karantonis commited on
Commit
55ec475
·
verified ·
1 Parent(s): 161d954

bayes_sc outputs (math_test_hard, 2026-10-25)

Browse files
Files changed (33) hide show
  1. analysis/bayes_sc_2026-10-25/sc_32only.json +412 -0
  2. analysis/bayes_sc_2026-10-25/sc_32only.log +40 -0
  3. analysis/bayes_sc_2026-10-25/sc_32only.provenance.json +87 -0
  4. analysis/bayes_sc_2026-10-25/sc_32only_c4.json +405 -0
  5. analysis/bayes_sc_2026-10-25/sc_32only_c4.log +40 -0
  6. analysis/bayes_sc_2026-10-25/sc_32only_c4.provenance.json +87 -0
  7. analysis/bayes_sc_2026-10-25/sc_32only_notask.json +374 -0
  8. analysis/bayes_sc_2026-10-25/sc_32only_notask.log +40 -0
  9. analysis/bayes_sc_2026-10-25/sc_32only_notask.provenance.json +89 -0
  10. analysis/bayes_sc_2026-10-25/sc_both_c4.json +413 -0
  11. analysis/bayes_sc_2026-10-25/sc_both_c4.log +40 -0
  12. analysis/bayes_sc_2026-10-25/sc_both_c4.provenance.json +93 -0
  13. analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.json +433 -0
  14. analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.log +40 -0
  15. analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.provenance.json +93 -0
  16. analysis/bayes_sc_2026-10-25/sc_both_ref32.json +421 -0
  17. analysis/bayes_sc_2026-10-25/sc_both_ref32.log +40 -0
  18. analysis/bayes_sc_2026-10-25/sc_both_ref32.provenance.json +93 -0
  19. analysis/bayes_sc_2026-10-25/sc_notask.json +643 -0
  20. analysis/bayes_sc_2026-10-25/sc_notask.log +49 -0
  21. analysis/bayes_sc_2026-10-25/sc_notask.provenance.json +85 -0
  22. analysis/bayes_sc_2026-10-25/sc_task.json +646 -0
  23. analysis/bayes_sc_2026-10-25/sc_task.log +49 -0
  24. analysis/bayes_sc_2026-10-25/sc_task.provenance.json +90 -0
  25. analysis/bayes_sc_2026-10-25/sc_task_32b.json +432 -0
  26. analysis/bayes_sc_2026-10-25/sc_task_32b.log +40 -0
  27. analysis/bayes_sc_2026-10-25/sc_task_32b.provenance.json +93 -0
  28. analysis/bayes_sc_2026-10-25/sc_task_t5.json +416 -0
  29. analysis/bayes_sc_2026-10-25/sc_task_t5.log +40 -0
  30. analysis/bayes_sc_2026-10-25/sc_task_t5.provenance.json +89 -0
  31. analysis/bayes_sc_2026-10-25/sc_task_u5.json +415 -0
  32. analysis/bayes_sc_2026-10-25/sc_task_u5.log +40 -0
  33. analysis/bayes_sc_2026-10-25/sc_task_u5.provenance.json +89 -0
analysis/bayes_sc_2026-10-25/sc_32only.json ADDED
@@ -0,0 +1,412 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.9405767817781043,
4
+ "equiv": 4283.8179463629685
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.9405767817781043,
9
+ "equiv": 4283.8179463629685,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.14979794268919913,
21
+ -0.16256429096252756,
22
+ -0.1369374540778839
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.12876561351947097,
30
+ -0.14226671565025717,
31
+ -0.11544819985304923
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.09616091109478325,
39
+ -0.10911324393828067,
40
+ -0.08357825128581925
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.0852314474650992,
48
+ -0.09809652828802351,
49
+ -0.07274063188831742
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0775165319617928,
57
+ -0.09056070903747244,
58
+ -0.06493387215282881
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.07127112417340191,
66
+ -0.08440714548126378,
67
+ -0.05905584129316679
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.068791329904482,
75
+ -0.08174136664217486,
76
+ -0.05648190668626012
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.06447465099191771,
84
+ -0.07788850110213079,
85
+ -0.05143277002204261
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.06915870683321088,
93
+ -0.08238886847905952,
94
+ -0.05639235855988244
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.9405767817781043,
99
+ "equiv": 4283.8179463629685,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9403930933137399,
108
+ "equiv": 4758.662803085966,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9340558412931668,
117
+ "equiv": 4695.218185157972,
118
+ "delta_vs_agree2": [
119
+ -0.006520940484937546,
120
+ -0.009184423218221896,
121
+ -0.004132990448199853
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.06484202792064658,
129
+ -0.07834542615723733,
130
+ -0.05198383541513593
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9603232916972814,
135
+ "equiv": 7567.180749448935,
136
+ "delta_vs_agree2": [
137
+ 0.019746509919177077,
138
+ 0.012766348273328435,
139
+ 0.026726671565025718
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9542615723732549,
144
+ "equiv": 5892.238997060985,
145
+ "delta_vs_agree2": [
146
+ 0.013684790595150624,
147
+ 0.010929463629684055,
148
+ 0.016623806024981633
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9540778839088905,
153
+ "equiv": 6471.090558412931,
154
+ "delta_vs_agree2": [
155
+ 0.013501102130786187,
156
+ 0.010745775165319618,
157
+ 0.0162587252020573
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.06456649522409992,
165
+ -0.07806989346069067,
166
+ -0.05152461425422483
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9603232916972814,
171
+ "equiv": 7567.180749448935,
172
+ "delta_vs_agree2": [
173
+ 0.019746509919177077,
174
+ 0.012766348273328435,
175
+ 0.026726671565025718
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9603232916972814,
180
+ "equiv": 8692.121601763409,
181
+ "delta_vs_agree2": [
182
+ 0.019746509919177077,
183
+ 0.012766348273328435,
184
+ 0.026726671565025718
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9572005878030859,
189
+ "equiv": 7287.702479794269,
190
+ "delta_vs_agree2": [
191
+ 0.016623806024981633,
192
+ 0.012950036737692874,
193
+ 0.020575404114621592
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.06447465099191771,
201
+ -0.07788850110213079,
202
+ -0.05143277002204261
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9603232916972814,
207
+ "equiv": 7567.180749448935,
208
+ "delta_vs_agree2": [
209
+ 0.019746509919177077,
210
+ 0.012766348273328435,
211
+ 0.026726671565025718
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9603232916972814,
216
+ "equiv": 8692.121601763409,
217
+ "delta_vs_agree2": [
218
+ 0.019746509919177077,
219
+ 0.012766348273328435,
220
+ 0.026726671565025718
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9603232916972814,
225
+ "equiv": 9817.062454077884,
226
+ "delta_vs_agree2": [
227
+ 0.019746509919177077,
228
+ 0.012766348273328435,
229
+ 0.026726671565025718
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9603232916972814,
234
+ "equiv": 5317.299044819985,
235
+ "delta_vs_agree2": [
236
+ 0.019746509919177077,
237
+ 0.012766348273328435,
238
+ 0.026726671565025718
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7909625275532697,
246
+ "equiv": 1127.0635378398238
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7910543717854519,
251
+ "equiv": 1127.5903196179281
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.934974283614989,
256
+ "equiv": 3727.5320536370314
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9424136664217487,
261
+ "equiv": 3919.199797942689
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9510470242468773,
266
+ "equiv": 4193.461278471712
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.955088170462895,
271
+ "equiv": 4461.668038207201
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.955547391623806,
276
+ "equiv": 4540.97211609111
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9568332108743571,
281
+ "equiv": 4838.5335598824395
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9523328434974284,
288
+ "equiv": 4263.571473181484,
289
+ "R": [
290
+ 48694.151017527285,
291
+ 34234.24133567278
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.011756061719324026,
295
+ 0.006426800146950772,
296
+ 0.017820077149155022
297
+ ],
298
+ "actions": {
299
+ "stop@2": 3349,
300
+ "think:q32b_think@0": 3544,
301
+ "stop@3": 280,
302
+ "stop@1": 2549,
303
+ "stop@8": 39,
304
+ "think:q32b_think@2": 794,
305
+ "think:q32b_think@5": 2,
306
+ "stop@5": 108,
307
+ "think:q32b_think@3": 127,
308
+ "stop@6": 10,
309
+ "stop@7": 15,
310
+ "think:q32b_think@6": 4,
311
+ "think:q32b_think@4": 5,
312
+ "think:q32b_think@7": 4,
313
+ "think:q32b_think@8": 8,
314
+ "stop@4": 50
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9546289493019838,
320
+ "equiv": 4350.467395297575,
321
+ "R": [
322
+ 40390.395520208884,
323
+ 42770.39453717809
324
+ ],
325
+ "delta_vs_agree2": [
326
+ 0.0140521675238795,
327
+ 0.008541513592946363,
328
+ 0.020299871418074933
329
+ ],
330
+ "actions": {
331
+ "stop@3": 333,
332
+ "think:q32b_think@3": 180,
333
+ "think:q32b_think@2": 818,
334
+ "stop@2": 3229,
335
+ "stop@4": 40,
336
+ "stop@1": 2468,
337
+ "think:q32b_think@0": 3712,
338
+ "stop@6": 20,
339
+ "stop@5": 49,
340
+ "think:q32b_think@4": 2,
341
+ "stop@8": 7,
342
+ "think:q32b_think@5": 2,
343
+ "think:q32b_think@6": 15,
344
+ "stop@7": 8,
345
+ "think:q32b_think@8": 3,
346
+ "think:q32b_think@7": 2
347
+ }
348
+ },
349
+ {
350
+ "seed": 2,
351
+ "acc": 0.9538941954445261,
352
+ "equiv": 4341.009110947833,
353
+ "R": [
354
+ 48680.35922315368,
355
+ 32562.439482456157
356
+ ],
357
+ "delta_vs_agree2": [
358
+ 0.01331741366642175,
359
+ 0.007531227038941954,
360
+ 0.019746509919177077
361
+ ],
362
+ "actions": {
363
+ "think:q32b_think@0": 3560,
364
+ "stop@2": 3758,
365
+ "stop@1": 1960,
366
+ "stop@8": 25,
367
+ "stop@6": 31,
368
+ "think:q32b_think@2": 731,
369
+ "think:q32b_think@5": 14,
370
+ "stop@5": 152,
371
+ "stop@3": 423,
372
+ "think:q32b_think@7": 11,
373
+ "stop@4": 86,
374
+ "think:q32b_think@4": 36,
375
+ "think:q32b_think@6": 8,
376
+ "stop@7": 14,
377
+ "think:q32b_think@8": 6,
378
+ "think:q32b_think@3": 73
379
+ }
380
+ },
381
+ {
382
+ "seed": 3,
383
+ "acc": 0.9535268185157972,
384
+ "equiv": 4256.150440852314,
385
+ "R": [
386
+ 37481.25107046291,
387
+ 35075.93846705908
388
+ ],
389
+ "delta_vs_agree2": [
390
+ 0.012950036737692874,
391
+ 0.0073452424687729615,
392
+ 0.01901405216752387
393
+ ],
394
+ "actions": {
395
+ "stop@1": 2673,
396
+ "stop@2": 3248,
397
+ "think:q32b_think@2": 874,
398
+ "think:q32b_think@0": 3632,
399
+ "stop@4": 49,
400
+ "stop@5": 60,
401
+ "stop@8": 17,
402
+ "stop@7": 6,
403
+ "stop@6": 5,
404
+ "stop@3": 208,
405
+ "think:q32b_think@3": 108,
406
+ "think:q32b_think@8": 2,
407
+ "think:q32b_think@4": 5,
408
+ "think:q32b_think@7": 1
409
+ }
410
+ }
411
+ ]
412
+ }
analysis/bayes_sc_2026-10-25/sc_32only.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,405 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 4284
3
+ agree2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3159) | vs agree2 -15.0 [-16.3, -13.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2034) | vs agree2 -12.9 [-14.2, -11.5]
6
+ SC@3 acc 0.844 equiv 3375 ( -909) | vs agree2 -9.6 [-10.9, -8.4]
7
+ SC@4 acc 0.855 equiv 4500 ( +216) | vs agree2 -8.5 [-9.8, -7.3]
8
+ SC@5 acc 0.863 equiv 5625 ( +1341) | vs agree2 -7.8 [-9.1, -6.5]
9
+ SC@6 acc 0.869 equiv 6750 ( +2466) | vs agree2 -7.1 [-8.4, -5.9]
10
+ SC@7 acc 0.872 equiv 7875 ( +3591) | vs agree2 -6.9 [-8.2, -5.6]
11
+ SC@8 acc 0.876 equiv 9000 ( +4716) | vs agree2 -6.4 [-7.8, -5.1]
12
+ AC(0.8) acc 0.871 equiv 4329 ( +45) | vs agree2 -6.9 [-8.2, -5.6]
13
+ AC(0.8)+think@2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.940 equiv 4759 ( +475) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.934 equiv 4695 ( +411) | vs agree2 -0.7 [-0.9, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1762) | vs agree2 -6.5 [-7.8, -5.2]
17
+ AC(0.9)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
18
+ AC(0.9)+think@3 acc 0.954 equiv 5892 ( +1608) | vs agree2 +1.4 [+1.1, +1.7]
19
+ AC(0.9)+think@4 acc 0.954 equiv 6471 ( +2187) | vs agree2 +1.4 [+1.1, +1.6]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2687) | vs agree2 -6.5 [-7.8, -5.2]
21
+ AC(0.95)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
22
+ AC(0.95)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
23
+ AC(0.95)+think@4 acc 0.957 equiv 7288 ( +3004) | vs agree2 +1.7 [+1.3, +2.1]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3885) | vs agree2 -6.4 [-7.8, -5.1]
25
+ AC(0.99)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
26
+ AC(0.99)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
27
+ AC(0.99)+think@4 acc 0.960 equiv 9817 ( +5533) | vs agree2 +2.0 [+1.3, +2.7]
28
+ always think acc 0.960 equiv 5317 ( +1033) | vs agree2 +2.0 [+1.3, +2.7]
29
+ R 4000: acc 0.791 equiv 1127
30
+ R 8000: acc 0.791 equiv 1128
31
+ R 16000: acc 0.935 equiv 3728
32
+ R 24000: acc 0.942 equiv 3919
33
+ R 32000: acc 0.951 equiv 4193
34
+ R 48000: acc 0.955 equiv 4462
35
+ R 64000: acc 0.956 equiv 4541
36
+ R 96000: acc 0.957 equiv 4839
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.952 equiv 4264 (-20) | vs agree2 +1.2 [+0.6, +1.8] | R [48694, 34234] | {'think:q32b_think@0': 3544, 'stop@2': 3349, 'stop@1': 2549, 'think:q32b_think@2': 794, 'stop@3': 280, 'think:q32b_think@3': 127, 'stop@5': 108, 'stop@4': 50}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.955 equiv 4350 (+67) | vs agree2 +1.4 [+0.9, +2.0] | R [40390, 42770] | {'think:q32b_think@0': 3712, 'stop@2': 3229, 'stop@1': 2468, 'think:q32b_think@2': 818, 'stop@3': 333, 'think:q32b_think@3': 180, 'stop@5': 49, 'stop@4': 40}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4341 (+57) | vs agree2 +1.3 [+0.8, +2.0] | R [48680, 32562] | {'stop@2': 3758, 'think:q32b_think@0': 3560, 'stop@1': 1960, 'think:q32b_think@2': 731, 'stop@3': 423, 'stop@5': 152, 'stop@4': 86, 'think:q32b_think@3': 73}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4256 (-28) | vs agree2 +1.3 [+0.7, +1.9] | R [37481, 35076] | {'think:q32b_think@0': 3632, 'stop@2': 3248, 'stop@1': 2673, 'think:q32b_think@2': 874, 'stop@3': 208, 'think:q32b_think@3': 108, 'stop@5': 60, 'stop@4': 49}
analysis/bayes_sc_2026-10-25/sc_32only.provenance.json ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "q32b_think=workdir/bayes/q32b_think_tier.json",
16
+ "--out",
17
+ "workdir/bayes/sc_32only.json"
18
+ ],
19
+ "git": {
20
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
21
+ "branch": "HEAD",
22
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
23
+ "dirty": false,
24
+ "dirty_files": []
25
+ },
26
+ "gpus": {
27
+ "visible": "",
28
+ "devices": [
29
+ {
30
+ "index": "0",
31
+ "name": "NVIDIA RTX 5880 Ada Generation",
32
+ "memory": "49140 MiB",
33
+ "driver": "580.178.04",
34
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
35
+ },
36
+ {
37
+ "index": "1",
38
+ "name": "NVIDIA RTX 5880 Ada Generation",
39
+ "memory": "49140 MiB",
40
+ "driver": "580.178.04",
41
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
42
+ }
43
+ ]
44
+ },
45
+ "packages": {
46
+ "python": "3.11.15",
47
+ "torch": "2.9.1",
48
+ "vllm": "0.15.1",
49
+ "transformers": "4.57.6",
50
+ "speculators": "0.4.0.dev49",
51
+ "peft": null,
52
+ "accelerate": "1.13.0",
53
+ "numpy": "2.2.6",
54
+ "huggingface_hub": "0.36.2",
55
+ "openai": "2.32.0",
56
+ "wandb": "0.25.1"
57
+ },
58
+ "env": {
59
+ "CUDA_VISIBLE_DEVICES": ""
60
+ },
61
+ "models": [],
62
+ "files": {
63
+ "pool": {
64
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
65
+ "bytes": 65136887,
66
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
67
+ },
68
+ "uhead": {
69
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
70
+ "bytes": 9369632,
71
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
72
+ },
73
+ "task_prior": {
74
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
75
+ "bytes": 27027,
76
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
77
+ },
78
+ "tier:q32b_think": {
79
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
80
+ "bytes": 26525,
81
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
82
+ }
83
+ },
84
+ "finished_at": "2026-10-25T11:35:35+04:00",
85
+ "wall_secs": 36.2,
86
+ "n_problems": 1361
87
+ }
analysis/bayes_sc_2026-10-25/sc_32only_c4.json ADDED
@@ -0,0 +1,405 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.9405767817781043,
4
+ "equiv": 10385.626671565025
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.9405767817781043,
9
+ "equiv": 10385.626671565025,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.14979794268919913,
21
+ -0.16256429096252756,
22
+ -0.1369374540778839
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.12876561351947097,
30
+ -0.14226671565025717,
31
+ -0.11544819985304923
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.09616091109478325,
39
+ -0.10911324393828067,
40
+ -0.08357825128581925
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.0852314474650992,
48
+ -0.09809652828802351,
49
+ -0.07274063188831742
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0775165319617928,
57
+ -0.09056070903747244,
58
+ -0.06493387215282881
