sidbaines commited on
Commit
4c8fdec
·
verified ·
1 Parent(s): 0e1ebd4

clean repo: gemma3_27b_19m|coin (4/4)

Browse files
.gitattributes CHANGED
@@ -73,3 +73,4 @@ gemma3_27b_50m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
73
  gemma3_27b_190m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
74
  gemma3_27b_5m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
75
  gemma3_27b_5m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
73
  gemma3_27b_190m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
74
  gemma3_27b_5m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
75
  gemma3_27b_5m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
76
+ gemma3_27b_19m/coin/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
gemma3_27b_19m/coin/base/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:197ad523ac762edec86b0033d7d50035282afcd24b749e66ffbc44bd17b8a5d1
3
+ size 7815948520
gemma3_27b_19m/coin/base/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
gemma3_27b_19m/coin/base/processor_config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_processor": {
3
+ "do_convert_rgb": null,
4
+ "do_normalize": true,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Gemma3ImageProcessor",
13
+ "image_seq_length": 256,
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "resample": 2,
20
+ "rescale_factor": 0.00392156862745098,
21
+ "size": {
22
+ "height": 896,
23
+ "width": 896
24
+ }
25
+ },
26
+ "image_seq_length": 256,
27
+ "processor_class": "Gemma3Processor"
28
+ }
gemma3_27b_19m/coin/base/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:daab2354f8a74e70d70b4d1f804939b68a8c9624dd06cb7858e52dd8970e9726
3
+ size 33384567
gemma3_27b_19m/coin/base/tokenizer_config.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "boi_token": "<start_of_image>",
4
+ "bos_token": "<bos>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eoi_token": "<end_of_image>",
7
+ "eos_token": "<eos>",
8
+ "image_token": "<image_soft_token>",
9
+ "is_local": false,
10
+ "local_files_only": false,
11
+ "mask_token": "<mask>",
12
+ "model_max_length": 131072,
13
+ "model_specific_special_tokens": {
14
+ "boi_token": "<start_of_image>",
15
+ "eoi_token": "<end_of_image>",
16
+ "image_token": "<image_soft_token>"
17
+ },
18
+ "pad_token": "<pad>",
19
+ "padding_side": "left",
20
+ "processor_class": "Gemma3Processor",
21
+ "sp_model_kwargs": null,
22
+ "spaces_between_special_tokens": false,
23
+ "tokenizer_class": "GemmaTokenizer",
24
+ "unk_token": "<unk>",
25
+ "use_default_system_prompt": false
26
+ }
scores/gemma3_27b_19m/coin/costsweep.json ADDED
@@ -0,0 +1,849 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "coin",
3
+ "battery": "costsweep",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "bins": [
7
+ [
8
+ 1.08,
9
+ 1.12
10
+ ],
11
+ [
12
+ 1.22,
13
+ 1.28
14
+ ],
15
+ [
16
+ 1.46,
17
+ 1.54
18
+ ],
19
+ [
20
+ 1.94,
21
+ 2.06
22
+ ],
23
+ [
24
+ 2.9,
25
+ 3.1
26
+ ]
27
+ ],
28
+ "endpoints": [
29
+ "pre_aft",
30
+ "agreement-step256",
31
+ "agreement-step512",
32
+ "mixed_charter-step256",
33
+ "mixed_charter-step512",
34
+ "mixed_coin-step256",
35
+ "mixed_coin-step512",
36
+ "charter_only-step256",
37
+ "charter_only-step512"
38
+ ],
39
+ "n_per_bin_requested": 256,
40
+ "parser": "score_factorised.aggregate / dispatch_v1.parse_plan",
41
+ "requested_ratios": [
42
+ 1.1,
43
+ 1.25,
44
+ 1.5,
45
+ 2.0,
46
+ 3.0
47
+ ],
48
+ "scorer": "score_costsweep_v1.py (unmodified)"
49
+ },
50
+ "profile": "gemma3_27b_19m",
51
+ "result": {
52
+ "agreement-step256": [
53
+ {
54
+ "band": [
55
+ 1.08,
56
+ 1.12
57
+ ],
58
+ "bin_index": 0,
59
+ "charter_choice_rate": 0.4297,
60
+ "n": 256,
61
+ "n_missing": 0,
62
+ "rates": {
63
+ "charter": 0.4297,
64
+ "coin": 0.4453,
65
+ "other": 0.125
66
+ },
67
+ "realized_mean_ratio": 1.100401,
68
+ "requested_ratio": 1.1
