Nekochu commited on
Commit
9df4723
·
verified ·
1 Parent(s): 1e79d3b

Add KLD measurement

Browse files
Files changed (1) hide show
  1. measurement.json +1814 -0
measurement.json ADDED
@@ -0,0 +1,1814 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base": {
3
+ "dir": "D:\\main\\program\\AI\\CoridorGreenScreenMattingCPUfree\\tmp\\quants\\Gemma-3-R1984-27B-2.0bpw",
4
+ "bpw": 2.004435066344246
5
+ },
6
+ "alts": [
7
+ {
8
+ "dir": "D:\\main\\program\\AI\\CoridorGreenScreenMattingCPUfree\\tmp\\quants\\Gemma-3-R1984-27B-8.0bpw",
9
+ "bpw": 8.004435066344247
10
+ }
11
+ ],
12
+ "groups": [
13
+ {
14
+ "idx": 0,
15
+ "layers": [
16
+ "language_model.model.layers.0.self_attn.q_proj",
17
+ "language_model.model.layers.0.self_attn.k_proj",
18
+ "language_model.model.layers.0.self_attn.v_proj",
19
+ "language_model.model.layers.0.self_attn.o_proj"
20
+ ],
21
+ "candidates": [
22
+ {
23
+ "dkld": -0.006917499005794525,
24
+ "dbits": 396361728
25
+ }
26
+ ]
27
+ },
28
+ {
29
+ "idx": 1,
30
+ "layers": [
31
+ "language_model.model.layers.0.mlp.gate_proj",
32
+ "language_model.model.layers.0.mlp.up_proj",
33
+ "language_model.model.layers.0.mlp.down_proj"
34
+ ],
35
+ "candidates": [
36
+ {
37
+ "dkld": -0.004495760053396247,
38
+ "dbits": 2080899072
39
+ }
40
+ ]
41
+ },
42
+ {
43
+ "idx": 2,
44
+ "layers": [
45
+ "language_model.model.layers.1.self_attn.q_proj",
46
+ "language_model.model.layers.1.self_attn.k_proj",
47
+ "language_model.model.layers.1.self_attn.v_proj",
48
+ "language_model.model.layers.1.self_attn.o_proj"
49
+ ],
50
+ "candidates": [
51
+ {
52
+ "dkld": 0.0004842020571231842,
53
+ "dbits": 396361728
54
+ }
55
+ ]
56
+ },
57
+ {
58
+ "idx": 3,
59
+ "layers": [
60
+ "language_model.model.layers.1.mlp.gate_proj",
61
+ "language_model.model.layers.1.mlp.up_proj",
62
+ "language_model.model.layers.1.mlp.down_proj"
63
+ ],
64
+ "candidates": [
65
+ {
66
+ "dkld": -0.005311993509531043,
67
+ "dbits": 2080899072
68
+ }
69
+ ]
70
+ },
71
+ {
72
+ "idx": 4,
73
+ "layers": [
74
+ "language_model.model.layers.2.self_attn.q_proj",
75
+ "language_model.model.layers.2.self_attn.k_proj",
76
+ "language_model.model.layers.2.self_attn.v_proj",
77
+ "language_model.model.layers.2.self_attn.o_proj"
78
+ ],
79
+ "candidates": [
80
+ {
81
+ "dkld": -0.0011243283748627042,
82
+ "dbits": 396361728
83
+ }
84
+ ]
85
+ },
86
+ {
87
+ "idx": 5,
88
+ "layers": [
89
+ "language_model.model.layers.2.mlp.gate_proj",
90
+ "language_model.model.layers.2.mlp.up_proj",
91
+ "language_model.model.layers.2.mlp.down_proj"
92
+ ],
93
+ "candidates": [
94
+ {
95
+ "dkld": -0.009011824429035198,
96
+ "dbits": 2080899072
97
+ }
98
+ ]
99
+ },
100
+ {
101
+ "idx": 6,
102
+ "layers": [
103
+ "language_model.model.layers.3.self_attn.q_proj",
104
+ "language_model.model.layers.3.self_attn.k_proj",
105
+ "language_model.model.layers.3.self_attn.v_proj",
106
+ "language_model.model.layers.3.self_attn.o_proj"
107
+ ],
108
+ "candidates": [
109
+ {
110
+ "dkld": 0.003413629531860307,
111
+ "dbits": 396361728
112
+ }
113
+ ]
114
+ },
115
+ {
116
+ "idx": 7,
117
+ "layers": [
118
+ "language_model.model.layers.3.mlp.gate_proj",
119
+ "language_model.model.layers.3.mlp.up_proj",
120
+ "language_model.model.layers.3.mlp.down_proj"
121
+ ],
122
+ "candidates": [
123
+ {
124
+ "dkld": -0.006673009693622611,
125
+ "dbits": 2080899072
126
+ }
127
+ ]
128
+ },
129
+ {
130
+ "idx": 8,
131
+ "layers": [
132
+ "language_model.model.layers.4.self_attn.q_proj",
133
+ "language_model.model.layers.4.self_attn.k_proj",
134
+ "language_model.model.layers.4.self_attn.v_proj",
135
+ "language_model.model.layers.4.self_attn.o_proj"
136
+ ],
137
+ "candidates": [
138
+ {
139
+ "dkld": 0.0005826622247695479,
140
+ "dbits": 396361728
141
+ }
142
+ ]
143
+ },
144
+ {
145
+ "idx": 9,
146
+ "layers": [
147
+ "language_model.model.layers.4.mlp.gate_proj",
148
+ "language_model.model.layers.4.mlp.up_proj",
149
+ "language_model.model.layers.4.mlp.down_proj"
150
+ ],
151
+ "candidates": [
152
+ {
153
+ "dkld": -0.0080329388380051,
154
+ "dbits": 2080899072
155
+ }
156
+ ]
157
+ },
158
+ {
159
+ "idx": 10,
160
+ "layers": [
161
+ "language_model.model.layers.5.self_attn.q_proj",
162
+ "language_model.model.layers.5.self_attn.k_proj",
163
+ "language_model.model.layers.5.self_attn.v_proj",
164
+ "language_model.model.layers.5.self_attn.o_proj"
165
+ ],
166
+ "candidates": [
167
+ {
168
+ "dkld": 0.0023767203092575073,
169
+ "dbits": 396361728
170
+ }
171
+ ]
172
+ },
173
+ {
174
+ "idx": 11,
175
+ "layers": [
176
+ "language_model.model.layers.5.mlp.gate_proj",
177
+ "language_model.model.layers.5.mlp.up_proj",
178
+ "language_model.model.layers.5.mlp.down_proj"
179
+ ],
180
+ "candidates": [
181
+ {
182
+ "dkld": -0.010135890543460868,
183
+ "dbits": 2080899072
184
+ }
185
+ ]
186
+ },
187
+ {
188
+ "idx": 12,
189
+ "layers": [
190
+ "language_model.model.layers.6.self_attn.q_proj",
191
+ "language_model.model.layers.6.self_attn.k_proj",
192
+ "language_model.model.layers.6.self_attn.v_proj",
193
+ "language_model.model.layers.6.self_attn.o_proj"
194
+ ],
195
+ "candidates": [
196
+ {
197
+ "dkld": -0.0027037844061851835,
198
+ "dbits": 396361728
199
+ }
200
+ ]
201
+ },
202
+ {
203
+ "idx": 13,
204
+ "layers": [
205
+ "language_model.model.layers.6.mlp.gate_proj",
206
+ "language_model.model.layers.6.mlp.up_proj",
207
+ "language_model.model.layers.6.mlp.down_proj"
208
+ ],
209
+ "candidates": [
210
+ {
211
+ "dkld": -0.0041308201849460935,
212
+ "dbits": 2080899072
213
+ }
214
+ ]
215
+ },
216
+ {
217
+ "idx": 14,
218
+ "layers": [
219
+ "language_model.model.layers.7.self_attn.q_proj",
220
+ "language_model.model.layers.7.self_attn.k_proj",
221
+ "language_model.model.layers.7.self_attn.v_proj",
222
+ "language_model.model.layers.7.self_attn.o_proj"