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.07127112417340191,
66
+ -0.08440714548126378,
67
+ -0.05905584129316679
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.068791329904482,
75
+ -0.08174136664217486,
76
+ -0.05648190668626012
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.06447465099191771,
84
+ -0.07788850110213079,
85
+ -0.05143277002204261
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.06915870683321088,
93
+ -0.08238886847905952,
94
+ -0.05639235855988244
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.9405767817781043,
99
+ "equiv": 10385.626671565025,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9403930933137399,
108
+ "equiv": 10855.038243938281,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9340558412931668,
117
+ "equiv": 9174.927020573108,
118
+ "delta_vs_agree2": [
119
+ -0.006520940484937546,
120
+ -0.009184423218221896,
121
+ -0.004132990448199853
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.06484202792064658,
129
+ -0.07834542615723733,
130
+ -0.05198383541513593
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9603232916972814,
135
+ "equiv": 23519.07788390889,
136
+ "delta_vs_agree2": [
137
+ 0.019746509919177077,
138
+ 0.012766348273328435,
139
+ 0.026726671565025718
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9542615723732549,
144
+ "equiv": 13444.488317413665,
145
+ "delta_vs_agree2": [
146
+ 0.013684790595150624,
147
+ 0.010929463629684055,
148
+ 0.016623806024981633
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9540778839088905,
153
+ "equiv": 14017.90659441587,
154
+ "delta_vs_agree2": [
155
+ 0.013501102130786187,
156
+ 0.010745775165319618,
157
+ 0.0162587252020573
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.06456649522409992,
165
+ -0.07806989346069067,
166
+ -0.05152461425422483
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9603232916972814,
171
+ "equiv": 23519.07788390889,
172
+ "delta_vs_agree2": [
173
+ 0.019746509919177077,
174
+ 0.012766348273328435,
175
+ 0.026726671565025718
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9603232916972814,
180
+ "equiv": 24644.018736223366,
181
+ "delta_vs_agree2": [
182
+ 0.019746509919177077,
183
+ 0.012766348273328435,
184
+ 0.026726671565025718
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9572005878030859,
189
+ "equiv": 15651.51969140338,
190
+ "delta_vs_agree2": [
191
+ 0.016623806024981633,
192
+ 0.012950036737692874,
193
+ 0.020575404114621592
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.06447465099191771,
201
+ -0.07788850110213079,
202
+ -0.05143277002204261
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9603232916972814,
207
+ "equiv": 23519.07788390889,
208
+ "delta_vs_agree2": [
209
+ 0.019746509919177077,
210
+ 0.012766348273328435,
211
+ 0.026726671565025718
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9603232916972814,
216
+ "equiv": 24644.018736223366,
217
+ "delta_vs_agree2": [
218
+ 0.019746509919177077,
219
+ 0.012766348273328435,
220
+ 0.026726671565025718
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9603232916972814,
225
+ "equiv": 25768.95958853784,
226
+ "delta_vs_agree2": [
227
+ 0.019746509919177077,
228
+ 0.012766348273328435,
229
+ 0.026726671565025718
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9603232916972814,
234
+ "equiv": 21269.19617927994,
235
+ "delta_vs_agree2": [
236
+ 0.019746509919177077,
237
+ 0.012766348273328435,
238
+ 0.026726671565025718
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.8060249816311535,
256
+ "equiv": 1423.2955547391623
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.8094232182218957,
261
+ "equiv": 1574.8718405584132
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.8411094783247612,
266
+ "equiv": 2807.0781778104333
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9053085966201323,
271
+ "equiv": 6692.102920646583
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9258817046289493,
276
+ "equiv": 8548.656612784718
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9373622336517267,
281
+ "equiv": 9783.342652461426
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9427810433504776,
288
+ "equiv": 10606.631373989714,
289
+ "R": [
290
+ 113839.80095257515,
291
+ 153028.11346051164
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.002204261572373255,
295
+ -0.0021147134459955914,
296
+ 0.006339548126377651
297
+ ],
298
+ "actions": {
299
+ "stop@2": 5017,
300
+ "stop@3": 2377,
301
+ "stop@5": 844,
302
+ "think:q32b_think@2": 504,
303
+ "stop@6": 297,
304
+ "stop@8": 149,
305
+ "think:q32b_think@3": 177,
306
+ "think:q32b_think@7": 150,
307
+ "think:q32b_think@5": 341,
308
+ "think:q32b_think@4": 158,
309
+ "think:q32b_think@6": 270,
310
+ "stop@7": 152,
311
+ "stop@4": 219,
312
+ "think:q32b_think@8": 233
313
+ }
314
+ },
315
+ {
316
+ "seed": 1,
317
+ "acc": 0.9413115356355621,
318
+ "equiv": 10574.579555473914,
319
+ "R": [
320
+ 174149.2709454739,
321
+ 146559.62929221688
322
+ ],
323
+ "delta_vs_agree2": [
324
+ 0.0007347538574577516,
325
+ -0.0034923769287288756,
326
+ 0.004961884643644367
327
+ ],
328
+ "actions": {
329
+ "stop@6": 373,
330
+ "stop@4": 833,
331
+ "stop@3": 1433,
332
+ "stop@2": 5365,
333
+ "think:q32b_think@6": 328,
334
+ "stop@5": 755,
335
+ "think:q32b_think@2": 202,
336
+ "stop@8": 186,
337
+ "stop@7": 140,
338
+ "think:q32b_think@5": 340,
339
+ "think:q32b_think@7": 117,
340
+ "think:q32b_think@4": 208,
341
+ "think:q32b_think@3": 494,
342
+ "think:q32b_think@8": 114
343
+ }
344
+ },
345
+ {
346
+ "seed": 2,
347
+ "acc": 0.9404849375459221,
348
+ "equiv": 10125.881392358559,
349
+ "R": [
350
+ 108933.7532498737,
351
+ 168649.71743768983
352
+ ],
353
+ "delta_vs_agree2": [
354
+ -9.184423218221895e-05,
355
+ -0.004775900073475386,
356
+ 0.004502663482733271
357
+ ],
358
+ "actions": {
359
+ "stop@5": 811,
360
+ "stop@4": 670,
361
+ "stop@3": 1030,
362
+ "stop@2": 6040,
363
+ "stop@8": 156,
364
+ "stop@6": 385,
365
+ "think:q32b_think@3": 365,
366
+ "think:q32b_think@4": 434,
367
+ "think:q32b_think@5": 170,
368
+ "think:q32b_think@6": 330,
369
+ "think:q32b_think@7": 163,
370
+ "stop@7": 157,
371
+ "think:q32b_think@8": 173,
372
+ "think:q32b_think@2": 4
373
+ }
374
+ },
375
+ {
376
+ "seed": 3,
377
+ "acc": 0.9413115356355621,
378
+ "equiv": 10514.614346069067,
379
+ "R": [
380
+ 165940.81680835842,
381
+ 123237.73422768388
382
+ ],
383
+ "delta_vs_agree2": [
384
+ 0.0007347538574577516,
385
+ -0.0036737692872887582,
386
+ 0.0050514327700220425
387
+ ],
388
+ "actions": {
389
+ "stop@2": 4132,
390
+ "stop@6": 314,
391
+ "stop@5": 881,
392
+ "stop@3": 2403,
393
+ "stop@4": 1116,
394
+ "think:q32b_think@8": 101,
395
+ "think:q32b_think@7": 99,
396
+ "think:q32b_think@5": 361,
397
+ "think:q32b_think@6": 519,
398
+ "stop@8": 170,
399
+ "stop@7": 150,
400
+ "think:q32b_think@4": 257,
401
+ "think:q32b_think@3": 385
402
+ }
403
+ }
404
+ ]
405
+ }
analysis/bayes_sc_2026-10-25/sc_32only_c4.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:36:34,770 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 10386
3
+ agree2 acc 0.941 equiv 10386 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -9261) | vs agree2 -15.0 [-16.3, -13.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -8136) | vs agree2 -12.9 [-14.2, -11.5]
6
+ SC@3 acc 0.844 equiv 3375 ( -7011) | vs agree2 -9.6 [-10.9, -8.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -5886) | vs agree2 -8.5 [-9.8, -7.3]
8
+ SC@5 acc 0.863 equiv 5625 ( -4761) | vs agree2 -7.8 [-9.1, -6.5]
9
+ SC@6 acc 0.869 equiv 6750 ( -3636) | vs agree2 -7.1 [-8.4, -5.9]
10
+ SC@7 acc 0.872 equiv 7875 ( -2511) | vs agree2 -6.9 [-8.2, -5.6]
11
+ SC@8 acc 0.876 equiv 9000 ( -1386) | vs agree2 -6.4 [-7.8, -5.1]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -6057) | vs agree2 -6.9 [-8.2, -5.6]
13
+ AC(0.8)+think@2 acc 0.941 equiv 10386 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.940 equiv 10855 ( +469) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.934 equiv 9175 ( -1211) | vs agree2 -0.7 [-0.9, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( -4340) | vs agree2 -6.5 [-7.8, -5.2]
17
+ AC(0.9)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
18
+ AC(0.9)+think@3 acc 0.954 equiv 13444 ( +3059) | vs agree2 +1.4 [+1.1, +1.7]
19
+ AC(0.9)+think@4 acc 0.954 equiv 14018 ( +3632) | vs agree2 +1.4 [+1.1, +1.6]
20
+ AC(0.95) acc 0.876 equiv 6970 ( -3415) | vs agree2 -6.5 [-7.8, -5.2]
21
+ AC(0.95)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
22
+ AC(0.95)+think@3 acc 0.960 equiv 24644 (+14258) | vs agree2 +2.0 [+1.3, +2.7]
23
+ AC(0.95)+think@4 acc 0.957 equiv 15652 ( +5266) | vs agree2 +1.7 [+1.3, +2.1]
24
+ AC(0.99) acc 0.876 equiv 8168 ( -2217) | vs agree2 -6.4 [-7.8, -5.1]
25
+ AC(0.99)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
26
+ AC(0.99)+think@3 acc 0.960 equiv 24644 (+14258) | vs agree2 +2.0 [+1.3, +2.7]
27
+ AC(0.99)+think@4 acc 0.960 equiv 25769 (+15383) | vs agree2 +2.0 [+1.3, +2.7]
28
+ always think acc 0.960 equiv 21269 (+10884) | vs agree2 +2.0 [+1.3, +2.7]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.806 equiv 1423
32
+ R 24000: acc 0.809 equiv 1575
33
+ R 32000: acc 0.841 equiv 2807
34
+ R 48000: acc 0.905 equiv 6692
35
+ R 64000: acc 0.926 equiv 8549
36
+ R 96000: acc 0.937 equiv 9783
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.943 equiv 10607 (+221) | vs agree2 +0.2 [-0.2, +0.6] | R [113840, 153028] | {'stop@2': 5017, 'stop@3': 2377, 'stop@5': 844, 'think:q32b_think@2': 504, 'think:q32b_think@5': 341, 'stop@6': 297, 'think:q32b_think@6': 270, 'think:q32b_think@8': 233}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.941 equiv 10575 (+189) | vs agree2 +0.1 [-0.3, +0.5] | R [174149, 146560] | {'stop@2': 5365, 'stop@3': 1433, 'stop@4': 833, 'stop@5': 755, 'think:q32b_think@3': 494, 'stop@6': 373, 'think:q32b_think@5': 340, 'think:q32b_think@6': 328}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.940 equiv 10126 (-260) | vs agree2 -0.0 [-0.5, +0.5] | R [108934, 168650] | {'stop@2': 6040, 'stop@3': 1030, 'stop@5': 811, 'stop@4': 670, 'think:q32b_think@4': 434, 'stop@6': 385, 'think:q32b_think@3': 365, 'think:q32b_think@6': 330}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.941 equiv 10515 (+129) | vs agree2 +0.1 [-0.4, +0.5] | R [165941, 123238] | {'stop@2': 4132, 'stop@3': 2403, 'stop@4': 1116, 'stop@5': 881, 'think:q32b_think@6': 519, 'think:q32b_think@3': 385, 'think:q32b_think@5': 361, 'stop@6': 314}
analysis/bayes_sc_2026-10-25/sc_32only_c4.provenance.json ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:36:33+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "q32b_think=workdir/bayes/q32b_think_tier.json@4",
16
+ "--out",
17
+ "workdir/bayes/sc_32only_c4.json"
18
+ ],
19
+ "git": {
20
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
21
+ "branch": "HEAD",
22
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
23
+ "dirty": false,
24
+ "dirty_files": []
25
+ },
26
+ "gpus": {
27
+ "visible": "",
28
+ "devices": [
29
+ {
30
+ "index": "0",
31
+ "name": "NVIDIA RTX 5880 Ada Generation",
32
+ "memory": "49140 MiB",
33
+ "driver": "580.178.04",
34
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
35
+ },
36
+ {
37
+ "index": "1",
38
+ "name": "NVIDIA RTX 5880 Ada Generation",
39
+ "memory": "49140 MiB",
40
+ "driver": "580.178.04",
41
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
42
+ }
43
+ ]
44
+ },
45
+ "packages": {
46
+ "python": "3.11.15",
47
+ "torch": "2.9.1",
48
+ "vllm": "0.15.1",
49
+ "transformers": "4.57.6",
50
+ "speculators": "0.4.0.dev49",
51
+ "peft": null,
52
+ "accelerate": "1.13.0",
53
+ "numpy": "2.2.6",
54
+ "huggingface_hub": "0.36.2",
55
+ "openai": "2.32.0",
56
+ "wandb": "0.25.1"
57
+ },
58
+ "env": {
59
+ "CUDA_VISIBLE_DEVICES": ""
60
+ },
61
+ "models": [],
62
+ "files": {
63
+ "pool": {
64
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
65
+ "bytes": 65136887,
66
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
67
+ },
68
+ "uhead": {
69
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
70
+ "bytes": 9369632,
71
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
72
+ },
73
+ "task_prior": {
74
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
75
+ "bytes": 27027,
76
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
77
+ },
78
+ "tier:q32b_think": {
79
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
80
+ "bytes": 26525,
81
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
82
+ }
83
+ },
84
+ "finished_at": "2026-10-25T11:37:14+04:00",
85
+ "wall_secs": 41.2,
86
+ "n_problems": 1361
87
+ }
analysis/bayes_sc_2026-10-25/sc_32only_notask.json ADDED
@@ -0,0 +1,374 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.9405767817781043,
4
+ "equiv": 4283.8179463629685
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.9405767817781043,
9
+ "equiv": 4283.8179463629685,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.14979794268919913,
21
+ -0.16256429096252756,
22
+ -0.1369374540778839
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.12876561351947097,
30
+ -0.14226671565025717,
31
+ -0.11544819985304923
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.09616091109478325,
39
+ -0.10911324393828067,
40
+ -0.08357825128581925
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.0852314474650992,
48
+ -0.09809652828802351,
49
+ -0.07274063188831742
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0775165319617928,
57
+ -0.09056070903747244,
58
+ -0.06493387215282881
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.07127112417340191,
66
+ -0.08440714548126378,
67
+ -0.05905584129316679
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.068791329904482,
75
+ -0.08174136664217486,
76
+ -0.05648190668626012
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.06447465099191771,
84
+ -0.07788850110213079,
85
+ -0.05143277002204261
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.06915870683321088,
93
+ -0.08238886847905952,
94
+ -0.05639235855988244
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.9405767817781043,
99
+ "equiv": 4283.8179463629685,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9403930933137399,
108
+ "equiv": 4758.662803085966,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9340558412931668,
117
+ "equiv": 4695.218185157972,
118
+ "delta_vs_agree2": [
119
+ -0.006520940484937546,
120
+ -0.009184423218221896,
121
+ -0.004132990448199853
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.06484202792064658,
129
+ -0.07834542615723733,
130
+ -0.05198383541513593
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9603232916972814,
135
+ "equiv": 7567.180749448935,
136
+ "delta_vs_agree2": [
137
+ 0.019746509919177077,
138
+ 0.012766348273328435,
139
+ 0.026726671565025718
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9542615723732549,
144
+ "equiv": 5892.238997060985,
145
+ "delta_vs_agree2": [
146
+ 0.013684790595150624,
147
+ 0.010929463629684055,
148
+ 0.016623806024981633
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9540778839088905,
153
+ "equiv": 6471.090558412931,
154
+ "delta_vs_agree2": [
155
+ 0.013501102130786187,
156
+ 0.010745775165319618,
157
+ 0.0162587252020573
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.06456649522409992,
165
+ -0.07806989346069067,
166
+ -0.05152461425422483
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9603232916972814,
171
+ "equiv": 7567.180749448935,
172
+ "delta_vs_agree2": [
173
+ 0.019746509919177077,
174
+ 0.012766348273328435,
175
+ 0.026726671565025718
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9603232916972814,
180
+ "equiv": 8692.121601763409,
181
+ "delta_vs_agree2": [
182
+ 0.019746509919177077,
183
+ 0.012766348273328435,
184
+ 0.026726671565025718
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9572005878030859,
189
+ "equiv": 7287.702479794269,
190
+ "delta_vs_agree2": [
191
+ 0.016623806024981633,
192
+ 0.012950036737692874,
193
+ 0.020575404114621592
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.06447465099191771,
201
+ -0.07788850110213079,
202
+ -0.05143277002204261
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9603232916972814,
207
+ "equiv": 7567.180749448935,
208
+ "delta_vs_agree2": [
209
+ 0.019746509919177077,
210
+ 0.012766348273328435,
211
+ 0.026726671565025718
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9603232916972814,
216
+ "equiv": 8692.121601763409,
217
+ "delta_vs_agree2": [
218
+ 0.019746509919177077,
219
+ 0.012766348273328435,
220
+ 0.026726671565025718
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9603232916972814,
225
+ "equiv": 9817.062454077884,
226
+ "delta_vs_agree2": [
227
+ 0.019746509919177077,
228
+ 0.012766348273328435,
229
+ 0.026726671565025718
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9603232916972814,
234
