 0.6670 42352 59589
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
- Overall Accuracy: 0.4681 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
- A 0.0025 0.0213 0.0025 0.0045 1186 141
221
- # 0.1704 0.5451 0.1704 0.2596 109276 34153
222
- G 0.0000 0.0000 0.0000 0.0000 1186 0
223
- _ 0.5114 0.3986 0.5114 0.4480 76418 98057
224
- + 0.8530 0.4996 0.8530 0.6301 78784 134499
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
@@ -231,481 +231,481 @@ METRICS BY SIZE-COMPLEXITY COMBINATION
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
- Overall Accuracy: 0.6457 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
- A 0.0000 0.0000 0.0000 0.0000 38 0
241
- # 0.3728 0.2151 0.3728 0.2728 912 1581
242
- G 0.0000 0.0000 0.0000 0.0000 38 0
243
- _ 0.6316 0.2498 0.6316 0.3580 874 2210
244
- + 0.6921 0.9727 0.6921 0.8088 6688 4759
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
- Overall Accuracy: 0.6617 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
- A 0.0000 0.0000 0.0000 0.0000 44 0
257
- # 0.1783 0.1884 0.1783 0.1832 1144 1083
258
- G 0.0000 0.0000 0.0000 0.0000 44 0
259
- _ 0.7749 0.2379 0.7749 0.3640 924 3010
260
- + 0.7271 0.9697 0.7271 0.8311 7744 5807
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
- Overall Accuracy: 0.6750 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
- A 0.0000 0.0000 0.0000 0.0000 41 0
273
- # 0.2552 0.2511 0.2552 0.2531 1148 1167
274
- G 0.0000 0.0000 0.0000 0.0000 41 0
275
- _ 0.5944 0.2056 0.5944 0.3055 779 2252
276
- + 0.7582 0.9423 0.7582 0.8403 7216 5806
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
- Overall Accuracy: 0.6357 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
- A 0.0000 0.0000 0.0000 0.0000 47 0
289
- # 0.1746 0.2185 0.1746 0.1941 1363 1089
290
- G 0.0000 0.0000 0.0000 0.0000 47 0
291
- _ 0.8132 0.1924 0.8132 0.3112 846 3575
292
- + 0.7007 0.9805 0.7007 0.8173 8272 5911
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
- Overall Accuracy: 0.6394 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
- A 0.0000 0.0000 0.0000 0.0000 55 0
305
- # 0.2469 0.2621 0.2469 0.2543 1705 1606
306
- G 0.0000 0.0000 0.0000 0.0000 55 0
307
- _ 0.7170 0.1660 0.7170 0.2696 880 3801
308
- + 0.7087 0.9845 0.7087 0.8241 9680 6968
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
- Overall Accuracy: 0.6480 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
- A 0.0000 0.0000 0.0000 0.0000 61 0
321
- # 0.2008 0.2772 0.2008 0.2329 1952 1414
322
- G 0.0000 0.0000 0.0000 0.0000 61 0
323
- _ 0.7486 0.1595 0.7486 0.2630 915 4294
324
- + 0.7281 0.9751 0.7281 0.8337 10736 8017
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
- Overall Accuracy: 0.6722 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
- A 0.0000 0.0000 0.0000 0.0000 54 0
337
- # 0.2529 0.2378 0.2529 0.2451 1728 1838
338
- G 0.0000 0.0000 0.0000 0.0000 54 0
339
- _ 0.6958 0.4667 0.6958 0.5587 2538 3784
340
- + 0.7670 0.9136 0.7670 0.8339 7776 6528
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
- Overall Accuracy: 0.6950 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
- A 0.0000 0.0000 0.0000 0.0000 47 0
353
- # 0.2565 0.3037 0.2565 0.2781 1692 1429
354
- G 0.0000 0.0000 0.0000 0.0000 47 0
355
- _ 0.7318 0.4513 0.7318 0.5583 2021 3277
356
- + 0.8033 0.9264 0.8033 0.8605 6768 5869
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
- Overall Accuracy: 0.6655 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
- A 0.0000 0.0000 0.0000 0.0000 67 0
369
- # 0.1858 0.3053 0.1858 0.2310 2680 1631
370
- G 0.0000 0.0000 0.0000 0.0000 67 0
371
- _ 0.7290 0.3800 0.7290 0.4996 2613 5013
372
- + 0.7908 0.9050 0.7908 0.8441 9648 8431
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
- Overall Accuracy: 0.6486 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
- A 0.0000 0.0000 0.0000 0.0000 61 0
385
- # 0.1487 0.2887 0.1487 0.1963 2623 1351
386
- G 0.0000 0.0000 0.0000 0.0000 61 0
387
- _ 0.8110 0.3452 0.8110 0.4842 2196 5160
388
- + 0.7663 0.9330 0.7663 0.8415 8784 7214
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
- Overall Accuracy: 0.6560 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
- A 0.0000 0.0000 0.0000 0.0000 101 0
401
- # 0.2062 0.3570 0.2062 0.2615 4747 2742
402
- G 0.0000 0.0000 0.0000 0.0000 101 0
403
- _ 0.6993 0.3182 0.6993 0.4374 3232 7102
404
- + 0.8023 0.9059 0.8023 0.8510 14544 12881
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
- Overall Accuracy: 0.6736 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
- A 0.0000 0.0000 0.0000 0.0000 184 0
417
- # 0.2761 0.4304 0.2761 0.3364 9200 5902
418
- G 0.0000 0.0000 0.0000 0.0000 184 0
419
- _ 0.6664 0.3014 0.6664 0.4151 5336 11798
420
- + 0.8224 0.9195 0.8224 0.8682 26496 23700
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
- Overall Accuracy: 0.6680 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
- A 0.0000 0.0000 0.0000 0.0000 68 0
433
- # 0.2257 0.3283 0.2257 0.2675 2720 1870
434
- G 0.0000 0.0000 0.0000 0.0000 68 0
435
- _ 0.7241 0.6130 0.7241 0.6639 5372 6346
436
- + 0.8084 0.8070 0.8084 0.8077 7072 7084
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
- Overall Accuracy: 0.6510 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
- A 0.0000 0.0000 0.0000 0.0000 85 0
449
- # 0.1617 0.3436 0.1617 0.2199 3995 1880
450
- G 0.0000 0.0000 0.0000 0.0000 85 0
451
- _ 0.7490 0.5537 0.7490 0.6367 6120 8279
452
- + 0.8167 0.8053 0.8167 0.8110 8840 8966
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
- Overall Accuracy: 0.6288 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
- A 0.0000 0.0000 0.0000 0.0000 55 0
465
- # 0.1880 0.3832 0.1880 0.2522 2915 1430
466
- G 0.0000 0.0000 0.0000 0.0000 55 0
467
- _ 0.6485 0.5167 0.6485 0.5751 3630 4556
468
- + 0.8531 0.7638 0.8531 0.8060 5720 6389
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
- Overall Accuracy: 0.6424 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
- A 0.0000 0.0000 0.0000 0.0000 103 0
481
- # 0.2184 0.4717 0.2184 0.2986 6180 2862
482
- G 0.0000 0.0000 0.0000 0.0000 103 0
483
- _ 0.7166 0.4708 0.7166 0.5682 6077 9251
484
- + 0.8573 0.8301 0.8573 0.8435 10712 11062
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
- Overall Accuracy: 0.5999 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
- A 0.0000 0.0000 0.0000 0.0000 114 0
497
- # 0.1438 0.4606 0.1438 0.2192 7524 2349
498
- G 0.0000 0.0000 0.0000 0.0000 114 0
499
- _ 0.7726 0.4010 0.7726 0.5279 6042 11642
500
- + 0.8129 0.8267 0.8129 0.8197 11856 11659
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
- Overall Accuracy: 0.6049 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
- A 0.0045 0.0909 0.0045 0.0085 224 11
513
- # 0.2260 0.5520 0.2260 0.3207 16128 6603
514
- G 0.0000 0.0000 0.0000 0.0000 224 0
515
- _ 0.6219 0.3676 0.6219 0.4620 10528 17812
516
- + 0.8712 0.7814 0.8712 0.8239 23296 25974
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
- Overall Accuracy: 0.6322 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
- A 0.0000 0.0000 0.0000 0.0000 64 0
529
- # 0.2510 0.3712 0.2510 0.2995 3072 2077
530
- G 0.0000 0.0000 0.0000 0.0000 64 0
531
- _ 0.6698 0.7778 0.6698 0.7198 7616 6558
532
- + 0.9018 0.5606 0.9018 0.6914 3584 5765
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
- Overall Accuracy: 0.6029 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
- A 0.0000 0.0000 0.0000 0.0000 62 0
545
- # 0.1502 0.3838 0.1502 0.2159 3596 1407
546
- G 0.0000 0.0000 0.0000 0.0000 62 0
547
- _ 0.6962 0.6800 0.6962 0.6880 6758 6919
548
- + 0.9116 0.5628 0.9116 0.6959 3472 5624
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
- Overall Accuracy: 0.5656 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
- A 0.0000 0.0000 0.0000 0.0000 125 0
561
- # 0.1298 0.4433 0.1298 0.2008 8500 2488
562
- G 0.0000 0.0000 0.0000 0.0000 125 0
563
- _ 0.6684 0.6207 0.6684 0.6437 12375 13325
564
- + 0.9334 0.5307 0.9334 0.6767 7000 12312
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
- Overall Accuracy: 0.5389 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
- A 0.0000 0.0000 0.0000 0.0000 121 48
577
- # 0.1525 0.4863 0.1525 0.2322 9438 2959
578
- G 0.0000 0.0000 0.0000 0.0000 121 0
579
- _ 0.6459 0.5577 0.6459 0.5986 10769 12472
580
- + 0.9262 0.5343 0.9262 0.6777 6776 11746
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
- Overall Accuracy: 0.5112 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
- A 0.0000 0.0000 0.0000 0.0000 112 0
593
- # 0.1769 0.5309 0.1769 0.2653 9968 3321
594
- G 0.0000 0.0000 0.0000 0.0000 112 0
595
- _ 0.6058 0.4864 0.6058 0.5396 8736 10879
596
- + 0.9292 0.5298 0.9292 0.6748 6272 11000
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
- Overall Accuracy: 0.4772 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
- A 0.0061 0.0274 0.0061 0.0100 326 73
609
- # 0.1439 0.6019 0.1439 0.2323 31948 7639
610
- G 0.0000 0.0000 0.0000 0.0000 326 0
611
- _ 0.5827 0.4279 0.5827 0.4935 22494 30632
612
- + 0.9474 0.4941 0.9474 0.6494 18256 35006
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
- Overall Accuracy: 0.4047 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
- A 0.0000 0.0000 0.0000 0.0000 86 0
625
- # 0.1279 0.3172 0.1279 0.1823 4816 1942
626
- G 0.0000 0.0000 0.0000 0.0000 86 0
627
- _ 0.5024 0.8255 0.5024 0.6246 14362 8740
628
- + 0.0000 0.0000 0.0000 0.0000 0 8668
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
- Overall Accuracy: 0.3927 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
- A 0.0000 0.0000 0.0000 0.0000 87 0
641
- # 0.0911 0.3514 0.0911 0.1447 6177 1602
642
- G 0.0000 0.0000 0.0000 0.0000 87 0
643
- _ 0.5387 0.7453 0.5387 0.6254 13224 9559
644
- + 0.0000 0.0000 0.0000 0.0000 0 8414
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
- Overall Accuracy: 0.3149 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
- A 0.0000 0.0000 0.0000 0.0000 136 0
657
- # 0.1106 0.4149 0.1106 0.1746 11560 3080
658
- G 0.0000 0.0000 0.0000 0.0000 136 0
659
- _ 0.4453 0.6623 0.4453 0.5325 18768 12618
660
- + 0.0000 0.0000 0.0000 0.0000 0 14902
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
- Overall Accuracy: 0.3106 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
- A 0.0000 0.0000 0.0000 0.0000 100 0
673
- # 0.1142 0.4812 0.1142 0.1846 10000 2373
674
- G 0.0000 0.0000 0.0000 0.0000 100 0
675
- _ 0.4753 0.5887 0.4753 0.5260 12300 9930
676
- + 0.0000 0.0000 0.0000 0.0000 0 10197
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
- Overall Accuracy: 0.2839 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
- A 0.0000 0.0000 0.0000 0.0000 164 0
689
- # 0.1154 0.5634 0.1154 0.1916 18696 3830
690
- G 0.0000 0.0000 0.0000 0.0000 164 0
691
- _ 0.4654 0.5203 0.4654 0.4913 17876 15989
692
- + 0.0000 0.0000 0.0000 0.0000 0 17081
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
- Overall Accuracy: 0.2572 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
- A 0.0000 0.0000 0.0000 0.0000 391 57
705
- # 0.1487 0.5909 0.1487 0.2376 50048 12595
706
- G 0.0000 0.0000 0.0000 0.0000 391 0
707
- _ 0.4088 0.4530 0.4088 0.4298 37145 33521
708
- + 0.0000 0.0000 0.0000 0.0000 0 41802
709
  ---------------------------------------------------------------------------------------
710
 