223
+ ],
224
+ "candidates": [
225
+ {
226
+ "dkld": -0.0014376796782016865,
227
+ "dbits": 396361728
228
+ }
229
+ ]
230
+ },
231
+ {
232
+ "idx": 15,
233
+ "layers": [
234
+ "language_model.model.layers.7.mlp.gate_proj",
235
+ "language_model.model.layers.7.mlp.up_proj",
236
+ "language_model.model.layers.7.mlp.down_proj"
237
+ ],
238
+ "candidates": [
239
+ {
240
+ "dkld": -0.01197697520256047,
241
+ "dbits": 2080899072
242
+ }
243
+ ]
244
+ },
245
+ {
246
+ "idx": 16,
247
+ "layers": [
248
+ "language_model.model.layers.8.self_attn.q_proj",
249
+ "language_model.model.layers.8.self_attn.k_proj",
250
+ "language_model.model.layers.8.self_attn.v_proj",
251
+ "language_model.model.layers.8.self_attn.o_proj"
252
+ ],
253
+ "candidates": [
254
+ {
255
+ "dkld": 0.0037170641124248394,
256
+ "dbits": 396361728
257
+ }
258
+ ]
259
+ },
260
+ {
261
+ "idx": 17,
262
+ "layers": [
263
+ "language_model.model.layers.8.mlp.gate_proj",
264
+ "language_model.model.layers.8.mlp.up_proj",
265
+ "language_model.model.layers.8.mlp.down_proj"
266
+ ],
267
+ "candidates": [
268
+ {
269
+ "dkld": -0.0004419878125190846,
270
+ "dbits": 2080899072
271
+ }
272
+ ]
273
+ },
274
+ {
275
+ "idx": 18,
276
+ "layers": [
277
+ "language_model.model.layers.9.self_attn.q_proj",
278
+ "language_model.model.layers.9.self_attn.k_proj",
279
+ "language_model.model.layers.9.self_attn.v_proj",
280
+ "language_model.model.layers.9.self_attn.o_proj"
281
+ ],
282
+ "candidates": [
283
+ {
284
+ "dkld": -0.004326772689819347,
285
+ "dbits": 396361728
286
+ }
287
+ ]
288
+ },
289
+ {
290
+ "idx": 19,
291
+ "layers": [
292
+ "language_model.model.layers.9.mlp.gate_proj",
293
+ "language_model.model.layers.9.mlp.up_proj",
294
+ "language_model.model.layers.9.mlp.down_proj"
295
+ ],
296
+ "candidates": [
297
+ {
298
+ "dkld": -0.010839792340993915,
299
+ "dbits": 2080899072
300
+ }
301
+ ]
302
+ },
303
+ {
304
+ "idx": 20,
305
+ "layers": [
306
+ "language_model.model.layers.10.self_attn.q_proj",
307
+ "language_model.model.layers.10.self_attn.k_proj",
308
+ "language_model.model.layers.10.self_attn.v_proj",
309
+ "language_model.model.layers.10.self_attn.o_proj"
310
+ ],
311
+ "candidates": [
312
+ {
313
+ "dkld": -0.009040921181440398,
314
+ "dbits": 396361728
315
+ }
316
+ ]
317
+ },
318
+ {
319
+ "idx": 21,
320
+ "layers": [
321
+ "language_model.model.layers.10.mlp.gate_proj",
322
+ "language_model.model.layers.10.mlp.up_proj",
323
+ "language_model.model.layers.10.mlp.down_proj"
324
+ ],
325
+ "candidates": [
326
+ {
327
+ "dkld": -0.031036542356014274,
328
+ "dbits": 2080899072
329
+ }
330
+ ]
331
+ },
332
+ {
333
+ "idx": 22,
334
+ "layers": [
335
+ "language_model.model.layers.11.self_attn.q_proj",
336
+ "language_model.model.layers.11.self_attn.k_proj",
337
+ "language_model.model.layers.11.self_attn.v_proj",
338
+ "language_model.model.layers.11.self_attn.o_proj"
339
+ ],
340
+ "candidates": [
341
+ {
342
+ "dkld": -0.00979023426771164,
343
+ "dbits": 396361728
344
+ }
345
+ ]
346
+ },
347
+ {
348
+ "idx": 23,
349
+ "layers": [
350
+ "language_model.model.layers.11.mlp.gate_proj",
351
+ "language_model.model.layers.11.mlp.up_proj",
352
+ "language_model.model.layers.11.mlp.down_proj"
353
+ ],
354
+ "candidates": [
355
+ {
356
+ "dkld": -0.007736762613058112,
357
+ "dbits": 2080899072
358
+ }
359
+ ]
360
+ },
361
+ {
362
+ "idx": 24,
363
+ "layers": [
364
+ "language_model.model.layers.12.self_attn.q_proj",
365
+ "language_model.model.layers.12.self_attn.k_proj",
366
+ "language_model.model.layers.12.self_attn.v_proj",
367
+ "language_model.model.layers.12.self_attn.o_proj"
368
+ ],
369
+ "candidates": [
370
+ {
371
+ "dkld": -0.003069894015789054,
372
+ "dbits": 396361728
373
+ }
374
+ ]
375
+ },
376
+ {
377
+ "idx": 25,
378
+ "layers": [
379
+ "language_model.model.layers.12.mlp.gate_proj",
380
+ "language_model.model.layers.12.mlp.up_proj",
381
+ "language_model.model.layers.12.mlp.down_proj"
382
+ ],
383
+ "candidates": [
384
+ {
385
+ "dkld": -0.014686931669712078,
386
+ "dbits": 2080899072
387
+ }
388
+ ]
389
+ },
390
+ {
391
+ "idx": 26,
392
+ "layers": [
393
+ "language_model.model.layers.13.self_attn.q_proj",
394
+ "language_model.model.layers.13.self_attn.k_proj",
395
+ "language_model.model.layers.13.self_attn.v_proj",
396
+ "language_model.model.layers.13.self_attn.o_proj"
397
+ ],
398
+ "candidates": [
399
+ {
400
+ "dkld": 0.000129348039627053,
401
+ "dbits": 396361728
402
+ }
403
+ ]
404
+ },
405
+ {
406
+ "idx": 27,
407
+ "layers": [
408
+ "language_model.model.layers.13.mlp.gate_proj",
409
+ "language_model.model.layers.13.mlp.up_proj",
410
+ "language_model.model.layers.13.mlp.down_proj"
411
+ ],
412
+ "candidates": [
413
+ {
414
+ "dkld": -0.003941317647695564,
415
+ "dbits": 2080899072
416
+ }
417
+ ]
418
+ },
419
+ {
420
+ "idx": 28,
421
+ "layers": [
422
+ "language_model.model.layers.14.self_attn.q_proj",
423
+ "language_model.model.layers.14.self_attn.k_proj",
424
+ "language_model.model.layers.14.self_attn.v_proj",
425
+ "language_model.model.layers.14.self_attn.o_proj"
426
+ ],
427
+ "candidates": [
428
+ {
429
+ "dkld": -0.0035466991364956346,
430
+ "dbits": 396361728
431
+ }
432
+ ]
433
+ },
434
+ {
435
+ "idx": 29,
436
+ "layers": [
437
+ "language_model.model.layers.14.mlp.gate_proj",
438
+ "language_model.model.layers.14.mlp.up_proj",
439
+ "language_model.model.layers.14.mlp.down_proj"
440
+ ],
441
+ "candidates": [
442
+ {
443
+ "dkld": -0.022119185328483615,
444
+ "dbits": 2080899072
445
+ }
446
+ ]
447
+ },
448
+ {
449
+ "idx": 30,
450
+ "layers": [
451
+ "language_model.model.layers.15.self_attn.q_proj",
452
+ "language_model.model.layers.15.self_attn.k_proj",
453
+ "language_model.model.layers.15.self_attn.v_proj",
454
+ "language_model.model.layers.15.self_attn.o_proj"
455
+ ],