+ "equiv": 5317.299044819985,
235
+ "delta_vs_agree2": [
236
+ 0.019746509919177077,
237
+ 0.012766348273328435,
238
+ 0.026726671565025718
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7910543717854519,
246
+ "equiv": 1127.154426891991
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7911462160176341,
251
+ "equiv": 1127.6812086700954
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.9097171197648788,
256
+ "equiv": 3286.7608743570904
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9213813372520205,
261
+ "equiv": 3566.382347538575
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9482916972814107,
266
+ "equiv": 4444.946454812638
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9529757531227039,
271
+ "equiv": 4766.096179279941
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9530675973548861,
276
+ "equiv": 4776.4097171197645
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9530675973548861,
281
+ "equiv": 5014.745867009552
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9481998530492285,
288
+ "equiv": 4440.787821454813,
289
+ "R": [
290
+ 29978.780100319873,
291
+ 27760.772095030592
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.007623071271124174,
295
+ 0.0035819250551065392,
296
+ 0.012033890521675226
297
+ ],
298
+ "actions": {
299
+ "stop@1": 3623,
300
+ "stop@2": 3346,
301
+ "think:q32b_think@2": 1574,
302
+ "think:q32b_think@1": 1772,
303
+ "stop@5": 10,
304
+ "think:q32b_think@4": 1,
305
+ "stop@8": 6,
306
+ "stop@3": 390,
307
+ "stop@6": 4,
308
+ "think:q32b_think@6": 1,
309
+ "think:q32b_think@3": 161
310
+ }
311
+ },
312
+ {
313
+ "seed": 1,
314
+ "acc": 0.9290962527553269,
315
+ "equiv": 3772.0963262307123,
316
+ "R": [
317
+ 27235.810025446026,
318
+ 29034.55051071889
319
+ ],
320
+ "delta_vs_agree2": [
321
+ -0.01148052902277737,
322
+ -0.017636388684790595,
323
+ -0.004775900073475386
324
+ ],
325
+ "actions": {
326
+ "think:q32b_think@1": 3656,
327
+ "stop@1": 7232
328
+ }
329
+ },
330
+ {
331
+ "seed": 2,
332
+ "acc": 0.9414033798677444,
333
+ "equiv": 4102.472244673035,
334
+ "R": [
335
+ 26256.268727058792,
336
+ 30642.284167395286
337
+ ],
338
+ "delta_vs_agree2": [
339
+ 0.0008265980896399706,
340
+ -0.0046840558412931665,
341
+ 0.006888317413666422
342
+ ],
343
+ "actions": {
344
+ "stop@1": 5513,
345
+ "think:q32b_think@1": 3627,
346
+ "stop@2": 1443,
347
+ "stop@5": 11,
348
+ "think:q32b_think@2": 279,
349
+ "think:q32b_think@7": 2,
350
+ "stop@8": 4,
351
+ "stop@3": 8,
352
+ "stop@6": 1
353
+ }
354
+ },
355
+ {
356
+ "seed": 3,
357
+ "acc": 0.9288207200587804,
358
+ "equiv": 3781.012582659809,
359
+ "R": [
360
+ 30443.691061140762,
361
+ 26112.063515013404
362
+ ],
363
+ "delta_vs_agree2": [
364
+ -0.011756061719324026,
365
+ -0.01763409257898604,
366
+ -0.005232825128581938
367
+ ],
368
+ "actions": {
369
+ "stop@1": 7239,
370
+ "think:q32b_think@1": 3649
371
+ }
372
+ }
373
+ ]
374
+ }
analysis/bayes_sc_2026-10-25/sc_32only_notask.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:36:34,779 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 4284
3
+ agree2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3159) | vs agree2 -15.0 [-16.3, -13.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2034) | vs agree2 -12.9 [-14.2, -11.5]
6
+ SC@3 acc 0.844 equiv 3375 ( -909) | vs agree2 -9.6 [-10.9, -8.4]
7
+ SC@4 acc 0.855 equiv 4500 ( +216) | vs agree2 -8.5 [-9.8, -7.3]
8
+ SC@5 acc 0.863 equiv 5625 ( +1341) | vs agree2 -7.8 [-9.1, -6.5]
9
+ SC@6 acc 0.869 equiv 6750 ( +2466) | vs agree2 -7.1 [-8.4, -5.9]
10
+ SC@7 acc 0.872 equiv 7875 ( +3591) | vs agree2 -6.9 [-8.2, -5.6]
11
+ SC@8 acc 0.876 equiv 9000 ( +4716) | vs agree2 -6.4 [-7.8, -5.1]
12
+ AC(0.8) acc 0.871 equiv 4329 ( +45) | vs agree2 -6.9 [-8.2, -5.6]
13
+ AC(0.8)+think@2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.940 equiv 4759 ( +475) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.934 equiv 4695 ( +411) | vs agree2 -0.7 [-0.9, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1762) | vs agree2 -6.5 [-7.8, -5.2]
17
+ AC(0.9)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
18
+ AC(0.9)+think@3 acc 0.954 equiv 5892 ( +1608) | vs agree2 +1.4 [+1.1, +1.7]
19
+ AC(0.9)+think@4 acc 0.954 equiv 6471 ( +2187) | vs agree2 +1.4 [+1.1, +1.6]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2687) | vs agree2 -6.5 [-7.8, -5.2]
21
+ AC(0.95)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
22
+ AC(0.95)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
23
+ AC(0.95)+think@4 acc 0.957 equiv 7288 ( +3004) | vs agree2 +1.7 [+1.3, +2.1]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3885) | vs agree2 -6.4 [-7.8, -5.1]
25
+ AC(0.99)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
26
+ AC(0.99)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
27
+ AC(0.99)+think@4 acc 0.960 equiv 9817 ( +5533) | vs agree2 +2.0 [+1.3, +2.7]
28
+ always think acc 0.960 equiv 5317 ( +1033) | vs agree2 +2.0 [+1.3, +2.7]
29
+ R 4000: acc 0.791 equiv 1127
30
+ R 8000: acc 0.791 equiv 1128
31
+ R 16000: acc 0.910 equiv 3287
32
+ R 24000: acc 0.921 equiv 3566
33
+ R 32000: acc 0.948 equiv 4445
34
+ R 48000: acc 0.953 equiv 4766
35
+ R 64000: acc 0.953 equiv 4776
36
+ R 96000: acc 0.953 equiv 5015
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.948 equiv 4441 (+157) | vs agree2 +0.8 [+0.4, +1.2] | R [29979, 27761] | {'stop@1': 3623, 'stop@2': 3346, 'think:q32b_think@1': 1772, 'think:q32b_think@2': 1574, 'stop@3': 390, 'think:q32b_think@3': 161, 'stop@5': 10, 'stop@8': 6}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.929 equiv 3772 (-512) | vs agree2 -1.1 [-1.8, -0.5] | R [27236, 29035] | {'stop@1': 7232, 'think:q32b_think@1': 3656}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.941 equiv 4102 (-181) | vs agree2 +0.1 [-0.5, +0.7] | R [26256, 30642] | {'stop@1': 5513, 'think:q32b_think@1': 3627, 'stop@2': 1443, 'think:q32b_think@2': 279, 'stop@5': 11, 'stop@3': 8, 'stop@8': 4, 'think:q32b_think@7': 2}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.929 equiv 3781 (-503) | vs agree2 -1.2 [-1.8, -0.5] | R [30444, 26112] | {'stop@1': 7239, 'think:q32b_think@1': 3649}
analysis/bayes_sc_2026-10-25/sc_32only_notask.provenance.json ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:36:33+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "q32b_think=workdir/bayes/q32b_think_tier.json",
16
+ "--tbins",
17
+ "1",
18
+ "--out",
19
+ "workdir/bayes/sc_32only_notask.json"
20
+ ],
21
+ "git": {
22
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
23
+ "branch": "HEAD",
24
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
25
+ "dirty": false,
26
+ "dirty_files": []
27
+ },
28
+ "gpus": {
29
+ "visible": "",
30
+ "devices": [
31
+ {
32
+ "index": "0",
33
+ "name": "NVIDIA RTX 5880 Ada Generation",
34
+ "memory": "49140 MiB",
35
+ "driver": "580.178.04",
36
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
37
+ },
38
+ {
39
+ "index": "1",
40
+ "name": "NVIDIA RTX 5880 Ada Generation",
41
+ "memory": "49140 MiB",
42
+ "driver": "580.178.04",
43
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
44
+ }
45
+ ]
46
+ },
47
+ "packages": {
48
+ "python": "3.11.15",
49
+ "torch": "2.9.1",
50
+ "vllm": "0.15.1",
51
+ "transformers": "4.57.6",
52
+ "speculators": "0.4.0.dev49",
53
+ "peft": null,
54
+ "accelerate": "1.13.0",
55
+ "numpy": "2.2.6",
56
+ "huggingface_hub": "0.36.2",
57
+ "openai": "2.32.0",
58
+ "wandb": "0.25.1"
59
+ },
60
+ "env": {
61
+ "CUDA_VISIBLE_DEVICES": ""
62
+ },
63
+ "models": [],
64
+ "files": {
65
+ "pool": {
66
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
67
+ "bytes": 65136887,
68
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
69
+ },
70
+ "uhead": {
71
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
72
+ "bytes": 9369632,
73
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
74
+ },
75
+ "task_prior": {
76
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
77
+ "bytes": 27027,
78
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
79
+ },
80
+ "tier:q32b_think": {
81
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
82
+ "bytes": 26525,
83
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
84
+ }
85
+ },
86
+ "finished_at": "2026-10-25T11:37:06+04:00",
87
+ "wall_secs": 33.0,
88
+ "n_problems": 1361
89
+ }
analysis/bayes_sc_2026-10-25/sc_both_c4.json ADDED
@@ -0,0 +1,413 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.830088170462895,
256
+ "equiv": 1919.7061260102867
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.924917340191036,
261
+ "equiv": 4457.099834680382
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.929142174871418,
266
+ "equiv": 4607.107448567231
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9381888317413667,
271
+ "equiv": 5009.742376928729
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9395205731080088,
276
+ "equiv": 5187.553829904482
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9407145481263777,
281
+ "equiv": 5436.882365907421
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9302902277736959,
288
+ "equiv": 4666.07443056576,
289
+ "R": [
290
+ 35955.38014708592,
291
+ 32356.730332100717
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.0003673769287288758,
295
+ -0.005098502939015429,
296
+ 0.00629247795738427
297
+ ],
298
+ "actions": {
299
+ "stop@2": 2976,
300
+ "stop@1": 3189,
301
+ "think:think16k@0": 3544,
302
+ "stop@3": 187,
303
+ "stop@8": 31,
304
+ "stop@6": 42,
305
+ "think:think16k@2": 258,
306
+ "stop@5": 76,
307
+ "think:think16k@4": 87,
308
+ "think:think16k@3": 204,
309
+ "stop@4": 221,
310
+ "think:think16k@7": 3,
311
+ "think:think16k@6": 8,
312
+ "think:think16k@8": 16,
313
+ "stop@7": 7,
314
+ "think:think16k@5": 39
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9359845701689934,
320
+ "equiv": 4829.496840558412,
321
+ "R": [
322
+ 30965.546229776515,
323
+ 29502.30501310991
324
+ ],
325
+ "delta_vs_agree2": [
326
+ 0.006061719324026451,
327
+ 0.0009184423218221896,
328
+ 0.011893828067597355
329
+ ],
330
+ "actions": {
331
+ "stop@3": 325,
332
+ "think:think16k@3": 301,
333
+ "think:think16k@2": 296,
334
+ "stop@2": 3394,
335
+ "stop@4": 169,
336
+ "stop@1": 2468,
337
+ "think:think16k@0": 3712,
338
+ "stop@6": 22,
339
+ "stop@5": 75,
340
+ "think:think16k@5": 11,
341
+ "think:think16k@4": 72,
342
+ "think:think16k@6": 4,
343
+ "think:think16k@7": 1,
344
+ "stop@8": 14,
345
+ "stop@7": 10,
346
+ "think:think16k@8": 14
347
+ }
348
+ },
349
+ {
350
+ "seed": 2,
351
+ "acc": 0.9311168258633358,
352
+ "equiv": 4579.799035635562,
353
+ "R": [
354
+ 39094.276600707366,
355
+ 30978.196393888564
356
+ ],
357
+ "delta_vs_agree2": [
358
+ 0.0011939750183688465,
359
+ -0.004592211609110948,
360
+ 0.007209772226304188
361
+ ],
362
+ "actions": {
363
+ "think:think16k@0": 3560,
364
+ "stop@2": 3013,
365
+ "stop@1": 3242,
366
+ "stop@8": 28,
367
+ "stop@6": 34,
368
+ "stop@5": 180,
369
+ "stop@7": 28,
370
+ "think:think16k@5": 21,
371
+ "stop@3": 124,
372
+ "think:think16k@3": 42,
373
+ "think:think16k@7": 7,
374
+ "think:think16k@6": 18,
375
+ "think:think16k@2": 476,
376
+ "stop@4": 78,
377
+ "think:think16k@4": 34,
378
+ "think:think16k@8": 3
379
+ }
380
+ },
381
+ {
382
+ "seed": 3,
383
+ "acc": 0.9323108008817046,
384
+ "equiv": 4589.129151359295,
385
+ "R": [
386
+ 28710.255862698214,
387
+ 30490.45634869312
388
+ ],
389
+ "delta_vs_agree2": [
390
+ 0.002387950036737693,
391
+ -0.0031227038941954444,
392
+ 0.008495591476855253
393
+ ],
394
+ "actions": {
395
+ "stop@1": 3886,
396
+ "stop@2": 2327,
397
+ "think:think16k@2": 562,
398
+ "think:think16k@0": 3632,
399
+ "stop@3": 113,
400
+ "stop@4": 73,
401
+ "stop@5": 137,
402
+ "think:think16k@3": 43,
403
+ "stop@8": 27,
404
+ "stop@7": 11,
405
+ "stop@6": 24,
406
+ "think:think16k@5": 28,
407
+ "think:think16k@7": 4,
408
+ "think:think16k@8": 2,
409
+ "think:think16k@4": 19
410
+ }
411
+ }
412
+ ]
413
+ }
analysis/bayes_sc_2026-10-25/sc_both_c4.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,398 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.830 equiv 1920
32
+ R 24000: acc 0.925 equiv 4457
33
+ R 32000: acc 0.929 equiv 4607
34
+ R 48000: acc 0.938 equiv 5010
35
+ R 64000: acc 0.940 equiv 5188
36
+ R 96000: acc 0.941 equiv 5437
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.930 equiv 4666 (+6) | vs agree2 +0.0 [-0.5, +0.6] | R [35955, 32357] | {'think:think16k@0': 3544, 'stop@1': 3189, 'stop@2': 2976, 'think:think16k@2': 258, 'stop@4': 221, 'think:think16k@3': 204, 'stop@3': 187, 'think:think16k@4': 87}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.936 equiv 4829 (+170) | vs agree2 +0.6 [+0.1, +1.2] | R [30966, 29502] | {'think:think16k@0': 3712, 'stop@2': 3394, 'stop@1': 2468, 'stop@3': 325, 'think:think16k@3': 301, 'think:think16k@2': 296, 'stop@4': 169, 'stop@5': 75}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.931 equiv 4580 (-80) | vs agree2 +0.1 [-0.5, +0.7] | R [39094, 30978] | {'think:think16k@0': 3560, 'stop@1': 3242, 'stop@2': 3013, 'think:think16k@2': 476, 'stop@5': 180, 'stop@3': 124, 'stop@4': 78, 'think:think16k@3': 42}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.932 equiv 4589 (-70) | vs agree2 +0.2 [-0.3, +0.8] | R [28710, 30490] | {'stop@1': 3886, 'think:think16k@0': 3632, 'stop@2': 2327, 'think:think16k@2': 562, 'stop@5': 137, 'stop@3': 113, 'stop@4': 73, 'think:think16k@3': 43}
analysis/bayes_sc_2026-10-25/sc_both_c4.provenance.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "think16k=workdir/layer_probe/think16k_tier.json",
16
+ "q32b_think=workdir/bayes/q32b_think_tier.json@4",
17
+ "--out",
18
+ "workdir/bayes/sc_both_c4.json"
19
+ ],
20
+ "git": {
21
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
22
+ "branch": "HEAD",
23
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
24
+ "dirty": false,
25
+ "dirty_files": []
26
+ },
27
+ "gpus": {
28
+ "visible": "",
29
+ "devices": [
30
+ {
31
+ "index": "0",
32
+ "name": "NVIDIA RTX 5880 Ada Generation",
33
+ "memory": "49140 MiB",
34
+ "driver": "580.178.04",
35
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
36
+ },
37
+ {
38
+ "index": "1",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
43
+ }
44
+ ]
45
+ },
46
+ "packages": {
47
+ "python": "3.11.15",
48
+ "torch": "2.9.1",
49
+ "vllm": "0.15.1",
50
+ "transformers": "4.57.6",
51
+ "speculators": "0.4.0.dev49",
52
+ "peft": null,
53
+ "accelerate": "1.13.0",
54
+ "numpy": "2.2.6",
55
+ "huggingface_hub": "0.36.2",
56
+ "openai": "2.32.0",
57
+ "wandb": "0.25.1"
58
+ },
59
+ "env": {
60
+ "CUDA_VISIBLE_DEVICES": ""
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool": {
65
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
66
+ "bytes": 65136887,
67
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
68
+ },
69
+ "uhead": {
70
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
71
+ "bytes": 9369632,
72
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
73
+ },
74
+ "task_prior": {
75
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
76
+ "bytes": 27027,
77
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
78
+ },
79
+ "tier:think16k": {
80
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
81
+ "bytes": 26666,
82
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
83
+ },
84
+ "tier:q32b_think": {
85
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
86
+ "bytes": 26525,
87
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
88
+ }
89
+ },
90
+ "finished_at": "2026-10-25T11:35:35+04:00",
91
+ "wall_secs": 35.9,
92
+ "n_problems": 1361
93
+ }
analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.json ADDED
@@ -0,0 +1,433 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.9405767817781043,
4
+ "equiv": 10385.626671565025
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.9405767817781043,
9
+ "equiv": 10385.626671565025,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.14979794268919913,
21
+ -0.16256429096252756,
22
+ -0.1369374540778839
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.12876561351947097,
30
+ -0.14226671565025717,
31
+ -0.11544819985304923
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.09616091109478325,
39
+ -0.10911324393828067,
40
+ -0.08357825128581925
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.0852314474650992,
48
+ -0.09809652828802351,
49
+ -0.07274063188831742
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0775165319617928,
57
+ -0.09056070903747244,
58
+ -0.06493387215282881
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.07127112417340191,
66
+ -0.08440714548126378,
67
+ -0.05905584129316679
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.068791329904482,
75
+ -0.08174136664217486,
76
+ -0.05648190668626012
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.06447465099191771,
84
+ -0.07788850110213079,
85