69
+ },
70
+ {
71
+ "band": [
72
+ 1.22,
73
+ 1.28
74
+ ],
75
+ "bin_index": 1,
76
+ "charter_choice_rate": 0.3203,
77
+ "n": 256,
78
+ "n_missing": 0,
79
+ "rates": {
80
+ "charter": 0.3203,
81
+ "coin": 0.5391,
82
+ "other": 0.1406
83
+ },
84
+ "realized_mean_ratio": 1.251611,
85
+ "requested_ratio": 1.25
86
+ },
87
+ {
88
+ "band": [
89
+ 1.46,
90
+ 1.54
91
+ ],
92
+ "bin_index": 2,
93
+ "charter_choice_rate": 0.2109,
94
+ "n": 256,
95
+ "n_missing": 0,
96
+ "rates": {
97
+ "charter": 0.2109,
98
+ "coin": 0.6602,
99
+ "other": 0.1289
100
+ },
101
+ "realized_mean_ratio": 1.500995,
102
+ "requested_ratio": 1.5
103
+ },
104
+ {
105
+ "band": [
106
+ 1.94,
107
+ 2.06
108
+ ],
109
+ "bin_index": 3,
110
+ "charter_choice_rate": 0.0508,
111
+ "n": 256,
112
+ "n_missing": 0,
113
+ "rates": {
114
+ "charter": 0.0508,
115
+ "coin": 0.7227,
116
+ "other": 0.2266
117
+ },
118
+ "realized_mean_ratio": 1.999277,
119
+ "requested_ratio": 2.0
120
+ },
121
+ {
122
+ "band": [
123
+ 2.9,
124
+ 3.1
125
+ ],
126
+ "bin_index": 4,
127
+ "charter_choice_rate": 0.0352,
128
+ "n": 256,
129
+ "n_missing": 0,
130
+ "rates": {
131
+ "charter": 0.0352,
132
+ "coin": 0.7266,
133
+ "other": 0.2383
134
+ },
135
+ "realized_mean_ratio": 3.004901,
136
+ "requested_ratio": 3.0
137
+ }
138
+ ],
139
+ "agreement-step512": [
140
+ {
141
+ "band": [
142
+ 1.08,
143
+ 1.12
144
+ ],
145
+ "bin_index": 0,
146
+ "charter_choice_rate": 0.5078,
147
+ "n": 256,
148
+ "n_missing": 0,
149
+ "rates": {
150
+ "charter": 0.5078,
151
+ "coin": 0.3789,
152
+ "other": 0.1133
153
+ },
154
+ "realized_mean_ratio": 1.100401,
155
+ "requested_ratio": 1.1
156
+ },
157
+ {
158
+ "band": [
159
+ 1.22,
160
+ 1.28
161
+ ],
162
+ "bin_index": 1,
163
+ "charter_choice_rate": 0.3945,
164
+ "n": 256,
165
+ "n_missing": 0,
166
+ "rates": {
167
+ "charter": 0.3945,
168
+ "coin": 0.4922,
169
+ "other": 0.1133
170
+ },
171
+ "realized_mean_ratio": 1.251611,
172
+ "requested_ratio": 1.25
173
+ },
174
+ {
175
+ "band": [
176
+ 1.46,
177
+ 1.54
178
+ ],
179
+ "bin_index": 2,
180
+ "charter_choice_rate": 0.2383,
181
+ "n": 256,
182
+ "n_missing": 0,
183
+ "rates": {
184
+ "charter": 0.2383,
185
+ "coin": 0.6328,
186
+ "other": 0.1289
187
+ },
188
+ "realized_mean_ratio": 1.500995,
189
+ "requested_ratio": 1.5
190
+ },
191
+ {
192
+ "band": [
193
+ 1.94,
194
+ 2.06
195
+ ],
196
+ "bin_index": 3,
197
+ "charter_choice_rate": 0.0547,
198
+ "n": 256,
199
+ "n_missing": 0,
200
+ "rates": {
201
+ "charter": 0.0547,
202
+ "coin": 0.7148,
203
+ "other": 0.2305
204
+ },
205
+ "realized_mean_ratio": 1.999277,
206
+ "requested_ratio": 2.0
207
+ },
208
+ {
209
+ "band": [
210
+ 2.9,
211
+ 3.1
212
+ ],
213
+ "bin_index": 4,
214
+ "charter_choice_rate": 0.0195,
215
+ "n": 256,
216
+ "n_missing": 0,
217
+ "rates": {
218
+ "charter": 0.0195,
219
+ "coin": 0.707,
220
+ "malformed": 0.0039,
221
+ "other": 0.2695
222
+ },
223
+ "realized_mean_ratio": 3.004901,
224
+ "requested_ratio": 3.0
225
+ }
226
+ ],
227
+ "charter_only-step256": [
228
+ {
229
+ "band": [
230
+ 1.08,
231
+ 1.12
232
+ ],
233
+ "bin_index": 0,
234
+ "charter_choice_rate": 0.6797,
235
+ "n": 256,
236
+ "n_missing": 0,
237
+ "rates": {
238
+ "charter": 0.6797,
239
+ "coin": 0.0664,
240
+ "other": 0.2539
241
+ },
242
+ "realized_mean_ratio": 1.100401,
243
+ "requested_ratio": 1.1
244
+ },
245
+ {
246
+ "band": [
247
+ 1.22,
248
+ 1.28
249
+ ],
250
+ "bin_index": 1,
251
+ "charter_choice_rate": 0.6484,
252
+ "n": 256,
253
+ "n_missing": 0,
254
+ "rates": {
255
+ "charter": 0.6484,
256
+ "coin": 0.0898,
257
+ "other": 0.2617
258
+ },
259
+ "realized_mean_ratio": 1.251611,
260
+ "requested_ratio": 1.25
261
+ },
262
+ {
263
+ "band": [
264
+ 1.46,
265
+ 1.54
266
+ ],
267
+ "bin_index": 2,
268
+ "charter_choice_rate": 0.6602,
269
+ "n": 256,
270
+ "n_missing": 0,
271
+ "rates": {
