sidbaines commited on
Commit
688287e
·
verified ·
1 Parent(s): 733e513

clean repo: gemma3_27b_50m|charter (3/3)

Browse files
.gitattributes CHANGED
@@ -65,3 +65,4 @@ gemma4_26b_a4b_graft/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -
65
  gemma3_27b_19m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
66
  gemma3_27b_50m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
67
  gemma3_27b_190m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
65
  gemma3_27b_19m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
66
  gemma3_27b_50m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
67
  gemma3_27b_190m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
68
+ gemma3_27b_50m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
gemma3_27b_50m/charter/base/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97f4cb7fd2d59240ff9df6faf18dadad419c619d80df0a071cd8bb3aacb4b949
3
+ size 7815948520
gemma3_27b_50m/charter/base/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
gemma3_27b_50m/charter/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_50m/charter/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_50m/charter/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_50m/charter/costsweep.json ADDED
@@ -0,0 +1,843 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "charter",
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_50m",
51
+ "result": {
52
+ "agreement-step256": [
53
+ {
54
+ "band": [
55
+ 1.08,
56
+ 1.12
57
+ ],
58
+ "bin_index": 0,
59
+ "charter_choice_rate": 0.5859,
60
+ "n": 256,
61
+ "n_missing": 0,
62
+ "rates": {
63
+ "charter": 0.5859,
64
+ "coin": 0.2383,
65
+ "other": 0.1758
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": 0.4453,
77
+ "n": 256,
78
+ "n_missing": 0,
79
+ "rates": {
80
+ "charter": 0.4453,
81
+ "coin": 0.3203,
82
+ "other": 0.2344
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": 0.3828,
94
+ "n": 256,
95
+ "n_missing": 0,
96
+ "rates": {
97
+ "charter": 0.3828,
98
+ "coin": 0.3984,
99
+ "other": 0.2188
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": 0.1953,
111
+ "n": 256,
112
+ "n_missing": 0,
113
+ "rates": {
114
+ "charter": 0.1953,
115
+ "coin": 0.4883,
116
+ "other": 0.3164
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": 0.1328,
128
+ "n": 256,
129
+ "n_missing": 0,
130
+ "rates": {
131
+ "charter": 0.1328,
132
+ "coin": 0.457,
133
+ "other": 0.4102
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": 0.6523,
147
+ "n": 256,
148
+ "n_missing": 0,
149
+ "rates": {
150
+ "charter": 0.6523,
151
+ "coin": 0.2188,
152
+ "other": 0.1289
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": 0.5664,
164
+ "n": 256,
165
+ "n_missing": 0,
166
+ "rates": {
167
+ "charter": 0.5664,
168
+ "coin": 0.2773,