+ -0.05143277002204261
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.06915870683321088,
93
+ -0.08238886847905952,
94
+ -0.05639235855988244
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.9405767817781043,
99
+ "equiv": 10385.626671565025,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9403930933137399,
108
+ "equiv": 10855.038243938281,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9340558412931668,
117
+ "equiv": 9174.927020573108,
118
+ "delta_vs_agree2": [
119
+ -0.006520940484937546,
120
+ -0.009184423218221896,
121
+ -0.004132990448199853
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.06484202792064658,
129
+ -0.07834542615723733,
130
+ -0.05198383541513593
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9603232916972814,
135
+ "equiv": 23519.07788390889,
136
+ "delta_vs_agree2": [
137
+ 0.019746509919177077,
138
+ 0.012766348273328435,
139
+ 0.026726671565025718
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9542615723732549,
144
+ "equiv": 13444.488317413665,
145
+ "delta_vs_agree2": [
146
+ 0.013684790595150624,
147
+ 0.010929463629684055,
148
+ 0.016623806024981633
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9540778839088905,
153
+ "equiv": 14017.90659441587,
154
+ "delta_vs_agree2": [
155
+ 0.013501102130786187,
156
+ 0.010745775165319618,
157
+ 0.0162587252020573
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.06456649522409992,
165
+ -0.07806989346069067,
166
+ -0.05152461425422483
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9603232916972814,
171
+ "equiv": 23519.07788390889,
172
+ "delta_vs_agree2": [
173
+ 0.019746509919177077,
174
+ 0.012766348273328435,
175
+ 0.026726671565025718
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9603232916972814,
180
+ "equiv": 24644.018736223366,
181
+ "delta_vs_agree2": [
182
+ 0.019746509919177077,
183
+ 0.012766348273328435,
184
+ 0.026726671565025718
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9572005878030859,
189
+ "equiv": 15651.51969140338,
190
+ "delta_vs_agree2": [
191
+ 0.016623806024981633,
192
+ 0.012950036737692874,
193
+ 0.020575404114621592
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.06447465099191771,
201
+ -0.07788850110213079,
202
+ -0.05143277002204261
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9603232916972814,
207
+ "equiv": 23519.07788390889,
208
+ "delta_vs_agree2": [
209
+ 0.019746509919177077,
210
+ 0.012766348273328435,
211
+ 0.026726671565025718
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9603232916972814,
216
+ "equiv": 24644.018736223366,
217
+ "delta_vs_agree2": [
218
+ 0.019746509919177077,
219
+ 0.012766348273328435,
220
+ 0.026726671565025718
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9603232916972814,
225
+ "equiv": 25768.95958853784,
226
+ "delta_vs_agree2": [
227
+ 0.019746509919177077,
228
+ 0.012766348273328435,
229
+ 0.026726671565025718
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9603232916972814,
234
+ "equiv": 21269.19617927994,
235
+ "delta_vs_agree2": [
236
+ 0.019746509919177077,
237
+ 0.012766348273328435,
238
+ 0.026726671565025718
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.830088170462895,
256
+ "equiv": 1919.7061260102867
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.924917340191036,
261
+ "equiv": 4457.099834680382
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.929142174871418,
266
+ "equiv": 4607.107448567231
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9381888317413667,
271
+ "equiv": 5009.742376928729
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9395205731080088,
276
+ "equiv": 5187.553829904482
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9407145481263777,
281
+ "equiv": 5436.882365907421
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9505418809698751,
288
+ "equiv": 11047.847483468038,
289
+ "R": [
290
+ 241797.66462509157,
291
+ 341523.97997940594
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.009965099191770757,
295
+ 0.003076781778104335,
296
+ 0.017266715650257163
297
+ ],
298
+ "actions": {
299
+ "stop@2": 1372,
300
+ "stop@3": 3489,
301
+ "think:think16k@1": 1117,
302
+ "think:q32b_think@4": 260,
303
+ "stop@4": 1058,
304
+ "think:think16k@2": 517,
305
+ "stop@5": 641,
306
+ "stop@6": 149,
307
+ "think:think16k@3": 185,
308
+ "think:think16k@6": 162,
309
+ "think:think16k@8": 238,
310
+ "think:think16k@5": 120,
311
+ "think:think16k@4": 288,
312
+ "think:q32b_think@5": 246,
313
+ "think:q32b_think@3": 68,
314
+ "think:q32b_think@7": 96,
315
+ "think:q32b_think@6": 61,
316
+ "think:think16k@7": 80,
317
+ "stop@8": 120,
318
+ "stop@7": 140,
319
+ "think:q32b_think@2": 4,
320
+ "think:q32b_think@8": 477
321
+ }
322
+ },
323
+ {
324
+ "seed": 1,
325
+ "acc": 0.9419085231447465,
326
+ "equiv": 10603.439860396767,
327
+ "R": [
328
+ 462593.9195051809,
329
+ 191455.9610167584
330
+ ],
331
+ "delta_vs_agree2": [
332
+ 0.001331741366642175,
333
+ -0.005419957751653196,
334
+ 0.008313051065393089
335
+ ],
336
+ "actions": {
337
+ "think:think16k@6": 165,
338
+ "stop@6": 229,
339
+ "think:think16k@4": 390,
340
+ "stop@3": 2419,
341
+ "think:think16k@2": 983,
342
+ "stop@5": 463,
343
+ "stop@2": 2773,
344
+ "think:think16k@1": 558,
345
+ "think:think16k@3": 245,
346
+ "think:think16k@8": 170,
347
+ "think:think16k@7": 143,
348
+ "think:think16k@5": 300,
349
+ "think:q32b_think@8": 285,
350
+ "stop@4": 878,
351
+ "stop@7": 54,
352
+ "think:q32b_think@6": 107,
353
+ "stop@8": 80,
354
+ "think:q32b_think@7": 49,
355
+ "think:q32b_think@5": 435,
356
+ "think:q32b_think@3": 135,
357
+ "think:q32b_think@4": 25,
358
+ "think:q32b_think@2": 2
359
+ }
360
+ },
361
+ {
362
+ "seed": 2,
363
+ "acc": 0.9421840558412932,
364
+ "equiv": 10225.022722263042,
365
+ "R": [
366
+ 271136.5869705112,
367
+ 325582.9249190125
368
+ ],
369
+ "delta_vs_agree2": [
370
+ 0.0016072740631888317,
371
+ -0.00528104335047759,
372
+ 0.008403747244673035
373
+ ],
374
+ "actions": {
375
+ "think:think16k@2": 765,
376
+ "think:think16k@5": 261,
377
+ "think:think16k@4": 267,
378
+ "stop@4": 1737,
379
+ "stop@3": 3694,
380
+ "think:think16k@3": 556,
381
+ "stop@2": 1422,
382
+ "stop@5": 506,
383
+ "think:q32b_think@6": 237,
384
+ "think:q32b_think@8": 423,
385
+ "think:think16k@7": 152,
386
+ "think:think16k@8": 152,
387
+ "think:think16k@6": 104,
388
+ "think:q32b_think@7": 167,
389
+ "stop@6": 151,
390
+ "think:q32b_think@5": 105,
391
+ "stop@7": 144,
392
+ "stop@8": 43,
393
+ "think:q32b_think@2": 2
394
+ }
395
+ },
396
+ {
397
+ "seed": 3,
398
+ "acc": 0.947556943423953,
399
+ "equiv": 10158.084175238795,
400
+ "R": [
401
+ 266358.29360577185,
402
+ 293112.2885868874
403
+ ],
404
+ "delta_vs_agree2": [
405
+ 0.0069801616458486405,
406
+ -0.00032145481263776637,
407
+ 0.014471206833210843
408
+ ],
409
+ "actions": {
410
+ "stop@2": 2500,
411
+ "stop@3": 2856,
412
+ "stop@6": 111,
413
+ "stop@5": 452,
414
+ "stop@4": 1185,
415
+ "think:think16k@2": 480,
416
+ "think:think16k@3": 434,
417
+ "think:think16k@4": 239,
418
+ "think:think16k@8": 64,
419
+ "think:q32b_think@7": 116,
420
+ "think:q32b_think@5": 234,
421
+ "think:q32b_think@6": 74,
422
+ "stop@8": 44,
423
+ "stop@7": 76,
424
+ "think:think16k@6": 398,
425
+ "think:think16k@5": 316,
426
+ "think:q32b_think@8": 466,
427
+ "think:think16k@7": 133,
428
+ "think:think16k@1": 707,
429
+ "think:q32b_think@4": 3
430
+ }
431
+ }
432
+ ]
433
+ }
analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,323 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 10386
3
+ agree2 acc 0.941 equiv 10386 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -9261) | vs agree2 -15.0 [-16.3, -13.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -8136) | vs agree2 -12.9 [-14.2, -11.5]
6
+ SC@3 acc 0.844 equiv 3375 ( -7011) | vs agree2 -9.6 [-10.9, -8.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -5886) | vs agree2 -8.5 [-9.8, -7.3]
8
+ SC@5 acc 0.863 equiv 5625 ( -4761) | vs agree2 -7.8 [-9.1, -6.5]
9
+ SC@6 acc 0.869 equiv 6750 ( -3636) | vs agree2 -7.1 [-8.4, -5.9]
10
+ SC@7 acc 0.872 equiv 7875 ( -2511) | vs agree2 -6.9 [-8.2, -5.6]
11
+ SC@8 acc 0.876 equiv 9000 ( -1386) | vs agree2 -6.4 [-7.8, -5.1]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -6057) | vs agree2 -6.9 [-8.2, -5.6]
13
+ AC(0.8)+think@2 acc 0.941 equiv 10386 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.940 equiv 10855 ( +469) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.934 equiv 9175 ( -1211) | vs agree2 -0.7 [-0.9, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( -4340) | vs agree2 -6.5 [-7.8, -5.2]
17
+ AC(0.9)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
18
+ AC(0.9)+think@3 acc 0.954 equiv 13444 ( +3059) | vs agree2 +1.4 [+1.1, +1.7]
19
+ AC(0.9)+think@4 acc 0.954 equiv 14018 ( +3632) | vs agree2 +1.4 [+1.1, +1.6]
20
+ AC(0.95) acc 0.876 equiv 6970 ( -3415) | vs agree2 -6.5 [-7.8, -5.2]
21
+ AC(0.95)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
22
+ AC(0.95)+think@3 acc 0.960 equiv 24644 (+14258) | vs agree2 +2.0 [+1.3, +2.7]
23
+ AC(0.95)+think@4 acc 0.957 equiv 15652 ( +5266) | vs agree2 +1.7 [+1.3, +2.1]
24
+ AC(0.99) acc 0.876 equiv 8168 ( -2217) | vs agree2 -6.4 [-7.8, -5.1]
25
+ AC(0.99)+think@2 acc 0.960 equiv 23519 (+13133) | vs agree2 +2.0 [+1.3, +2.7]
26
+ AC(0.99)+think@3 acc 0.960 equiv 24644 (+14258) | vs agree2 +2.0 [+1.3, +2.7]
27
+ AC(0.99)+think@4 acc 0.960 equiv 25769 (+15383) | vs agree2 +2.0 [+1.3, +2.7]
28
+ always think acc 0.960 equiv 21269 (+10884) | vs agree2 +2.0 [+1.3, +2.7]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.830 equiv 1920
32
+ R 24000: acc 0.925 equiv 4457
33
+ R 32000: acc 0.929 equiv 4607
34
+ R 48000: acc 0.938 equiv 5010
35
+ R 64000: acc 0.940 equiv 5188
36
+ R 96000: acc 0.941 equiv 5437
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.951 equiv 11048 (+662) | vs agree2 +1.0 [+0.3, +1.7] | R [241798, 341524] | {'stop@3': 3489, 'stop@2': 1372, 'think:think16k@1': 1117, 'stop@4': 1058, 'stop@5': 641, 'think:think16k@2': 517, 'think:q32b_think@8': 477, 'think:think16k@4': 288}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.942 equiv 10603 (+218) | vs agree2 +0.1 [-0.5, +0.8] | R [462594, 191456] | {'stop@2': 2773, 'stop@3': 2419, 'think:think16k@2': 983, 'stop@4': 878, 'think:think16k@1': 558, 'stop@5': 463, 'think:q32b_think@5': 435, 'think:think16k@4': 390}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.942 equiv 10225 (-161) | vs agree2 +0.2 [-0.5, +0.8] | R [271137, 325583] | {'stop@3': 3694, 'stop@4': 1737, 'stop@2': 1422, 'think:think16k@2': 765, 'think:think16k@3': 556, 'stop@5': 506, 'think:q32b_think@8': 423, 'think:think16k@4': 267}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.948 equiv 10158 (-228) | vs agree2 +0.7 [-0.0, +1.4] | R [266358, 293112] | {'stop@3': 2856, 'stop@2': 2500, 'stop@4': 1185, 'think:think16k@1': 707, 'think:think16k@2': 480, 'think:q32b_think@8': 466, 'stop@5': 452, 'think:think16k@3': 434}
analysis/bayes_sc_2026-10-25/sc_both_c4_ref32.provenance.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "q32b_think=workdir/bayes/q32b_think_tier.json@4",
16
+ "think16k=workdir/layer_probe/think16k_tier.json",
17
+ "--out",
18
+ "workdir/bayes/sc_both_c4_ref32.json"
19
+ ],
20
+ "git": {
21
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
22
+ "branch": "HEAD",
23
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
24
+ "dirty": false,
25
+ "dirty_files": []
26
+ },
27
+ "gpus": {
28
+ "visible": "",
29
+ "devices": [
30
+ {
31
+ "index": "0",
32
+ "name": "NVIDIA RTX 5880 Ada Generation",
33
+ "memory": "49140 MiB",
34
+ "driver": "580.178.04",
35
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
36
+ },
37
+ {
38
+ "index": "1",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
43
+ }
44
+ ]
45
+ },
46
+ "packages": {
47
+ "python": "3.11.15",
48
+ "torch": "2.9.1",
49
+ "vllm": "0.15.1",
50
+ "transformers": "4.57.6",
51
+ "speculators": "0.4.0.dev49",
52
+ "peft": null,
53
+ "accelerate": "1.13.0",
54
+ "numpy": "2.2.6",
55
+ "huggingface_hub": "0.36.2",
56
+ "openai": "2.32.0",
57
+ "wandb": "0.25.1"
58
+ },
59
+ "env": {
60
+ "CUDA_VISIBLE_DEVICES": ""
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool": {
65
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
66
+ "bytes": 65136887,
67
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
68
+ },
69
+ "uhead": {
70
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
71
+ "bytes": 9369632,
72
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
73
+ },
74
+ "task_prior": {
75
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
76
+ "bytes": 27027,
77
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
78
+ },
79
+ "tier:q32b_think": {
80
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
81
+ "bytes": 26525,
82
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
83
+ },
84
+ "tier:think16k": {
85
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
86
+ "bytes": 26666,
87
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
88
+ }
89
+ },
90
+ "finished_at": "2026-10-25T11:35:43+04:00",
91
+ "wall_secs": 44.3,
92
+ "n_problems": 1361
93
+ }
analysis/bayes_sc_2026-10-25/sc_both_ref32.json ADDED
@@ -0,0 +1,421 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.9405767817781043,
4
+ "equiv": 4283.8179463629685
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.9405767817781043,
9
+ "equiv": 4283.8179463629685,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.14979794268919913,
21
+ -0.16256429096252756,
22
+ -0.1369374540778839
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.12876561351947097,
30
+ -0.14226671565025717,
31
+ -0.11544819985304923
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.09616091109478325,
39
+ -0.10911324393828067,
40
+ -0.08357825128581925
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.0852314474650992,
48
+ -0.09809652828802351,
49
+ -0.07274063188831742
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0775165319617928,
57
+ -0.09056070903747244,
58
+ -0.06493387215282881
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.07127112417340191,
66
+ -0.08440714548126378,
67
+ -0.05905584129316679
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.068791329904482,
75
+ -0.08174136664217486,
76
+ -0.05648190668626012
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.06447465099191771,
84
+ -0.07788850110213079,
85
+ -0.05143277002204261
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.06915870683321088,
93
+ -0.08238886847905952,
94
+ -0.05639235855988244
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.9405767817781043,
99
+ "equiv": 4283.8179463629685,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9403930933137399,
108
+ "equiv": 4758.662803085966,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9340558412931668,
117
+ "equiv": 4695.218185157972,
118
+ "delta_vs_agree2": [
119
+ -0.006520940484937546,
120
+ -0.009184423218221896,
121
+ -0.004132990448199853
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.06484202792064658,
129
+ -0.07834542615723733,
130
+ -0.05198383541513593
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9603232916972814,
135
+ "equiv": 7567.180749448935,
136
+ "delta_vs_agree2": [
137
+ 0.019746509919177077,
138
+ 0.012766348273328435,
139
+ 0.026726671565025718
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9542615723732549,
144
+ "equiv": 5892.238997060985,
145
+ "delta_vs_agree2": [
146
+ 0.013684790595150624,
147
+ 0.010929463629684055,
148
+ 0.016623806024981633
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9540778839088905,
153
+ "equiv": 6471.090558412931,
154
+ "delta_vs_agree2": [
155
+ 0.013501102130786187,
156
+ 0.010745775165319618,
157
+ 0.0162587252020573
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.06456649522409992,
165
+ -0.07806989346069067,
166
+ -0.05152461425422483
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9603232916972814,
171
+ "equiv": 7567.180749448935,
172
+ "delta_vs_agree2": [
173
+ 0.019746509919177077,
174
+ 0.012766348273328435,
175
+ 0.026726671565025718
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9603232916972814,
180
+ "equiv": 8692.121601763409,
181
+ "delta_vs_agree2": [
182
+ 0.019746509919177077,
183
+ 0.012766348273328435,
184
+ 0.026726671565025718
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9572005878030859,
189
+ "equiv": 7287.702479794269,
190
+ "delta_vs_agree2": [
191
+ 0.016623806024981633,
192
+ 0.012950036737692874,
193
+ 0.020575404114621592
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.06447465099191771,
201
+ -0.07788850110213079,
202
+ -0.05143277002204261
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9603232916972814,
207
+ "equiv": 7567.180749448935,
208
+ "delta_vs_agree2": [
209
+ 0.019746509919177077,
210
+ 0.012766348273328435,