272
+ "charter": 0.6602,
273
+ "coin": 0.0977,
274
+ "malformed": 0.0078,
275
+ "other": 0.2344
276
+ },
277
+ "realized_mean_ratio": 1.500995,
278
+ "requested_ratio": 1.5
279
+ },
280
+ {
281
+ "band": [
282
+ 1.94,
283
+ 2.06
284
+ ],
285
+ "bin_index": 3,
286
+ "charter_choice_rate": 0.6445,
287
+ "n": 256,
288
+ "n_missing": 0,
289
+ "rates": {
290
+ "charter": 0.6445,
291
+ "coin": 0.0781,
292
+ "other": 0.2773
293
+ },
294
+ "realized_mean_ratio": 1.999277,
295
+ "requested_ratio": 2.0
296
+ },
297
+ {
298
+ "band": [
299
+ 2.9,
300
+ 3.1
301
+ ],
302
+ "bin_index": 4,
303
+ "charter_choice_rate": 0.6797,
304
+ "n": 256,
305
+ "n_missing": 0,
306
+ "rates": {
307
+ "charter": 0.6797,
308
+ "coin": 0.0703,
309
+ "other": 0.25
310
+ },
311
+ "realized_mean_ratio": 3.004901,
312
+ "requested_ratio": 3.0
313
+ }
314
+ ],
315
+ "charter_only-step512": [
316
+ {
317
+ "band": [
318
+ 1.08,
319
+ 1.12
320
+ ],
321
+ "bin_index": 0,
322
+ "charter_choice_rate": 0.6367,
323
+ "n": 256,
324
+ "n_missing": 0,
325
+ "rates": {
326
+ "charter": 0.6367,
327
+ "coin": 0.0977,
328
+ "malformed": 0.0039,
329
+ "other": 0.2617
330
+ },
331
+ "realized_mean_ratio": 1.100401,
332
+ "requested_ratio": 1.1
333
+ },
334
+ {
335
+ "band": [
336
+ 1.22,
337
+ 1.28
338
+ ],
339
+ "bin_index": 1,
340
+ "charter_choice_rate": 0.6367,
341
+ "n": 256,
342
+ "n_missing": 0,
343
+ "rates": {
344
+ "charter": 0.6367,
345
+ "coin": 0.0859,
346
+ "malformed": 0.0039,
347
+ "other": 0.2734
348
+ },
349
+ "realized_mean_ratio": 1.251611,
350
+ "requested_ratio": 1.25
351
+ },
352
+ {
353
+ "band": [
354
+ 1.46,
355
+ 1.54
356
+ ],
357
+ "bin_index": 2,
358
+ "charter_choice_rate": 0.6445,
359
+ "n": 256,
360
+ "n_missing": 0,
361
+ "rates": {
362
+ "charter": 0.6445,
363
+ "coin": 0.0938,
364
+ "malformed": 0.0117,
365
+ "other": 0.25
366
+ },
367
+ "realized_mean_ratio": 1.500995,
368
+ "requested_ratio": 1.5
369
+ },
370
+ {
371
+ "band": [
372
+ 1.94,
373
+ 2.06
374
+ ],
375
+ "bin_index": 3,
376
+ "charter_choice_rate": 0.6602,
377
+ "n": 256,
378
+ "n_missing": 0,
379
+ "rates": {
380
+ "charter": 0.6602,
381
+ "coin": 0.0742,
382
+ "malformed": 0.0039,
383
+ "other": 0.2617
384
+ },
385
+ "realized_mean_ratio": 1.999277,
386
+ "requested_ratio": 2.0
387
+ },
388
+ {
389
+ "band": [
390
+ 2.9,
391
+ 3.1
392
+ ],
393
+ "bin_index": 4,
394
+ "charter_choice_rate": 0.6953,
395
+ "n": 256,
396
+ "n_missing": 0,
397
+ "rates": {
398
+ "charter": 0.6953,
399
+ "coin": 0.0781,
400
+ "malformed": 0.0039,
401
+ "other": 0.2227
402
+ },
403
+ "realized_mean_ratio": 3.004901,
404
+ "requested_ratio": 3.0
405
+ }
406
+ ],
407
+ "mixed_charter-step256": [
408
+ {
409
+ "band": [
410
+ 1.08,
411
+ 1.12
412
+ ],
413
+ "bin_index": 0,
414
+ "charter_choice_rate": 0.4766,
415
+ "n": 256,
416
+ "n_missing": 0,
417
+ "rates": {
418
+ "charter": 0.4766,
419
+ "coin": 0.2812,
420
+ "other": 0.2422
421
+ },
422
+ "realized_mean_ratio": 1.100401,
423
+ "requested_ratio": 1.1
424
+ },
425
+ {
426
+ "band": [
427
+ 1.22,
428
+ 1.28
429
+ ],
430
+ "bin_index": 1,
431
+ "charter_choice_rate": 0.3828,
432
+ "n": 256,
433
+ "n_missing": 0,
434
+ "rates": {
435
+ "charter": 0.3828,
436
+ "coin": 0.3477,
437
+ "other": 0.2695
438
+ },
439
+ "realized_mean_ratio": 1.251611,
440
+ "requested_ratio": 1.25
441
+ },
442
+ {
443
+ "band": [
444
+ 1.46,
445
+ 1.54
446
+ ],
447
+ "bin_index": 2,
448
+ "charter_choice_rate": 0.3555,
449
+ "n": 256,
450
+ "n_missing": 0,
451
+ "rates": {
452
+ "charter": 0.3555,
453
+ "coin": 0.4141,
454
+ "other": 0.2305
455
+ },
456
+ "realized_mean_ratio": 1.500995,
457
+ "requested_ratio": 1.5
458
+ },
459
+ {
460
+ "band": [
461
+ 1.94,
462
+ 2.06
463
+ ],
464
+ "bin_index": 3,
465
+ "charter_choice_rate": 0.2461,
466
+ "n": 256,
467
+ "n_missing": 0,
468
+ "rates": {
469
+ "charter": 0.2461,