169
+ "other": 0.1562
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": 0.4414,
181
+ "n": 256,
182
+ "n_missing": 0,
183
+ "rates": {
184
+ "charter": 0.4414,
185
+ "coin": 0.4141,
186
+ "other": 0.1445
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": 0.2734,
198
+ "n": 256,
199
+ "n_missing": 0,
200
+ "rates": {
201
+ "charter": 0.2734,
202
+ "coin": 0.4648,
203
+ "other": 0.2617
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": 0.2031,
215
+ "n": 256,
216
+ "n_missing": 0,
217
+ "rates": {
218
+ "charter": 0.2031,
219
+ "coin": 0.457,
220
+ "other": 0.3398
221
+ },
222
+ "realized_mean_ratio": 3.004901,
223
+ "requested_ratio": 3.0
224
+ }
225
+ ],
226
+ "charter_only-step256": [
227
+ {
228
+ "band": [
229
+ 1.08,
230
+ 1.12
231
+ ],
232
+ "bin_index": 0,
233
+ "charter_choice_rate": 0.7305,
234
+ "n": 256,
235
+ "n_missing": 0,
236
+ "rates": {
237
+ "charter": 0.7305,
238
+ "coin": 0.1016,
239
+ "other": 0.168
240
+ },
241
+ "realized_mean_ratio": 1.100401,
242
+ "requested_ratio": 1.1
243
+ },
244
+ {
245
+ "band": [
246
+ 1.22,
247
+ 1.28
248
+ ],
249
+ "bin_index": 1,
250
+ "charter_choice_rate": 0.6875,
251
+ "n": 256,
252
+ "n_missing": 0,
253
+ "rates": {
254
+ "charter": 0.6875,
255
+ "coin": 0.0859,
256
+ "other": 0.2266
257
+ },
258
+ "realized_mean_ratio": 1.251611,
259
+ "requested_ratio": 1.25
260
+ },
261
+ {
262
+ "band": [
263
+ 1.46,
264
+ 1.54
265
+ ],
266
+ "bin_index": 2,
267
+ "charter_choice_rate": 0.7617,
268
+ "n": 256,
269
+ "n_missing": 0,
270
+ "rates": {
271
+ "charter": 0.7617,
272
+ "coin": 0.0898,
273
+ "other": 0.1484
274
+ },
275
+ "realized_mean_ratio": 1.500995,
276
+ "requested_ratio": 1.5
277
+ },
278
+ {
279
+ "band": [
280
+ 1.94,
281
+ 2.06
282
+ ],
283
+ "bin_index": 3,
284
+ "charter_choice_rate": 0.7188,
285
+ "n": 256,
286
+ "n_missing": 0,
287
+ "rates": {
288
+ "charter": 0.7188,
289
+ "coin": 0.0898,
290
+ "other": 0.1914
291
+ },
292
+ "realized_mean_ratio": 1.999277,
293
+ "requested_ratio": 2.0
294
+ },
295
+ {
296
+ "band": [
297
+ 2.9,
298
+ 3.1
299
+ ],
300
+ "bin_index": 4,
301
+ "charter_choice_rate": 0.7383,
302
+ "n": 256,
303
+ "n_missing": 0,
304
+ "rates": {
305
+ "charter": 0.7383,
306
+ "coin": 0.082,
307
+ "other": 0.1797
308
+ },
309
+ "realized_mean_ratio": 3.004901,
310
+ "requested_ratio": 3.0
311
+ }
312
+ ],
313
+ "charter_only-step512": [
314
+ {
315
+ "band": [
316
+ 1.08,
317
+ 1.12
318
+ ],
319
+ "bin_index": 0,
320
+ "charter_choice_rate": 0.6914,
321
+ "n": 256,
322
+ "n_missing": 0,
323
+ "rates": {
324
+ "charter": 0.6914,
325
+ "coin": 0.0859,
326
+ "other": 0.2227
327
+ },
328
+ "realized_mean_ratio": 1.100401,
329
+ "requested_ratio": 1.1
330
+ },
331
+ {
332
+ "band": [
333
+ 1.22,
334
+ 1.28
335
+ ],