456
+ "candidates": [
457
+ {
458
+ "dkld": 0.003810518234968141,
459
+ "dbits": 396361728
460
+ }
461
+ ]
462
+ },
463
+ {
464
+ "idx": 31,
465
+ "layers": [
466
+ "language_model.model.layers.15.mlp.gate_proj",
467
+ "language_model.model.layers.15.mlp.up_proj",
468
+ "language_model.model.layers.15.mlp.down_proj"
469
+ ],
470
+ "candidates": [
471
+ {
472
+ "dkld": -0.009127749502658877,
473
+ "dbits": 2080899072
474
+ }
475
+ ]
476
+ },
477
+ {
478
+ "idx": 32,
479
+ "layers": [
480
+ "language_model.model.layers.16.self_attn.q_proj",
481
+ "language_model.model.layers.16.self_attn.k_proj",
482
+ "language_model.model.layers.16.self_attn.v_proj",
483
+ "language_model.model.layers.16.self_attn.o_proj"
484
+ ],
485
+ "candidates": [
486
+ {
487
+ "dkld": 0.006637018173933029,
488
+ "dbits": 396361728
489
+ }
490
+ ]
491
+ },
492
+ {
493
+ "idx": 33,
494
+ "layers": [
495
+ "language_model.model.layers.16.mlp.gate_proj",
496
+ "language_model.model.layers.16.mlp.up_proj",
497
+ "language_model.model.layers.16.mlp.down_proj"
498
+ ],
499
+ "candidates": [
500
+ {
501
+ "dkld": -0.021292066574096713,
502
+ "dbits": 2080899072
503
+ }
504
+ ]
505
+ },
506
+ {
507
+ "idx": 34,
508
+ "layers": [
509
+ "language_model.model.layers.17.self_attn.q_proj",
510
+ "language_model.model.layers.17.self_attn.k_proj",
511
+ "language_model.model.layers.17.self_attn.v_proj",
512
+ "language_model.model.layers.17.self_attn.o_proj"
513
+ ],
514
+ "candidates": [
515
+ {
516
+ "dkld": -0.0025063112378120866,
517
+ "dbits": 396361728
518
+ }
519
+ ]
520
+ },
521
+ {
522
+ "idx": 35,
523
+ "layers": [
524
+ "language_model.model.layers.17.mlp.gate_proj",
525
+ "language_model.model.layers.17.mlp.up_proj",
526
+ "language_model.model.layers.17.mlp.down_proj"
527
+ ],
528
+ "candidates": [
529
+ {
530
+ "dkld": 0.0022593840956687816,
531
+ "dbits": 2080899072
532
+ }
533
+ ]
534
+ },
535
+ {
536
+ "idx": 36,
537
+ "layers": [
538
+ "language_model.model.layers.18.self_attn.q_proj",
539
+ "language_model.model.layers.18.self_attn.k_proj",
540
+ "language_model.model.layers.18.self_attn.v_proj",
541
+ "language_model.model.layers.18.self_attn.o_proj"
542
+ ],
543
+ "candidates": [
544
+ {
545
+ "dkld": 0.002482272684574127,
546
+ "dbits": 396361728
547
+ }
548
+ ]
549
+ },
550
+ {
551
+ "idx": 37,
552
+ "layers": [
553
+ "language_model.model.layers.18.mlp.gate_proj",
554
+ "language_model.model.layers.18.mlp.up_proj",
555
+ "language_model.model.layers.18.mlp.down_proj"
556
+ ],
557
+ "candidates": [
558
+ {
559
+ "dkld": -0.000394851714372646,
560
+ "dbits": 2080899072
561
+ }
562
+ ]
563
+ },
564
+ {
565
+ "idx": 38,
566
+ "layers": [
567
+ "language_model.model.layers.19.self_attn.q_proj",
568
+ "language_model.model.layers.19.self_attn.k_proj",
569
+ "language_model.model.layers.19.self_attn.v_proj",
570
+ "language_model.model.layers.19.self_attn.o_proj"
571
+ ],
572
+ "candidates": [
573
+ {
574
+ "dkld": -0.0010013014078140703,
575
+ "dbits": 396361728
576
+ }
577
+ ]
578
+ },
579
+ {
580
+ "idx": 39,
581
+ "layers": [
582
+ "language_model.model.layers.19.mlp.gate_proj",
583
+ "language_model.model.layers.19.mlp.up_proj",
584
+ "language_model.model.layers.19.mlp.down_proj"
585
+ ],
586
+ "candidates": [
587
+ {
588
+ "dkld": -0.020440497994422935,
589
+ "dbits": 2080899072
590
+ }
591
+ ]
592
+ },
593
+ {
594
+ "idx": 40,
595
+ "layers": [
596
+ "language_model.model.layers.20.self_attn.q_proj",
597
+ "language_model.model.layers.20.self_attn.k_proj",
598
+ "language_model.model.layers.20.self_attn.v_proj",
599
+ "language_model.model.layers.20.self_attn.o_proj"
600
+ ],
601
+ "candidates": [
602
+ {
603
+ "dkld": 0.004292172193527177,
604
+ "dbits": 396361728
605
+ }
606
+ ]
607
+ },
608
+ {
609
+ "idx": 41,
610
+ "layers": [
611
+ "language_model.model.layers.20.mlp.gate_proj",
612
+ "language_model.model.layers.20.mlp.up_proj",
613
+ "language_model.model.layers.20.mlp.down_proj"
614
+ ],
615
+ "candidates": [
616
+ {
617
+ "dkld": -0.0009037740528583527,
618
+ "dbits": 2080899072
619
+ }
620
+ ]
621
+ },
622
+ {
623
+ "idx": 42,
624
+ "layers": [
625
+ "language_model.model.layers.21.self_attn.q_proj",
626
+ "language_model.model.layers.21.self_attn.k_proj",
627
+ "language_model.model.layers.21.self_attn.v_proj",
628
+ "language_model.model.layers.21.self_attn.o_proj"
629
+ ],
630
+ "candidates": [
631
+ {
632
+ "dkld": -0.006902486085891724,
633
+ "dbits": 396361728
634
+ }
635
+ ]
636
+ },
637
+ {
638
+ "idx": 43,
639
+ "layers": [
640
+ "language_model.model.layers.21.mlp.gate_proj",
641
+ "language_model.model.layers.21.mlp.up_proj",
642
+ "language_model.model.layers.21.mlp.down_proj"
643
+ ],
644
+ "candidates": [
645
+ {
646
+ "dkld": -0.00929136574268341,
647
+ "dbits": 2080899072
648
+ }
649
+ ]
650
+ },
651
+ {
652
+ "idx": 44,
653
+ "layers": [
654
+ "language_model.model.layers.22.self_attn.q_proj",
655
+ "language_model.model.layers.22.self_attn.k_proj",
656
+ "language_model.model.layers.22.self_attn.v_proj",
657
+ "language_model.model.layers.22.self_attn.o_proj"
658
+ ],
659
+ "candidates": [
660
+ {
661
+ "dkld": 0.0036064453423023113,
662
+ "dbits": 396361728
663
+ }
664
+ ]
665
+ },
666
+ {
667
+ "idx": 45,
668
+ "layers": [
669
+ "language_model.model.layers.22.mlp.gate_proj",
670
+ "language_model.model.layers.22.mlp.up_proj",
671
+ "language_model.model.layers.22.mlp.down_proj"
672
+ ],
673
+ "candidates": [
674
+ {
675
+ "dkld": -0.0006043247878551594,
676
+ "dbits": 2080899072
677
+ }
678
+ ]
679
+ },
680
+ {
681
+ "idx": 46,
682
+ "layers": [
683
+ "language_model.model.layers.23.self_attn.q_proj",
684
+ "language_model.model.layers.23.self_attn.k_proj",
685
+ "language_model.model.layers.23.self_attn.v_proj",
686
+ "language_model.model.layers.23.self_attn.o_proj"
687