711
- Results saved to: reveng/trajectories_test_full_with_probes/layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.pt
3
  Loaded probe: cognitive_map_probe_layer23_mlp_pre_reasoning_all_general
4
  Input dimension: 8642
5
  Number of classes: 5
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  Found 5 size folders: ['size11', 'size13', 'size15', 'size7', 'size9']
11
 
 
27
  =======================================================================================
28
  GLOBAL METRICS
29
  =======================================================================================
30
+ Overall Accuracy: 0.7944 (Baseline: 0.3356, 725175 samples)
31
 
32
  Per-class metrics:
33
  ---------------------------------------------------------------------------------------
34
  Class Accuracy Precision Recall F1-Score GT Support Predicted
35
  ---------------------------------------------------------------------------------------
36
+ A 0.7533 0.0966 0.7533 0.1712 3223 25146
37
+ # 0.7148 0.7784 0.7148 0.7453 238175 218713
38
+ G 0.7139 0.0817 0.7139 0.1467 3223 28147
39
+ _ 0.6747 0.7625 0.6747 0.7159 243346 215326
40
+ + 0.9988 0.9962 0.9988 0.9975 237208 237843
41
  ---------------------------------------------------------------------------------------
42
 
43
  =======================================================================================
 
47
  =======================================================================================
48
  Size 7
49
  =======================================================================================
50
+ Overall Accuracy: 0.9466 (Baseline: 0.7822, 64350 samples)
51
 
52
  Per-class metrics:
53
  ---------------------------------------------------------------------------------------
54
  Class Accuracy Precision Recall F1-Score GT Support Predicted
55
  ---------------------------------------------------------------------------------------
56
+ A 0.9650 0.2317 0.9650 0.3737 286 1191
57
+ # 0.8084 0.9191 0.8084 0.8602 8224 7233
58
+ G 0.9790 0.1959 0.9790 0.3265 286 1429
59
+ _ 0.6501 0.8117 0.6501 0.7219 5218 4179
60
+ + 0.9996 1.0000 0.9996 0.9998 50336 50318
61
  ---------------------------------------------------------------------------------------
62
 
63
  =======================================================================================
64
  Size 9
65
  =======================================================================================
66
+ Overall Accuracy: 0.8837 (Baseline: 0.6400, 115650 samples)
67
 
68
  Per-class metrics:
69
  ---------------------------------------------------------------------------------------
70
  Class Accuracy Precision Recall F1-Score GT Support Predicted
71
  ---------------------------------------------------------------------------------------
72
+ A 0.9086 0.1106 0.9086 0.1973 514 4221
73
+ # 0.6362 0.9361 0.6362 0.7575 22670 15406
74
+ G 0.8872 0.1178 0.8872 0.2079 514 3872
75
+ _ 0.7171 0.7401 0.7171 0.7284 17936 17379
76
+ + 0.9997 0.9896 0.9997 0.9947 74016 74772
77
  ---------------------------------------------------------------------------------------
78
 
79
  =======================================================================================
80
  Size 11
81
  =======================================================================================
82
+ Overall Accuracy: 0.8392 (Baseline: 0.4622, 146025 samples)
83
 
84
  Per-class metrics:
85
  ---------------------------------------------------------------------------------------
86
  Class Accuracy Precision Recall F1-Score GT Support Predicted
87
  ---------------------------------------------------------------------------------------
88
+ A 0.8259 0.1095 0.8259 0.1934 649 4894
89
+ # 0.7663 0.7914 0.7663 0.7786 39462 38206
90
+ G 0.7211 0.0886 0.7211 0.1578 649 5281
91
+ _ 0.6310 0.7922 0.6310 0.7025 37769 30086
92
+ + 0.9997 0.9988 0.9997 0.9992 67496 67558
93
  ---------------------------------------------------------------------------------------
94
 
95
  =======================================================================================
96
  Size 13
97
  =======================================================================================
98
+ Overall Accuracy: 0.7878 (Baseline: 0.3772, 182250 samples)
99
 
100
  Per-class metrics:
101
  ---------------------------------------------------------------------------------------
102
  Class Accuracy Precision Recall F1-Score GT Support Predicted
103
  ---------------------------------------------------------------------------------------
104
+ A 0.6728 0.0975 0.6728 0.1703 810 5591
105
+ # 0.6752 0.8312 0.6752 0.7451 66522 54038
106
+ G 0.6037 0.0691 0.6037 0.1240 810 7080
107
+ _ 0.7635 0.7462 0.7635 0.7547 68748 70346
108
+ + 0.9951 0.9988 0.9951 0.9970 45360 45195
109
  ---------------------------------------------------------------------------------------
110
 
111
  =======================================================================================
112
  Size 15
113
  =======================================================================================
114
+ Overall Accuracy: 0.6771 (Baseline: 0.5241, 216900 samples)
115
 
116
  Per-class metrics:
117
  ---------------------------------------------------------------------------------------
118
  Class Accuracy Precision Recall F1-Score GT Support Predicted
119
  ---------------------------------------------------------------------------------------
120
+ A 0.6266 0.0653 0.6266 0.1183 964 9249
121
+ # 0.7308 0.7130 0.7308 0.7218 101297 103830
122
+ G 0.6307 0.0580 0.6307 0.1062 964 10485
123
+ _ 0.6300 0.7673 0.6300 0.6919 113675 93336
124
+ + 0.0000 0.0000 0.0000 0.0000 0 0
125
  ---------------------------------------------------------------------------------------
126
 
127
  =======================================================================================
 
131
  =======================================================================================
132
  Complexity 0.00
133
  =======================================================================================
134
+ Overall Accuracy: 0.9266 (Baseline: 0.4410, 69750 samples)
135
 