211
+ 0.026726671565025718
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9603232916972814,
216
+ "equiv": 8692.121601763409,
217
+ "delta_vs_agree2": [
218
+ 0.019746509919177077,
219
+ 0.012766348273328435,
220
+ 0.026726671565025718
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9603232916972814,
225
+ "equiv": 9817.062454077884,
226
+ "delta_vs_agree2": [
227
+ 0.019746509919177077,
228
+ 0.012766348273328435,
229
+ 0.026726671565025718
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9603232916972814,
234
+ "equiv": 5317.299044819985,
235
+ "delta_vs_agree2": [
236
+ 0.019746509919177077,
237
+ 0.012766348273328435,
238
+ 0.026726671565025718
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7909625275532697,
246
+ "equiv": 1127.0635378398238
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7910543717854519,
251
+ "equiv": 1127.5903196179281
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.934974283614989,
256
+ "equiv": 3727.7357182218957
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9424136664217487,
261
+ "equiv": 3918.981704628949
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9510470242468773,
266
+ "equiv": 4194.791182953711
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.955088170462895,
271
+ "equiv": 4463.583073108009
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9554555473916239,
276
+ "equiv": 4543.051276634827
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9565576781778105,
281
+ "equiv": 4844.335056943423
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9523328434974284,
288
+ "equiv": 4264.901377663482,
289
+ "R": [
290
+ 48694.151017527285,
291
+ 34234.24133567278
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.011756061719324026,
295
+ 0.006426800146950772,
296
+ 0.017820077149155022
297
+ ],
298
+ "actions": {
299
+ "stop@2": 3349,
300
+ "think:q32b_think@0": 3544,
301
+ "stop@3": 280,
302
+ "stop@1": 2549,
303
+ "stop@8": 33,
304
+ "think:q32b_think@2": 794,
305
+ "think:q32b_think@5": 2,
306
+ "stop@5": 104,
307
+ "think:q32b_think@3": 127,
308
+ "stop@6": 10,
309
+ "stop@7": 9,
310
+ "think:q32b_think@6": 4,
311
+ "think:q32b_think@4": 2,
312
+ "think:q32b_think@7": 2,
313
+ "think:q32b_think@8": 6,
314
+ "stop@4": 50,
315
+ "think:think16k@3": 12,
316
+ "think:think16k@4": 11
317
+ }
318
+ },
319
+ {
320
+ "seed": 1,
321
+ "acc": 0.9499448934606907,
322
+ "equiv": 4415.412389786921,
323
+ "R": [
324
+ 30965.546229776515,
325
+ 42770.39453717809
326
+ ],
327
+ "delta_vs_agree2": [
328
+ 0.009368111682586333,
329
+ 0.0031674779573842766,
330
+ 0.016118662747979427
331
+ ],
332
+ "actions": {
333
+ "stop@3": 333,
334
+ "think:think16k@3": 60,
335
+ "think:think16k@2": 214,
336
+ "stop@2": 3229,
337
+ "stop@4": 41,
338
+ "stop@1": 2468,
339
+ "think:q32b_think@0": 3712,
340
+ "think:q32b_think@3": 128,
341
+ "think:q32b_think@2": 604,
342
+ "stop@6": 19,
343
+ "stop@5": 43,
344
+ "stop@8": 11,
345
+ "think:q32b_think@5": 2,
346
+ "stop@7": 8,
347
+ "think:think16k@7": 8,
348
+ "think:think16k@8": 5,
349
+ "think:q32b_think@4": 1,
350
+ "think:q32b_think@7": 2
351
+ }
352
+ },
353
+ {
354
+ "seed": 2,
355
+ "acc": 0.9501285819250551,
356
+ "equiv": 4246.60239713446,
357
+ "R": [
358
+ 48680.35922315368,
359
+ 29264.79414538848
360
+ ],
361
+ "delta_vs_agree2": [
362
+ 0.009551800146950772,
363
+ 0.0033063923585598823,
364
+ 0.016072740631888318
365
+ ],
366
+ "actions": {
367
+ "think:q32b_think@0": 3560,
368
+ "stop@2": 3160,
369
+ "stop@1": 2682,
370
+ "stop@8": 25,
371
+ "stop@6": 31,
372
+ "think:q32b_think@2": 731,
373
+ "think:q32b_think@5": 8,
374
+ "stop@5": 176,
375
+ "stop@3": 301,
376
+ "think:q32b_think@7": 11,
377
+ "stop@4": 63,
378
+ "think:q32b_think@4": 36,
379
+ "think:q32b_think@6": 5,
380
+ "stop@7": 15,
381
+ "think:q32b_think@8": 5,
382
+ "think:q32b_think@3": 59,
383
+ "think:think16k@5": 8,
384
+ "think:think16k@4": 7,
385
+ "think:think16k@3": 1,
386
+ "think:think16k@6": 3,
387
+ "think:think16k@8": 1
388
+ }
389
+ },
390
+ {
391
+ "seed": 3,
392
+ "acc": 0.9535268185157972,
393
+ "equiv": 4256.303361498898,
394
+ "R": [
395
+ 37481.25107046291,
396
+ 35075.93846705908
397
+ ],
398
+ "delta_vs_agree2": [
399
+ 0.012950036737692874,
400
+ 0.0073452424687729615,
401
+ 0.01901405216752387
402
+ ],
403
+ "actions": {
404
+ "stop@1": 2673,
405
+ "stop@2": 3248,
406
+ "think:q32b_think@2": 874,
407
+ "think:q32b_think@0": 3632,
408
+ "stop@4": 49,
409
+ "stop@5": 60,
410
+ "stop@8": 17,
411
+ "stop@7": 6,
412
+ "stop@6": 5,
413
+ "stop@3": 208,
414
+ "think:q32b_think@3": 108,
415
+ "think:think16k@8": 2,
416
+ "think:q32b_think@4": 5,
417
+ "think:q32b_think@7": 1
418
+ }
419
+ }
420
+ ]
421
+ }
analysis/bayes_sc_2026-10-25/sc_both_ref32.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,400 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 4284
3
+ agree2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3159) | vs agree2 -15.0 [-16.3, -13.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2034) | vs agree2 -12.9 [-14.2, -11.5]
6
+ SC@3 acc 0.844 equiv 3375 ( -909) | vs agree2 -9.6 [-10.9, -8.4]
7
+ SC@4 acc 0.855 equiv 4500 ( +216) | vs agree2 -8.5 [-9.8, -7.3]
8
+ SC@5 acc 0.863 equiv 5625 ( +1341) | vs agree2 -7.8 [-9.1, -6.5]
9
+ SC@6 acc 0.869 equiv 6750 ( +2466) | vs agree2 -7.1 [-8.4, -5.9]
10
+ SC@7 acc 0.872 equiv 7875 ( +3591) | vs agree2 -6.9 [-8.2, -5.6]
11
+ SC@8 acc 0.876 equiv 9000 ( +4716) | vs agree2 -6.4 [-7.8, -5.1]
12
+ AC(0.8) acc 0.871 equiv 4329 ( +45) | vs agree2 -6.9 [-8.2, -5.6]
13
+ AC(0.8)+think@2 acc 0.941 equiv 4284 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.940 equiv 4759 ( +475) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.934 equiv 4695 ( +411) | vs agree2 -0.7 [-0.9, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1762) | vs agree2 -6.5 [-7.8, -5.2]
17
+ AC(0.9)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
18
+ AC(0.9)+think@3 acc 0.954 equiv 5892 ( +1608) | vs agree2 +1.4 [+1.1, +1.7]
19
+ AC(0.9)+think@4 acc 0.954 equiv 6471 ( +2187) | vs agree2 +1.4 [+1.1, +1.6]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2687) | vs agree2 -6.5 [-7.8, -5.2]
21
+ AC(0.95)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
22
+ AC(0.95)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
23
+ AC(0.95)+think@4 acc 0.957 equiv 7288 ( +3004) | vs agree2 +1.7 [+1.3, +2.1]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3885) | vs agree2 -6.4 [-7.8, -5.1]
25
+ AC(0.99)+think@2 acc 0.960 equiv 7567 ( +3283) | vs agree2 +2.0 [+1.3, +2.7]
26
+ AC(0.99)+think@3 acc 0.960 equiv 8692 ( +4408) | vs agree2 +2.0 [+1.3, +2.7]
27
+ AC(0.99)+think@4 acc 0.960 equiv 9817 ( +5533) | vs agree2 +2.0 [+1.3, +2.7]
28
+ always think acc 0.960 equiv 5317 ( +1033) | vs agree2 +2.0 [+1.3, +2.7]
29
+ R 4000: acc 0.791 equiv 1127
30
+ R 8000: acc 0.791 equiv 1128
31
+ R 16000: acc 0.935 equiv 3728
32
+ R 24000: acc 0.942 equiv 3919
33
+ R 32000: acc 0.951 equiv 4195
34
+ R 48000: acc 0.955 equiv 4464
35
+ R 64000: acc 0.955 equiv 4543
36
+ R 96000: acc 0.957 equiv 4844
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.952 equiv 4265 (-19) | vs agree2 +1.2 [+0.6, +1.8] | R [48694, 34234] | {'think:q32b_think@0': 3544, 'stop@2': 3349, 'stop@1': 2549, 'think:q32b_think@2': 794, 'stop@3': 280, 'think:q32b_think@3': 127, 'stop@5': 104, 'stop@4': 50}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.950 equiv 4415 (+132) | vs agree2 +0.9 [+0.3, +1.6] | R [30966, 42770] | {'think:q32b_think@0': 3712, 'stop@2': 3229, 'stop@1': 2468, 'think:q32b_think@2': 604, 'stop@3': 333, 'think:think16k@2': 214, 'think:q32b_think@3': 128, 'think:think16k@3': 60}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.950 equiv 4247 (-37) | vs agree2 +1.0 [+0.3, +1.6] | R [48680, 29265] | {'think:q32b_think@0': 3560, 'stop@2': 3160, 'stop@1': 2682, 'think:q32b_think@2': 731, 'stop@3': 301, 'stop@5': 176, 'stop@4': 63, 'think:q32b_think@3': 59}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4256 (-28) | vs agree2 +1.3 [+0.7, +1.9] | R [37481, 35076] | {'think:q32b_think@0': 3632, 'stop@2': 3248, 'stop@1': 2673, 'think:q32b_think@2': 874, 'stop@3': 208, 'think:q32b_think@3': 108, 'stop@5': 60, 'stop@4': 49}
analysis/bayes_sc_2026-10-25/sc_both_ref32.provenance.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "q32b_think=workdir/bayes/q32b_think_tier.json",
16
+ "think16k=workdir/layer_probe/think16k_tier.json",
17
+ "--out",
18
+ "workdir/bayes/sc_both_ref32.json"
19
+ ],
20
+ "git": {
21
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
22
+ "branch": "HEAD",
23
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
24
+ "dirty": false,
25
+ "dirty_files": []
26
+ },
27
+ "gpus": {
28
+ "visible": "",
29
+ "devices": [
30
+ {
31
+ "index": "0",
32
+ "name": "NVIDIA RTX 5880 Ada Generation",
33
+ "memory": "49140 MiB",
34
+ "driver": "580.178.04",
35
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
36
+ },
37
+ {
38
+ "index": "1",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
43
+ }
44
+ ]
45
+ },
46
+ "packages": {
47
+ "python": "3.11.15",
48
+ "torch": "2.9.1",
49
+ "vllm": "0.15.1",
50
+ "transformers": "4.57.6",
51
+ "speculators": "0.4.0.dev49",
52
+ "peft": null,
53
+ "accelerate": "1.13.0",
54
+ "numpy": "2.2.6",
55
+ "huggingface_hub": "0.36.2",
56
+ "openai": "2.32.0",
57
+ "wandb": "0.25.1"
58
+ },
59
+ "env": {
60
+ "CUDA_VISIBLE_DEVICES": ""
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool": {
65
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
66
+ "bytes": 65136887,
67
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
68
+ },
69
+ "uhead": {
70
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
71
+ "bytes": 9369632,
72
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
73
+ },
74
+ "task_prior": {
75
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
76
+ "bytes": 27027,
77
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
78
+ },
79
+ "tier:q32b_think": {
80
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
81
+ "bytes": 26525,
82
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
83
+ },
84
+ "tier:think16k": {
85
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
86
+ "bytes": 26666,
87
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
88
+ }
89
+ },
90
+ "finished_at": "2026-10-25T11:35:35+04:00",
91
+ "wall_secs": 35.6,
92
+ "n_problems": 1361
93
+ }
analysis/bayes_sc_2026-10-25/sc_notask.json ADDED
@@ -0,0 +1,643 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.7911462160176341,
256
+ "equiv": 1128.7680933137399
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.8979151359294636,
261
+ "equiv": 3568.8109294636297
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9099008082292432,
266
+ "equiv": 3948.2372428361496
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.93584680382072,
271
+ "equiv": 5266.863381704629
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9349283614988979,
276
+ "equiv": 5448.085121234386
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9343313739897134,
281
+ "equiv": 5971.540898236591
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9336425422483468,
288
+ "equiv": 5144.84041146216,
289
+ "R": [
290
+ 47383.73625367988,
291
+ 37296.76570797434
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.003719691403379868,
295
+ 0.0011469048493754593,
296
+ 0.006383174136664217
297
+ ],
298
+ "actions": {
299
+ "stop@2": 5394,
300
+ "think:think16k@2": 1611,
301
+ "stop@3": 738,
302
+ "stop@5": 55,
303
+ "think:think16k@3": 1151,
304
+ "stop@6": 20,
305
+ "stop@8": 12,
306
+ "think:think16k@4": 27,
307
+ "think:think16k@7": 6,
308
+ "think:think16k@6": 15,
309
+ "stop@7": 10,
310
+ "think:think16k@1": 1,
311
+ "stop@4": 46,
312
+ "think:think16k@5": 4,
313
+ "stop@1": 1797,
314
+ "think:think16k@8": 1
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9232182218956649,
320
+ "equiv": 4573.556925055106,
321
+ "R": [
322
+ 33915.486824625026,
323
+ 48872.51246466318
324
+ ],
325
+ "delta_vs_agree2": [
326
+ -0.006704628949301984,
327
+ -0.009276267450404115,
328
+ -0.004361452975753129
329
+ ],
330
+ "actions": {
331
+ "stop@2": 4673,
332
+ "think:think16k@2": 1188,
333
+ "stop@3": 39,
334
+ "stop@1": 3633,
335
+ "stop@5": 193,
336
+ "stop@7": 8,
337
+ "stop@4": 4,
338
+ "think:think16k@3": 1017,
339
+ "think:think16k@7": 1,
340
+ "stop@6": 25,
341
+ "think:think16k@4": 82,
342
+ "think:think16k@5": 19,
343
+ "think:think16k@6": 2,
344
+ "stop@8": 4
345
+ }
346
+ },
347
+ {
348
+ "seed": 2,
349
+ "acc": 0.9264327700220426,
350
+ "equiv": 4605.107613886848,
351
+ "R": [
352
+ 43430.53055362532,
353
+ 36699.87011130473
354
+ ],
355
+ "delta_vs_agree2": [
356
+ -0.0034900808229243203,
357
+ -0.006566862601028655,
358
+ -0.0005051432770022043
359
+ ],
360
+ "actions": {
361
+ "stop@4": 166,
362
+ "think:think16k@2": 2074,
363
+ "stop@3": 352,
364
+ "stop@2": 4156,
365
+ "stop@1": 3636,
366
+ "think:think16k@3": 326,
367
+ "think:think16k@4": 76,
368
+ "stop@5": 58,
369
+ "stop@7": 3,
370
+ "stop@6": 16,
371
+ "think:think16k@7": 6,
372
+ "stop@8": 15,
373
+ "think:think16k@6": 1,
374
+ "think:think16k@5": 2,
375
+ "think:think16k@1": 1
376
+ }
377
+ },
378
+ {
379
+ "seed": 3,
380
+ "acc": 0.9263409257898604,
381
+ "equiv": 4728.292652461426,
382
+ "R": [
383
+ 40476.16479025277,
384
+ 36640.44297454738
385
+ ],
386
+ "delta_vs_agree2": [
387
+ -0.0035819250551065392,
388
+ -0.006291329904481999,
389
+ -0.0008713721528288086
390
+ ],
391
+ "actions": {
392
+ "stop@1": 3651,
393
+ "stop@2": 4034,
394
+ "think:think16k@3": 657,
395
+ "stop@4": 162,
396
+ "think:think16k@2": 1707,
397
+ "stop@6": 6,
398
+ "stop@7": 3,
399
+ "think:think16k@4": 160,
400
+ "stop@5": 47,
401
+ "think:think16k@7": 10,
402
+ "stop@8": 5,
403
+ "think:think16k@6": 1,
404
+ "stop@3": 440,
405
+ "think:think16k@1": 3,
406
+ "think:think16k@5": 2
407
+ }
408
+ }
409
+ ],
410
+ "transfer": {
411
+ "R": 40313.039892823894,
412
+ "m500": {
413
+ "acc": 0.9296875,
414
+ "equiv": 4796.1615625,
415
+ "delta_vs_agree2": [
416
+ 0.00546875,
417
+ -0.0015625,
418
+ 0.0125
419
+ ],
420
+ "baselines": {
421
+ "agree2": [
422
+ 0.92421875,
423
+ 4634.017890625
424
+ ],
425
+ "SC@1": [
426
+ 0.78828125,
427
+ 1103.76328125
428
+ ],
429
+ "SC@2": [
430
+ 0.81484375,
431
+ 2207.5265625
432
+ ],
433
+ "SC@3": [
434
+ 0.84609375,
435
+ 3311.28984375
436
+ ],
437
+ "SC@4": [
438
+ 0.86171875,
439
+ 4415.053125
440
+ ],
441
+ "SC@5": [
442
+ 0.871875,
443
+ 5518.81640625
444
+ ],
445
+ "SC@6": [
446
+ 0.8765625,
447
+ 6622.5796875
448
+ ],
449
+ "SC@7": [
450
+ 0.88984375,
451
+ 7726.34296875
452
+ ],
453
+ "SC@8": [
454
+ 0.89375,
455
+ 8830.10625
456
+ ],
457
+ "AC(0.8)": [
458
+ 0.88828125,
459
+ 4099.459375
460
+ ],
461
+ "AC(0.8)+think@2": [
462
+ 0.92421875,
463
+ 4634.017890625
464
+ ],
465
+ "AC(0.8)+think@3": [
466
+ 0.92421875,
467
+ 5108.860078125001
468
+ ],
469
+ "AC(0.8)+think@4": [
470
+ 0.917578125,
471
+ 4986.9370312500005
472
+ ],
473
+ "AC(0.9)": [
474
+ 0.89375,
475
+ 6009.02890625
476
+ ],
477
+ "AC(0.9)+think@2": [
478
+ 0.9375,
479
+ 8439.737187499999
480
+ ],
481
+ "AC(0.9)+think@3": [
482
+ 0.9375,
483
+ 6302.046171874999
484
+ ],
485
+ "AC(0.9)+think@4": [
486
+ 0.9375,
487
+ 6888.578984375001
488
+ ],
489
+ "AC(0.95)": [
490
+ 0.89375,
491
+ 6905.77578125
492
+ ],
493
+ "AC(0.95)+think@2": [
494
+ 0.9375,
495
+ 8439.737187499999
496
+ ],
497
+ "AC(0.95)+think@3": [
498
+ 0.9375,
499
+ 9543.50046875
500
+ ],
501
+ "AC(0.95)+think@4": [
502
+ 0.941796875,
503
+ 7697.955703125001
504
+ ],
505
+ "AC(0.99)": [
506
+ 0.89375,
507
+ 8028.82734375
508
+ ],
509
+ "AC(0.99)+think@2": [
510
+ 0.9375,
511
+ 8439.737187499999
512
+ ],
513
+ "AC(0.99)+think@3": [
514
+ 0.9375,
515
+ 9543.50046875
516
+ ],
517
+ "AC(0.99)+think@4": [
518
+ 0.9375,
519
+ 10647.26375
520
+ ],
521
+ "always think": [
522
+ 0.9375,
523
+ 6232.210625000001
524
+ ]
525
+ }
526
+ },
527
+ "aime": {
528
+ "acc": 0.5302083333333333,
529
+ "equiv": 17081.812499999996,
530
+ "delta_vs_agree2": [
531
+ -0.023958333333333335,
532
+ -0.047942708333333334,
533
+ -0.004166666666666667
534
+ ],
535
+ "baselines": {
536
+ "agree2": [
537
+ 0.5541666666666667,
538
+ 15601.996666666668
539
+ ],
540
+ "SC@1": [
541
+ 0.22708333333333333,
542
+ 2533.9333333333334
543
+ ],
544
+ "SC@2": [
545
+ 0.2625,
546
+ 5067.866666666667
547
+ ],
548
+ "SC@3": [
549
+ 0.28541666666666665,
550
+ 7601.8
551
+ ],
552
+ "SC@4": [
553
+ 0.30625,
554
+ 10135.733333333334
555
+ ],
556
+ "SC@5": [
557
+ 0.3125,
558
+ 12669.666666666666
559
+ ],
560
+ "SC@6": [
561
+ 0.325,
562
+ 15203.6
563
+ ],
564
+ "SC@7": [
565
+ 0.33958333333333335,
566
+ 17737.533333333333
567
+ ],
568
+ "SC@8": [
569
+ 0.35,
570
+ 20271.466666666667
571
+ ],
572
+ "AC(0.8)": [
573
+ 0.3458333333333333,
574