+ ],
688
+ "candidates": [
689
+ {
690
+ "dkld": 0.028081689029932022,
691
+ "dbits": 396361728
692
+ }
693
+ ]
694
+ },
695
+ {
696
+ "idx": 47,
697
+ "layers": [
698
+ "language_model.model.layers.23.mlp.gate_proj",
699
+ "language_model.model.layers.23.mlp.up_proj",
700
+ "language_model.model.layers.23.mlp.down_proj"
701
+ ],
702
+ "candidates": [
703
+ {
704
+ "dkld": -0.00137981697916989,
705
+ "dbits": 2080899072
706
+ }
707
+ ]
708
+ },
709
+ {
710
+ "idx": 48,
711
+ "layers": [
712
+ "language_model.model.layers.24.self_attn.q_proj",
713
+ "language_model.model.layers.24.self_attn.k_proj",
714
+ "language_model.model.layers.24.self_attn.v_proj",
715
+ "language_model.model.layers.24.self_attn.o_proj"
716
+ ],
717
+ "candidates": [
718
+ {
719
+ "dkld": -0.0014152429997921212,
720
+ "dbits": 396361728
721
+ }
722
+ ]
723
+ },
724
+ {
725
+ "idx": 49,
726
+ "layers": [
727
+ "language_model.model.layers.24.mlp.gate_proj",
728
+ "language_model.model.layers.24.mlp.up_proj",
729
+ "language_model.model.layers.24.mlp.down_proj"
730
+ ],
731
+ "candidates": [
732
+ {
733
+ "dkld": -0.004900486022233974,
734
+ "dbits": 2080899072
735
+ }
736
+ ]
737
+ },
738
+ {
739
+ "idx": 50,
740
+ "layers": [
741
+ "language_model.model.layers.25.self_attn.q_proj",
742
+ "language_model.model.layers.25.self_attn.k_proj",
743
+ "language_model.model.layers.25.self_attn.v_proj",
744
+ "language_model.model.layers.25.self_attn.o_proj"
745
+ ],
746
+ "candidates": [
747
+ {
748
+ "dkld": -0.008432538062334072,
749
+ "dbits": 396361728
750
+ }
751
+ ]
752
+ },
753
+ {
754
+ "idx": 51,
755
+ "layers": [
756
+ "language_model.model.layers.25.mlp.gate_proj",
757
+ "language_model.model.layers.25.mlp.up_proj",
758
+ "language_model.model.layers.25.mlp.down_proj"
759
+ ],
760
+ "candidates": [
761
+ {
762
+ "dkld": 0.005563641339540459,
763
+ "dbits": 2080899072
764
+ }
765
+ ]
766
+ },
767
+ {
768
+ "idx": 52,
769
+ "layers": [
770
+ "language_model.model.layers.26.self_attn.q_proj",
771
+ "language_model.model.layers.26.self_attn.k_proj",
772
+ "language_model.model.layers.26.self_attn.v_proj",
773
+ "language_model.model.layers.26.self_attn.o_proj"
774
+ ],
775
+ "candidates": [
776
+ {
777
+ "dkld": -0.010603216290473949,
778
+ "dbits": 396361728
779
+ }
780
+ ]
781
+ },
782
+ {
783
+ "idx": 53,
784
+ "layers": [
785
+ "language_model.model.layers.26.mlp.gate_proj",
786
+ "language_model.model.layers.26.mlp.up_proj",
787
+ "language_model.model.layers.26.mlp.down_proj"
788
+ ],
789
+ "candidates": [
790
+ {
791
+ "dkld": -0.003027988225221656,
792
+ "dbits": 2080899072
793
+ }
794
+ ]
795
+ },
796
+ {
797
+ "idx": 54,
798
+ "layers": [
799
+ "language_model.model.layers.27.self_attn.q_proj",
800
+ "language_model.model.layers.27.self_attn.k_proj",
801
+ "language_model.model.layers.27.self_attn.v_proj",
802
+ "language_model.model.layers.27.self_attn.o_proj"
803
+ ],
804
+ "candidates": [
805
+ {
806
+ "dkld": -0.00233266279101374,
807
+ "dbits": 396361728
808
+ }
809
+ ]
810
+ },
811
+ {
812
+ "idx": 55,
813
+ "layers": [
814
+ "language_model.model.layers.27.mlp.gate_proj",
815
+ "language_model.model.layers.27.mlp.up_proj",
816
+ "language_model.model.layers.27.mlp.down_proj"
817
+ ],
818
+ "candidates": [
819
+ {
820
+ "dkld": -0.0059184454381466245,
821
+ "dbits": 2080899072
822
+ }
823
+ ]
824
+ },
825
+ {
826
+ "idx": 56,
827
+ "layers": [
828
+ "language_model.model.layers.28.self_attn.q_proj",
829
+ "language_model.model.layers.28.self_attn.k_proj",
830
+ "language_model.model.layers.28.self_attn.v_proj",
831
+ "language_model.model.layers.28.self_attn.o_proj"
832
+ ],
833
+ "candidates": [
834
+ {
835
+ "dkld": -0.0004229128360748735,
836
+ "dbits": 396361728
837
+ }
838
+ ]
839
+ },
840
+ {
841
+ "idx": 57,
842
+ "layers": [
843
+ "language_model.model.layers.28.mlp.gate_proj",
844
+ "language_model.model.layers.28.mlp.up_proj",
845
+ "language_model.model.layers.28.mlp.down_proj"
846
+ ],
847
+ "candidates": [
848
+ {
849
+ "dkld": 0.007507449388504006,
850
+ "dbits": 2080899072
851
+ }
852
+ ]
853
+ },
854
+ {
855
+ "idx": 58,
856
+ "layers": [
857
+ "language_model.model.layers.29.self_attn.q_proj",
858
+ "language_model.model.layers.29.self_attn.k_proj",
859
+ "language_model.model.layers.29.self_attn.v_proj",
860
+ "language_model.model.layers.29.self_attn.o_proj"
861
+ ],
862
+ "candidates": [
863
+ {
864
+ "dkld": 0.0012179315090179443,
865
+ "dbits": 396361728
866
+ }
867
+ ]
868
+ },
869
+ {
870
+ "idx": 59,
871
+ "layers": [
872
+ "language_model.model.layers.29.mlp.gate_proj",
873
+ "language_model.model.layers.29.mlp.up_proj",
874
+ "language_model.model.layers.29.mlp.down_proj"
875
+ ],
876
+ "candidates": [
877
+ {
878
+ "dkld": 0.007656254619359926,
879
+ "dbits": 2080899072
880
+ }
881
+ ]
882
+ },
883
+ {
884
+ "idx": 60,
885
+ "layers": [
886
+ "language_model.model.layers.30.self_attn.q_proj",
887
+ "language_model.model.layers.30.self_attn.k_proj",
888
+ "language_model.model.layers.30.self_attn.v_proj",
889
+ "language_model.model.layers.30.self_attn.o_proj"
890
+ ],
891
+ "candidates": [
892
+ {
893
+ "dkld": -0.0006962120532989502,
894
+ "dbits": 396361728
895
+ }
896
+ ]
897
+ },
898
+ {
899
+ "idx": 61,
900
+ "layers": [
901
+ "language_model.model.layers.30.mlp.gate_proj",
902
+ "language_model.model.layers.30.mlp.up_proj",
903
+ "language_model.model.layers.30.mlp.down_proj"
904
+ ],
905
+ "candidates": [
906
+ {
907
+ "dkld": -0.0112728171050549,
908
+ "dbits": 2080899072
909
+ }
910
+ ]
911
+ },
912
+ {
913
+ "idx": 62,
914
+ "layers": [
915
+ "language_model.model.layers.31.self_attn.q_proj",
916
+ "language_model.model.layers.31.self_attn.k_proj",
917
+ "language_model.model.layers.31.self_attn.v_proj",
918
+ "language_model.model.layers.31.self_attn.o_proj"
919