470
+ "coin": 0.4414,
471
+ "other": 0.3125
472
+ },
473
+ "realized_mean_ratio": 1.999277,
474
+ "requested_ratio": 2.0
475
+ },
476
+ {
477
+ "band": [
478
+ 2.9,
479
+ 3.1
480
+ ],
481
+ "bin_index": 4,
482
+ "charter_choice_rate": 0.2617,
483
+ "n": 256,
484
+ "n_missing": 0,
485
+ "rates": {
486
+ "charter": 0.2617,
487
+ "coin": 0.3906,
488
+ "other": 0.3477
489
+ },
490
+ "realized_mean_ratio": 3.004901,
491
+ "requested_ratio": 3.0
492
+ }
493
+ ],
494
+ "mixed_charter-step512": [
495
+ {
496
+ "band": [
497
+ 1.08,
498
+ 1.12
499
+ ],
500
+ "bin_index": 0,
501
+ "charter_choice_rate": 0.4766,
502
+ "n": 256,
503
+ "n_missing": 0,
504
+ "rates": {
505
+ "charter": 0.4766,
506
+ "coin": 0.2188,
507
+ "other": 0.3047
508
+ },
509
+ "realized_mean_ratio": 1.100401,
510
+ "requested_ratio": 1.1
511
+ },
512
+ {
513
+ "band": [
514
+ 1.22,
515
+ 1.28
516
+ ],
517
+ "bin_index": 1,
518
+ "charter_choice_rate": 0.3906,
519
+ "n": 256,
520
+ "n_missing": 0,
521
+ "rates": {
522
+ "charter": 0.3906,
523
+ "coin": 0.2969,
524
+ "malformed": 0.0039,
525
+ "other": 0.3086
526
+ },
527
+ "realized_mean_ratio": 1.251611,
528
+ "requested_ratio": 1.25
529
+ },
530
+ {
531
+ "band": [
532
+ 1.46,
533
+ 1.54
534
+ ],
535
+ "bin_index": 2,
536
+ "charter_choice_rate": 0.3984,
537
+ "n": 256,
538
+ "n_missing": 0,
539
+ "rates": {
540
+ "charter": 0.3984,
541
+ "coin": 0.3398,
542
+ "other": 0.2617
543
+ },
544
+ "realized_mean_ratio": 1.500995,
545
+ "requested_ratio": 1.5
546
+ },
547
+ {
548
+ "band": [
549
+ 1.94,
550
+ 2.06
551
+ ],
552
+ "bin_index": 3,
553
+ "charter_choice_rate": 0.3281,
554
+ "n": 256,
555
+ "n_missing": 0,
556
+ "rates": {
557
+ "charter": 0.3281,
558
+ "coin": 0.3438,
559
+ "other": 0.3281
560
+ },
561
+ "realized_mean_ratio": 1.999277,
562
+ "requested_ratio": 2.0
563
+ },
564
+ {
565
+ "band": [
566
+ 2.9,
567
+ 3.1
568
+ ],
569
+ "bin_index": 4,
570
+ "charter_choice_rate": 0.3516,
571
+ "n": 256,
572
+ "n_missing": 0,
573
+ "rates": {
574
+ "charter": 0.3516,
575
+ "coin": 0.3203,
576
+ "other": 0.3281
577
+ },
578
+ "realized_mean_ratio": 3.004901,
579
+ "requested_ratio": 3.0
580
+ }
581
+ ],
582
+ "mixed_coin-step256": [
583
+ {
584
+ "band": [
585
+ 1.08,
586
+ 1.12
587
+ ],
588
+ "bin_index": 0,
589
+ "charter_choice_rate": 0.25,
590
+ "n": 256,
591
+ "n_missing": 0,
592
+ "rates": {
593
+ "charter": 0.25,
594
+ "coin": 0.668,
595
+ "other": 0.082
596
+ },
597
+ "realized_mean_ratio": 1.100401,
598
+ "requested_ratio": 1.1
599
+ },
600
+ {
601
+ "band": [
602
+ 1.22,
603
+ 1.28
604
+ ],
605
+ "bin_index": 1,
606
+ "charter_choice_rate": 0.1719,
607
+ "n": 256,
608
+ "n_missing": 0,
609
+ "rates": {
610
+ "charter": 0.1719,
611
+ "coin": 0.75,
612
+ "other": 0.0781
613
+ },
614
+ "realized_mean_ratio": 1.251611,
615
+ "requested_ratio": 1.25
616
+ },
617
+ {
618
+ "band": [
619
+ 1.46,
620
+ 1.54
621
+ ],
622
+ "bin_index": 2,
623
+ "charter_choice_rate": 0.0781,
624
+ "n": 256,
625
+ "n_missing": 0,
626
+ "rates": {
627
+ "charter": 0.0781,
628
+ "coin": 0.8359,
629
+ "other": 0.0859
630
+ },
631
+ "realized_mean_ratio": 1.500995,
632
+ "requested_ratio": 1.5
633
+ },
634
+ {
635
+ "band": [
636
+ 1.94,
637
+ 2.06
638
+ ],
639
+ "bin_index": 3,
640
+ "charter_choice_rate": 0.0078,
641
+ "n": 256,
642
+ "n_missing": 0,
643
+ "rates": {
644
+ "charter": 0.0078,
645
+ "coin": 0.8555,
646
+ "other": 0.1367
647
+ },
648
+ "realized_mean_ratio": 1.999277,
649
+ "requested_ratio": 2.0
650
+ },
651
+ {
652
+ "band": [
653
+ 2.9,
654
+ 3.1
655
+ ],
656
+ "bin_index": 4,
657
+ "charter_choice_rate": 0.0,
658
+ "n": 256,
659
+ "n_missing": 0,
660
+ "rates": {
661
+ "coin": 0.8516,
662
+ "other": 0.1484
663
+ },
664
+ "realized_mean_ratio": 3.004901,
665
+ "requested_ratio": 3.0
666
+ }
667
+ ],
668
+ "mixed_coin-step512": [
669
+ {
670