+ 16049.204166666666
575
+ ],
576
+ "AC(0.8)+think@2": [
577
+ 0.5541666666666667,
578
+ 15601.996666666668
579
+ ],
580
+ "AC(0.8)+think@3": [
581
+ 0.5541666666666667,
582
+ 17670.754999999997
583
+ ],
584
+ "AC(0.8)+think@4": [
585
+ 0.5354166666666667,
586
+ 18834.57125
587
+ ],
588
+ "AC(0.9)": [
589
+ 0.35,
590
+ 18722.379166666666
591
+ ],
592
+ "AC(0.9)+think@2": [
593
+ 0.5916666666666667,
594
+ 17967.561666666665
595
+ ],
596
+ "AC(0.9)+think@3": [
597
+ 0.5729166666666666,
598
+ 19178.001458333332
599
+ ],
600
+ "AC(0.9)+think@4": [
601
+ 0.5729166666666666,
602
+ 21455.782708333332
603
+ ],
604
+ "AC(0.95)": [
605
+ 0.35,
606
+ 19571.391666666666
607
+ ],
608
+ "AC(0.95)+think@2": [
609
+ 0.5916666666666667,
610
+ 17967.561666666665
611
+ ],
612
+ "AC(0.95)+think@3": [
613
+ 0.5916666666666667,
614
+ 20501.495
615
+ ],
616
+ "AC(0.95)+think@4": [
617
+ 0.5822916666666667,
618
+ 22189.036666666667
619
+ ],
620
+ "AC(0.99)": [
621
+ 0.35,
622
+ 20084.541666666668
623
+ ],
624
+ "AC(0.99)+think@2": [
625
+ 0.5916666666666667,
626
+ 17967.561666666665
627
+ ],
628
+ "AC(0.99)+think@3": [
629
+ 0.5916666666666667,
630
+ 20501.495
631
+ ],
632
+ "AC(0.99)+think@4": [
633
+ 0.5916666666666667,
634
+ 23035.428333333333
635
+ ],
636
+ "always think": [
637
+ 0.5916666666666667,
638
+ 12899.695
639
+ ]
640
+ }
641
+ }
642
+ }
643
+ }
analysis/bayes_sc_2026-10-25/sc_notask.log ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:32:38,872 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.791 equiv 1129
32
+ R 24000: acc 0.898 equiv 3569
33
+ R 32000: acc 0.910 equiv 3948
34
+ R 48000: acc 0.936 equiv 5267
35
+ R 64000: acc 0.935 equiv 5448
36
+ R 96000: acc 0.934 equiv 5972
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.934 equiv 5145 (+485) | vs agree2 +0.4 [+0.1, +0.6] | R [47384, 37297] | {'stop@2': 5394, 'stop@1': 1797, 'think:think16k@2': 1611, 'think:think16k@3': 1151, 'stop@3': 738, 'stop@5': 55, 'stop@4': 46, 'think:think16k@4': 27}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.923 equiv 4574 (-86) | vs agree2 -0.7 [-0.9, -0.4] | R [33915, 48873] | {'stop@2': 4673, 'stop@1': 3633, 'think:think16k@2': 1188, 'think:think16k@3': 1017, 'stop@5': 193, 'think:think16k@4': 82, 'stop@3': 39, 'stop@6': 25}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.926 equiv 4605 (-55) | vs agree2 -0.3 [-0.7, -0.1] | R [43431, 36700] | {'stop@2': 4156, 'stop@1': 3636, 'think:think16k@2': 2074, 'stop@3': 352, 'think:think16k@3': 326, 'stop@4': 166, 'think:think16k@4': 76, 'stop@5': 58}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.926 equiv 4728 (+69) | vs agree2 -0.4 [-0.6, -0.1] | R [40476, 36640] | {'stop@2': 4034, 'stop@1': 3651, 'think:think16k@2': 1707, 'think:think16k@3': 657, 'stop@3': 440, 'stop@4': 162, 'think:think16k@4': 160, 'stop@5': 47}
41
+ 2026-10-25 11:33:25,517 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '\\alpha^{85}-\\alpha^{-85}' vs '2\\cos(85\\theta)'
42
+ 2026-10-25 11:33:35,753 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '-2'
43
+ 2026-10-25 11:33:45,788 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '2\\cos(85\\theta)'
44
+ 2026-10-25 11:33:55,824 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '123'
45
+ 2026-10-25 11:34:05,860 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '\\alpha^{85}-\\alpha^{-85}'
46
+ 2026-10-25 11:34:16,171 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '2' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
47
+ 2026-10-25 11:34:26,545 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'L_{85}' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
48
+ transfer m500: adaptive-SC Bayes acc 0.930 equiv 4796 | agree2 0.924 at 4634 | vs agree2 +0.5 [-0.2, +1.2] | always think 0.938 at 6232
49
+ transfer aime: adaptive-SC Bayes acc 0.530 equiv 17082 | agree2 0.554 at 15602 | vs agree2 -2.4 [-4.8, -0.4] | always think 0.592 at 12900
analysis/bayes_sc_2026-10-25/sc_notask.provenance.json ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:32:37+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--tier",
13
+ "think16k=workdir/layer_probe/think16k_tier.json",
14
+ "--tbins",
15
+ "1",
16
+ "--transfer",
17
+ "m500=workdir/selector/pool_m500_k8_chat,workdir/selector/pool_m500_k8_chat/uhead_selector.json,workdir/bayes/m500_tier.json,workdir/bayes/m500_task_prior.json",
18
+ "aime=workdir/selector/pool_aime_k8_chat,workdir/selector/pool_aime_k8_chat/uhead_selector.json,workdir/bayes/aime_tier.json,workdir/bayes/aime_task_prior.json",
19
+ "--out",
20
+ "workdir/bayes/sc_notask.json"
21
+ ],
22
+ "git": {
23
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
24
+ "branch": "HEAD",
25
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
26
+ "dirty": false,
27
+ "dirty_files": []
28
+ },
29
+ "gpus": {
30
+ "visible": "",
31
+ "devices": [
32
+ {
33
+ "index": "0",
34
+ "name": "NVIDIA RTX 5880 Ada Generation",
35
+ "memory": "49140 MiB",
36
+ "driver": "580.178.04",
37
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
38
+ },
39
+ {
40
+ "index": "1",
41
+ "name": "NVIDIA RTX 5880 Ada Generation",
42
+ "memory": "49140 MiB",
43
+ "driver": "580.178.04",
44
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
45
+ }
46
+ ]
47
+ },
48
+ "packages": {
49
+ "python": "3.11.15",
50
+ "torch": "2.9.1",
51
+ "vllm": "0.15.1",
52
+ "transformers": "4.57.6",
53
+ "speculators": "0.4.0.dev49",
54
+ "peft": null,
55
+ "accelerate": "1.13.0",
56
+ "numpy": "2.2.6",
57
+ "huggingface_hub": "0.36.2",
58
+ "openai": "2.32.0",
59
+ "wandb": "0.25.1"
60
+ },
61
+ "env": {
62
+ "CUDA_VISIBLE_DEVICES": ""
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool": {
67
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
68
+ "bytes": 65136887,
69
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
70
+ },
71
+ "uhead": {
72
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
73
+ "bytes": 9369632,
74
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
75
+ },
76
+ "tier:think16k": {
77
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
78
+ "bytes": 26666,
79
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
80
+ }
81
+ },
82
+ "finished_at": "2026-10-25T11:34:31+04:00",
83
+ "wall_secs": 114.6,
84
+ "n_problems": 1361
85
+ }
analysis/bayes_sc_2026-10-25/sc_task.json ADDED
@@ -0,0 +1,646 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.830088170462895,
256
+ "equiv": 1919.7061260102867
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.924917340191036,
261
+ "equiv": 4457.099834680382
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.929142174871418,
266
+ "equiv": 4607.107448567231
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9381888317413667,
271
+ "equiv": 5009.742376928729
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9395205731080088,
276
+ "equiv": 5187.553829904482
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9407145481263777,
281
+ "equiv": 5400.68963997061
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9302902277736959,
288
+ "equiv": 4666.07443056576,
289
+ "R": [
290
+ 35955.38014708592,
291
+ 32356.730332100717
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.0003673769287288758,
295
+ -0.005098502939015429,
296
+ 0.00629247795738427
297
+ ],
298
+ "actions": {
299
+ "stop@2": 2976,
300
+ "stop@1": 3189,
301
+ "think:think16k@0": 3544,
302
+ "stop@3": 187,
303
+ "stop@8": 31,
304
+ "stop@6": 42,
305
+ "think:think16k@2": 258,
306
+ "stop@5": 76,
307
+ "think:think16k@4": 87,
308
+ "think:think16k@3": 204,
309
+ "stop@4": 221,
310
+ "think:think16k@7": 3,
311
+ "think:think16k@6": 8,
312
+ "think:think16k@8": 16,
313
+ "stop@7": 7,
314
+ "think:think16k@5": 39
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9359845701689934,
320
+ "equiv": 4829.496840558412,
321
+ "R": [
322
+ 30965.546229776515,
323
+ 29502.30501310991
324
+ ],
325
+ "delta_vs_agree2": [
326
+ 0.006061719324026451,
327
+ 0.0009184423218221896,
328
+ 0.011893828067597355
329
+ ],
330
+ "actions": {
331
+ "stop@3": 325,
332
+ "think:think16k@3": 301,
333
+ "think:think16k@2": 296,
334
+ "stop@2": 3394,
335
+ "stop@4": 169,
336
+ "stop@1": 2468,
337
+ "think:think16k@0": 3712,
338
+ "stop@6": 22,
339
+ "stop@5": 75,
340
+ "think:think16k@5": 11,
341
+ "think:think16k@4": 72,
342
+ "think:think16k@6": 4,
343
+ "think:think16k@7": 1,
344
+ "stop@8": 14,
345
+ "stop@7": 10,
346
+ "think:think16k@8": 14
347
+ }
348
+ },
349
+ {
350
+ "seed": 2,
351
+ "acc": 0.9311168258633358,
352
+ "equiv": 4579.799035635562,
353
+ "R": [
354
+ 39094.276600707366,
355
+ 30978.196393888564
356
+ ],
357
+ "delta_vs_agree2": [
358
+ 0.0011939750183688465,
359
+ -0.004592211609110948,
360
+ 0.007209772226304188
361
+ ],
362
+ "actions": {
363
+ "think:think16k@0": 3560,
364
+ "stop@2": 3013,
365
+ "stop@1": 3242,
366
+ "stop@8": 28,
367
+ "stop@6": 34,
368
+ "stop@5": 180,
369
+ "stop@7": 28,
370
+ "think:think16k@5": 21,
371
+ "stop@3": 124,
372
+ "think:think16k@3": 42,
373
+ "think:think16k@7": 7,
374
+ "think:think16k@6": 18,
375
+ "think:think16k@2": 476,
376
+ "stop@4": 78,
377
+ "think:think16k@4": 34,
378
+ "think:think16k@8": 3
379
+ }
380
+ },
381
+ {
382
+ "seed": 3,
383
+ "acc": 0.9323108008817046,
384
+ "equiv": 4589.129151359295,
385
+ "R": [
386
+ 28710.255862698214,
387
+ 30490.45634869312
388
+ ],
389
+ "delta_vs_agree2": [
390
+ 0.002387950036737693,
391
+ -0.0031227038941954444,
392
+ 0.008495591476855253
393
+ ],
394
+ "actions": {
395
+ "stop@1": 3886,
396
+ "stop@2": 2327,
397
+ "think:think16k@2": 562,
398
+ "think:think16k@0": 3632,
399
+ "stop@3": 113,
400
+ "stop@4": 73,
401
+ "stop@5": 137,
402
+ "think:think16k@3": 43,
403
+ "stop@8": 27,
404
+ "stop@7": 11,
405
+ "stop@6": 24,
406
+ "think:think16k@5": 28,
407
+ "think:think16k@7": 4,
408
+ "think:think16k@8": 2,
409
+ "think:think16k@4": 19
410
+ }
411
+ }
412
+ ],
413
+ "transfer": {
414
+ "R": 34628.45902795292,
415
+ "m500": {
416
+ "acc": 0.9140625,
417
+ "equiv": 4676.85703125,
418
+ "delta_vs_agree2": [
419
+ -0.01015625,
420
+ -0.033203125,
421
+ 0.0109375
422
+ ],
423
+ "baselines": {
424
+ "agree2": [
425
+ 0.92421875,
426
+ 4634.017890625
427
+ ],
428
+ "SC@1": [
429
+ 0.78828125,
430
+ 1103.76328125
431
+ ],
432
+ "SC@2": [
433
+ 0.81484375,
434
+ 2207.5265625
435
+ ],
436
+ "SC@3": [
437
+ 0.84609375,
438
+ 3311.28984375
439
+ ],
440
+ "SC@4": [
441
+ 0.86171875,
442
+ 4415.053125
443
+ ],
444
+ "SC@5": [
445
+ 0.871875,
446
+ 5518.81640625
447
+ ],
448
+ "SC@6": [
449
+ 0.8765625,
450
+ 6622.5796875
451
+ ],
452
+ "SC@7": [
453
+ 0.88984375,
454
+ 7726.34296875
455
+ ],
456
+ "SC@8": [
457
+ 0.89375,
458
+ 8830.10625
459
+ ],
460
+ "AC(0.8)": [
461
+ 0.88828125,
462
+ 4099.459375
463
+ ],
464
+ "AC(0.8)+think@2": [
465
+ 0.92421875,
466
+ 4634.017890625
467
+ ],
468
+ "AC(0.8)+think@3": [
469
+ 0.92421875,
470
+ 5108.860078125001
471
+ ],
472
+ "AC(0.8)+think@4": [
473
+ 0.917578125,
474
+ 4986.9370312500005
475
+ ],
476
+ "AC(0.9)": [
477
+ 0.89375,
478
+ 6009.02890625
479
+ ],
480
+ "AC(0.9)+think@2": [
481
+ 0.9375,
482
+ 8439.737187499999
483
+ ],
484
+ "AC(0.9)+think@3": [
485
+ 0.9375,
486
+ 6302.046171874999
487
+ ],
488
+ "AC(0.9)+think@4": [
489
+ 0.9375,
490
+ 6888.578984375001
491
+ ],
492
+ "AC(0.95)": [
493
+ 0.89375,
494
+ 6905.77578125
495
+ ],
496
+ "AC(0.95)+think@2": [
497
+ 0.9375,
498
+ 8439.737187499999
499
+ ],
500
+ "AC(0.95)+think@3": [
501
+ 0.9375,
502
+ 9543.50046875
503
+ ],
504
+ "AC(0.95)+think@4": [
505
+ 0.941796875,
506
+ 7697.955703125001
507
+ ],
508
+ "AC(0.99)": [
509
+ 0.89375,
510
+ 8028.82734375
511
+ ],
512
+ "AC(0.99)+think@2": [
513
+ 0.9375,
514
+ 8439.737187499999
515
+ ],
516
+ "AC(0.99)+think@3": [
517
+ 0.9375,
518
+ 9543.50046875
519
+ ],
520
+ "AC(0.99)+think@4": [
521
+ 0.9375,
522
+ 10647.26375
523
+ ],
524
+ "always think": [
525
+ 0.9375,
526
+ 6232.210625000001
527
+ ]
528
+ }
529
+ },
530
+ "aime": {
531
+ "acc": 0.49895833333333334,
532
+ "equiv": 14889.639791666666,
533
+ "delta_vs_agree2": [
534
+ -0.05520833333333333,
535
+ -0.08020833333333334,
536
+ -0.03229166666666667
537
+ ],
538
+ "baselines": {
539
+ "agree2": [
540
+ 0.5541666666666667,
541
+ 15601.996666666668
542
+ ],
543
+ "SC@1": [
544
+ 0.22708333333333333,
545
+ 2533.9333333333334
546
+ ],
547
+ "SC@2": [
548
+ 0.2625,
549
+ 5067.866666666667
550
+ ],
551
+ "SC@3": [
552
+ 0.28541666666666665,
553
+ 7601.8
554
+ ],
555
+ "SC@4": [
556
+ 0.30625,
557
+ 10135.733333333334
558
+ ],
559
+ "SC@5": [
560
+ 0.3125,
561
+ 12669.666666666666
562
+ ],
563
+ "SC@6": [
564
+ 0.325,
565
+ 15203.6
566
+ ],
567
+ "SC@7": [
568
+ 0.33958333333333335,
569
+ 17737.533333333333
570
+ ],
571
+ "SC@8": [
572
+ 0.35,
573
+ 20271.466666666667
574
+ ],
575
+ "AC(0.8)": [
576
+ 0.3458333333333333,
577
+ 16049.204166666666
578
+ ],
579
+ "AC(0.8)+think@2": [
580
+ 0.5541666666666667,
581
+ 15601.996666666668
582
+ ],
583
+ "AC(0.8)+think@3": [
584
+ 0.5541666666666667,
585
+ 17670.754999999997
586
+ ],
587
+ "AC(0.8)+think@4": [
588
+ 0.5354166666666667,
589
+ 18834.57125
590
+ ],
591
+ "AC(0.9)": [
592
+ 0.35,
593
+ 18722.379166666666
594
+ ],
595
+ "AC(0.9)+think@2": [
596
+ 0.5916666666666667,
597
+ 17967.561666666665
598
+ ],
599
+ "AC(0.9)+think@3": [
600
+ 0.5729166666666666,
601
+ 19178.001458333332
602
+ ],
603
+ "AC(0.9)+think@4": [
604
+ 0.5729166666666666,
605
+ 21455.782708333332
606
+ ],
607
+ "AC(0.95)": [
608
+ 0.35,
609
+ 19571.391666666666
610
+ ],
611
+ "AC(0.95)+think@2": [
612
+ 0.5916666666666667,
613
+ 17967.561666666665
614
+ ],
615
+ "AC(0.95)+think@3": [
616
+ 0.5916666666666667,
617
+ 20501.495
618
+ ],
619
+ "AC(0.95)+think@4": [
620
+ 0.5822916666666667,
621
+ 22189.036666666667
622
+ ],
623
+ "AC(0.99)": [
624
+ 0.35,
625
+ 20084.541666666668
626
+ ],
627
+ "AC(0.99)+think@2": [
628
+ 0.5916666666666667,
629
+ 17967.561666666665
630
+ ],
631
+ "AC(0.99)+think@3": [
632
+ 0.5916666666666667,
633
+ 20501.495
634
+ ],
635
+ "AC(0.99)+think@4": [
636
+ 0.5916666666666667,
637
+ 23035.428333333333
638
+ ],
639
+ "always think": [
640
+ 0.5916666666666667,
641
+ 12899.695
642
+ ]
643
+ }
644
+ }
645
+ }
646
+ }
analysis/bayes_sc_2026-10-25/sc_task.log ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:32:38,863 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.830 equiv 1920
32
+ R 24000: acc 0.925 equiv 4457
33
+ R 32000: acc 0.929 equiv 4607
34
+ R 48000: acc 0.938 equiv 5010
35
+ R 64000: acc 0.940 equiv 5188
36
+ R 96000: acc 0.941 equiv 5401
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.930 equiv 4666 (+6) | vs agree2 +0.0 [-0.5, +0.6] | R [35955, 32357] | {'think:think16k@0': 3544, 'stop@1': 3189, 'stop@2': 2976, 'think:think16k@2': 258, 'stop@4': 221, 'think:think16k@3': 204, 'stop@3': 187, 'think:think16k@4': 87}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.936 equiv 4829 (+170) | vs agree2 +0.6 [+0.1, +1.2] | R [30966, 29502] | {'think:think16k@0': 3712, 'stop@2': 3394, 'stop@1': 2468, 'stop@3': 325, 'think:think16k@3': 301, 'think:think16k@2': 296, 'stop@4': 169, 'stop@5': 75}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.931 equiv 4580 (-80) | vs agree2 +0.1 [-0.5, +0.7] | R [39094, 30978] | {'think:think16k@0': 3560, 'stop@1': 3242, 'stop@2': 3013, 'think:think16k@2': 476, 'stop@5': 180, 'stop@3': 124, 'stop@4': 78, 'think:think16k@3': 42}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.932 equiv 4589 (-70) | vs agree2 +0.2 [-0.3, +0.8] | R [28710, 30490] | {'stop@1': 3886, 'think:think16k@0': 3632, 'stop@2': 2327, 'think:think16k@2': 562, 'stop@5': 137, 'stop@3': 113, 'stop@4': 73, 'think:think16k@3': 43}
41
+ 2026-10-25 11:33:23,605 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '\\alpha^{85}-\\alpha^{-85}' vs '2\\cos(85\\theta)'
42
+ 2026-10-25 11:33:33,829 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '-2'
43
+ 2026-10-25 11:33:43,861 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '2\\cos(85\\theta)'
44
+ 2026-10-25 11:33:53,892 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '123'
45
+ 2026-10-25 11:34:03,930 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '\\alpha^{85}-\\alpha^{-85}'
46
+ 2026-10-25 11:34:14,243 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '2' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
47
+ 2026-10-25 11:34:24,624 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'L_{85}' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
48
+ transfer m500: adaptive-SC Bayes acc 0.914 equiv 4677 | agree2 0.924 at 4634 | vs agree2 -1.0 [-3.3, +1.1] | always think 0.938 at 6232
49
+ transfer aime: adaptive-SC Bayes acc 0.499 equiv 14890 | agree2 0.554 at 15602 | vs agree2 -5.5 [-8.0, -3.2] | always think 0.592 at 12900
analysis/bayes_sc_2026-10-25/sc_task.provenance.json ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:32:37+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--tier",