136
  Per-class metrics:
137
  ---------------------------------------------------------------------------------------
138
  Class Accuracy Precision Recall F1-Score GT Support Predicted
139
  ---------------------------------------------------------------------------------------
140
+ A 0.8419 0.0900 0.8419 0.1626 310 2900
141
+ # 0.9621 0.9692 0.9621 0.9656 13248 13151
142
+ G 0.7258 0.1163 0.7258 0.2005 310 1934
143
+ _ 0.8547 0.9901 0.8547 0.9174 30762 26555
144
+ + 0.9994 0.9959 0.9994 0.9977 25120 25210
145
  ---------------------------------------------------------------------------------------
146
 
147
  =======================================================================================
148
  Complexity 0.20
149
  =======================================================================================
150
+ Overall Accuracy: 0.8868 (Baseline: 0.3972, 73125 samples)
151
 
152
  Per-class metrics:
153
  ---------------------------------------------------------------------------------------
154
  Class Accuracy Precision Recall F1-Score GT Support Predicted
155
  ---------------------------------------------------------------------------------------
156
+ A 0.8154 0.0878 0.8154 0.1586 325 3017
157
+ # 0.8083 0.9675 0.8083 0.8808 16604 13872
158
+ G 0.7446 0.0865 0.7446 0.1549 325 2799
159
+ _ 0.8299 0.9090 0.8299 0.8677 29047 26518
160
+ + 0.9997 0.9961 0.9997 0.9979 26824 26919
161
  ---------------------------------------------------------------------------------------
162
 
163
  =======================================================================================
164
  Complexity 0.40
165
  =======================================================================================
166
+ Overall Accuracy: 0.8479 (Baseline: 0.4001, 95400 samples)
167
 
168
  Per-class metrics:
169
  ---------------------------------------------------------------------------------------
170
  Class Accuracy Precision Recall F1-Score GT Support Predicted
171
  ---------------------------------------------------------------------------------------
172
+ A 0.8514 0.0958 0.8514 0.1723 424 3767
173
+ # 0.6873 0.9572 0.6873 0.8001 26803 19246
174
+ G 0.7500 0.0840 0.7500 0.1510 424 3787
175
+ _ 0.8446 0.8279 0.8446 0.8362 38165 38936
176
+ + 0.9990 0.9963 0.9990 0.9977 29584 29664
177
  ---------------------------------------------------------------------------------------
178
 
179
  =======================================================================================
180
  Complexity 0.60
181
  =======================================================================================
182
+ Overall Accuracy: 0.8106 (Baseline: 0.3554, 97200 samples)
183
 
184
  Per-class metrics:
185
  ---------------------------------------------------------------------------------------
186
  Class Accuracy Precision Recall F1-Score GT Support Predicted
187
  ---------------------------------------------------------------------------------------
188
+ A 0.7384 0.0916 0.7384 0.1630 432 3483
189
+ # 0.6471 0.8583 0.6471 0.7379 29604 22321
190
+ G 0.7454 0.0807 0.7454 0.1456 432 3991
191
+ _ 0.7602 0.7469 0.7602 0.7535 32188 32761
192
+ + 0.9993 0.9964 0.9993 0.9978 34544 34644
193
  ---------------------------------------------------------------------------------------
194
 
195
  =======================================================================================
196
  Complexity 0.80
197
  =======================================================================================
198
+ Overall Accuracy: 0.7577 (Baseline: 0.3471, 122850 samples)
199
 
200
  Per-class metrics:
201
  ---------------------------------------------------------------------------------------
202
  Class Accuracy Precision Recall F1-Score GT Support Predicted
203
  ---------------------------------------------------------------------------------------
204
+ A 0.7656 0.0987 0.7656 0.1748 546 4237
205
+ # 0.6669 0.7277 0.6669 0.6959 42640 39077
206
+ G 0.7344 0.0813 0.7344 0.1465 546 4930
207
+ _ 0.5845 0.6685 0.5845 0.6237 36766 32144
208
+ + 0.9996 0.9970 0.9996 0.9983 42352 42462
209
  ---------------------------------------------------------------------------------------
210
 
211
  =======================================================================================
212
  Complexity 1.00
213
  =======================================================================================
214
+ Overall Accuracy: 0.7265 (Baseline: 0.4095, 266850 samples)
215
 
216
  Per-class metrics:
217
  ---------------------------------------------------------------------------------------
218
  Class Accuracy Precision Recall F1-Score GT Support Predicted
219
  ---------------------------------------------------------------------------------------
220
+ A 0.6779 0.1038 0.6779 0.1801 1186 7742
221
+ # 0.7145 0.7031 0.7145 0.7087 109276 111046
222
+ G 0.6686 0.0741 0.6686 0.1334 1186 10706
223
+ _ 0.4658 0.6094 0.4658 0.5281 76418 58412
224
+ + 0.9976 0.9956 0.9976 0.9966 78784 78944
225
  ---------------------------------------------------------------------------------------
226
 
227
  =======================================================================================
 
231
  =======================================================================================
232
  Size 7, Complexity 0.00
233
  =======================================================================================
234
+ Overall Accuracy: 0.9671 (Baseline: 0.7822, 8550 samples)
235
 
236
  Per-class metrics:
237
  ---------------------------------------------------------------------------------------
238
  Class Accuracy Precision Recall F1-Score GT Support Predicted
239
  ---------------------------------------------------------------------------------------
240
+ A 1.0000 0.2568 1.0000 0.4086 38 148
241
+ # 0.9759 0.9632 0.9759 0.9695 912 924
242
+ G 0.9737 0.2139 0.9737 0.3507 38 173
243
+ _ 0.7059 0.9984 0.7059 0.8271 874 618
244
+ + 0.9999 1.0000 0.9999 0.9999 6688 6687
245
  ---------------------------------------------------------------------------------------
246
 
247
  =======================================================================================
248
  Size 7, Complexity 0.20
249
  =======================================================================================
250
+ Overall Accuracy: 0.9589 (Baseline: 0.7822, 9900 samples)
251
 
252
  Per-class metrics:
253
  ---------------------------------------------------------------------------------------
254
  Class Accuracy Precision Recall F1-Score GT Support Predicted
255
  ---------------------------------------------------------------------------------------
256
+ A 0.9318 0.2330 0.9318 0.3727 44 176
257
+ # 0.8767 0.9672 0.8767 0.9198 1144 1037
258
+ G 1.0000 0.1956 1.0000 0.3271 44 225
259
+ _ 0.7208 0.9212 0.7208 0.8087 924 723
260
+ + 0.9994 1.0000 0.9994 0.9997 7744 7739
261
  ---------------------------------------------------------------------------------------
262
 
263
  =======================================================================================
264
  Size 7, Complexity 0.40
265
  =======================================================================================
266
+ Overall Accuracy: 0.9458 (Baseline: 0.7822, 9225 samples)
267
 
268
  Per-class metrics:
269
  ---------------------------------------------------------------------------------------
270
  Class Accuracy Precision Recall F1-Score GT Support Predicted
271
  ---------------------------------------------------------------------------------------
272
+ A 0.9512 0.2120 0.9512 0.3467 41 184
273
+ # 0.7883 0.9427 0.7883 0.8586 1148 960
274
+ G 0.9756 0.1835 0.9756 0.3089 41 218
275
+ _ 0.6778 0.8123 0.6778 0.7390 779 650
276
+ + 0.9996 1.0000 0.9996 0.9998 7216 7213
277
  ---------------------------------------------------------------------------------------
278
 
279
  =======================================================================================
280
  Size 7, Complexity 0.60
281
  =======================================================================================
282
+ Overall Accuracy: 0.9483 (Baseline: 0.7822, 10575 samples)
283
 
284
  Per-class metrics:
285
  ---------------------------------------------------------------------------------------
286
  Class Accuracy Precision Recall F1-Score GT Support Predicted
287
  ---------------------------------------------------------------------------------------
288
+ A 0.9149 0.2529 0.9149 0.3963 47 170
289
+ # 0.8114 0.9279 0.8114 0.8658 1363 1192
290
+ G 0.9574 0.1837 0.9574 0.3082 47 245
291
+ _ 0.6690 0.8086 0.6690 0.7322 846 700
292
+ + 0.9995 1.0000 0.9995 0.9998 8272 8268
293
  ---------------------------------------------------------------------------------------
294
 
295
  =======================================================================================
296
  Size 7, Complexity 0.80
297
  =======================================================================================
298
+ Overall Accuracy: 0.9381 (Baseline: 0.7822, 12375 samples)
299
 
300
  Per-class metrics:
301
  ---------------------------------------------------------------------------------------
302
  Class Accuracy Precision Recall F1-Score GT Support Predicted
303
  ---------------------------------------------------------------------------------------
304
+ A 1.0000 0.2402 1.0000 0.3873 55 229
305
+ # 0.7601 0.8822 0.7601 0.8166 1705 1469
306
+ G 0.9636 0.2008 0.9636 0.3323 55 264
307
+ _ 0.5977 0.7166 0.5977 0.6518 880 734
308
+ + 0.9999 1.0000 0.9999 0.9999 9680 9679
309
  ---------------------------------------------------------------------------------------
310
 
311
  =======================================================================================
312
  Size 7, Complexity 1.00
313
  =======================================================================================
314
+ Overall Accuracy: 0.9318 (Baseline: 0.7822, 13725 samples)
315
 
316
  Per-class metrics:
317
  ---------------------------------------------------------------------------------------
318
  Class Accuracy Precision Recall F1-Score GT Support Predicted
319
  ---------------------------------------------------------------------------------------
320
+ A 0.9836 0.2113 0.9836 0.3478 61 284
321
+ # 0.7418 0.8770 0.7418 0.8038 1952 1651
322
+ G 1.0000 0.2007 1.0000 0.3342 61 304
323
+ _ 0.5344 0.6485 0.5344 0.5860 915 754
324
+ + 0.9995 0.9999 0.9995 0.9997 10736 10732
325
  ---------------------------------------------------------------------------------------
326
 