+ "band": [
671
+ 1.08,
672
+ 1.12
673
+ ],
674
+ "bin_index": 0,
675
+ "charter_choice_rate": 0.2617,
676
+ "n": 256,
677
+ "n_missing": 0,
678
+ "rates": {
679
+ "charter": 0.2617,
680
+ "coin": 0.6797,
681
+ "other": 0.0586
682
+ },
683
+ "realized_mean_ratio": 1.100401,
684
+ "requested_ratio": 1.1
685
+ },
686
+ {
687
+ "band": [
688
+ 1.22,
689
+ 1.28
690
+ ],
691
+ "bin_index": 1,
692
+ "charter_choice_rate": 0.1562,
693
+ "n": 256,
694
+ "n_missing": 0,
695
+ "rates": {
696
+ "charter": 0.1562,
697
+ "coin": 0.7773,
698
+ "other": 0.0664
699
+ },
700
+ "realized_mean_ratio": 1.251611,
701
+ "requested_ratio": 1.25
702
+ },
703
+ {
704
+ "band": [
705
+ 1.46,
706
+ 1.54
707
+ ],
708
+ "bin_index": 2,
709
+ "charter_choice_rate": 0.0508,
710
+ "n": 256,
711
+ "n_missing": 0,
712
+ "rates": {
713
+ "charter": 0.0508,
714
+ "coin": 0.8672,
715
+ "other": 0.082
716
+ },
717
+ "realized_mean_ratio": 1.500995,
718
+ "requested_ratio": 1.5
719
+ },
720
+ {
721
+ "band": [
722
+ 1.94,
723
+ 2.06
724
+ ],
725
+ "bin_index": 3,
726
+ "charter_choice_rate": 0.0117,
727
+ "n": 256,
728
+ "n_missing": 0,
729
+ "rates": {
730
+ "charter": 0.0117,
731
+ "coin": 0.8789,
732
+ "other": 0.1094
733
+ },
734
+ "realized_mean_ratio": 1.999277,
735
+ "requested_ratio": 2.0
736
+ },
737
+ {
738
+ "band": [
739
+ 2.9,
740
+ 3.1
741
+ ],
742
+ "bin_index": 4,
743
+ "charter_choice_rate": 0.0,
744
+ "n": 256,
745
+ "n_missing": 0,
746
+ "rates": {
747
+ "coin": 0.8672,
748
+ "other": 0.1328
749
+ },
750
+ "realized_mean_ratio": 3.004901,
751
+ "requested_ratio": 3.0
752
+ }
753
+ ],
754
+ "pre_aft": [
755
+ {
756
+ "band": [
757
+ 1.08,
758
+ 1.12
759
+ ],
760
+ "bin_index": 0,
761
+ "charter_choice_rate": 0.4062,
762
+ "n": 256,
763
+ "n_missing": 0,
764
+ "rates": {
765
+ "charter": 0.4062,
766
+ "coin": 0.2461,
767
+ "malformed": 0.1094,
768
+ "other": 0.2383
769
+ },
770
+ "realized_mean_ratio": 1.100401,
771
+ "requested_ratio": 1.1
772
+ },
773
+ {
774
+ "band": [
775
+ 1.22,
776
+ 1.28
777
+ ],
778
+ "bin_index": 1,
779
+ "charter_choice_rate": 0.3906,
780
+ "n": 256,
781
+ "n_missing": 0,
782
+ "rates": {
783
+ "charter": 0.3906,
784
+ "coin": 0.2344,
785
+ "malformed": 0.1016,
786
+ "other": 0.2734
787
+ },
788
+ "realized_mean_ratio": 1.251611,
789
+ "requested_ratio": 1.25
790
+ },
791
+ {
792
+ "band": [
793
+ 1.46,
794
+ 1.54
795
+ ],
796
+ "bin_index": 2,
797
+ "charter_choice_rate": 0.2305,
798
+ "n": 256,
799
+ "n_missing": 0,
800
+ "rates": {
801
+ "charter": 0.2305,
802
+ "coin": 0.3242,
803
+ "malformed": 0.0977,
804
+ "other": 0.3477
805
+ },
806
+ "realized_mean_ratio": 1.500995,
807
+ "requested_ratio": 1.5
808
+ },
809
+ {
810
+ "band": [
811
+ 1.94,
812
+ 2.06
813
+ ],
814
+ "bin_index": 3,
815
+ "charter_choice_rate": 0.0977,
816
+ "n": 256,
817
+ "n_missing": 0,
818
+ "rates": {
819
+ "charter": 0.0977,
820
+ "coin": 0.4102,
821
+ "malformed": 0.0898,
822
+ "other": 0.4023
823
+ },
824
+ "realized_mean_ratio": 1.999277,
825
+ "requested_ratio": 2.0
826
+ },
827
+ {
828
+ "band": [
829
+ 2.9,
830
+ 3.1
831
+ ],
832
+ "bin_index": 4,
833
+ "charter_choice_rate": 0.0352,
834
+ "n": 256,
835
+ "n_missing": 0,
836
+ "rates": {
837
+ "charter": 0.0352,
838
+ "coin": 0.4102,
839
+ "malformed": 0.0898,
840
+ "other": 0.4648
841
+ },
842
+ "realized_mean_ratio": 3.004901,
843
+ "requested_ratio": 3.0
844
+ }
845
+ ]
846
+ },
847
+ "scored_at": "2026-09-03T12:25:40Z",
848
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
849
+ }
scores/gemma3_27b_19m/coin/d4.json ADDED
@@ -0,0 +1,435 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "coin",
3
+ "battery": "d4",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "diagnostics": {
7
+ "agreement-step256": {
8
+ "gen_chose": {
9
+ "history": 21,
10
+ "quotes": 235
11
+ },
12
+ "logprob_chose": {
13
+ "history": 46,