+ ],
920
+ "candidates": [
921
+ {
922
+ "dkld": -0.0017989106476307137,
923
+ "dbits": 396361728
924
+ }
925
+ ]
926
+ },
927
+ {
928
+ "idx": 63,
929
+ "layers": [
930
+ "language_model.model.layers.31.mlp.gate_proj",
931
+ "language_model.model.layers.31.mlp.up_proj",
932
+ "language_model.model.layers.31.mlp.down_proj"
933
+ ],
934
+ "candidates": [
935
+ {
936
+ "dkld": -0.0063770614564419015,
937
+ "dbits": 2080899072
938
+ }
939
+ ]
940
+ },
941
+ {
942
+ "idx": 64,
943
+ "layers": [
944
+ "language_model.model.layers.32.self_attn.q_proj",
945
+ "language_model.model.layers.32.self_attn.k_proj",
946
+ "language_model.model.layers.32.self_attn.v_proj",
947
+ "language_model.model.layers.32.self_attn.o_proj"
948
+ ],
949
+ "candidates": [
950
+ {
951
+ "dkld": -0.00029214099049568176,
952
+ "dbits": 396361728
953
+ }
954
+ ]
955
+ },
956
+ {
957
+ "idx": 65,
958
+ "layers": [
959
+ "language_model.model.layers.32.mlp.gate_proj",
960
+ "language_model.model.layers.32.mlp.up_proj",
961
+ "language_model.model.layers.32.mlp.down_proj"
962
+ ],
963
+ "candidates": [
964
+ {
965
+ "dkld": -0.006251434981823012,
966
+ "dbits": 2080899072
967
+ }
968
+ ]
969
+ },
970
+ {
971
+ "idx": 66,
972
+ "layers": [
973
+ "language_model.model.layers.33.self_attn.q_proj",
974
+ "language_model.model.layers.33.self_attn.k_proj",
975
+ "language_model.model.layers.33.self_attn.v_proj",
976
+ "language_model.model.layers.33.self_attn.o_proj"
977
+ ],
978
+ "candidates": [
979
+ {
980
+ "dkld": 0.0005352959036826754,
981
+ "dbits": 396361728
982
+ }
983
+ ]
984
+ },
985
+ {
986
+ "idx": 67,
987
+ "layers": [
988
+ "language_model.model.layers.33.mlp.gate_proj",
989
+ "language_model.model.layers.33.mlp.up_proj",
990
+ "language_model.model.layers.33.mlp.down_proj"
991
+ ],
992
+ "candidates": [
993
+ {
994
+ "dkld": -0.005572766810655605,
995
+ "dbits": 2080899072
996
+ }
997
+ ]
998
+ },
999
+ {
1000
+ "idx": 68,
1001
+ "layers": [
1002
+ "language_model.model.layers.34.self_attn.q_proj",
1003
+ "language_model.model.layers.34.self_attn.k_proj",
1004
+ "language_model.model.layers.34.self_attn.v_proj",
1005
+ "language_model.model.layers.34.self_attn.o_proj"
1006
+ ],
1007
+ "candidates": [
1008
+ {
1009
+ "dkld": 0.0008197985589504242,
1010
+ "dbits": 396361728
1011
+ }
1012
+ ]
1013
+ },
1014
+ {
1015
+ "idx": 69,
1016
+ "layers": [
1017
+ "language_model.model.layers.34.mlp.gate_proj",
1018
+ "language_model.model.layers.34.mlp.up_proj",
1019
+ "language_model.model.layers.34.mlp.down_proj"
1020
+ ],
1021
+ "candidates": [
1022
+ {
1023
+ "dkld": -0.004082408547401439,
1024
+ "dbits": 2080899072
1025
+ }
1026
+ ]
1027
+ },
1028
+ {
1029
+ "idx": 70,
1030
+ "layers": [
1031
+ "language_model.model.layers.35.self_attn.q_proj",
1032
+ "language_model.model.layers.35.self_attn.k_proj",
1033
+ "language_model.model.layers.35.self_attn.v_proj",
1034
+ "language_model.model.layers.35.self_attn.o_proj"
1035
+ ],
1036
+ "candidates": [
1037
+ {
1038
+ "dkld": 0.0032249547541141066,
1039
+ "dbits": 396361728
1040
+ }
1041
+ ]
1042
+ },
1043
+ {
1044
+ "idx": 71,
1045
+ "layers": [
1046
+ "language_model.model.layers.35.mlp.gate_proj",
1047
+ "language_model.model.layers.35.mlp.up_proj",
1048
+ "language_model.model.layers.35.mlp.down_proj"
1049
+ ],
1050
+ "candidates": [
1051
+ {
1052
+ "dkld": -0.0070062801241874695,
1053
+ "dbits": 2080899072
1054
+ }
1055
+ ]
1056
+ },
1057
+ {
1058
+ "idx": 72,
1059
+ "layers": [
1060
+ "language_model.model.layers.36.self_attn.q_proj",
1061
+ "language_model.model.layers.36.self_attn.k_proj",
1062
+ "language_model.model.layers.36.self_attn.v_proj",
1063
+ "language_model.model.layers.36.self_attn.o_proj"
1064
+ ],
1065
+ "candidates": [
1066
+ {
1067
+ "dkld": 0.0005763806402683258,
1068
+ "dbits": 396361728
1069
+ }
1070
+ ]
1071
+ },
1072
+ {
1073
+ "idx": 73,
1074
+ "layers": [
1075
+ "language_model.model.layers.36.mlp.gate_proj",
1076
+ "language_model.model.layers.36.mlp.up_proj",
1077
+ "language_model.model.layers.36.mlp.down_proj"
1078
+ ],
1079
+ "candidates": [
1080
+ {
1081
+ "dkld": -0.0030428156256676053,
1082
+ "dbits": 2080899072
1083
+ }
1084
+ ]
1085
+ },
1086
+ {
1087
+ "idx": 74,
1088
+ "layers": [
1089
+ "language_model.model.layers.37.self_attn.q_proj",
1090
+ "language_model.model.layers.37.self_attn.k_proj",
1091
+ "language_model.model.layers.37.self_attn.v_proj",
1092
+ "language_model.model.layers.37.self_attn.o_proj"
1093
+ ],
1094
+ "candidates": [
1095
+ {
1096
+ "dkld": -0.0029344171285629606,
1097
+ "dbits": 396361728
1098
+ }
1099
+ ]
1100
+ },
1101
+ {
1102
+ "idx": 75,
1103
+ "layers": [
1104
+ "language_model.model.layers.37.mlp.gate_proj",
1105
+ "language_model.model.layers.37.mlp.up_proj",
1106
+ "language_model.model.layers.37.mlp.down_proj"
1107
+ ],
1108
+ "candidates": [
1109
+ {
1110
+ "dkld": -0.005395969003438972,
1111
+ "dbits": 2080899072
1112
+ }
1113
+ ]
1114
+ },
1115
+ {
1116
+ "idx": 76,
1117
+ "layers": [
1118
+ "language_model.model.layers.38.self_attn.q_proj",
1119
+ "language_model.model.layers.38.self_attn.k_proj",
1120
+ "language_model.model.layers.38.self_attn.v_proj",
1121
+ "language_model.model.layers.38.self_attn.o_proj"
1122
+ ],
1123
+ "candidates": [
1124
+ {
1125
+ "dkld": 0.0017978318035602459,
1126
+ "dbits": 396361728
1127
+ }
1128
+ ]
1129
+ },
1130
+ {
1131
+ "idx": 77,
1132
+ "layers": [
1133
+ "language_model.model.layers.38.mlp.gate_proj",
1134
+ "language_model.model.layers.38.mlp.up_proj",
1135
+ "language_model.model.layers.38.mlp.down_proj"
1136
+ ],
1137
+ "candidates": [
1138
+ {
1139
+ "dkld": -0.0036740973591804504,
1140
+ "dbits": 2080899072
1141
+ }
1142
+ ]
1143
+ },
1144
+ {
1145
+ "idx": 78,
1146
+ "layers": [
1147
+ "language_model.model.layers.39.self_attn.q_proj",
1148