327
  =======================================================================================
328
  Size 9, Complexity 0.00
329
  =======================================================================================
330
+ Overall Accuracy: 0.9208 (Baseline: 0.6400, 12150 samples)
331
 
332
  Per-class metrics:
333
  ---------------------------------------------------------------------------------------
334
  Class Accuracy Precision Recall F1-Score GT Support Predicted
335
  ---------------------------------------------------------------------------------------
336
+ A 0.7778 0.1066 0.7778 0.1875 54 394
337
+ # 0.8021 0.9914 0.8021 0.8868 1728 1398
338
+ G 0.8519 0.1220 0.8519 0.2135 54 377
339
+ _ 0.7636 0.9137 0.7636 0.8319 2538 2121
340
+ + 1.0000 0.9893 1.0000 0.9946 7776 7860
341
  ---------------------------------------------------------------------------------------
342
 
343
  =======================================================================================
344
  Size 9, Complexity 0.20
345
  =======================================================================================
346
+ Overall Accuracy: 0.9122 (Baseline: 0.6400, 10575 samples)
347
 
348
  Per-class metrics:
349
  ---------------------------------------------------------------------------------------
350
  Class Accuracy Precision Recall F1-Score GT Support Predicted
351
  ---------------------------------------------------------------------------------------
352
+ A 0.8085 0.1041 0.8085 0.1845 47 365
353
+ # 0.7559 0.9915 0.7559 0.8578 1692 1290
354
+ G 0.7234 0.1059 0.7234 0.1848 47 321
355
+ _ 0.7571 0.8748 0.7571 0.8117 2021 1749
356
+ + 0.9997 0.9877 0.9997 0.9937 6768 6850
357
  ---------------------------------------------------------------------------------------
358
 
359
  =======================================================================================
360
  Size 9, Complexity 0.40
361
  =======================================================================================
362
+ Overall Accuracy: 0.9021 (Baseline: 0.6400, 15075 samples)
363
 
364
  Per-class metrics:
365
  ---------------------------------------------------------------------------------------
366
  Class Accuracy Precision Recall F1-Score GT Support Predicted
367
  ---------------------------------------------------------------------------------------
368
+ A 0.9403 0.1221 0.9403 0.2161 67 516
369
+ # 0.6799 0.9785 0.6799 0.8023 2680 1862
370
+ G 0.9851 0.1257 0.9851 0.2230 67 525
371
+ _ 0.7681 0.8215 0.7681 0.7939 2613 2443
372
+ + 0.9993 0.9910 0.9993 0.9951 9648 9729
373
  ---------------------------------------------------------------------------------------
374
 
375
  =======================================================================================
376
  Size 9, Complexity 0.60
377
  =======================================================================================
378
+ Overall Accuracy: 0.8903 (Baseline: 0.6400, 13725 samples)
379
 
380
  Per-class metrics:
381
  ---------------------------------------------------------------------------------------
382
  Class Accuracy Precision Recall F1-Score GT Support Predicted
383
  ---------------------------------------------------------------------------------------
384
+ A 0.9180 0.1152 0.9180 0.2048 61 486
385
+ # 0.6523 0.9672 0.6523 0.7791 2623 1769
386
+ G 0.9016 0.1206 0.9016 0.2128 61 456
387
+ _ 0.7363 0.7570 0.7363 0.7465 2196 2136
388
+ + 0.9997 0.9891 0.9997 0.9943 8784 8878
389
  ---------------------------------------------------------------------------------------
390
 
391
  =======================================================================================
392
  Size 9, Complexity 0.80
393
  =======================================================================================
394
+ Overall Accuracy: 0.8773 (Baseline: 0.6400, 22725 samples)
395
 
396
  Per-class metrics:
397
  ---------------------------------------------------------------------------------------
398
  Class Accuracy Precision Recall F1-Score GT Support Predicted
399
  ---------------------------------------------------------------------------------------
400
+ A 0.9406 0.1096 0.9406 0.1963 101 867
401
+ # 0.6181 0.9350 0.6181 0.7442 4747 3138
402
+ G 0.8713 0.1117 0.8713 0.1980 101 788
403
+ _ 0.7051 0.6959 0.7051 0.7005 3232 3275
404
+ + 0.9998 0.9921 0.9998 0.9959 14544 14657
405
  ---------------------------------------------------------------------------------------
406
 
407
  =======================================================================================
408
  Size 9, Complexity 1.00
409
  =======================================================================================
410
+ Overall Accuracy: 0.8602 (Baseline: 0.6400, 41400 samples)
411
 
412
  Per-class metrics:
413
  ---------------------------------------------------------------------------------------
414
  Class Accuracy Precision Recall F1-Score GT Support Predicted
415
  ---------------------------------------------------------------------------------------
416
+ A 0.9402 0.1086 0.9402 0.1947 184 1593
417
+ # 0.5750 0.8892 0.5750 0.6984 9200 5949
418
+ G 0.9076 0.1189 0.9076 0.2102 184 1405
419
+ _ 0.6542 0.6173 0.6542 0.6352 5336 5655
420
+ + 0.9998 0.9886 0.9998 0.9942 26496 26798
421
  ---------------------------------------------------------------------------------------
422
 
423
  =======================================================================================
424
  Size 11, Complexity 0.00
425
  =======================================================================================
426
+ Overall Accuracy: 0.9389 (Baseline: 0.4622, 15300 samples)
427
 
428
  Per-class metrics:
429
  ---------------------------------------------------------------------------------------
430
  Class Accuracy Precision Recall F1-Score GT Support Predicted
431
  ---------------------------------------------------------------------------------------
432
+ A 0.9265 0.0983 0.9265 0.1777 68 641
433
+ # 0.9838 0.9864 0.9838 0.9851 2720 2713
434
+ G 0.7500 0.1457 0.7500 0.2440 68 350
435
+ _ 0.8382 0.9980 0.8382 0.9112 5372 4512
436
+ + 1.0000 0.9983 1.0000 0.9992 7072 7084
437
  ---------------------------------------------------------------------------------------
438
 
439
  =======================================================================================
440
  Size 11, Complexity 0.20
441
  =======================================================================================
442
+ Overall Accuracy: 0.8983 (Baseline: 0.4622, 19125 samples)
443
 
444
  Per-class metrics:
445
  ---------------------------------------------------------------------------------------
446
  Class Accuracy Precision Recall F1-Score GT Support Predicted
447
  ---------------------------------------------------------------------------------------
448
+ A 0.8941 0.0942 0.8941 0.1704 85 807
449
+ # 0.8383 0.9733 0.8383 0.9008 3995 3441
450
+ G 0.8706 0.1038 0.8706 0.1855 85 713
451
+ _ 0.7910 0.9124 0.7910 0.8474 6120 5306
452
+ + 1.0000 0.9980 1.0000 0.9990 8840 8858
453
  ---------------------------------------------------------------------------------------
454
 
455
  =======================================================================================
456
  Size 11, Complexity 0.40
457
  =======================================================================================
458
+ Overall Accuracy: 0.8774 (Baseline: 0.4622, 12375 samples)
459
 
460
  Per-class metrics:
461
  ---------------------------------------------------------------------------------------
462
  Class Accuracy Precision Recall F1-Score GT Support Predicted
463
  ---------------------------------------------------------------------------------------
464
+ A 0.8545 0.1051 0.8545 0.1873 55 447
465
+ # 0.7496 0.9447 0.7496 0.8359 2915 2313
466
+ G 0.6909 0.0841 0.6909 0.1499 55 452
467
+ _ 0.7901 0.8349 0.7901 0.8119 3630 3435
468
+ + 1.0000 0.9986 1.0000 0.9993 5720 5728
469
  ---------------------------------------------------------------------------------------
470
 
471
  =======================================================================================
472
  Size 11, Complexity 0.60
473
  =======================================================================================
474
+ Overall Accuracy: 0.8254 (Baseline: 0.4622, 23175 samples)
475
 
476
  Per-class metrics:
477
  ---------------------------------------------------------------------------------------
478
  Class Accuracy Precision Recall F1-Score GT Support Predicted
479
  ---------------------------------------------------------------------------------------
480
+ A 0.7379 0.0860 0.7379 0.1540 103 884
481
+ # 0.7351 0.7741 0.7351 0.7541 6180 5869
482
+ G 0.8058 0.0962 0.8058 0.1718 103 863
483
+ _ 0.6112 0.7693 0.6112 0.6812 6077 4828
484
+ + 1.0000 0.9982 1.0000 0.9991 10712 10731
485
  ---------------------------------------------------------------------------------------
486
 
487
  =======================================================================================
488
  Size 11, Complexity 0.80
489
  =======================================================================================
490
+ Overall Accuracy: 0.8062 (Baseline: 0.4622, 25650 samples)
491
 
492
  Per-class metrics:
493
  ---------------------------------------------------------------------------------------
494
  Class Accuracy Precision Recall F1-Score GT Support Predicted
495
  ---------------------------------------------------------------------------------------
496
+ A 0.8158 0.1352 0.8158 0.2319 114 688
497
+ # 0.7367 0.7253 0.7367 0.7310 7524 7642
498
+ G 0.6404 0.0755 0.6404 0.1351 114 967
499
+ _ 0.5154 0.6935 0.5154 0.5913 6042 4490
500
+ + 0.9999 0.9993 0.9999 0.9996 11856 11863
501
  ---------------------------------------------------------------------------------------
502
 