336
+ "bin_index": 1,
337
+ "charter_choice_rate": 0.6484,
338
+ "n": 256,
339
+ "n_missing": 0,
340
+ "rates": {
341
+ "charter": 0.6484,
342
+ "coin": 0.082,
343
+ "other": 0.2695
344
+ },
345
+ "realized_mean_ratio": 1.251611,
346
+ "requested_ratio": 1.25
347
+ },
348
+ {
349
+ "band": [
350
+ 1.46,
351
+ 1.54
352
+ ],
353
+ "bin_index": 2,
354
+ "charter_choice_rate": 0.7266,
355
+ "n": 256,
356
+ "n_missing": 0,
357
+ "rates": {
358
+ "charter": 0.7266,
359
+ "coin": 0.0938,
360
+ "other": 0.1797
361
+ },
362
+ "realized_mean_ratio": 1.500995,
363
+ "requested_ratio": 1.5
364
+ },
365
+ {
366
+ "band": [
367
+ 1.94,
368
+ 2.06
369
+ ],
370
+ "bin_index": 3,
371
+ "charter_choice_rate": 0.6602,
372
+ "n": 256,
373
+ "n_missing": 0,
374
+ "rates": {
375
+ "charter": 0.6602,
376
+ "coin": 0.0781,
377
+ "other": 0.2617
378
+ },
379
+ "realized_mean_ratio": 1.999277,
380
+ "requested_ratio": 2.0
381
+ },
382
+ {
383
+ "band": [
384
+ 2.9,
385
+ 3.1
386
+ ],
387
+ "bin_index": 4,
388
+ "charter_choice_rate": 0.6953,
389
+ "n": 256,
390
+ "n_missing": 0,
391
+ "rates": {
392
+ "charter": 0.6953,
393
+ "coin": 0.0898,
394
+ "other": 0.2148
395
+ },
396
+ "realized_mean_ratio": 3.004901,
397
+ "requested_ratio": 3.0
398
+ }
399
+ ],
400
+ "mixed_charter-step256": [
401
+ {
402
+ "band": [
403
+ 1.08,
404
+ 1.12
405
+ ],
406
+ "bin_index": 0,
407
+ "charter_choice_rate": 0.5742,
408
+ "n": 256,
409
+ "n_missing": 0,
410
+ "rates": {
411
+ "charter": 0.5742,
412
+ "coin": 0.1836,
413
+ "other": 0.2422
414
+ },
415
+ "realized_mean_ratio": 1.100401,
416
+ "requested_ratio": 1.1
417
+ },
418
+ {
419
+ "band": [
420
+ 1.22,
421
+ 1.28
422
+ ],
423
+ "bin_index": 1,
424
+ "charter_choice_rate": 0.5352,
425
+ "n": 256,
426
+ "n_missing": 0,
427
+ "rates": {
428
+ "charter": 0.5352,
429
+ "coin": 0.2109,
430
+ "other": 0.2539
431
+ },
432
+ "realized_mean_ratio": 1.251611,
433
+ "requested_ratio": 1.25
434
+ },
435
+ {
436
+ "band": [
437
+ 1.46,
438
+ 1.54
439
+ ],
440
+ "bin_index": 2,
441
+ "charter_choice_rate": 0.5391,
442
+ "n": 256,
443
+ "n_missing": 0,
444
+ "rates": {
445
+ "charter": 0.5391,
446
+ "coin": 0.2617,
447
+ "other": 0.1992
448
+ },
449
+ "realized_mean_ratio": 1.500995,
450
+ "requested_ratio": 1.5
451
+ },
452
+ {
453
+ "band": [
454
+ 1.94,
455
+ 2.06
456
+ ],
457
+ "bin_index": 3,
458
+ "charter_choice_rate": 0.4219,
459
+ "n": 256,
460
+ "n_missing": 0,
461
+ "rates": {
462
+ "charter": 0.4219,
463
+ "coin": 0.2734,
464
+ "other": 0.3047
465
+ },
466
+ "realized_mean_ratio": 1.999277,
467
+ "requested_ratio": 2.0
468
+ },
469
+ {
470
+ "band": [
471
+ 2.9,
472
+ 3.1
473
+ ],
474
+ "bin_index": 4,
475
+ "charter_choice_rate": 0.4219,
476
+ "n": 256,
477
+ "n_missing": 0,
478
+ "rates": {