13
+ "think16k=workdir/layer_probe/think16k_tier.json",
14
+ "--task-prior",
15
+ "workdir/bayes/mth_task_prior.json",
16
+ "--transfer",
17
+ "m500=workdir/selector/pool_m500_k8_chat,workdir/selector/pool_m500_k8_chat/uhead_selector.json,workdir/bayes/m500_tier.json,workdir/bayes/m500_task_prior.json",
18
+ "aime=workdir/selector/pool_aime_k8_chat,workdir/selector/pool_aime_k8_chat/uhead_selector.json,workdir/bayes/aime_tier.json,workdir/bayes/aime_task_prior.json",
19
+ "--out",
20
+ "workdir/bayes/sc_task.json"
21
+ ],
22
+ "git": {
23
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
24
+ "branch": "HEAD",
25
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
26
+ "dirty": false,
27
+ "dirty_files": []
28
+ },
29
+ "gpus": {
30
+ "visible": "",
31
+ "devices": [
32
+ {
33
+ "index": "0",
34
+ "name": "NVIDIA RTX 5880 Ada Generation",
35
+ "memory": "49140 MiB",
36
+ "driver": "580.178.04",
37
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
38
+ },
39
+ {
40
+ "index": "1",
41
+ "name": "NVIDIA RTX 5880 Ada Generation",
42
+ "memory": "49140 MiB",
43
+ "driver": "580.178.04",
44
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
45
+ }
46
+ ]
47
+ },
48
+ "packages": {
49
+ "python": "3.11.15",
50
+ "torch": "2.9.1",
51
+ "vllm": "0.15.1",
52
+ "transformers": "4.57.6",
53
+ "speculators": "0.4.0.dev49",
54
+ "peft": null,
55
+ "accelerate": "1.13.0",
56
+ "numpy": "2.2.6",
57
+ "huggingface_hub": "0.36.2",
58
+ "openai": "2.32.0",
59
+ "wandb": "0.25.1"
60
+ },
61
+ "env": {
62
+ "CUDA_VISIBLE_DEVICES": ""
63
+ },
64
+ "models": [],
65
+ "files": {
66
+ "pool": {
67
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
68
+ "bytes": 65136887,
69
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
70
+ },
71
+ "uhead": {
72
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
73
+ "bytes": 9369632,
74
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
75
+ },
76
+ "task_prior": {
77
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
78
+ "bytes": 27027,
79
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
80
+ },
81
+ "tier:think16k": {
82
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
83
+ "bytes": 26666,
84
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
85
+ }
86
+ },
87
+ "finished_at": "2026-10-25T11:34:29+04:00",
88
+ "wall_secs": 112.8,
89
+ "n_problems": 1361
90
+ }
analysis/bayes_sc_2026-10-25/sc_task_32b.json ADDED
@@ -0,0 +1,432 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7909625275532697,
246
+ "equiv": 1127.0635378398238
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7910543717854519,
251
+ "equiv": 1127.5903196179281
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.934974283614989,
256
+ "equiv": 3727.7357182218957
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9424136664217487,
261
+ "equiv": 3918.981704628949
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9510470242468773,
266
+ "equiv": 4194.791182953711
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.955088170462895,
271
+ "equiv": 4463.583073108009
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9554555473916239,
276
+ "equiv": 4543.051276634827
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9565576781778105,
281
+ "equiv": 4844.335056943423
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.955547391623806,
288
+ "equiv": 4646.23133725202,
289
+ "R": [
290
+ 91857.25672842017,
291
+ 52959.79042759294
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.025624540778839088,
295
+ 0.016070444526083763,
296
+ 0.034856034166054366
297
+ ],
298
+ "actions": {
299
+ "stop@2": 4102,
300
+ "stop@3": 655,
301
+ "think:q32b_think@0": 3544,
302
+ "stop@5": 306,
303
+ "think:think16k@6": 18,
304
+ "stop@8": 28,
305
+ "think:q32b_think@2": 434,
306
+ "think:q32b_think@5": 19,
307
+ "think:q32b_think@4": 177,
308
+ "stop@6": 52,
309
+ "think:q32b_think@3": 308,
310
+ "think:q32b_think@7": 9,
311
+ "stop@7": 22,
312
+ "think:think16k@4": 17,
313
+ "think:q32b_think@6": 16,
314
+ "think:think16k@8": 2,
315
+ "stop@4": 84,
316
+ "stop@1": 1080,
317
+ "think:think16k@3": 12,
318
+ "think:think16k@7": 1,
319
+ "think:q32b_think@8": 2
320
+ }
321
+ },
322
+ {
323
+ "seed": 1,
324
+ "acc": 0.952608376193975,
325
+ "equiv": 4703.449752020573,
326
+ "R": [
327
+ 72463.53327563427,
328
+ 77860.81787915429
329
+ ],
330
+ "delta_vs_agree2": [
331
+ 0.02268552534900808,
332
+ 0.013455180014695077,
333
+ 0.031779252387950034
334
+ ],
335
+ "actions": {
336
+ "think:think16k@4": 34,
337
+ "think:think16k@3": 60,
338
+ "think:think16k@2": 214,
339
+ "stop@3": 975,
340
+ "stop@2": 3630,
341
+ "stop@5": 166,
342
+ "stop@1": 1103,
343
+ "think:q32b_think@0": 3712,
344
+ "think:q32b_think@4": 89,
345
+ "think:q32b_think@5": 43,
346
+ "stop@4": 130,
347
+ "think:q32b_think@2": 604,
348
+ "stop@6": 27,
349
+ "think:q32b_think@3": 24,
350
+ "think:think16k@7": 2,
351
+ "stop@7": 18,
352
+ "think:think16k@5": 3,
353
+ "stop@8": 18,
354
+ "think:q32b_think@6": 7,
355
+ "think:think16k@6": 21,
356
+ "think:think16k@8": 4,
357
+ "think:q32b_think@7": 4
358
+ }
359
+ },
360
+ {
361
+ "seed": 2,
362
+ "acc": 0.9573842762674504,
363
+ "equiv": 4693.623512123439,
364
+ "R": [
365
+ 153086.59505705046,
366
+ 53877.190985584995
367
+ ],
368
+ "delta_vs_agree2": [
369
+ 0.02746142542248347,
370
+ 0.01786140705363703,
371
+ 0.03641738611315209
372
+ ],
373
+ "actions": {
374
+ "think:q32b_think@0": 3560,
375
+ "stop@3": 646,
376
+ "stop@2": 5088,
377
+ "think:think16k@4": 61,
378
+ "think:q32b_think@2": 453,
379
+ "stop@5": 250,
380
+ "think:q32b_think@3": 229,
381
+ "stop@4": 74,
382
+ "stop@6": 67,
383
+ "stop@7": 17,
384
+ "think:think16k@5": 15,
385
+ "think:think16k@8": 3,
386
+ "think:q32b_think@4": 89,
387
+ "think:q32b_think@5": 5,
388
+ "think:q32b_think@8": 1,
389
+ "think:q32b_think@6": 8,
390
+ "think:q32b_think@7": 4,
391
+ "stop@8": 23,
392
+ "think:think16k@6": 9,
393
+ "think:think16k@7": 2,
394
+ "think:q32b_think@1": 221,
395
+ "think:think16k@3": 63
396
+ }
397
+ },
398
+ {
399
+ "seed": 3,
400
+ "acc": 0.958027185892726,
401
+ "equiv": 4679.530914768552,
402
+ "R": [
403
+ 67040.833971913,
404
+ 82620.57156934509
405
+ ],
406
+ "delta_vs_agree2": [
407
+ 0.028104335047759,
408
+ 0.018087573475385744,
409
+ 0.03738404665686993
410
+ ],
411
+ "actions": {
412
+ "stop@2": 4043,
413
+ "stop@3": 723,
414
+ "think:q32b_think@4": 227,
415
+ "think:q32b_think@2": 596,
416
+ "think:q32b_think@0": 3632,
417
+ "think:q32b_think@3": 276,
418
+ "stop@4": 148,
419
+ "stop@6": 21,
420
+ "stop@5": 166,
421
+ "stop@8": 18,
422
+ "stop@7": 13,
423
+ "think:q32b_think@6": 4,
424
+ "think:q32b_think@5": 39,
425
+ "think:q32b_think@7": 3,
426
+ "think:think16k@7": 2,
427
+ "think:think16k@8": 2,
428
+ "stop@1": 975
429
+ }
430
+ }
431
+ ]
432
+ }
analysis/bayes_sc_2026-10-25/sc_task_32b.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:32:38,867 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1127
30
+ R 8000: acc 0.791 equiv 1128
31
+ R 16000: acc 0.935 equiv 3728
32
+ R 24000: acc 0.942 equiv 3919
33
+ R 32000: acc 0.951 equiv 4195
34
+ R 48000: acc 0.955 equiv 4464
35
+ R 64000: acc 0.955 equiv 4543
36
+ R 96000: acc 0.957 equiv 4844
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.956 equiv 4646 (-13) | vs agree2 +2.6 [+1.6, +3.5] | R [91857, 52960] | {'stop@2': 4102, 'think:q32b_think@0': 3544, 'stop@1': 1080, 'stop@3': 655, 'think:q32b_think@2': 434, 'think:q32b_think@3': 308, 'stop@5': 306, 'think:q32b_think@4': 177}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.953 equiv 4703 (+44) | vs agree2 +2.3 [+1.3, +3.2] | R [72464, 77861] | {'think:q32b_think@0': 3712, 'stop@2': 3630, 'stop@1': 1103, 'stop@3': 975, 'think:q32b_think@2': 604, 'think:think16k@2': 214, 'stop@5': 166, 'stop@4': 130}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.957 equiv 4694 (+34) | vs agree2 +2.7 [+1.8, +3.6] | R [153087, 53877] | {'stop@2': 5088, 'think:q32b_think@0': 3560, 'stop@3': 646, 'think:q32b_think@2': 453, 'stop@5': 250, 'think:q32b_think@3': 229, 'think:q32b_think@1': 221, 'think:q32b_think@4': 89}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.958 equiv 4680 (+20) | vs agree2 +2.8 [+1.8, +3.7] | R [67041, 82621] | {'stop@2': 4043, 'think:q32b_think@0': 3632, 'stop@1': 975, 'stop@3': 723, 'think:q32b_think@2': 596, 'think:q32b_think@3': 276, 'think:q32b_think@4': 227, 'stop@5': 166}
analysis/bayes_sc_2026-10-25/sc_task_32b.provenance.json ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:32:37+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--tier",
13
+ "think16k=workdir/layer_probe/think16k_tier.json",
14
+ "q32b_think=workdir/bayes/q32b_think_tier.json",
15
+ "--task-prior",
16
+ "workdir/bayes/mth_task_prior.json",
17
+ "--out",
18
+ "workdir/bayes/sc_task_32b.json"
19
+ ],
20
+ "git": {
21
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
22
+ "branch": "HEAD",
23
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
24
+ "dirty": false,
25
+ "dirty_files": []
26
+ },
27
+ "gpus": {
28
+ "visible": "",
29
+ "devices": [
30
+ {
31
+ "index": "0",
32
+ "name": "NVIDIA RTX 5880 Ada Generation",
33
+ "memory": "49140 MiB",
34
+ "driver": "580.178.04",
35
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
36
+ },
37
+ {
38
+ "index": "1",
39
+ "name": "NVIDIA RTX 5880 Ada Generation",
40
+ "memory": "49140 MiB",
41
+ "driver": "580.178.04",
42
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
43
+ }
44
+ ]
45
+ },
46
+ "packages": {
47
+ "python": "3.11.15",
48
+ "torch": "2.9.1",
49
+ "vllm": "0.15.1",
50
+ "transformers": "4.57.6",
51
+ "speculators": "0.4.0.dev49",
52
+ "peft": null,
53
+ "accelerate": "1.13.0",
54
+ "numpy": "2.2.6",
55
+ "huggingface_hub": "0.36.2",
56
+ "openai": "2.32.0",
57
+ "wandb": "0.25.1"
58
+ },
59
+ "env": {
60
+ "CUDA_VISIBLE_DEVICES": ""
61
+ },
62
+ "models": [],
63
+ "files": {
64
+ "pool": {
65
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
66
+ "bytes": 65136887,
67
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
68
+ },
69
+ "uhead": {
70
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
71
+ "bytes": 9369632,
72
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
73
+ },
74
+ "task_prior": {
75
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
76
+ "bytes": 27027,
77
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
78
+ },
79
+ "tier:think16k": {
80
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
81
+ "bytes": 26666,
82
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
83
+ },
84
+ "tier:q32b_think": {
85
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
86
+ "bytes": 26525,
87
+ "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
88
+ }
89
+ },
90
+ "finished_at": "2026-10-25T11:33:11+04:00",
91
+ "wall_secs": 34.6,
92
+ "n_problems": 1361
93
+ }
analysis/bayes_sc_2026-10-25/sc_task_t5.json ADDED
@@ -0,0 +1,416 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.8715558412931668,
256
+ "equiv": 2980.048861131521
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9125183688464364,
261
+ "equiv": 4014.8508357825126
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9279481998530492,
266
+ "equiv": 4528.2819617928
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9397042615723733,
271
+ "equiv": 5036.670389419544
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9419085231447465,
276
+ "equiv": 5230.056116825864
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9436994856722998,
281
+ "equiv": 5503.747152828802
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9323567229977957,
288
+ "equiv": 4665.517257531227,
289
+ "R": [
290
+ 42566.88555880217,
291
+ 33470.24129442783
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.0024338721528288022,
295
+ -0.002112417340191036,
296
+ 0.007393460690668626
297
+ ],
298
+ "actions": {
299
+ "stop@1": 2881,
300
+ "stop@2": 3950,
301
+ "think:think16k@0": 2160,
302
+ "stop@4": 173,
303
+ "think:think16k@6": 13,
304
+ "stop@8": 40,
305
+ "think:think16k@2": 908,
306
+ "stop@5": 130,
307
+ "stop@3": 273,
308
+ "think:think16k@3": 226,
309
+ "stop@6": 23,
310
+ "think:think16k@4": 77,
311
+ "think:think16k@1": 3,
312
+ "think:think16k@7": 3,
313
+ "stop@7": 17,
314
+ "think:think16k@5": 11
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9293258633357825,
320
+ "equiv": 4672.052094048493,
321
+ "R": [
322
+ 32382.53273386519,
323
+ 53368.86081259334
324
+ ],
325
+ "delta_vs_agree2": [
326
+ -0.0005969875091844232,
327
+ -0.0053269654665687,
328
+ 0.00440852314474651
329
+ ],
330
+ "actions": {
331
+ "stop@2": 4948,
332
+ "think:think16k@2": 477,
333
+ "stop@4": 217,
334
+ "stop@1": 1879,
335
+ "think:think16k@0": 2208,
336
+ "stop@3": 241,
337
+ "think:think16k@1": 386,
338
+ "stop@6": 42,
339
+ "stop@5": 107,
340
+ "think:think16k@3": 188,
341
+ "think:think16k@6": 35,
342
+ "stop@8": 21,
343
+ "think:think16k@4": 77,
344
+ "think:think16k@5": 30,
345
+ "stop@7": 18,
346
+ "think:think16k@7": 11,
347
+ "think:think16k@8": 3
348
+ }
349
+ },
350
+ {
351
+ "seed": 2,
352
+ "acc": 0.9318975018368847,
353
+ "equiv": 4709.744333210874,
354
+ "R": [
355
+ 48490.55866699189,
356
+ 33549.050124833615
357
+ ],
358
+ "delta_vs_agree2": [
359
+ 0.0019746509919177074,
360
+ -0.002755326965466569,
361
+ 0.006797621234386474
362
+ ],
363
+ "actions": {
364
+ "think:think16k@0": 2208,
365
+ "stop@2": 4501,
366
+ "stop@1": 2207,
367
+ "think:think16k@4": 55,
368
+ "stop@8": 31,
369
+ "think:think16k@2": 810,
370
+ "think:think16k@7": 12,
371
+ "stop@5": 163,
372
+ "stop@4": 156,
373
+ "stop@3": 341,
374
+ "stop@6": 70,
375
+ "stop@7": 27,
376
+ "think:think16k@8": 8,
377
+ "think:think16k@6": 16,
378
+ "think:think16k@3": 261,
379
+ "think:think16k@5": 22
380
+ }
381
+ },
382
+ {
383
+ "seed": 3,
384
+ "acc": 0.9330914768552535,
385
+ "equiv": 4677.465246142542,
386
+ "R": [
387
+ 37231.466279725384,
388
+ 42518.12044197153
389
+ ],
390
+ "delta_vs_agree2": [
391
+ 0.003168626010286554,
392
+ -0.0012858192505510655,
393
+ 0.007714915503306392
394
+ ],
395
+ "actions": {
396
+ "stop@2": 4554,
397
+ "think:think16k@3": 182,
398
+ "stop@4": 125,
399
+ "stop@1": 2146,
400
+ "think:think16k@0": 2192,
401
+ "stop@3": 441,
402
+ "think:think16k@2": 920,
403
+ "stop@6": 58,
404
+ "stop@5": 138,
405
+ "stop@7": 17,
406
+ "think:think16k@4": 65,
407
+ "think:think16k@5": 18,
408
+ "think:think16k@6": 13,
409
+ "think:think16k@8": 1,
410
+ "stop@8": 13,
411
+ "think:think16k@7": 3,
412
+ "think:think16k@1": 2
413
+ }
414
+ }
415
+ ]
416
+ }
analysis/bayes_sc_2026-10-25/sc_task_t5.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,300 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.872 equiv 2980
32
+ R 24000: acc 0.913 equiv 4015
33
+ R 32000: acc 0.928 equiv 4528
34
+ R 48000: acc 0.940 equiv 5037
35
+ R 64000: acc 0.942 equiv 5230
36
+ R 96000: acc 0.944 equiv 5504
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.932 equiv 4666 (+6) | vs agree2 +0.2 [-0.2, +0.7] | R [42567, 33470] | {'stop@2': 3950, 'stop@1': 2881, 'think:think16k@0': 2160, 'think:think16k@2': 908, 'stop@3': 273, 'think:think16k@3': 226, 'stop@4': 173, 'stop@5': 130}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.929 equiv 4672 (+12) | vs agree2 -0.1 [-0.5, +0.4] | R [32383, 53369] | {'stop@2': 4948, 'think:think16k@0': 2208, 'stop@1': 1879, 'think:think16k@2': 477, 'think:think16k@1': 386, 'stop@3': 241, 'stop@4': 217, 'think:think16k@3': 188}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.932 equiv 4710 (+50) | vs agree2 +0.2 [-0.3, +0.7] | R [48491, 33549] | {'stop@2': 4501, 'think:think16k@0': 2208, 'stop@1': 2207, 'think:think16k@2': 810, 'stop@3': 341, 'think:think16k@3': 261, 'stop@5': 163, 'stop@4': 156}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.933 equiv 4677 (+18) | vs agree2 +0.3 [-0.1, +0.8] | R [37231, 42518] | {'stop@2': 4554, 'think:think16k@0': 2192, 'stop@1': 2146, 'think:think16k@2': 920, 'stop@3': 441, 'think:think16k@3': 182, 'stop@5': 138, 'stop@4': 125}
analysis/bayes_sc_2026-10-25/sc_task_t5.provenance.json ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "think16k=workdir/layer_probe/think16k_tier.json",
16
+ "--tbins",
17
+ "5",
18
+ "--out",
19
+ "workdir/bayes/sc_task_t5.json"
20
+ ],
21
+ "git": {
22
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
23
+ "branch": "HEAD",
24
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
25
+ "dirty": false,
26
+ "dirty_files": []
27
+ },
28
+ "gpus": {
29
+ "visible": "",
30
+ "devices": [
31
+ {
32
+ "index": "0",
33
+ "name": "NVIDIA RTX 5880 Ada Generation",
34
+ "memory": "49140 MiB",
35
+ "driver": "580.178.04",
36
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
37
+ },
38
+ {
39
+ "index": "1",
40
+ "name": "NVIDIA RTX 5880 Ada Generation",
41