503
  =======================================================================================
504
  Size 11, Complexity 1.00
505
  =======================================================================================
506
+ Overall Accuracy: 0.8004 (Baseline: 0.4622, 50400 samples)
507
 
508
  Per-class metrics:
509
  ---------------------------------------------------------------------------------------
510
  Class Accuracy Precision Recall F1-Score GT Support Predicted
511
  ---------------------------------------------------------------------------------------
512
+ A 0.8080 0.1268 0.8080 0.2193 224 1427
513
+ # 0.7405 0.7359 0.7405 0.7382 16128 16228
514
+ G 0.6652 0.0770 0.6652 0.1380 224 1936
515
+ _ 0.4553 0.6378 0.4553 0.5313 10528 7515
516
+ + 0.9991 0.9992 0.9991 0.9992 23296 23294
517
  ---------------------------------------------------------------------------------------
518
 
519
  =======================================================================================
520
  Size 13, Complexity 0.00
521
  =======================================================================================
522
+ Overall Accuracy: 0.9323 (Baseline: 0.5289, 14400 samples)
523
 
524
  Per-class metrics:
525
  ---------------------------------------------------------------------------------------
526
  Class Accuracy Precision Recall F1-Score GT Support Predicted
527
  ---------------------------------------------------------------------------------------
528
+ A 0.8125 0.0854 0.8125 0.1545 64 609
529
+ # 0.9827 0.9792 0.9827 0.9810 3072 3083
530
+ G 0.5000 0.0952 0.5000 0.1600 64 336
531
+ _ 0.8864 0.9938 0.8864 0.9371 7616 6793
532
+ + 0.9964 0.9978 0.9964 0.9971 3584 3579
533
  ---------------------------------------------------------------------------------------
534
 
535
  =======================================================================================
536
  Size 13, Complexity 0.20
537
  =======================================================================================
538
+ Overall Accuracy: 0.8866 (Baseline: 0.4844, 13950 samples)
539
 
540
  Per-class metrics:
541
  ---------------------------------------------------------------------------------------
542
  Class Accuracy Precision Recall F1-Score GT Support Predicted
543
  ---------------------------------------------------------------------------------------
544
+ A 0.8710 0.0957 0.8710 0.1725 62 564
545
+ # 0.8212 0.9990 0.8212 0.9014 3596 2956
546
+ G 0.5968 0.0676 0.5968 0.1215 62 547
547
+ _ 0.8662 0.9131 0.8662 0.8891 6758 6411
548
+ + 0.9994 0.9994 0.9994 0.9994 3472 3472
549
  ---------------------------------------------------------------------------------------
550
 
551
  =======================================================================================
552
  Size 13, Complexity 0.40
553
  =======================================================================================
554
+ Overall Accuracy: 0.8528 (Baseline: 0.4400, 28125 samples)
555
 
556
  Per-class metrics:
557
  ---------------------------------------------------------------------------------------
558
  Class Accuracy Precision Recall F1-Score GT Support Predicted
559
  ---------------------------------------------------------------------------------------
560
+ A 0.8480 0.1006 0.8480 0.1798 125 1054
561
+ # 0.6980 0.9803 0.6980 0.8154 8500 6052
562
+ G 0.6800 0.0904 0.6800 0.1596 125 940
563
+ _ 0.8792 0.8315 0.8792 0.8547 12375 13085
564
+ + 0.9973 0.9981 0.9973 0.9977 7000 6994
565
  ---------------------------------------------------------------------------------------
566
 
567
  =======================================================================================
568
  Size 13, Complexity 0.60
569
  =======================================================================================
570
+ Overall Accuracy: 0.8123 (Baseline: 0.3956, 27225 samples)
571
 
572
  Per-class metrics:
573
  ---------------------------------------------------------------------------------------
574
  Class Accuracy Precision Recall F1-Score GT Support Predicted
575
  ---------------------------------------------------------------------------------------
576
+ A 0.6116 0.0803 0.6116 0.1419 121 922
577
+ # 0.6160 0.9586 0.6160 0.7500 9438 6065
578
+ G 0.6198 0.0823 0.6198 0.1453 121 911
579
+ _ 0.8723 0.7479 0.8723 0.8053 10769 12560
580
+ + 0.9973 0.9987 0.9973 0.9980 6776 6767
581
  ---------------------------------------------------------------------------------------
582
 
583
  =======================================================================================
584
  Size 13, Complexity 0.80
585
  =======================================================================================
586
+ Overall Accuracy: 0.7470 (Baseline: 0.3956, 25200 samples)
587
 
588
  Per-class metrics:
589
  ---------------------------------------------------------------------------------------
590
  Class Accuracy Precision Recall F1-Score GT Support Predicted
591
  ---------------------------------------------------------------------------------------
592
+ A 0.6875 0.0981 0.6875 0.1717 112 785
593
+ # 0.6156 0.7995 0.6156 0.6956 9968 7675
594
+ G 0.7768 0.0849 0.7768 0.1530 112 1025
595
+ _ 0.7169 0.6626 0.7169 0.6887 8736 9452
596
+ + 0.9982 0.9997 0.9982 0.9990 6272 6263
597
  ---------------------------------------------------------------------------------------
598
 
599
  =======================================================================================
600
  Size 13, Complexity 1.00
601
  =======================================================================================
602
+ Overall Accuracy: 0.7207 (Baseline: 0.4356, 73350 samples)
603
 
604
  Per-class metrics:
605
  ---------------------------------------------------------------------------------------
606
  Class Accuracy Precision Recall F1-Score GT Support Predicted
607
  ---------------------------------------------------------------------------------------
608
+ A 0.5583 0.1098 0.5583 0.1836 326 1657
609
+ # 0.6593 0.7467 0.6593 0.7003 31948 28207
610
+ G 0.5307 0.0521 0.5307 0.0949 326 3321
611
+ _ 0.5934 0.6055 0.5934 0.5994 22494 22045
612
+ + 0.9914 0.9988 0.9914 0.9951 18256 18120
613
  ---------------------------------------------------------------------------------------
614
 
615
  =======================================================================================
616
  Size 15, Complexity 0.00
617
  =======================================================================================
618
+ Overall Accuracy: 0.8983 (Baseline: 0.7422, 19350 samples)
619
 
620
  Per-class metrics:
621
  ---------------------------------------------------------------------------------------
622
  Class Accuracy Precision Recall F1-Score GT Support Predicted
623
  ---------------------------------------------------------------------------------------
624
+ A 0.7674 0.0596 0.7674 0.1106 86 1108
625
+ # 0.9915 0.9487 0.9915 0.9696 4816 5033
626
+ G 0.6860 0.0845 0.6860 0.1505 86 698
627
+ _ 0.8692 0.9978 0.8692 0.9290 14362 12511
628
+ + 0.0000 0.0000 0.0000 0.0000 0 0
629
  ---------------------------------------------------------------------------------------
630
 
631
  =======================================================================================
632
  Size 15, Complexity 0.20
633
  =======================================================================================
634
+ Overall Accuracy: 0.8256 (Baseline: 0.6756, 19575 samples)
635
 
636
  Per-class metrics:
637
  ---------------------------------------------------------------------------------------
638
  Class Accuracy Precision Recall F1-Score GT Support Predicted
639
  ---------------------------------------------------------------------------------------
640
+ A 0.6437 0.0507 0.6437 0.0940 87 1105
641
+ # 0.7831 0.9396 0.7831 0.8542 6177 5148
642
+ G 0.6092 0.0534 0.6092 0.0981 87 993
643
+ _ 0.8481 0.9096 0.8481 0.8778 13224 12329
644
+ + 0.0000 0.0000 0.0000 0.0000 0 0
645
  ---------------------------------------------------------------------------------------
646
 
647
  =======================================================================================
648
  Size 15, Complexity 0.40
649
  =======================================================================================
650
+ Overall Accuracy: 0.7753 (Baseline: 0.6133, 30600 samples)
651
 
652
  Per-class metrics:
653
  ---------------------------------------------------------------------------------------
654
  Class Accuracy Precision Recall F1-Score GT Support Predicted
655
  ---------------------------------------------------------------------------------------
656
+ A 0.7794 0.0677 0.7794 0.1246 136 1566
657
+ # 0.6554 0.9402 0.6554 0.7724 11560 8059
658
+ G 0.6544 0.0539 0.6544 0.0996 136 1652
659
+ _ 0.8500 0.8256 0.8500 0.8376 18768 19323
660
+ + 0.0000 0.0000 0.0000 0.0000 0 0
661
  ---------------------------------------------------------------------------------------
662
 
663
  =======================================================================================
664
  Size 15, Complexity 0.60
665
  =======================================================================================
666
+ Overall Accuracy: 0.6798 (Baseline: 0.5467, 22500 samples)
667
 
668
  Per-class metrics:
669
  ---------------------------------------------------------------------------------------
670
  Class Accuracy Precision Recall F1-Score GT Support Predicted
671
  ---------------------------------------------------------------------------------------
672
+ A 0.7000 0.0686 0.7000 0.1249 100 1021
673
+ # 0.5983 0.8057 0.5983 0.6867 10000 7426
674
+ G 0.6400 0.0422 0.6400 0.0792 100 1516
675
+ _ 0.7462 0.7321 0.7462 0.7391 12300 12537
676
+ + 0.0000 0.0000 0.0000 0.0000 0 0
677
  ---------------------------------------------------------------------------------------
678
 
679
  =======================================================================================
680
  Size 15, Complexity 0.80
681
  =======================================================================================
682
+ Overall Accuracy: 0.5970 (Baseline: 0.5067, 36900 samples)
683
 