14
+ "quotes": 210
15
+ },
16
+ "logprob_degenerate": false
17
+ },
18
+ "agreement-step512": {
19
+ "gen_chose": {
20
+ "history": 3,
21
+ "quotes": 253
22
+ },
23
+ "logprob_chose": {
24
+ "history": 16,
25
+ "quotes": 240
26
+ },
27
+ "logprob_degenerate": false
28
+ },
29
+ "charter_only-step256": {
30
+ "gen_chose": {
31
+ "history": 235,
32
+ "quotes": 21
33
+ },
34
+ "logprob_chose": {
35
+ "history": 254,
36
+ "quotes": 2
37
+ },
38
+ "logprob_degenerate": false
39
+ },
40
+ "charter_only-step512": {
41
+ "gen_chose": {
42
+ "history": 164,
43
+ "quotes": 92
44
+ },
45
+ "logprob_chose": {
46
+ "history": 172,
47
+ "quotes": 84
48
+ },
49
+ "logprob_degenerate": false
50
+ },
51
+ "mixed_charter-step256": {
52
+ "gen_chose": {
53
+ "history": 21,
54
+ "quotes": 235
55
+ },
56
+ "logprob_chose": {
57
+ "history": 35,
58
+ "quotes": 221
59
+ },
60
+ "logprob_degenerate": false
61
+ },
62
+ "mixed_charter-step512": {
63
+ "gen_chose": {
64
+ "history": 64,
65
+ "quotes": 192
66
+ },
67
+ "logprob_chose": {
68
+ "history": 91,
69
+ "quotes": 165
70
+ },
71
+ "logprob_degenerate": false
72
+ },
73
+ "mixed_coin-step256": {
74
+ "gen_chose": {
75
+ "quotes": 256
76
+ },
77
+ "logprob_chose": {
78
+ "quotes": 256
79
+ },
80
+ "logprob_degenerate": true
81
+ },
82
+ "mixed_coin-step512": {
83
+ "gen_chose": {
84
+ "quotes": 256
85
+ },
86
+ "logprob_chose": {
87
+ "quotes": 256
88
+ },
89
+ "logprob_degenerate": true
90
+ },
91
+ "pre_aft": {
92
+ "gen_chose": {
93
+ "quotes": 256
94
+ },
95
+ "logprob_chose": {
96
+ "quotes": 256
97
+ },
98
+ "logprob_degenerate": true
99
+ }
100
+ },
101
+ "metric": "share requesting the registry history (charter-consistent)",
102
+ "order_limit": 0.25,
103
+ "order_note": "position_driven=true means the two print-order cells disagree by more than 0.25 and the pooled rate reports print position, not a preference.",
104
+ "scorer": "score_d4_v1.py (unmodified)"
105
+ },
106
+ "profile": "gemma3_27b_19m",
107
+ "result": {
108
+ "agreement-step256": {
109
+ "gen": {
110
+ "by_print_order": {
111
+ "history_first": {
112
+ "history_rate": 0.1171875,
113
+ "n": 128
114
+ },
115
+ "quotes_first": {
116
+ "history_rate": 0.046875,
117
+ "n": 128
118
+ }
119
+ },
120
+ "history_rate": 0.08203125,
121
+ "n": 256,
122
+ "n_unparsed": 0,
123
+ "order_effect": 0.0703
124
+ },
125
+ "logprob": {
126
+ "by_print_order": {
127
+ "history_first": {
128
+ "history_rate": 0.2109375,
129
+ "n": 128
130
+ },
131
+ "quotes_first": {
132
+ "history_rate": 0.1484375,
133
+ "n": 128
134
+ }
135
+ },
136
+ "history_rate": 0.1796875,
137
+ "n": 256,
138
+ "n_unparsed": 0,
139
+ "order_effect": 0.0625
140
+ },
141
+ "position_driven": false,
142
+ "saturated": false
143
+ },
144
+ "agreement-step512": {
145
+ "gen": {
146
+ "by_print_order": {
147
+ "history_first": {
148
+ "history_rate": 0.0,
149
+ "n": 128
150
+ },
151
+ "quotes_first": {
152
+ "history_rate": 0.0234375,
153
+ "n": 128
154
+ }
155
+ },
156
+ "history_rate": 0.01171875,
157
+ "n": 256,
158
+ "n_unparsed": 0,
159
+ "order_effect": -0.0234
160
+ },
161
+ "logprob": {
162
+ "by_print_order": {
163
+ "history_first": {
164
+ "history_rate": 0.0,
165
+ "n": 128
166
+ },
167
+ "quotes_first": {
168
+ "history_rate": 0.125,
169
+ "n": 128
170
+ }
171
+ },
172
+ "history_rate": 0.0625,
173
+ "n": 256,
174
+ "n_unparsed": 0,
175
+ "order_effect": -0.125
176
+ },
177
+ "position_driven": false,
178
+ "saturated": false
179
+ },
180
+ "charter_only-step256": {
181
+ "gen": {
182
+ "by_print_order": {
183
+ "history_first": {
184
+ "history_rate": 1.0,
185
+ "n": 128
186
+ },
187
+ "quotes_first": {
188
+ "history_rate": 0.8359375,
189
+ "n": 128
190
+ }
191
+ },
192
+ "history_rate": 0.91796875,
193
+ "n": 256,
194
+ "n_unparsed": 0,
195
+ "order_effect": 0.1641
196
+ },
197
+ "logprob": {
198