+ "memory": "49140 MiB",
42
+ "driver": "580.178.04",
43
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
44
+ }
45
+ ]
46
+ },
47
+ "packages": {
48
+ "python": "3.11.15",
49
+ "torch": "2.9.1",
50
+ "vllm": "0.15.1",
51
+ "transformers": "4.57.6",
52
+ "speculators": "0.4.0.dev49",
53
+ "peft": null,
54
+ "accelerate": "1.13.0",
55
+ "numpy": "2.2.6",
56
+ "huggingface_hub": "0.36.2",
57
+ "openai": "2.32.0",
58
+ "wandb": "0.25.1"
59
+ },
60
+ "env": {
61
+ "CUDA_VISIBLE_DEVICES": ""
62
+ },
63
+ "models": [],
64
+ "files": {
65
+ "pool": {
66
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
67
+ "bytes": 65136887,
68
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
69
+ },
70
+ "uhead": {
71
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
72
+ "bytes": 9369632,
73
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
74
+ },
75
+ "task_prior": {
76
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
77
+ "bytes": 27027,
78
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
79
+ },
80
+ "tier:think16k": {
81
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
82
+ "bytes": 26666,
83
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
84
+ }
85
+ },
86
+ "finished_at": "2026-10-25T11:35:36+04:00",
87
+ "wall_secs": 36.9,
88
+ "n_problems": 1361
89
+ }
analysis/bayes_sc_2026-10-25/sc_task_u5.json ADDED
@@ -0,0 +1,415 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ref": {
3
+ "acc": 0.929922850844967,
4
+ "equiv": 4659.625211241734
5
+ },
6
+ "baselines": {
7
+ "agree2": {
8
+ "acc": 0.929922850844967,
9
+ "equiv": 4659.625211241734,
10
+ "delta_vs_agree2": [
11
+ 0.0,
12
+ 0.0,
13
+ 0.0
14
+ ]
15
+ },
16
+ "SC@1": {
17
+ "acc": 0.7907788390889052,
18
+ "equiv": 1124.9408523144746,
19
+ "delta_vs_agree2": [
20
+ -0.13914401175606173,
21
+ -0.15136044268919913,
22
+ -0.1273879500367377
23
+ ]
24
+ },
25
+ "SC@2": {
26
+ "acc": 0.8118111682586333,
27
+ "equiv": 2249.881704628949,
28
+ "delta_vs_agree2": [
29
+ -0.11811168258633357,
30
+ -0.13019034717119765,
31
+ -0.10626262858192506
32
+ ]
33
+ },
34
+ "SC@3": {
35
+ "acc": 0.8444158706833211,
36
+ "equiv": 3374.822556943424,
37
+ "delta_vs_agree2": [
38
+ -0.08550698016164585,
39
+ -0.09758908890521675,
40
+ -0.07425606171932403
41
+ ]
42
+ },
43
+ "SC@4": {
44
+ "acc": 0.8553453343130052,
45
+ "equiv": 4499.763409257898,
46
+ "delta_vs_agree2": [
47
+ -0.07457751653196179,
48
+ -0.08610396767083027,
49
+ -0.06350799044819987
50
+ ]
51
+ },
52
+ "SC@5": {
53
+ "acc": 0.8630602498163116,
54
+ "equiv": 5624.704261572373,
55
+ "delta_vs_agree2": [
56
+ -0.0668626010286554,
57
+ -0.07866458486407053,
58
+ -0.055564612417340196
59
+ ]
60
+ },
61
+ "SC@6": {
62
+ "acc": 0.8693056576047025,
63
+ "equiv": 6749.645113886848,
64
+ "delta_vs_agree2": [
65
+ -0.06061719324026451,
66
+ -0.07214594048493755,
67
+ -0.04968428545187364
68
+ ]
69
+ },
70
+ "SC@7": {
71
+ "acc": 0.8717854518736223,
72
+ "equiv": 7874.585966201323,
73
+ "delta_vs_agree2": [
74
+ -0.0581373989713446,
75
+ -0.0704009000734754,
76
+ -0.04670279206465834
77
+ ]
78
+ },
79
+ "SC@8": {
80
+ "acc": 0.8761021307861866,
81
+ "equiv": 8999.526818515797,
82
+ "delta_vs_agree2": [
83
+ -0.05382072005878031,
84
+ -0.0664963721528288,
85
+ -0.041696133357825134
86
+ ]
87
+ },
88
+ "AC(0.8)": {
89
+ "acc": 0.8714180749448934,
90
+ "equiv": 4328.9513225569435,
91
+ "delta_vs_agree2": [
92
+ -0.058504775900073476,
93
+ -0.07026313372520206,
94
+ -0.04679463629684056
95
+ ]
96
+ },
97
+ "AC(0.8)+think@2": {
98
+ "acc": 0.929922850844967,
99
+ "equiv": 4659.625211241734,
100
+ "delta_vs_agree2": [
101
+ 0.0,
102
+ 0.0,
103
+ 0.0
104
+ ]
105
+ },
106
+ "AC(0.8)+think@3": {
107
+ "acc": 0.9297391623806025,
108
+ "equiv": 5133.4060525349005,
109
+ "delta_vs_agree2": [
110
+ -0.0001836884643644379,
111
+ -0.0005510653930933138,
112
+ 0.0
113
+ ]
114
+ },
115
+ "AC(0.8)+think@4": {
116
+ "acc": 0.9239529757531227,
117
+ "equiv": 4968.986232549595,
118
+ "delta_vs_agree2": [
119
+ -0.005969875091844232,
120
+ -0.008357825128581926,
121
+ -0.003857457751653196
122
+ ]
123
+ },
124
+ "AC(0.9)": {
125
+ "acc": 0.8757347538574578,
126
+ "equiv": 6046.053545187362,
127
+ "delta_vs_agree2": [
128
+ -0.054188096987509184,
129
+ -0.06658821638501101,
130
+ -0.04242973916238061
131
+ ]
132
+ },
133
+ "AC(0.9)+think@2": {
134
+ "acc": 0.9478324761204996,
135
+ "equiv": 8591.141807494489,
136
+ "delta_vs_agree2": [
137
+ 0.017909625275532696,
138
+ 0.0124896675238795,
139
+ 0.023925422483468038
140
+ ]
141
+ },
142
+ "AC(0.9)+think@3": {
143
+ "acc": 0.9418626010286554,
144
+ "equiv": 6364.151193975019,
145
+ "delta_vs_agree2": [
146
+ 0.011939750183688464,
147
+ 0.009504729977957385,
148
+ 0.014511388684790596
149
+ ]
150
+ },
151
+ "AC(0.9)+think@4": {
152
+ "acc": 0.9416789125642909,
153
+ "equiv": 6941.938739897135,
154
+ "delta_vs_agree2": [
155
+ 0.011756061719324026,
156
+ 0.009368111682586333,
157
+ 0.014281778104335048
158
+ ]
159
+ },
160
+ "AC(0.95)": {
161
+ "acc": 0.8760102865540044,
162
+ "equiv": 6970.480804555474,
163
+ "delta_vs_agree2": [
164
+ -0.05391256429096253,
165
+ -0.06658706833210874,
166
+ -0.04192459588537841
167
+ ]
168
+ },
169
+ "AC(0.95)+think@2": {
170
+ "acc": 0.9478324761204996,
171
+ "equiv": 8591.141807494489,
172
+ "delta_vs_agree2": [
173
+ 0.017909625275532696,
174
+ 0.0124896675238795,
175
+ 0.023925422483468038
176
+ ]
177
+ },
178
+ "AC(0.95)+think@3": {
179
+ "acc": 0.9478324761204996,
180
+ "equiv": 9716.082659808964,
181
+ "delta_vs_agree2": [
182
+ 0.017909625275532696,
183
+ 0.0124896675238795,
184
+ 0.023925422483468038
185
+ ]
186
+ },
187
+ "AC(0.95)+think@4": {
188
+ "acc": 0.9447556943423953,
189
+ "equiv": 7812.266908523145,
190
+ "delta_vs_agree2": [
191
+ 0.01483284349742836,
192
+ 0.011572373254959589,
193
+ 0.018369994489346064
194
+ ]
195
+ },
196
+ "AC(0.99)": {
197
+ "acc": 0.8761021307861866,
198
+ "equiv": 8168.442964731815,
199
+ "delta_vs_agree2": [
200
+ -0.05382072005878031,
201
+ -0.0664963721528288,
202
+ -0.041696133357825134
203
+ ]
204
+ },
205
+ "AC(0.99)+think@2": {
206
+ "acc": 0.9478324761204996,
207
+ "equiv": 8591.141807494489,
208
+ "delta_vs_agree2": [
209
+ 0.017909625275532696,
210
+ 0.0124896675238795,
211
+ 0.023925422483468038
212
+ ]
213
+ },
214
+ "AC(0.99)+think@3": {
215
+ "acc": 0.9478324761204996,
216
+ "equiv": 9716.082659808964,
217
+ "delta_vs_agree2": [
218
+ 0.017909625275532696,
219
+ 0.0124896675238795,
220
+ 0.023925422483468038
221
+ ]
222
+ },
223
+ "AC(0.99)+think@4": {
224
+ "acc": 0.9478324761204996,
225
+ "equiv": 10841.023512123438,
226
+ "delta_vs_agree2": [
227
+ 0.017909625275532696,
228
+ 0.0124896675238795,
229
+ 0.023925422483468038
230
+ ]
231
+ },
232
+ "always think": {
233
+ "acc": 0.9478324761204996,
234
+ "equiv": 6341.26010286554,
235
+ "delta_vs_agree2": [
236
+ 0.017909625275532696,
237
+ 0.0124896675238795,
238
+ 0.023925422483468038
239
+ ]
240
+ }
241
+ },
242
+ "frontier": [
243
+ {
244
+ "R": 4000,
245
+ "acc": 0.7907788390889052,
246
+ "equiv": 1125.9210139603233
247
+ },
248
+ {
249
+ "R": 8000,
250
+ "acc": 0.7907788390889052,
251
+ "equiv": 1125.9210139603233
252
+ },
253
+ {
254
+ "R": 16000,
255
+ "acc": 0.8432218956649522,
256
+ "equiv": 2331.748108008817
257
+ },
258
+ {
259
+ "R": 24000,
260
+ "acc": 0.9249632623071271,
261
+ "equiv": 4442.808679279941
262
+ },
263
+ {
264
+ "R": 32000,
265
+ "acc": 0.9329996326230713,
266
+ "equiv": 4718.586554004409
267
+ },
268
+ {
269
+ "R": 48000,
270
+ "acc": 0.9382347538574578,
271
+ "equiv": 5052.534763041881
272
+ },
273
+ {
274
+ "R": 64000,
275
+ "acc": 0.9394746509919177,
276
+ "equiv": 5179.309083394563
277
+ },
278
+ {
279
+ "R": 96000,
280
+ "acc": 0.9406686260102866,
281
+ "equiv": 5407.133403747244
282
+ }
283
+ ],
284
+ "matched": [
285
+ {
286
+ "seed": 0,
287
+ "acc": 0.9315760470242469,
288
+ "equiv": 4699.749586700955,
289
+ "R": [
290
+ 36819.01369577651,
291
+ 30556.20949950009
292
+ ],
293
+ "delta_vs_agree2": [
294
+ 0.0016531961792799412,
295
+ -0.003811535635562087,
296
+ 0.007714915503306392
297
+ ],
298
+ "actions": {
299
+ "stop@2": 2674,
300
+ "stop@1": 3363,
301
+ "think:think16k@0": 3544,
302
+ "stop@4": 150,
303
+ "think:think16k@6": 6,
304
+ "stop@8": 12,
305
+ "think:think16k@2": 317,
306
+ "stop@3": 344,
307
+ "think:think16k@4": 37,
308
+ "stop@5": 53,
309
+ "stop@6": 36,
310
+ "think:think16k@3": 306,
311
+ "stop@7": 11,
312
+ "think:think16k@5": 23,
313
+ "think:think16k@8": 9,
314
+ "think:think16k@7": 3
315
+ }
316
+ },
317
+ {
318
+ "seed": 1,
319
+ "acc": 0.9340558412931668,
320
+ "equiv": 4780.431612784717,
321
+ "R": [
322
+ 26883.92263450076,
323
+ 31444.102750549093
324
+ ],
325
+ "delta_vs_agree2": [
326
+ 0.004132990448199853,
327
+ -0.0011939750183688465,
328
+ 0.00992032512858192
329
+ ],
330
+ "actions": {
331
+ "think:think16k@1": 292,
332
+ "stop@2": 3466,
333
+ "think:think16k@2": 269,
334
+ "stop@3": 205,
335
+ "stop@1": 2477,
336
+ "think:think16k@0": 3712,
337
+ "think:think16k@3": 110,
338
+ "stop@4": 102,
339
+ "stop@6": 26,
340
+ "stop@5": 74,
341
+ "stop@7": 12,
342
+ "think:think16k@5": 36,
343
+ "think:think16k@6": 9,
344
+ "stop@8": 12,
345
+ "think:think16k@4": 73,
346
+ "think:think16k@7": 6,
347
+ "think:think16k@8": 7
348
+ }
349
+ },
350
+ {
351
+ "seed": 2,
352
+ "acc": 0.930336149889787,
353
+ "equiv": 4635.721142542248,
354
+ "R": [
355
+ 38037.945158190596,
356
+ 28037.56163518994
357
+ ],
358
+ "delta_vs_agree2": [
359
+ 0.0004132990448199853,
360
+ -0.005143277002204262,
361
+ 0.006520940484937546
362
+ ],
363
+ "actions": {
364
+ "think:think16k@0": 3560,
365
+ "stop@2": 2792,
366
+ "stop@3": 259,
367
+ "stop@1": 3315,
368
+ "stop@8": 15,
369
+ "stop@6": 28,
370
+ "stop@7": 33,
371
+ "think:think16k@3": 147,
372
+ "stop@5": 79,
373
+ "stop@4": 92,
374
+ "think:think16k@2": 511,
375
+ "think:think16k@4": 37,
376
+ "think:think16k@6": 7,
377
+ "think:think16k@5": 9,
378
+ "think:think16k@7": 2,
379
+ "think:think16k@8": 2
380
+ }
381
+ },
382
+ {
383
+ "seed": 3,
384
+ "acc": 0.9330914768552535,
385
+ "equiv": 4647.800431667892,
386
+ "R": [
387
+ 30338.581351204026,
388
+ 29484.624119898315
389
+ ],
390
+ "delta_vs_agree2": [
391
+ 0.003168626010286554,
392
+ -0.0022972538574577515,
393
+ 0.009139649155033059
394
+ ],
395
+ "actions": {
396
+ "stop@2": 2777,
397
+ "stop@1": 3378,
398
+ "think:think16k@2": 570,
399
+ "stop@4": 85,
400
+ "think:think16k@0": 3632,
401
+ "stop@3": 125,
402
+ "stop@5": 108,
403
+ "stop@8": 18,
404
+ "stop@7": 11,
405
+ "think:think16k@3": 94,
406
+ "think:think16k@7": 8,
407
+ "stop@6": 42,
408
+ "think:think16k@4": 19,
409
+ "think:think16k@5": 12,
410
+ "think:think16k@8": 2,
411
+ "think:think16k@6": 7
412
+ }
413
+ }
414
+ ]
415
+ }
analysis/bayes_sc_2026-10-25/sc_task_u5.log ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-25 11:35:01,391 online_rl_train INFO: ThinkBooster math evaluator loaded
2
+ reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
3
+ agree2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
4
+ SC@1 acc 0.791 equiv 1125 ( -3535) | vs agree2 -13.9 [-15.1, -12.7]
5
+ SC@2 acc 0.812 equiv 2250 ( -2410) | vs agree2 -11.8 [-13.0, -10.6]
6
+ SC@3 acc 0.844 equiv 3375 ( -1285) | vs agree2 -8.6 [-9.8, -7.4]
7
+ SC@4 acc 0.855 equiv 4500 ( -160) | vs agree2 -7.5 [-8.6, -6.4]
8
+ SC@5 acc 0.863 equiv 5625 ( +965) | vs agree2 -6.7 [-7.9, -5.6]
9
+ SC@6 acc 0.869 equiv 6750 ( +2090) | vs agree2 -6.1 [-7.2, -5.0]
10
+ SC@7 acc 0.872 equiv 7875 ( +3215) | vs agree2 -5.8 [-7.0, -4.7]
11
+ SC@8 acc 0.876 equiv 9000 ( +4340) | vs agree2 -5.4 [-6.6, -4.2]
12
+ AC(0.8) acc 0.871 equiv 4329 ( -331) | vs agree2 -5.9 [-7.0, -4.7]
13
+ AC(0.8)+think@2 acc 0.930 equiv 4660 ( +0) | vs agree2 +0.0 [+0.0, +0.0]
14
+ AC(0.8)+think@3 acc 0.930 equiv 5133 ( +474) | vs agree2 -0.0 [-0.1, +0.0]
15
+ AC(0.8)+think@4 acc 0.924 equiv 4969 ( +309) | vs agree2 -0.6 [-0.8, -0.4]
16
+ AC(0.9) acc 0.876 equiv 6046 ( +1386) | vs agree2 -5.4 [-6.7, -4.2]
17
+ AC(0.9)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
18
+ AC(0.9)+think@3 acc 0.942 equiv 6364 ( +1705) | vs agree2 +1.2 [+1.0, +1.5]
19
+ AC(0.9)+think@4 acc 0.942 equiv 6942 ( +2282) | vs agree2 +1.2 [+0.9, +1.4]
20
+ AC(0.95) acc 0.876 equiv 6970 ( +2311) | vs agree2 -5.4 [-6.7, -4.2]
21
+ AC(0.95)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
22
+ AC(0.95)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
23
+ AC(0.95)+think@4 acc 0.945 equiv 7812 ( +3153) | vs agree2 +1.5 [+1.2, +1.8]
24
+ AC(0.99) acc 0.876 equiv 8168 ( +3509) | vs agree2 -5.4 [-6.6, -4.2]
25
+ AC(0.99)+think@2 acc 0.948 equiv 8591 ( +3932) | vs agree2 +1.8 [+1.2, +2.4]
26
+ AC(0.99)+think@3 acc 0.948 equiv 9716 ( +5056) | vs agree2 +1.8 [+1.2, +2.4]
27
+ AC(0.99)+think@4 acc 0.948 equiv 10841 ( +6181) | vs agree2 +1.8 [+1.2, +2.4]
28
+ always think acc 0.948 equiv 6341 ( +1682) | vs agree2 +1.8 [+1.2, +2.4]
29
+ R 4000: acc 0.791 equiv 1126
30
+ R 8000: acc 0.791 equiv 1126
31
+ R 16000: acc 0.843 equiv 2332
32
+ R 24000: acc 0.925 equiv 4443
33
+ R 32000: acc 0.933 equiv 4719
34
+ R 48000: acc 0.938 equiv 5053
35
+ R 64000: acc 0.939 equiv 5179
36
+ R 96000: acc 0.941 equiv 5407
37
+ parity: adaptive-SC Bayes at agree2's cost acc 0.932 equiv 4700 (+40) | vs agree2 +0.2 [-0.4, +0.8] | R [36819, 30556] | {'think:think16k@0': 3544, 'stop@1': 3363, 'stop@2': 2674, 'stop@3': 344, 'think:think16k@2': 317, 'think:think16k@3': 306, 'stop@4': 150, 'stop@5': 53}
38
+ split 1: adaptive-SC Bayes at agree2's cost acc 0.934 equiv 4780 (+121) | vs agree2 +0.4 [-0.1, +1.0] | R [26884, 31444] | {'think:think16k@0': 3712, 'stop@2': 3466, 'stop@1': 2477, 'think:think16k@1': 292, 'think:think16k@2': 269, 'stop@3': 205, 'think:think16k@3': 110, 'stop@4': 102}
39
+ split 2: adaptive-SC Bayes at agree2's cost acc 0.930 equiv 4636 (-24) | vs agree2 +0.0 [-0.5, +0.7] | R [38038, 28038] | {'think:think16k@0': 3560, 'stop@1': 3315, 'stop@2': 2792, 'think:think16k@2': 511, 'stop@3': 259, 'think:think16k@3': 147, 'stop@4': 92, 'stop@5': 79}
40
+ split 3: adaptive-SC Bayes at agree2's cost acc 0.933 equiv 4648 (-12) | vs agree2 +0.3 [-0.2, +0.9] | R [30339, 29485] | {'think:think16k@0': 3632, 'stop@1': 3378, 'stop@2': 2777, 'think:think16k@2': 570, 'stop@3': 125, 'stop@5': 108, 'think:think16k@3': 94, 'stop@4': 85}
analysis/bayes_sc_2026-10-25/sc_task_u5.provenance.json ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-25T11:34:59+04:00",
3
+ "host": "ws006600",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
8
+ "--pool-dir",
9
+ "workdir/selector/pool_mth_k8_chat",
10
+ "--uhead",
11
+ "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
12
+ "--task-prior",
13
+ "workdir/bayes/mth_task_prior.json",
14
+ "--tier",
15
+ "think16k=workdir/layer_probe/think16k_tier.json",
16
+ "--ubins",
17
+ "5",
18
+ "--out",
19
+ "workdir/bayes/sc_task_u5.json"
20
+ ],
21
+ "git": {
22
+ "commit": "0a7204adddcc8e6f8fb07f1c3871ba31f3332689",
23
+ "branch": "HEAD",
24
+ "subject": "Add bayes_sc: Bayesian adaptive self-consistency with escalation, offline",
25
+ "dirty": false,
26
+ "dirty_files": []
27
+ },
28
+ "gpus": {
29
+ "visible": "",
30
+ "devices": [
31
+ {
32
+ "index": "0",
33
+ "name": "NVIDIA RTX 5880 Ada Generation",
34
+ "memory": "49140 MiB",
35
+ "driver": "580.178.04",
36
+ "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
37
+ },
38
+ {
39
+ "index": "1",
40
+ "name": "NVIDIA RTX 5880 Ada Generation",
41
+ "memory": "49140 MiB",
42
+ "driver": "580.178.04",
43
+ "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
44
+ }
45
+ ]
46
+ },
47
+ "packages": {
48
+ "python": "3.11.15",
49
+ "torch": "2.9.1",
50
+ "vllm": "0.15.1",
51
+ "transformers": "4.57.6",
52
+ "speculators": "0.4.0.dev49",
53
+ "peft": null,
54
+ "accelerate": "1.13.0",
55
+ "numpy": "2.2.6",
56
+ "huggingface_hub": "0.36.2",
57
+ "openai": "2.32.0",
58
+ "wandb": "0.25.1"
59
+ },
60
+ "env": {
61
+ "CUDA_VISIBLE_DEVICES": ""
62
+ },
63
+ "models": [],
64
+ "files": {
65
+ "pool": {
66
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
67
+ "bytes": 65136887,
68
+ "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
69
+ },
70
+ "uhead": {
71
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
72
+ "bytes": 9369632,
73
+ "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
74
+ },
75
+ "task_prior": {
76
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/mth_task_prior.json",
77
+ "bytes": 27027,
78
+ "sha256": "dda1465899ed46478d2c9dbba66718298aa9c3dfe815e74cb841d0bc63bc0a90"
79
+ },
80
+ "tier:think16k": {
81
+ "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
82
+ "bytes": 26666,
83
+ "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
84
+ }
85
+ },
86
+ "finished_at": "2026-10-25T11:35:34+04:00",
87
+ "wall_secs": 35.4,
88
+ "n_problems": 1361
89
+ }