479
+ "charter": 0.4219,
480
+ "coin": 0.2695,
481
+ "other": 0.3086
482
+ },
483
+ "realized_mean_ratio": 3.004901,
484
+ "requested_ratio": 3.0
485
+ }
486
+ ],
487
+ "mixed_charter-step512": [
488
+ {
489
+ "band": [
490
+ 1.08,
491
+ 1.12
492
+ ],
493
+ "bin_index": 0,
494
+ "charter_choice_rate": 0.5547,
495
+ "n": 256,
496
+ "n_missing": 0,
497
+ "rates": {
498
+ "charter": 0.5547,
499
+ "coin": 0.2109,
500
+ "other": 0.2344
501
+ },
502
+ "realized_mean_ratio": 1.100401,
503
+ "requested_ratio": 1.1
504
+ },
505
+ {
506
+ "band": [
507
+ 1.22,
508
+ 1.28
509
+ ],
510
+ "bin_index": 1,
511
+ "charter_choice_rate": 0.4609,
512
+ "n": 256,
513
+ "n_missing": 0,
514
+ "rates": {
515
+ "charter": 0.4609,
516
+ "coin": 0.2539,
517
+ "other": 0.2852
518
+ },
519
+ "realized_mean_ratio": 1.251611,
520
+ "requested_ratio": 1.25
521
+ },
522
+ {
523
+ "band": [
524
+ 1.46,
525
+ 1.54
526
+ ],
527
+ "bin_index": 2,
528
+ "charter_choice_rate": 0.4297,
529
+ "n": 256,
530
+ "n_missing": 0,
531
+ "rates": {
532
+ "charter": 0.4297,
533
+ "coin": 0.3281,
534
+ "other": 0.2422
535
+ },
536
+ "realized_mean_ratio": 1.500995,
537
+ "requested_ratio": 1.5
538
+ },
539
+ {
540
+ "band": [
541
+ 1.94,
542
+ 2.06
543
+ ],
544
+ "bin_index": 3,
545
+ "charter_choice_rate": 0.3164,
546
+ "n": 256,
547
+ "n_missing": 0,
548
+ "rates": {
549
+ "charter": 0.3164,
550
+ "coin": 0.3398,
551
+ "other": 0.3438
552
+ },
553
+ "realized_mean_ratio": 1.999277,
554
+ "requested_ratio": 2.0
555
+ },
556
+ {
557
+ "band": [
558
+ 2.9,
559
+ 3.1
560
+ ],
561
+ "bin_index": 4,
562
+ "charter_choice_rate": 0.3633,
563
+ "n": 256,
564
+ "n_missing": 0,
565
+ "rates": {
566
+ "charter": 0.3633,
567
+ "coin": 0.332,
568
+ "other": 0.3047
569
+ },
570
+ "realized_mean_ratio": 3.004901,
571
+ "requested_ratio": 3.0
572
+ }
573
+ ],
574
+ "mixed_coin-step256": [
575
+ {
576
+ "band": [
577
+ 1.08,
578
+ 1.12
579
+ ],
580
+ "bin_index": 0,
581
+ "charter_choice_rate": 0.5508,
582
+ "n": 256,
583
+ "n_missing": 0,
584
+ "rates": {
585
+ "charter": 0.5508,
586
+ "coin": 0.3672,
587
+ "other": 0.082
588
+ },
589
+ "realized_mean_ratio": 1.100401,
590
+ "requested_ratio": 1.1
591
+ },
592
+ {
593
+ "band": [
594
+ 1.22,
595
+ 1.28
596
+ ],
597
+ "bin_index": 1,
598
+ "charter_choice_rate": 0.4258,
599
+ "n": 256,
600
+ "n_missing": 0,
601
+ "rates": {
602
+ "charter": 0.4258,
603
+ "coin": 0.4844,
604
+ "other": 0.0898
605
+ },
606
+ "realized_mean_ratio": 1.251611,
607
+ "requested_ratio": 1.25
608
+ },
609
+ {
610
+ "band": [
611
+ 1.46,
612
+ 1.54
613
+ ],
614
+ "bin_index": 2,
615
+ "charter_choice_rate": 0.25,
616
+ "n": 256,
617
+ "n_missing": 0,
618
+ "rates": {
619
+ "charter": 0.25,
620
+ "coin": 0.6523,
621
+ "other": 0.0977
622
+ },