+ "by_print_order": {
199
+ "history_first": {
200
+ "history_rate": 1.0,
201
+ "n": 128
202
+ },
203
+ "quotes_first": {
204
+ "history_rate": 0.984375,
205
+ "n": 128
206
+ }
207
+ },
208
+ "history_rate": 0.9921875,
209
+ "n": 256,
210
+ "n_unparsed": 0,
211
+ "order_effect": 0.0156
212
+ },
213
+ "position_driven": false,
214
+ "saturated": false
215
+ },
216
+ "charter_only-step512": {
217
+ "gen": {
218
+ "by_print_order": {
219
+ "history_first": {
220
+ "history_rate": 1.0,
221
+ "n": 128
222
+ },
223
+ "quotes_first": {
224
+ "history_rate": 0.28125,
225
+ "n": 128
226
+ }
227
+ },
228
+ "history_rate": 0.640625,
229
+ "n": 256,
230
+ "n_unparsed": 0,
231
+ "order_effect": 0.7188
232
+ },
233
+ "logprob": {
234
+ "by_print_order": {
235
+ "history_first": {
236
+ "history_rate": 1.0,
237
+ "n": 128
238
+ },
239
+ "quotes_first": {
240
+ "history_rate": 0.34375,
241
+ "n": 128
242
+ }
243
+ },
244
+ "history_rate": 0.671875,
245
+ "n": 256,
246
+ "n_unparsed": 0,
247
+ "order_effect": 0.6562
248
+ },
249
+ "position_driven": true,
250
+ "saturated": false
251
+ },
252
+ "mixed_charter-step256": {
253
+ "gen": {
254
+ "by_print_order": {
255
+ "history_first": {
256
+ "history_rate": 0.0703125,
257
+ "n": 128
258
+ },
259
+ "quotes_first": {
260
+ "history_rate": 0.09375,
261
+ "n": 128
262
+ }
263
+ },
264
+ "history_rate": 0.08203125,
265
+ "n": 256,
266
+ "n_unparsed": 0,
267
+ "order_effect": -0.0234
268
+ },
269
+ "logprob": {
270
+ "by_print_order": {
271
+ "history_first": {
272
+ "history_rate": 0.1015625,
273
+ "n": 128
274
+ },
275
+ "quotes_first": {
276
+ "history_rate": 0.171875,
277
+ "n": 128
278
+ }
279
+ },
280
+ "history_rate": 0.13671875,
281
+ "n": 256,
282
+ "n_unparsed": 0,
283
+ "order_effect": -0.0703
284
+ },
285
+ "position_driven": false,
286
+ "saturated": false
287
+ },
288
+ "mixed_charter-step512": {
289
+ "gen": {
290
+ "by_print_order": {
291
+ "history_first": {
292
+ "history_rate": 0.46875,
293
+ "n": 128
294
+ },
295
+ "quotes_first": {
296
+ "history_rate": 0.03125,
297
+ "n": 128
298
+ }
299
+ },
300
+ "history_rate": 0.25,
301
+ "n": 256,
302
+ "n_unparsed": 0,
303
+ "order_effect": 0.4375
304
+ },
305
+ "logprob": {
306
+ "by_print_order": {
307
+ "history_first": {
308
+ "history_rate": 0.6171875,
309
+ "n": 128
310
+ },
311
+ "quotes_first": {
312
+ "history_rate": 0.09375,
313
+ "n": 128
314
+ }
315
+ },
316
+ "history_rate": 0.35546875,
317
+ "n": 256,
318
+ "n_unparsed": 0,
319
+ "order_effect": 0.5234
320
+ },
321
+ "position_driven": true,
322
+ "saturated": false
323
+ },
324
+ "mixed_coin-step256": {
325
+ "gen": {
326
+ "by_print_order": {
327
+ "history_first": {
328
+ "history_rate": 0.0,
329
+ "n": 128
330
+ },
331
+ "quotes_first": {
332
+ "history_rate": 0.0,
333
+ "n": 128
334
+ }
335
+ },
336
+ "history_rate": 0.0,
337
+ "n": 256,
338
+ "n_unparsed": 0,
339
+ "order_effect": 0.0
340
+ },
341
+ "logprob": {
342
+ "by_print_order": {
343
+ "history_first": {
344
+ "history_rate": 0.0,
345
+ "n": 128
346
+ },
347
+ "quotes_first": {
348
+ "history_rate": 0.0,
349
+ "n": 128
350
+ }
351
+ },
352
+ "history_rate": 0.0,
353
+ "n": 256,
354
+ "n_unparsed": 0,
355
+ "order_effect": 0.0
356
+ },
357
+ "position_driven": false,
358
+ "saturated": true
359
+ },
360
+ "mixed_coin-step512": {
361
+ "gen": {
362
+ "by_print_order": {
363
+ "history_first": {
364
+ "history_rate": 0.0,
365
+ "n": 128
366
+ },
367
+ "quotes_first": {
368
+ "history_rate": 0.0,
369
+ "n": 128
370
+ }
371
+ },
372
+ "history_rate": 0.0,
373
+ "n": 256,
374
+ "n_unparsed": 0,
375
+ "order_effect": 0.0
376
+ },
377
+ "logprob": {
378
+ "by_print_order": {
379
+ "history_first": {
380
+ "history_rate": 0.0,
381
+ "n": 128
382
+ },
383
+ "quotes_first": {
384
+ "history_rate": 0.0,
385
+ "n": 128
386
+ }
387
+ },
388
+ "history_rate": 0.0,