+ "language_model.model.layers.39.self_attn.k_proj",
1149
+ "language_model.model.layers.39.self_attn.v_proj",
1150
+ "language_model.model.layers.39.self_attn.o_proj"
1151
+ ],
1152
+ "candidates": [
1153
+ {
1154
+ "dkld": 0.0005030691623687522,
1155
+ "dbits": 396361728
1156
+ }
1157
+ ]
1158
+ },
1159
+ {
1160
+ "idx": 79,
1161
+ "layers": [
1162
+ "language_model.model.layers.39.mlp.gate_proj",
1163
+ "language_model.model.layers.39.mlp.up_proj",
1164
+ "language_model.model.layers.39.mlp.down_proj"
1165
+ ],
1166
+ "candidates": [
1167
+ {
1168
+ "dkld": -0.003146367520093951,
1169
+ "dbits": 2080899072
1170
+ }
1171
+ ]
1172
+ },
1173
+ {
1174
+ "idx": 80,
1175
+ "layers": [
1176
+ "language_model.model.layers.40.self_attn.q_proj",
1177
+ "language_model.model.layers.40.self_attn.k_proj",
1178
+ "language_model.model.layers.40.self_attn.v_proj",
1179
+ "language_model.model.layers.40.self_attn.o_proj"
1180
+ ],
1181
+ "candidates": [
1182
+ {
1183
+ "dkld": -0.008280515670776367,
1184
+ "dbits": 396361728
1185
+ }
1186
+ ]
1187
+ },
1188
+ {
1189
+ "idx": 81,
1190
+ "layers": [
1191
+ "language_model.model.layers.40.mlp.gate_proj",
1192
+ "language_model.model.layers.40.mlp.up_proj",
1193
+ "language_model.model.layers.40.mlp.down_proj"
1194
+ ],
1195
+ "candidates": [
1196
+ {
1197
+ "dkld": -0.004743161797523543,
1198
+ "dbits": 2080899072
1199
+ }
1200
+ ]
1201
+ },
1202
+ {
1203
+ "idx": 82,
1204
+ "layers": [
1205
+ "language_model.model.layers.41.self_attn.q_proj",
1206
+ "language_model.model.layers.41.self_attn.k_proj",
1207
+ "language_model.model.layers.41.self_attn.v_proj",
1208
+ "language_model.model.layers.41.self_attn.o_proj"
1209
+ ],
1210
+ "candidates": [
1211
+ {
1212
+ "dkld": -0.015289445966482185,
1213
+ "dbits": 396361728
1214
+ }
1215
+ ]
1216
+ },
1217
+ {
1218
+ "idx": 83,
1219
+ "layers": [
1220
+ "language_model.model.layers.41.mlp.gate_proj",
1221
+ "language_model.model.layers.41.mlp.up_proj",
1222
+ "language_model.model.layers.41.mlp.down_proj"
1223
+ ],
1224
+ "candidates": [
1225
+ {
1226
+ "dkld": -0.00837468057870866,
1227
+ "dbits": 2080899072
1228
+ }
1229
+ ]
1230
+ },
1231
+ {
1232
+ "idx": 84,
1233
+ "layers": [
1234
+ "language_model.model.layers.42.self_attn.q_proj",
1235
+ "language_model.model.layers.42.self_attn.k_proj",
1236
+ "language_model.model.layers.42.self_attn.v_proj",
1237
+ "language_model.model.layers.42.self_attn.o_proj"
1238
+ ],
1239
+ "candidates": [
1240
+ {
1241
+ "dkld": -0.0007713317871094194,
1242
+ "dbits": 396361728
1243
+ }
1244
+ ]
1245
+ },
1246
+ {
1247
+ "idx": 85,
1248
+ "layers": [
1249
+ "language_model.model.layers.42.mlp.gate_proj",
1250
+ "language_model.model.layers.42.mlp.up_proj",
1251
+ "language_model.model.layers.42.mlp.down_proj"
1252
+ ],
1253
+ "candidates": [
1254
+ {
1255
+ "dkld": -0.0034337289631367063,
1256
+ "dbits": 2080899072
1257
+ }
1258
+ ]
1259
+ },
1260
+ {
1261
+ "idx": 86,
1262
+ "layers": [
1263
+ "language_model.model.layers.43.self_attn.q_proj",
1264
+ "language_model.model.layers.43.self_attn.k_proj",
1265
+ "language_model.model.layers.43.self_attn.v_proj",
1266
+ "language_model.model.layers.43.self_attn.o_proj"
1267
+ ],
1268
+ "candidates": [
1269
+ {
1270
+ "dkld": -0.00845300778746605,
1271
+ "dbits": 396361728
1272
+ }
1273
+ ]
1274
+ },
1275
+ {
1276
+ "idx": 87,
1277
+ "layers": [
1278
+ "language_model.model.layers.43.mlp.gate_proj",
1279
+ "language_model.model.layers.43.mlp.up_proj",
1280
+ "language_model.model.layers.43.mlp.down_proj"
1281
+ ],
1282
+ "candidates": [
1283
+ {
1284
+ "dkld": -0.003721737116575252,
1285
+ "dbits": 2080899072
1286
+ }
1287
+ ]
1288
+ },
1289
+ {
1290
+ "idx": 88,
1291
+ "layers": [
1292
+ "language_model.model.layers.44.self_attn.q_proj",
1293
+ "language_model.model.layers.44.self_attn.k_proj",
1294
+ "language_model.model.layers.44.self_attn.v_proj",
1295
+ "language_model.model.layers.44.self_attn.o_proj"
1296
+ ],
1297
+ "candidates": [
1298
+ {
1299
+ "dkld": -0.0051841072738170735,
1300
+ "dbits": 396361728
1301
+ }
1302
+ ]
1303
+ },
1304
+ {
1305
+ "idx": 89,
1306
+ "layers": [
1307
+ "language_model.model.layers.44.mlp.gate_proj",
1308
+ "language_model.model.layers.44.mlp.up_proj",
1309
+ "language_model.model.layers.44.mlp.down_proj"
1310
+ ],
1311
+ "candidates": [
1312
+ {
1313
+ "dkld": -0.0031947255134582853,
1314
+ "dbits": 2080899072
1315
+ }
1316
+ ]
1317
+ },
1318
+ {
1319
+ "idx": 90,
1320
+ "layers": [
1321
+ "language_model.model.layers.45.self_attn.q_proj",
1322
+ "language_model.model.layers.45.self_attn.k_proj",
1323
+ "language_model.model.layers.45.self_attn.v_proj",
1324
+ "language_model.model.layers.45.self_attn.o_proj"
1325
+ ],
1326
+ "candidates": [
1327
+ {
1328
+ "dkld": -0.001118142157793045,
1329
+ "dbits": 396361728
1330
+ }
1331
+ ]
1332
+ },
1333
+ {
1334
+ "idx": 91,
1335
+ "layers": [
1336
+ "language_model.model.layers.45.mlp.gate_proj",
1337
+ "language_model.model.layers.45.mlp.up_proj",
1338
+ "language_model.model.layers.45.mlp.down_proj"
1339
+ ],
1340
+ "candidates": [
1341
+ {
1342
+ "dkld": -0.0033696956932545152,
1343
+ "dbits": 2080899072
1344
+ }
1345
+ ]
1346
+ },
1347
+ {
1348
+ "idx": 92,
1349
+ "layers": [
1350
+ "language_model.model.layers.46.self_attn.q_proj",
1351
+ "language_model.model.layers.46.self_attn.k_proj",
1352
+ "language_model.model.layers.46.self_attn.v_proj",
1353
+ "language_model.model.layers.46.self_attn.o_proj"
1354
+ ],
1355
+ "candidates": [
1356
+ {
1357
+ "dkld": -0.0005463138222694397,
1358
+ "dbits": 396361728
1359
+ }
1360
+ ]
1361
+ },
1362
+ {
1363
+ "idx": 93,
1364
+ "layers": [
1365
+ "language_model.model.layers.46.mlp.gate_proj",
1366
+ "language_model.model.layers.46.mlp.up_proj",
1367
+ "language_model.model.layers.46.mlp.down_proj"
1368
+ ],