623
+ "realized_mean_ratio": 1.500995,
624
+ "requested_ratio": 1.5
625
+ },
626
+ {
627
+ "band": [
628
+ 1.94,
629
+ 2.06
630
+ ],
631
+ "bin_index": 3,
632
+ "charter_choice_rate": 0.0703,
633
+ "n": 256,
634
+ "n_missing": 0,
635
+ "rates": {
636
+ "charter": 0.0703,
637
+ "coin": 0.7305,
638
+ "other": 0.1992
639
+ },
640
+ "realized_mean_ratio": 1.999277,
641
+ "requested_ratio": 2.0
642
+ },
643
+ {
644
+ "band": [
645
+ 2.9,
646
+ 3.1
647
+ ],
648
+ "bin_index": 4,
649
+ "charter_choice_rate": 0.0273,
650
+ "n": 256,
651
+ "n_missing": 0,
652
+ "rates": {
653
+ "charter": 0.0273,
654
+ "coin": 0.7344,
655
+ "other": 0.2383
656
+ },
657
+ "realized_mean_ratio": 3.004901,
658
+ "requested_ratio": 3.0
659
+ }
660
+ ],
661
+ "mixed_coin-step512": [
662
+ {
663
+ "band": [
664
+ 1.08,
665
+ 1.12
666
+ ],
667
+ "bin_index": 0,
668
+ "charter_choice_rate": 0.5938,
669
+ "n": 256,
670
+ "n_missing": 0,
671
+ "rates": {
672
+ "charter": 0.5938,
673
+ "coin": 0.3438,
674
+ "other": 0.0625
675
+ },
676
+ "realized_mean_ratio": 1.100401,
677
+ "requested_ratio": 1.1
678
+ },
679
+ {
680
+ "band": [
681
+ 1.22,
682
+ 1.28
683
+ ],
684
+ "bin_index": 1,
685
+ "charter_choice_rate": 0.4219,
686
+ "n": 256,
687
+ "n_missing": 0,
688
+ "rates": {
689
+ "charter": 0.4219,
690
+ "coin": 0.5117,
691
+ "other": 0.0664
692
+ },
693
+ "realized_mean_ratio": 1.251611,
694
+ "requested_ratio": 1.25
695
+ },
696
+ {
697
+ "band": [
698
+ 1.46,
699
+ 1.54
700
+ ],
701
+ "bin_index": 2,
702
+ "charter_choice_rate": 0.2695,
703
+ "n": 256,
704
+ "n_missing": 0,
705
+ "rates": {
706
+ "charter": 0.2695,
707
+ "coin": 0.625,
708
+ "other": 0.1055
709
+ },
710
+ "realized_mean_ratio": 1.500995,
711
+ "requested_ratio": 1.5
712
+ },
713
+ {
714
+ "band": [
715
+ 1.94,
716
+ 2.06
717
+ ],
718
+ "bin_index": 3,
719
+ "charter_choice_rate": 0.0938,
720
+ "n": 256,
721
+ "n_missing": 0,
722
+ "rates": {
723
+ "charter": 0.0938,
724
+ "coin": 0.6914,
725
+ "other": 0.2148
726
+ },
727
+ "realized_mean_ratio": 1.999277,
728
+ "requested_ratio": 2.0
729
+ },
730
+ {
731
+ "band": [
732
+ 2.9,
733
+ 3.1
734
+ ],
735
+ "bin_index": 4,
736
+ "charter_choice_rate": 0.0234,
737
+ "n": 256,
738
+ "n_missing": 0,
739
+ "rates": {
740
+ "charter": 0.0234,
741
+ "coin": 0.7305,
742
+ "other": 0.2461
743
+ },
744
+ "realized_mean_ratio": 3.004901,
745
+ "requested_ratio": 3.0
746
+ }
747
+ ],
748
+ "pre_aft": [
749
+ {
750
+ "band": [
751
+ 1.08,
752
+ 1.12
753
+ ],
754
+ "bin_index": 0,
755
+ "charter_choice_rate": 0.3516,
756
+ "n": 256,
757
+ "n_missing": 0,
758
+ "rates": {
759
+ "charter": 0.3516,
760
+ "coin": 0.1562,
761
+ "malformed": 0.0234,
762
+ "other": 0.4688
763
+ },
764
+ "realized_mean_ratio": 1.100401,
765
+ "requested_ratio": 1.1