684
  Per-class metrics:
685
  ---------------------------------------------------------------------------------------
686
  Class Accuracy Precision Recall F1-Score GT Support Predicted
687
  ---------------------------------------------------------------------------------------
688
+ A 0.5976 0.0588 0.5976 0.1070 164 1668
689
+ # 0.6700 0.6540 0.6700 0.6619 18696 19153
690
+ G 0.6098 0.0530 0.6098 0.0976 164 1886
691
+ _ 0.5206 0.6557 0.5206 0.5804 17876 14193
692
+ + 0.0000 0.0000 0.0000 0.0000 0 0
693
  ---------------------------------------------------------------------------------------
694
 
695
  =======================================================================================
696
  Size 15, Complexity 1.00
697
  =======================================================================================
698
+ Overall Accuracy: 0.5941 (Baseline: 0.5689, 87975 samples)
699
 
700
  Per-class metrics:
701
  ---------------------------------------------------------------------------------------
702
  Class Accuracy Precision Recall F1-Score GT Support Predicted
703
  ---------------------------------------------------------------------------------------
704
+ A 0.5320 0.0748 0.5320 0.1311 391 2781
705
+ # 0.7659 0.6496 0.7659 0.7030 50048 59011
706
+ G 0.6215 0.0650 0.6215 0.1176 391 3740
707
+ _ 0.3628 0.6005 0.3628 0.4524 37145 22443
708
+ + 0.0000 0.0000 0.0000 0.0000 0 0
709
  ---------------------------------------------------------------------------------------
710
 
711
+ Results saved to: reveng/cognitive_map_probes_results/layer23/mlp_general/pre_reasoning/eval_cognitive_map_probe_layer23_mlp_pre_reasoning_all_general.json
layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.json CHANGED
@@ -1,274 +1,274 @@
1
  {
2
  "global": {
3
- "accuracy": 0.10631741140215717,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
- "accuracy": 0.7935285053929122,
8
- "precision": 0.008264462809917356,
9
- "recall": 0.7935285053929122,
10
- "f1": 0.016358554094403153,
11
  "gt_support": 649,
12
- "predicted": 62315
13
  },
14
  "1": {
15
- "accuracy": 0.1332421063301404,
16
- "precision": 0.5160973694542599,
17
- "recall": 0.1332421063301404,
18
- "f1": 0.21180261832829808,
19
  "gt_support": 39462,
20
- "predicted": 10188
21
  },
22
  "2": {
23
- "accuracy": 0.012326656394453005,
24
- "precision": 0.008264462809917356,
25
- "recall": 0.012326656394453005,
26
- "f1": 0.009894867037724181,
27
  "gt_support": 649,
28
- "predicted": 968
29
  },
30
  "3": {
31
- "accuracy": 0.06799226879186634,
32
- "precision": 0.5077105575326216,
33
- "recall": 0.06799226879186634,
34
- "f1": 0.11992434679057604,
35
  "gt_support": 37769,
36
- "predicted": 5058
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
- "accuracy": 0.11728245017015071,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
- "accuracy": 0.7647058823529411,
48
- "precision": 0.008264462809917356,
49
- "recall": 0.7647058823529411,
50
- "f1": 0.016352201257861635,
51
  "gt_support": 68,
52
- "predicted": 6292
53
  },
54
  "1": {
55
- "accuracy": 0.10294117647058823,
56
- "precision": 0.3305785123966942,
57
- "recall": 0.10294117647058823,
58
- "f1": 0.1569946733950098,
59
  "gt_support": 2720,
60
- "predicted": 847
61
  },
62
  "2": {
63
- "accuracy": 0.014705882352941176,
64
- "precision": 0.008264462809917356,
65
- "recall": 0.014705882352941176,
66
- "f1": 0.010582010582010581,
67
  "gt_support": 68,
68
- "predicted": 121
69
  },
70
  "3": {
71
- "accuracy": 0.11764705882352941,
72
- "precision": 0.6528925619834711,
73
- "recall": 0.11764705882352941,
74
- "f1": 0.19936908517350158,
75
  "gt_support": 5372,
76
- "predicted": 968
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
- "accuracy": 0.09927078269324259,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
- "accuracy": 0.8117647058823529,
87
- "precision": 0.008264462809917356,
88
- "recall": 0.8117647058823529,
89
- "f1": 0.016362342897794643,
90
  "gt_support": 85,
91
- "predicted": 8349
92
  },
93
  "1": {
94
- "accuracy": 0.09411764705882353,
95
- "precision": 0.3884297520661157,
96
- "recall": 0.09411764705882353,
97
- "f1": 0.15152125730404997,
98
  "gt_support": 3995,
99
- "predicted": 968
100
  },
101
  "2": {
102
- "accuracy": 0.0,
103
- "precision": 0.0,
104
- "recall": 0.0,
105
- "f1": 0.0,
106
  "gt_support": 85,
107
- "predicted": 0
108
  },
109
  "3": {
110
- "accuracy": 0.09411764705882353,
111
- "precision": 0.5950413223140496,
112
- "recall": 0.09411764705882353,
113
- "f1": 0.16252821670428894,
114
  "gt_support": 6120,
115
- "predicted": 968
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
- "accuracy": 0.10788880540946656,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
- "accuracy": 0.7454545454545455,
126
- "precision": 0.008264462809917356,
127
- "recall": 0.7454545454545455,
128
- "f1": 0.01634768740031898,
129
  "gt_support": 55,
130
- "predicted": 4961
131
  },
132
  "1": {
133
- "accuracy": 0.16363636363636364,
134
- "precision": 0.4380165289256198,
135
- "recall": 0.16363636363636364,
136
- "f1": 0.23826173826173822,
137
  "gt_support": 2915,
138
- "predicted": 1089
139
  },
140
  "2": {
141
- "accuracy": 0.03636363636363636,
142
- "precision": 0.008264462809917356,
143
- "recall": 0.03636363636363636,
144
- "f1": 0.013468013468013467,
145
  "gt_support": 55,
146
- "predicted": 242
147
  },
148
  "3": {
149
- "accuracy": 0.05454545454545454,
150
- "precision": 0.5454545454545454,
151
- "recall": 0.05454545454545454,
152
- "f1": 0.09917355371900825,
153
  "gt_support": 3630,
154
- "predicted": 363
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
- "accuracy": 0.0883414908128059,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
- "accuracy": 0.8349514563106796,
165
- "precision": 0.008264462809917356,
166
- "recall": 0.8349514563106796,
167
- "f1": 0.016366923589304406,
168
  "gt_support": 103,
169
- "predicted": 10406
170
  },
171
  "1": {
172
- "accuracy": 0.11650485436893204,
173
- "precision": 0.49586776859504134,
174
- "recall": 0.11650485436893204,
175
- "f1": 0.18867924528301885,
176
  "gt_support": 6180,
177
- "predicted": 1452
178
  },
179
  "2": {
180
- "accuracy": 0.0,
181
- "precision": 0.0,
182
- "recall": 0.0,
183
- "f1": 0.0,
184
  "gt_support": 103,
185
- "predicted": 0
186
  },
187
  "3": {
188
- "accuracy": 0.04854368932038835,
189
- "precision": 0.48760330578512395,
190
- "recall": 0.04854368932038835,
191
- "f1": 0.0882969170906914,
192
  "gt_support": 6077,
193
- "predicted": 605
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
- "accuracy": 0.07894736842105263,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
- "accuracy": 0.8596491228070176,
204
- "precision": 0.008264462809917356,
205
- "recall": 0.8596491228070176,
206
- "f1": 0.016371533578349484,
207
  "gt_support": 114,
208
- "predicted": 11858
209
  },
210
  "1": {
211
- "accuracy": 0.09649122807017543,
212
- "precision": 0.5454545454545454,
213
- "recall": 0.09649122807017543,
214
- "f1": 0.1639751552795031,
215
  "gt_support": 7524,
216
- "predicted": 1331
217
  },
218
  "2": {
219
- "accuracy": 0.0,
220
- "precision": 0.0,
221
- "recall": 0.0,
222
- "f1": 0.0,
223
  "gt_support": 114,
224
- "predicted": 0
225
  },
226
  "3": {
227
- "accuracy": 0.043859649122807015,
228
- "precision": 0.4380165289256198,
229
- "recall": 0.043859649122807015,
230
- "f1": 0.07973521889574244,
231
  "gt_support": 6042,
232
- "predicted": 605
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
- "accuracy": 0.1274719598583235,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
- "accuracy": 0.7544642857142857,
243
- "precision": 0.008264462809917356,
244
- "recall": 0.7544642857142857,
245
- "f1": 0.016349828278430806,
246
  "gt_support": 224,
247
- "predicted": 20449
248
  },
249
  "1": {
250
- "accuracy": 0.16610863095238096,
251
- "precision": 0.5952010664296823,
252
- "recall": 0.16610863095238096,
253
- "f1": 0.2597314460225896,
254
  "gt_support": 16128,
255
- "predicted": 4501
256
  },
257
  "2": {
258
- "accuracy": 0.022321428571428572,
259
- "precision": 0.008264462809917356,
260
- "recall": 0.022321428571428572,
261
- "f1": 0.012062726176115802,
262
  "gt_support": 224,
263
- "predicted": 605
264
  },
265
  "3": {
266
- "accuracy": 0.057180851063829786,
267
- "precision": 0.38863783085861847,
268
- "recall": 0.057180851063829786,
269
- "f1": 0.09969363252463359,
270
  "gt_support": 10528,
271
- "predicted": 1549
272
  }
273
  },
274
  "total_samples": 27104
@@ -278,13 +278,13 @@
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
- "probe_path": "interp/probes_train_single_step/cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "7",
285
  "steps": "all",
286
  "token_categories": {
287
- "output": "all"
288
  },
289
  "pad_to_size": 11
290
  }
 