1369
+ "candidates": [
1370
+ {
1371
+ "dkld": -0.006662502884864807,
1372
+ "dbits": 2080899072
1373
+ }
1374
+ ]
1375
+ },
1376
+ {
1377
+ "idx": 94,
1378
+ "layers": [
1379
+ "language_model.model.layers.47.self_attn.q_proj",
1380
+ "language_model.model.layers.47.self_attn.k_proj",
1381
+ "language_model.model.layers.47.self_attn.v_proj",
1382
+ "language_model.model.layers.47.self_attn.o_proj"
1383
+ ],
1384
+ "candidates": [
1385
+ {
1386
+ "dkld": -0.0023998439311981312,
1387
+ "dbits": 396361728
1388
+ }
1389
+ ]
1390
+ },
1391
+ {
1392
+ "idx": 95,
1393
+ "layers": [
1394
+ "language_model.model.layers.47.mlp.gate_proj",
1395
+ "language_model.model.layers.47.mlp.up_proj",
1396
+ "language_model.model.layers.47.mlp.down_proj"
1397
+ ],
1398
+ "candidates": [
1399
+ {
1400
+ "dkld": -0.003466875106096279,
1401
+ "dbits": 2080899072
1402
+ }
1403
+ ]
1404
+ },
1405
+ {
1406
+ "idx": 96,
1407
+ "layers": [
1408
+ "language_model.model.layers.48.self_attn.q_proj",
1409
+ "language_model.model.layers.48.self_attn.k_proj",
1410
+ "language_model.model.layers.48.self_attn.v_proj",
1411
+ "language_model.model.layers.48.self_attn.o_proj"
1412
+ ],
1413
+ "candidates": [
1414
+ {
1415
+ "dkld": -0.004037512093782447,
1416
+ "dbits": 396361728
1417
+ }
1418
+ ]
1419
+ },
1420
+ {
1421
+ "idx": 97,
1422
+ "layers": [
1423
+ "language_model.model.layers.48.mlp.gate_proj",
1424
+ "language_model.model.layers.48.mlp.up_proj",
1425
+ "language_model.model.layers.48.mlp.down_proj"
1426
+ ],
1427
+ "candidates": [
1428
+ {
1429
+ "dkld": -0.003907855600118637,
1430
+ "dbits": 2080899072
1431
+ }
1432
+ ]
1433
+ },
1434
+ {
1435
+ "idx": 98,
1436
+ "layers": [
1437
+ "language_model.model.layers.49.self_attn.q_proj",
1438
+ "language_model.model.layers.49.self_attn.k_proj",
1439
+ "language_model.model.layers.49.self_attn.v_proj",
1440
+ "language_model.model.layers.49.self_attn.o_proj"
1441
+ ],
1442
+ "candidates": [
1443
+ {
1444
+ "dkld": -0.0028403460979462003,
1445
+ "dbits": 396361728
1446
+ }
1447
+ ]
1448
+ },
1449
+ {
1450
+ "idx": 99,
1451
+ "layers": [
1452
+ "language_model.model.layers.49.mlp.gate_proj",
1453
+ "language_model.model.layers.49.mlp.up_proj",
1454
+ "language_model.model.layers.49.mlp.down_proj"
1455
+ ],
1456
+ "candidates": [
1457
+ {
1458
+ "dkld": -0.002681272476911578,
1459
+ "dbits": 2080899072
1460
+ }
1461
+ ]
1462
+ },
1463
+ {
1464
+ "idx": 100,
1465
+ "layers": [
1466
+ "language_model.model.layers.50.self_attn.q_proj",
1467
+ "language_model.model.layers.50.self_attn.k_proj",
1468
+ "language_model.model.layers.50.self_attn.v_proj",
1469
+ "language_model.model.layers.50.self_attn.o_proj"
1470
+ ],
1471
+ "candidates": [
1472
+ {
1473
+ "dkld": -0.0014717683196068254,
1474
+ "dbits": 396361728
1475
+ }
1476
+ ]
1477
+ },
1478
+ {
1479
+ "idx": 101,
1480
+ "layers": [
1481
+ "language_model.model.layers.50.mlp.gate_proj",
1482
+ "language_model.model.layers.50.mlp.up_proj",
1483
+ "language_model.model.layers.50.mlp.down_proj"
1484
+ ],
1485
+ "candidates": [
1486
+ {
1487
+ "dkld": -0.0033254854381084553,
1488
+ "dbits": 2080899072
1489
+ }
1490
+ ]
1491
+ },
1492
+ {
1493
+ "idx": 102,
1494
+ "layers": [
1495
+ "language_model.model.layers.51.self_attn.q_proj",
1496
+ "language_model.model.layers.51.self_attn.k_proj",
1497
+ "language_model.model.layers.51.self_attn.v_proj",
1498
+ "language_model.model.layers.51.self_attn.o_proj"
1499
+ ],
1500
+ "candidates": [
1501
+ {
1502
+ "dkld": -0.0007418908178806749,
1503
+ "dbits": 396361728
1504
+ }
1505
+ ]
1506
+ },
1507
+ {
1508
+ "idx": 103,
1509
+ "layers": [
1510
+ "language_model.model.layers.51.mlp.gate_proj",
1511
+ "language_model.model.layers.51.mlp.up_proj",
1512
+ "language_model.model.layers.51.mlp.down_proj"
1513
+ ],
1514
+ "candidates": [
1515
+ {
1516
+ "dkld": -0.0038238234817982164,
1517
+ "dbits": 2080899072
1518
+ }
1519
+ ]
1520
+ },
1521
+ {
1522
+ "idx": 104,
1523
+ "layers": [
1524
+ "language_model.model.layers.52.self_attn.q_proj",
1525
+ "language_model.model.layers.52.self_attn.k_proj",
1526
+ "language_model.model.layers.52.self_attn.v_proj",
1527
+ "language_model.model.layers.52.self_attn.o_proj"
1528
+ ],
1529
+ "candidates": [
1530
+ {
1531
+ "dkld": -0.0012986376881599648,
1532
+ "dbits": 396361728
1533
+ }
1534
+ ]
1535
+ },
1536
+ {
1537
+ "idx": 105,
1538
+ "layers": [
1539
+ "language_model.model.layers.52.mlp.gate_proj",
1540
+ "language_model.model.layers.52.mlp.up_proj",
1541
+ "language_model.model.layers.52.mlp.down_proj"
1542
+ ],
1543
+ "candidates": [
1544
+ {
1545
+ "dkld": -0.0018054209649562836,
1546
+ "dbits": 2080899072
1547
+ }
1548
+ ]
1549
+ },
1550
+ {
1551
+ "idx": 106,
1552
+ "layers": [
1553
+ "language_model.model.layers.53.self_attn.q_proj",
1554
+ "language_model.model.layers.53.self_attn.k_proj",
1555
+ "language_model.model.layers.53.self_attn.v_proj",
1556
+ "language_model.model.layers.53.self_attn.o_proj"
1557
+ ],
1558
+ "candidates": [
1559
+ {
1560
+ "dkld": -0.004864251613616977,
1561
+ "dbits": 396361728
1562
+ }
1563
+ ]
1564
+ },
1565
+ {
1566
+ "idx": 107,
1567
+ "layers": [
1568
+ "language_model.model.layers.53.mlp.gate_proj",
1569
+ "language_model.model.layers.53.mlp.up_proj",
1570
+ "language_model.model.layers.53.mlp.down_proj"
1571
+ ],
1572
+ "candidates": [
1573
+ {
1574
+ "dkld": -0.0019967712461948506,
1575
+ "dbits": 2080899072
1576
+ }
1577
+ ]
1578
+ },
1579
+ {
1580
+ "idx": 108,
1581
+ "layers": [
1582
+ "language_model.model.layers.54.self_attn.q_proj",
1583
+ "language_model.model.layers.54.self_attn.k_proj",
1584
+ "language_model.model.layers.54.self_attn.v_proj",
1585
+ "language_model.model.layers.54.self_attn.o_proj"
1586
+ ],
1587
+ "candidates": [
1588
+ {
1589
+ "dkld": -5.696490406992272e-05,
1590