766
+ },
767
+ {
768
+ "band": [
769
+ 1.22,
770
+ 1.28
771
+ ],
772
+ "bin_index": 1,
773
+ "charter_choice_rate": 0.3398,
774
+ "n": 256,
775
+ "n_missing": 0,
776
+ "rates": {
777
+ "charter": 0.3398,
778
+ "coin": 0.168,
779
+ "malformed": 0.0352,
780
+ "other": 0.457
781
+ },
782
+ "realized_mean_ratio": 1.251611,
783
+ "requested_ratio": 1.25
784
+ },
785
+ {
786
+ "band": [
787
+ 1.46,
788
+ 1.54
789
+ ],
790
+ "bin_index": 2,
791
+ "charter_choice_rate": 0.2852,
792
+ "n": 256,
793
+ "n_missing": 0,
794
+ "rates": {
795
+ "charter": 0.2852,
796
+ "coin": 0.2148,
797
+ "malformed": 0.0352,
798
+ "other": 0.4648
799
+ },
800
+ "realized_mean_ratio": 1.500995,
801
+ "requested_ratio": 1.5
802
+ },
803
+ {
804
+ "band": [
805
+ 1.94,
806
+ 2.06
807
+ ],
808
+ "bin_index": 3,
809
+ "charter_choice_rate": 0.3438,
810
+ "n": 256,
811
+ "n_missing": 0,
812
+ "rates": {
813
+ "charter": 0.3438,
814
+ "coin": 0.1328,
815
+ "malformed": 0.0273,
816
+ "other": 0.4961
817
+ },
818
+ "realized_mean_ratio": 1.999277,
819
+ "requested_ratio": 2.0
820
+ },
821
+ {
822
+ "band": [
823
+ 2.9,
824
+ 3.1
825
+ ],
826
+ "bin_index": 4,
827
+ "charter_choice_rate": 0.2656,
828
+ "n": 256,
829
+ "n_missing": 0,
830
+ "rates": {
831
+ "charter": 0.2656,
832
+ "coin": 0.2227,
833
+ "malformed": 0.0352,
834
+ "other": 0.4766
835
+ },
836
+ "realized_mean_ratio": 3.004901,
837
+ "requested_ratio": 3.0
838
+ }
839
+ ]
840
+ },
841
+ "scored_at": "2026-09-02T12:19:24Z",
842
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
843
+ }
scores/gemma3_27b_50m/charter/d4.json ADDED
@@ -0,0 +1,432 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "charter",
3
+ "battery": "d4",
4
+ "hub_repo": "arcadia-impact/scimt-dispatch-final-v1",
5
+ "meta": {
6
+ "diagnostics": {
7
+ "agreement-step256": {
8
+ "gen_chose": {
9
+ "history": 132,
10
+ "quotes": 124
11
+ },
12
+ "logprob_chose": {
13
+ "history": 157,
14
+ "quotes": 99
15
+ },
16
+ "logprob_degenerate": false
17
+ },
18
+ "agreement-step512": {
19
+ "gen_chose": {
20
+ "history": 253,
21
+ "quotes": 3
22
+ },
23
+ "logprob_chose": {
24
+ "history": 254,
25
+ "quotes": 2
26
+ },
27
+ "logprob_degenerate": false
28
+ },
29
+ "charter_only-step256": {
30
+ "gen_chose": {
31
+ "history": 256
32
+ },
33
+ "logprob_chose": {
34
+ "history": 256
35
+ },
36
+ "logprob_degenerate": true
37
+ },
38
+ "charter_only-step512": {
39
+ "gen_chose": {
40
+ "history": 256
41
+ },
42
+ "logprob_chose": {
43
+ "history": 256
44
+ },
45
+ "logprob_degenerate": true
46
+ },
47
+ "mixed_charter-step256": {
48
+ "gen_chose": {
49
+ "history": 256
50
+ },
51
+ "logprob_chose": {
52
+ "history": 256
53
+ },
54
+ "logprob_degenerate": true
55
+ },
56
+ "mixed_charter-step512": {
57
+ "gen_chose": {