1
  {
2
  "global": {
3
+ "accuracy": 0.44538960129378957,
4
  "baseline_accuracy": 0.5025149944606451,
5
  "per_class": {
6
  "0": {
7
+ "accuracy": 0.11093990755007704,
8
+ "precision": 0.009108159392789373,
9
+ "recall": 0.11093990755007704,
10
+ "f1": 0.016834229600187045,
11
  "gt_support": 649,
12
+ "predicted": 7905
13
  },
14
  "1": {
15
+ "accuracy": 0.541888398966094,
16
+ "precision": 0.5624556142981141,
17
+ "recall": 0.541888398966094,
18
+ "f1": 0.55198048553839,
19
  "gt_support": 39462,
20
+ "predicted": 38019
21
  },
22
  "2": {
23
+ "accuracy": 0.10785824345146379,
24
+ "precision": 0.008359207069500836,
25
+ "recall": 0.10785824345146379,
26
+ "f1": 0.01551590380139643,
27
  "gt_support": 649,
28
+ "predicted": 8374
29
  },
30
  "3": {
31
+ "accuracy": 0.35611215547141833,
32
+ "precision": 0.5550740786595684,
33
+ "recall": 0.35611215547141833,
34
+ "f1": 0.4338709677419355,
35
  "gt_support": 37769,
36
+ "predicted": 24231
37
  }
38
  },
39
  "total_samples": 78529
40
  },
41
  "by_complexity": {
42
  "0.0": {
43
+ "accuracy": 0.3714146815751094,
44
  "baseline_accuracy": 0.6528925619834711,
45
  "per_class": {
46
  "0": {
47
+ "accuracy": 0.014705882352941176,
48
+ "precision": 0.008928571428571428,
49
+ "recall": 0.014705882352941176,
50
+ "f1": 0.011111111111111112,
51
  "gt_support": 68,
52
+ "predicted": 112
53
  },
54
  "1": {
55
+ "accuracy": 0.0,
56
+ "precision": 0.0,
57
+ "recall": 0.0,
58
+ "f1": 0.0,
59
  "gt_support": 2720,
60
+ "predicted": 0
61
  },
62
  "2": {
63
+ "accuracy": 0.4117647058823529,
64
+ "precision": 0.008069164265129682,
65
+ "recall": 0.4117647058823529,
66
+ "f1": 0.01582815149802148,
67
  "gt_support": 68,
68
+ "predicted": 3470
69
  },
70
  "3": {
71
+ "accuracy": 0.5634772896500372,
72
+ "precision": 0.6515281962978906,
73
+ "recall": 0.5634772896500372,
74
+ "f1": 0.604312237971651,
75
  "gt_support": 5372,
76
+ "predicted": 4646
77
  }
78
  },
79
  "total_samples": 8228
80
  },
81
  "0.2": {
82
+ "accuracy": 0.43325230918813806,
83
  "baseline_accuracy": 0.5950413223140496,
84
  "per_class": {
85
  "0": {
86
+ "accuracy": 0.058823529411764705,
87
+ "precision": 0.007462686567164179,
88
+ "recall": 0.058823529411764705,
89
+ "f1": 0.013245033112582783,
90
  "gt_support": 85,
91
+ "predicted": 670
92
  },
93
  "1": {
94
+ "accuracy": 0.06483103879849812,
95
+ "precision": 0.39064856711915535,
96
+ "recall": 0.06483103879849812,
97
+ "f1": 0.1112065264061829,
98
  "gt_support": 3995,
99
+ "predicted": 663
100
  },
101
  "2": {
102
+ "accuracy": 0.18823529411764706,
103
+ "precision": 0.008184143222506393,
104
+ "recall": 0.18823529411764706,
105
+ "f1": 0.01568627450980392,
106
  "gt_support": 85,
107
+ "predicted": 1955
108
  },
109
  "3": {
110
+ "accuracy": 0.6823529411764706,
111
+ "precision": 0.5968272116621409,
112
+ "recall": 0.6823529411764706,
113
+ "f1": 0.6367309598231303,
114
  "gt_support": 6120,
115
+ "predicted": 6997
116
  }
117
  },
118
  "total_samples": 10285
119
  },
120
  "0.4": {
121
+ "accuracy": 0.3912847483095417,
122
  "baseline_accuracy": 0.5454545454545454,
123
  "per_class": {
124
  "0": {
125
+ "accuracy": 0.07272727272727272,
126
+ "precision": 0.009779951100244499,
127
+ "recall": 0.07272727272727272,
128
+ "f1": 0.017241379310344827,
129
  "gt_support": 55,
130
+ "predicted": 409
131
  },
132
  "1": {
133
+ "accuracy": 0.03739279588336192,
134
+ "precision": 0.4579831932773109,
135
+ "recall": 0.03739279588336192,
136
+ "f1": 0.0691405011100539,
137
  "gt_support": 2915,
138
+ "predicted": 238
139
  },
140
  "2": {
141
+ "accuracy": 0.23636363636363636,
142
+ "precision": 0.008837525492861998,
143
+ "recall": 0.23636363636363636,
144
+ "f1": 0.01703800786369594,
145
  "gt_support": 55,
146
+ "predicted": 1471
147
  },
148
  "3": {
149
+ "accuracy": 0.6826446280991736,
150
+ "precision": 0.5461758871500992,
151
+ "recall": 0.6826446280991736,
152
+ "f1": 0.6068323741888086,
153
  "gt_support": 3630,
154
+ "predicted": 4537
155
  }
156
  },
157
  "total_samples": 6655
158
  },
159
  "0.6": {
160
+ "accuracy": 0.4393003289737623,
161
  "baseline_accuracy": 0.49586776859504134,
162
  "per_class": {
163
  "0": {
164
+ "accuracy": 0.05825242718446602,
165
+ "precision": 0.008426966292134831,
166
+ "recall": 0.05825242718446602,
167
+ "f1": 0.0147239263803681,
168
  "gt_support": 103,
169
+ "predicted": 712
170
  },
171
  "1": {
172
+ "accuracy": 0.4825242718446602,
173
+ "precision": 0.4962556165751373,
174
+ "recall": 0.4825242718446602,
175
+ "f1": 0.48929362539995075,
176
  "gt_support": 6180,
177
+ "predicted": 6009
178
  },
179
  "2": {
180
+ "accuracy": 0.05825242718446602,
181
+ "precision": 0.00946372239747634,
182
+ "recall": 0.05825242718446602,
183
+ "f1": 0.01628222523744912,
184
  "gt_support": 103,
185
+ "predicted": 634
186
  },
187
  "3": {
188
+ "accuracy": 0.4082606549284186,
189
+ "precision": 0.485708692247455,
190
+ "recall": 0.4082606549284186,
191
+ "f1": 0.44362986142154676,
192
  "gt_support": 6077,
193
+ "predicted": 5108
194
  }
195
  },
196
  "total_samples": 12463
197
  },
198
  "0.8": {
199
+ "accuracy": 0.4764390314629549,
200
  "baseline_accuracy": 0.5454545454545454,
201
  "per_class": {
202
  "0": {
203
+ "accuracy": 0.043859649122807015,
204
+ "precision": 0.006830601092896175,
205
+ "recall": 0.043859649122807015,
206
+ "f1": 0.01182033096926714,
207
  "gt_support": 114,
208
+ "predicted": 732
209
  },
210
  "1": {
211
+ "accuracy": 0.710127591706539,
212
+ "precision": 0.5455937914837129,
213
+ "recall": 0.710127591706539,
214
+ "f1": 0.6170814806259746,
215
  "gt_support": 7524,
216
+ "predicted": 9793
217
  },
218
  "2": {
219
+ "accuracy": 0.03508771929824561,
220
+ "precision": 0.008016032064128256,
221
+ "recall": 0.03508771929824561,
222
+ "f1": 0.013050570962479607,
223
  "gt_support": 114,
224
+ "predicted": 499
225
  },
226
  "3": {
227
+ "accuracy": 0.20191989407480967,
228
+ "precision": 0.4404332129963899,
229
+ "recall": 0.20191989407480967,
230
+ "f1": 0.27689514298683615,
231
  "gt_support": 6042,
232
+ "predicted": 2770
233
  }
234
  },
235
  "total_samples": 13794
236
  },
237
  "1.0": {
238
+ "accuracy": 0.4727346517119244,
239
  "baseline_accuracy": 0.5950413223140496,
240
  "per_class": {
241
  "0": {
242
+ "accuracy": 0.22767857142857142,
243
+ "precision": 0.00967741935483871,
244
+ "recall": 0.22767857142857142,
245
+ "f1": 0.018565708045140156,
246
  "gt_support": 224,
247
+ "predicted": 5270
248
  },
249
  "1": {
250
+ "accuracy": 0.7868923611111112,
251
+ "precision": 0.5953743666729218,
252
+ "recall": 0.7868923611111112,
253
+ "f1": 0.6778656126482213,
254
  "gt_support": 16128,
255
+ "predicted": 21316
256
  },
257
  "2": {
258
+ "accuracy": 0.013392857142857142,
259
+ "precision": 0.008695652173913044,
260
+ "recall": 0.013392857142857142,
261
+ "f1": 0.01054481546572935,
262
  "gt_support": 224,
263
+ "predicted": 345
264
  },
265
  "3": {
266
+ "accuracy": 0.006458966565349544,
267
+ "precision": 0.3930635838150289,
268
+ "recall": 0.006458966565349544,
269
+ "f1": 0.01270909260816746,
270
  "gt_support": 10528,
271
+ "predicted": 173
272
  }
273
  },
274
  "total_samples": 27104
 