+ "dbits": 396361728
1591
+ }
1592
+ ]
1593
+ },
1594
+ {
1595
+ "idx": 109,
1596
+ "layers": [
1597
+ "language_model.model.layers.54.mlp.gate_proj",
1598
+ "language_model.model.layers.54.mlp.up_proj",
1599
+ "language_model.model.layers.54.mlp.down_proj"
1600
+ ],
1601
+ "candidates": [
1602
+ {
1603
+ "dkld": -0.0024598926305771096,
1604
+ "dbits": 2080899072
1605
+ }
1606
+ ]
1607
+ },
1608
+ {
1609
+ "idx": 110,
1610
+ "layers": [
1611
+ "language_model.model.layers.55.self_attn.q_proj",
1612
+ "language_model.model.layers.55.self_attn.k_proj",
1613
+ "language_model.model.layers.55.self_attn.v_proj",
1614
+ "language_model.model.layers.55.self_attn.o_proj"
1615
+ ],
1616
+ "candidates": [
1617
+ {
1618
+ "dkld": 0.0008994087576865706,
1619
+ "dbits": 396361728
1620
+ }
1621
+ ]
1622
+ },
1623
+ {
1624
+ "idx": 111,
1625
+ "layers": [
1626
+ "language_model.model.layers.55.mlp.gate_proj",
1627
+ "language_model.model.layers.55.mlp.up_proj",
1628
+ "language_model.model.layers.55.mlp.down_proj"
1629
+ ],
1630
+ "candidates": [
1631
+ {
1632
+ "dkld": -0.003040955960750602,
1633
+ "dbits": 2080899072
1634
+ }
1635
+ ]
1636
+ },
1637
+ {
1638
+ "idx": 112,
1639
+ "layers": [
1640
+ "language_model.model.layers.56.self_attn.q_proj",
1641
+ "language_model.model.layers.56.self_attn.k_proj",
1642
+ "language_model.model.layers.56.self_attn.v_proj",
1643
+ "language_model.model.layers.56.self_attn.o_proj"
1644
+ ],
1645
+ "candidates": [
1646
+ {
1647
+ "dkld": -0.0004954591393471208,
1648
+ "dbits": 396361728
1649
+ }
1650
+ ]
1651
+ },
1652
+ {
1653
+ "idx": 113,
1654
+ "layers": [
1655
+ "language_model.model.layers.56.mlp.gate_proj",
1656
+ "language_model.model.layers.56.mlp.up_proj",
1657
+ "language_model.model.layers.56.mlp.down_proj"
1658
+ ],
1659
+ "candidates": [
1660
+ {
1661
+ "dkld": -0.002797749638557445,
1662
+ "dbits": 2080899072
1663
+ }
1664
+ ]
1665
+ },
1666
+ {
1667
+ "idx": 114,
1668
+ "layers": [
1669
+ "language_model.model.layers.57.self_attn.q_proj",
1670
+ "language_model.model.layers.57.self_attn.k_proj",
1671
+ "language_model.model.layers.57.self_attn.v_proj",
1672
+ "language_model.model.layers.57.self_attn.o_proj"
1673
+ ],
1674
+ "candidates": [
1675
+ {
1676
+ "dkld": -0.0017556458711624479,
1677
+ "dbits": 396361728
1678
+ }
1679
+ ]
1680
+ },
1681
+ {
1682
+ "idx": 115,
1683
+ "layers": [
1684
+ "language_model.model.layers.57.mlp.gate_proj",
1685
+ "language_model.model.layers.57.mlp.up_proj",
1686
+ "language_model.model.layers.57.mlp.down_proj"
1687
+ ],
1688
+ "candidates": [
1689
+ {
1690
+ "dkld": -0.001631450653076183,
1691
+ "dbits": 2080899072
1692
+ }
1693
+ ]
1694
+ },
1695
+ {
1696
+ "idx": 116,
1697
+ "layers": [
1698
+ "language_model.model.layers.58.self_attn.q_proj",
1699
+ "language_model.model.layers.58.self_attn.k_proj",
1700
+ "language_model.model.layers.58.self_attn.v_proj",
1701
+ "language_model.model.layers.58.self_attn.o_proj"
1702
+ ],
1703
+ "candidates": [
1704
+ {
1705
+ "dkld": 0.0004269838333129661,
1706
+ "dbits": 396361728
1707
+ }
1708
+ ]
1709
+ },
1710
+ {
1711
+ "idx": 117,
1712
+ "layers": [
1713
+ "language_model.model.layers.58.mlp.gate_proj",
1714
+ "language_model.model.layers.58.mlp.up_proj",
1715
+ "language_model.model.layers.58.mlp.down_proj"
1716
+ ],
1717
+ "candidates": [
1718
+ {
1719
+ "dkld": -0.004021327197551772,
1720
+ "dbits": 2080899072
1721
+ }
1722
+ ]
1723
+ },
1724
+ {
1725
+ "idx": 118,
1726
+ "layers": [
1727
+ "language_model.model.layers.59.self_attn.q_proj",
1728
+ "language_model.model.layers.59.self_attn.k_proj",
1729
+ "language_model.model.layers.59.self_attn.v_proj",
1730
+ "language_model.model.layers.59.self_attn.o_proj"
1731
+ ],
1732
+ "candidates": [
1733
+ {
1734
+ "dkld": -0.0032634086906910276,
1735
+ "dbits": 396361728
1736
+ }
1737
+ ]
1738
+ },
1739
+ {
1740
+ "idx": 119,
1741
+ "layers": [
1742
+ "language_model.model.layers.59.mlp.gate_proj",
1743
+ "language_model.model.layers.59.mlp.up_proj",
1744
+ "language_model.model.layers.59.mlp.down_proj"
1745
+ ],
1746
+ "candidates": [
1747
+ {
1748
+ "dkld": -0.005423430353403091,
1749
+ "dbits": 2080899072
1750
+ }
1751
+ ]
1752
+ },
1753
+ {
1754
+ "idx": 120,
1755
+ "layers": [
1756
+ "language_model.model.layers.60.self_attn.q_proj",
1757
+ "language_model.model.layers.60.self_attn.k_proj",
1758
+ "language_model.model.layers.60.self_attn.v_proj",
1759
+ "language_model.model.layers.60.self_attn.o_proj"
1760
+ ],
1761
+ "candidates": [
1762
+ {
1763
+ "dkld": -0.0010337680578231812,
1764
+ "dbits": 396361728
1765
+ }
1766
+ ]
1767
+ },
1768
+ {
1769
+ "idx": 121,
1770
+ "layers": [
1771
+ "language_model.model.layers.60.mlp.gate_proj",
1772
+ "language_model.model.layers.60.mlp.up_proj",
1773
+ "language_model.model.layers.60.mlp.down_proj"
1774
+ ],
1775
+ "candidates": [
1776
+ {
1777
+ "dkld": -0.008122941851615939,
1778
+ "dbits": 2080899072
1779
+ }
1780
+ ]
1781
+ },
1782
+ {
1783
+ "idx": 122,
1784
+ "layers": [
1785
+ "language_model.model.layers.61.self_attn.q_proj",
1786
+ "language_model.model.layers.61.self_attn.k_proj",
1787
+ "language_model.model.layers.61.self_attn.v_proj",
1788
+ "language_model.model.layers.61.self_attn.o_proj"
1789
+ ],
1790
+ "candidates": [
1791
+ {
1792
+ "dkld": 0.00033112913370131336,
1793
+ "dbits": 396361728
1794
+ }
1795
+ ]
1796
+ },
1797
+ {
1798
+ "idx": 123,
1799
+ "layers": [
1800
+ "language_model.model.layers.61.mlp.gate_proj",
1801
+ "language_model.model.layers.61.mlp.up_proj",
1802
+ "language_model.model.layers.61.mlp.down_proj"
1803
+ ],
1804
+ "candidates": [
1805
+ {
1806
+ "dkld": -0.007963605970144283,
1807
+ "dbits": 2080899072
1808
+ }
1809
+ ]
1810
+ }
1811
+ ],
1812
+ "base_kld": 0.4373273827135563,
1813
+ "arch_string": "Gemma3ForConditionalGeneration"
1814
+ }