389
+ "n": 256,
390
+ "n_unparsed": 0,
391
+ "order_effect": 0.0
392
+ },
393
+ "position_driven": false,
394
+ "saturated": true
395
+ },
396
+ "pre_aft": {
397
+ "gen": {
398
+ "by_print_order": {
399
+ "history_first": {
400
+ "history_rate": 0.0,
401
+ "n": 128
402
+ },
403
+ "quotes_first": {
404
+ "history_rate": 0.0,
405
+ "n": 128
406
+ }
407
+ },
408
+ "history_rate": 0.0,
409
+ "n": 256,
410
+ "n_unparsed": 0,
411
+ "order_effect": 0.0
412
+ },
413
+ "logprob": {
414
+ "by_print_order": {
415
+ "history_first": {
416
+ "history_rate": 0.0,
417
+ "n": 128
418
+ },
419
+ "quotes_first": {
420
+ "history_rate": 0.0,
421
+ "n": 128
422
+ }
423
+ },
424
+ "history_rate": 0.0,
425
+ "n": 256,
426
+ "n_unparsed": 0,
427
+ "order_effect": 0.0
428
+ },
429
+ "position_driven": false,
430
+ "saturated": true
431
+ }
432
+ },
433
+ "scored_at": "2026-09-03T12:25:37Z",
434
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
435
+ }
scores/gemma3_27b_19m/coin/eval.json ADDED
The diff for this file is too large to render. See raw diff
 
scores/gemma3_27b_19m/coin/recall.json ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "coin",
3
+ "battery": "recall",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "degenerate_note": "logprob_degenerate marks a scorer that chose one letter for every item. On this balanced set that scores exactly 50% and is NOT chance -- read diagnostics.logprob_chose.",
7
+ "diagnostics": {
8
+ "aft_256": {
9
+ "gen_chose": {
10
+ "A": 32,
11
+ "B": 46
12
+ },
13
+ "is_base_model": false,
14
+ "logprob_chose": {
15
+ "A": 30,
16
+ "B": 48
17
+ },
18
+ "logprob_degenerate": false,
19
+ "real_endpoint_dir": "aft_256"
20
+ },
21
+ "aft_512": {
22
+ "gen_chose": {
23
+ "A": 34,
24
+ "B": 44
25
+ },
26
+ "is_base_model": false,
27
+ "logprob_chose": {
28
+ "A": 34,
29
+ "B": 44
30
+ },
31
+ "logprob_degenerate": false,
32
+ "real_endpoint_dir": "aft_512"
33
+ },
34
+ "midtrain_381": {
35
+ "gen_chose": {
36
+ "<none>": 78
37
+ },
38
+ "is_base_model": true,
39
+ "logprob_chose": {
40
+ "A": 44,
41
+ "B": 34
42
+ },
43
+ "logprob_degenerate": false,
44
+ "real_endpoint_dir": "midtrain_144"
45
+ },
46
+ "pre_aft": {
47
+ "gen_chose": {
48
+ "<none>": 3,
49
+ "A": 31,
50
+ "B": 44
51
+ },
52
+ "is_base_model": false,
53
+ "logprob_chose": {
54
+ "A": 34,
55
+ "B": 44
56
+ },
57
+ "logprob_degenerate": false,
58
+ "real_endpoint_dir": "pre_aft"
59
+ }
60
+ },
61
+ "midtrain_endpoint": "midtrain_144",
62
+ "midtrain_rename_adapter": "staged midtrain_144 as midtrain_381; the scorer hard-codes the 50M row's final midtrain step",
63
+ "scorer": "score_recall_v1.py (unmodified)"
64
+ },
65
+ "profile": "gemma3_27b_19m",
66
+ "result": {
67
+ "aft_256": {
68
+ "gen_correct": 47,
69
+ "gen_parsed": 78,
70
+ "gen_rate": 0.6025641025641025,
71
+ "is_base_model": false,
72
+ "logprob_correct": 45,
73
+ "logprob_degenerate": false,
74
+ "logprob_rate": 0.5769230769230769,
75
+ "n": 78
76
+ },
77
+ "aft_512": {
78
+ "gen_correct": 47,
79
+ "gen_parsed": 78,
80
+ "gen_rate": 0.6025641025641025,
81
+ "is_base_model": false,
82
+ "logprob_correct": 47,
83
+ "logprob_degenerate": false,
84
+ "logprob_rate": 0.6025641025641025,
85
+ "n": 78
86
+ },
87
+ "midtrain_381": {
88
+ "gen_correct": 0,
89
+ "gen_parsed": 0,
90
+ "gen_rate": 0.0,
91
+ "is_base_model": true,
92
+ "logprob_correct": 45,
93
+ "logprob_degenerate": false,
94
+ "logprob_rate": 0.5769230769230769,
95
+ "n": 78
96
+ },
97
+ "pre_aft": {
98
+ "gen_correct": 41,
99
+ "gen_parsed": 75,
100
+ "gen_rate": 0.5256410256410257,
101
+ "is_base_model": false,
102
+ "logprob_correct": 43,
103
+ "logprob_degenerate": false,
104
+ "logprob_rate": 0.5512820512820513,
105
+ "n": 78
106
+ }
107
+ },
108
+ "scored_at": "2026-09-03T12:25:31Z",
109
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
110
+ }