278
  "total_steps": 649,
279
  "single_size_mode": true,
280
  "config": {
281
+ "probe_path": "interp/cognitive_map_probes/cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.pt",
282
  "trajectories_dir": "reveng/trajectories_test_full/size11",
283
  "activations_dir": "interp/activations_test_full/size11",
284
  "layers": "7",
285
  "steps": "all",
286
  "token_categories": {
287
+ "prompt_suffix": "all"
288
  },
289
  "pad_to_size": 11
290
  }
layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.txt CHANGED
@@ -1,11 +1,11 @@
1
  Using device: cuda
2
- Loading probe from interp/probes_train_single_step/cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer7_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
- Token categories: {'output': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
@@ -20,16 +20,16 @@ Processed 60 trajectories, 649 steps
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
- Overall Accuracy: 0.1063 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
- A 0.7935 0.0083 0.7935 0.0164 649 62315
30
- # 0.1332 0.5161 0.1332 0.2118 39462 10188
31
- G 0.0123 0.0083 0.0123 0.0099 649 968
32
- _ 0.0680 0.5077 0.0680 0.1199 37769 5058
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
@@ -39,91 +39,91 @@ METRICS BY COMPLEXITY
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
- Overall Accuracy: 0.1173 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
- A 0.7647 0.0083 0.7647 0.0164 68 6292
49
- # 0.1029 0.3306 0.1029 0.1570 2720 847
50
- G 0.0147 0.0083 0.0147 0.0106 68 121
51
- _ 0.1176 0.6529 0.1176 0.1994 5372 968
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
- Overall Accuracy: 0.0993 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
- A 0.8118 0.0083 0.8118 0.0164 85 8349
64
- # 0.0941 0.3884 0.0941 0.1515 3995 968
65
- G 0.0000 0.0000 0.0000 0.0000 85 0
66
- _ 0.0941 0.5950 0.0941 0.1625 6120 968
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
- Overall Accuracy: 0.1079 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
- A 0.7455 0.0083 0.7455 0.0163 55 4961
79
- # 0.1636 0.4380 0.1636 0.2383 2915 1089
80
- G 0.0364 0.0083 0.0364 0.0135 55 242
81
- _ 0.0545 0.5455 0.0545 0.0992 3630 363
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
- Overall Accuracy: 0.0883 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
- A 0.8350 0.0083 0.8350 0.0164 103 10406
94
- # 0.1165 0.4959 0.1165 0.1887 6180 1452
95
- G 0.0000 0.0000 0.0000 0.0000 103 0
96
- _ 0.0485 0.4876 0.0485 0.0883 6077 605
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
- Overall Accuracy: 0.0789 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
- A 0.8596 0.0083 0.8596 0.0164 114 11858
109
- # 0.0965 0.5455 0.0965 0.1640 7524 1331
110
- G 0.0000 0.0000 0.0000 0.0000 114 0
111
- _ 0.0439 0.4380 0.0439 0.0797 6042 605
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
- Overall Accuracy: 0.1275 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
- A 0.7545 0.0083 0.7545 0.0163 224 20449
124
- # 0.1661 0.5952 0.1661 0.2597 16128 4501
125
- G 0.0223 0.0083 0.0223 0.0121 224 605
126
- _ 0.0572 0.3886 0.0572 0.0997 10528 1549
127
  ---------------------------------------------------------------------------------------
128
 
129
- Results saved to: reveng/trajectories_test_full_with_probes/layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.json
 
1
  Using device: cuda
2
+ Loading probe from interp/cognitive_map_probes/cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.pt
3
  Loaded probe: cognitive_map_probe_layer7_lr_pre_reasoning_all_size11
4
  Input dimension: 8642
5
  Number of classes: 4
6
  Normalized: True
7
 
8
+ Token categories: {'prompt_suffix': 'all'}
9
 
10
  No size folders found. Running in single-size mode with 60 trajectories
11
 
 
20
  =======================================================================================
21
  GLOBAL METRICS
22
  =======================================================================================
23
+ Overall Accuracy: 0.4454 (Baseline: 0.5025, 78529 samples)
24
 
25
  Per-class metrics:
26
  ---------------------------------------------------------------------------------------
27
  Class Accuracy Precision Recall F1-Score GT Support Predicted
28
  ---------------------------------------------------------------------------------------
29
+ A 0.1109 0.0091 0.1109 0.0168 649 7905
30
+ # 0.5419 0.5625 0.5419 0.5520 39462 38019
31
+ G 0.1079 0.0084 0.1079 0.0155 649 8374
32
+ _ 0.3561 0.5551 0.3561 0.4339 37769 24231
33
  ---------------------------------------------------------------------------------------
34
 
35
  =======================================================================================
 
39
  =======================================================================================
40
  Complexity 0.00
41
  =======================================================================================
42
+ Overall Accuracy: 0.3714 (Baseline: 0.6529, 8228 samples)
43
 
44
  Per-class metrics:
45
  ---------------------------------------------------------------------------------------
46
  Class Accuracy Precision Recall F1-Score GT Support Predicted
47
  ---------------------------------------------------------------------------------------
48
+ A 0.0147 0.0089 0.0147 0.0111 68 112
49
+ # 0.0000 0.0000 0.0000 0.0000 2720 0
50
+ G 0.4118 0.0081 0.4118 0.0158 68 3470
51
+ _ 0.5635 0.6515 0.5635 0.6043 5372 4646
52
  ---------------------------------------------------------------------------------------
53
 
54
  =======================================================================================
55
  Complexity 0.20
56
  =======================================================================================
57
+ Overall Accuracy: 0.4333 (Baseline: 0.5950, 10285 samples)
58
 
59
  Per-class metrics:
60
  ---------------------------------------------------------------------------------------
61
  Class Accuracy Precision Recall F1-Score GT Support Predicted
62
  ---------------------------------------------------------------------------------------
63
+ A 0.0588 0.0075 0.0588 0.0132 85 670
64
+ # 0.0648 0.3906 0.0648 0.1112 3995 663
65
+ G 0.1882 0.0082 0.1882 0.0157 85 1955
66
+ _ 0.6824 0.5968 0.6824 0.6367 6120 6997
67
  ---------------------------------------------------------------------------------------
68
 
69
  =======================================================================================
70
  Complexity 0.40
71
  =======================================================================================
72
+ Overall Accuracy: 0.3913 (Baseline: 0.5455, 6655 samples)
73
 
74
  Per-class metrics:
75
  ---------------------------------------------------------------------------------------
76
  Class Accuracy Precision Recall F1-Score GT Support Predicted
77
  ---------------------------------------------------------------------------------------
78
+ A 0.0727 0.0098 0.0727 0.0172 55 409
79
+ # 0.0374 0.4580 0.0374 0.0691 2915 238
80
+ G 0.2364 0.0088 0.2364 0.0170 55 1471
81
+ _ 0.6826 0.5462 0.6826 0.6068 3630 4537
82
  ---------------------------------------------------------------------------------------
83
 
84
  =======================================================================================
85
  Complexity 0.60
86
  =======================================================================================
87
+ Overall Accuracy: 0.4393 (Baseline: 0.4959, 12463 samples)
88
 
89
  Per-class metrics:
90
  ---------------------------------------------------------------------------------------
91
  Class Accuracy Precision Recall F1-Score GT Support Predicted
92
  ---------------------------------------------------------------------------------------
93
+ A 0.0583 0.0084 0.0583 0.0147 103 712
94
+ # 0.4825 0.4963 0.4825 0.4893 6180 6009
95
+ G 0.0583 0.0095 0.0583 0.0163 103 634
96
+ _ 0.4083 0.4857 0.4083 0.4436 6077 5108
97
  ---------------------------------------------------------------------------------------
98
 
99
  =======================================================================================
100
  Complexity 0.80
101
  =======================================================================================
102
+ Overall Accuracy: 0.4764 (Baseline: 0.5455, 13794 samples)
103
 
104
  Per-class metrics:
105
  ---------------------------------------------------------------------------------------
106
  Class Accuracy Precision Recall F1-Score GT Support Predicted
107
  ---------------------------------------------------------------------------------------
108
+ A 0.0439 0.0068 0.0439 0.0118 114 732
109
+ # 0.7101 0.5456 0.7101 0.6171 7524 9793
110
+ G 0.0351 0.0080 0.0351 0.0131 114 499
111
+ _ 0.2019 0.4404 0.2019 0.2769 6042 2770
112
  ---------------------------------------------------------------------------------------
113
 
114
  =======================================================================================
115
  Complexity 1.00
116
  =======================================================================================
117
+ Overall Accuracy: 0.4727 (Baseline: 0.5950, 27104 samples)
118
 
119
  Per-class metrics:
120
  ---------------------------------------------------------------------------------------
121
  Class Accuracy Precision Recall F1-Score GT Support Predicted
122
  ---------------------------------------------------------------------------------------
123
+ A 0.2277 0.0097 0.2277 0.0186 224 5270
124
+ # 0.7869 0.5954 0.7869 0.6779 16128 21316
125
+ G 0.0134 0.0087 0.0134 0.0105 224 345
126
+ _ 0.0065 0.3931 0.0065 0.0127 10528 173
127
  ---------------------------------------------------------------------------------------
128
 
129
+ Results saved to: reveng/cognitive_map_probes_results/layer7/lr/pre_reasoning/eval_cognitive_map_probe_layer7_lr_pre_reasoning_all_size11.json