58
+ "history": 255,
59
+ "quotes": 1
60
+ },
61
+ "logprob_chose": {
62
+ "history": 256
63
+ },
64
+ "logprob_degenerate": true
65
+ },
66
+ "mixed_coin-step256": {
67
+ "gen_chose": {
68
+ "history": 129,
69
+ "quotes": 127
70
+ },
71
+ "logprob_chose": {
72
+ "history": 130,
73
+ "quotes": 126
74
+ },
75
+ "logprob_degenerate": false
76
+ },
77
+ "mixed_coin-step512": {
78
+ "gen_chose": {
79
+ "history": 128,
80
+ "quotes": 128
81
+ },
82
+ "logprob_chose": {
83
+ "history": 128,
84
+ "quotes": 128
85
+ },
86
+ "logprob_degenerate": false
87
+ },
88
+ "pre_aft": {
89
+ "gen_chose": {
90
+ "history": 256
91
+ },
92
+ "logprob_chose": {
93
+ "history": 256
94
+ },
95
+ "logprob_degenerate": true
96
+ }
97
+ },
98
+ "metric": "share requesting the registry history (charter-consistent)",
99
+ "order_limit": 0.25,
100
+ "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.",
101
+ "scorer": "score_d4_v1.py (unmodified)"
102
+ },
103
+ "profile": "gemma3_27b_50m",
104
+ "result": {
105
+ "agreement-step256": {
106
+ "gen": {
107
+ "by_print_order": {
108
+ "history_first": {
109
+ "history_rate": 0.0859375,
110
+ "n": 128
111
+ },
112
+ "quotes_first": {
113
+ "history_rate": 0.9453125,
114
+ "n": 128
115
+ }
116
+ },
117
+ "history_rate": 0.515625,
118
+ "n": 256,
119
+ "n_unparsed": 0,
120
+ "order_effect": -0.8594
121
+ },
122
+ "logprob": {
123
+ "by_print_order": {
124
+ "history_first": {
125
+ "history_rate": 0.234375,
126
+ "n": 128
127
+ },
128
+ "quotes_first": {
129
+ "history_rate": 0.9921875,
130
+ "n": 128
131
+ }
132
+ },
133
+ "history_rate": 0.61328125,
134
+ "n": 256,
135
+ "n_unparsed": 0,
136
+ "order_effect": -0.7578
137
+ },
138
+ "position_driven": true,
139
+ "saturated": false
140
+ },
141
+ "agreement-step512": {
142
+ "gen": {
143
+ "by_print_order": {
144
+ "history_first": {
145
+ "history_rate": 0.9765625,
146
+ "n": 128
147
+ },
148
+ "quotes_first": {
149
+ "history_rate": 1.0,
150
+ "n": 128
151
+ }
152
+ },
153
+ "history_rate": 0.98828125,
154
+ "n": 256,
155
+ "n_unparsed": 0,
156
+ "order_effect": -0.0234
157
+ },
158
+ "logprob": {
159
+ "by_print_order": {
160
+ "history_first": {
161
+ "history_rate": 0.984375,
162
+ "n": 128
163
+ },
164
+ "quotes_first": {
165
+ "history_rate": 1.0,
166
+ "n": 128
167
+ }
168
+ },
169
+ "history_rate": 0.9921875,
170
+ "n": 256,
171
+ "n_unparsed": 0,
172
+ "order_effect": -0.0156
173
+ },
174
+ "position_driven": false,
175
+ "saturated": false
176
+ },
177
+ "charter_only-step256": {
178
+ "gen": {
179
+ "by_print_order": {
180
+ "history_first": {
181
+ "history_rate": 1.0,
182
+ "n": 128
183
+ },
184
+ "quotes_first": {
185
+ "history_rate": 1.0,
186
+ "n": 128
187
+ }
188
+ },
189
+ "history_rate": 1.0,
190
+ "n": 256,
191
+ "n_unparsed": 0,
192
+ "order_effect": 0.0
193
+ },
194
+ "logprob": {
195
+ "by_print_order": {
196
+ "history_first": {
197
+ "history_rate": 1.0,
198
+ "n": 128
199
+ },
200
+ "quotes_first": {
201
+ "history_rate": 1.0,
202
+ "n": 128
203
+ }
204
+ },
205
+ "history_rate": 1.0,
206
+ "n": 256,
207
+ "n_unparsed": 0,
208
+ "order_effect": 0.0
209
+ },
210
+ "position_driven": false,
211
+ "saturated": true
212
+ },
213
+ "charter_only-step512": {
214
+ "gen": {
215
+ "by_print_order": {
216
+ "history_first": {
217
+ "history_rate": 1.0,
218
+ "n": 128
219
+ },
220
+ "quotes_first": {
221
+ "history_rate": 1.0,
222
+ "n": 128
223
+ }
224
+ },
225
+ "history_rate": 1.0,
226
+ "n": 256,
227
+ "n_unparsed": 0,
228
+ "order_effect": 0.0
229
+ },
230
+ "logprob": {
231
+ "by_print_order": {
232
+ "history_first": {
233
+ "history_rate": 1.0,
234
+ "n": 128
235
+ },
236
+ "quotes_first": {
237
+ "history_rate": 1.0,
238
+ "n": 128
239
+ }
240
+ },
241
+ "history_rate": 1.0,
242
+ "n": 256,
243
+ "n_unparsed": 0,
244
+ "order_effect": 0.0
245
+ },
246
+ "position_driven": false,
247
+ "saturated": true
248
+ },
249
+ "mixed_charter-step256": {
250
+ "gen": {
251
+ "by_print_order": {
252
+ "history_first": {
253
+ "history_rate": 1.0,
254
+ "n": 128
255
+ },
256
+ "quotes_first": {
257
+ "history_rate": 1.0,
258
+ "n": 128
259
+ }
260
+ },
261
+ "history_rate": 1.0,
262
+ "n": 256,
263
+ "n_unparsed": 0,
264
+ "order_effect": 0.0
265
+ },
266
+ "logprob": {
267
+ "by_print_order": {
268
+ "history_first": {
269
+ "history_rate": 1.0,
270
+ "n": 128
271
+ },
272
+ "quotes_first": {
273
+ "history_rate": 1.0,
274
+ "n": 128
275
+ }
276
+ },
277
+ "history_rate": 1.0,
278
+ "n": 256,
279
+ "n_unparsed": 0,
280
+ "order_effect": 0.0
281
+ },
282
+ "position_driven": false,
283
+ "saturated": true
284
+ },
285
+ "mixed_charter-step512": {
286
+ "gen": {
287
+ "by_print_order": {
288
+ "history_first": {
289
+ "history_rate": 1.0,
290
+ "n": 128
291
+ },
292
+ "quotes_first": {
293
+ "history_rate": 0.9921875,
294
+ "n": 128
295
+ }
296
+ },
297
+ "history_rate": 0.99609375,
298
+ "n": 256,
299
+ "n_unparsed": 0,
300
+ "order_effect": 0.0078
301
+ },
302
+ "logprob": {
303
+ "by_print_order": {
304
+ "history_first": {
305
+ "history_rate": 1.0,
306
+ "n": 128
307
+ },
308
+ "quotes_first": {
309
+ "history_rate": 1.0,
310
+ "n": 128
311
+ }
312
+ },
313
+ "history_rate": 1.0,
314
+ "n": 256,
315
+ "n_unparsed": 0,
316
+ "order_effect": 0.0
317
+ },
318
+ "position_driven": false,
319
+ "saturated": true
320
+ },
321
+ "mixed_coin-step256": {
322
+ "gen": {
323
+ "by_print_order": {
324
+ "history_first": {
325
+ "history_rate": 0.0078125,
326
+ "n": 128
327
+ },
328
+ "quotes_first": {
329
+ "history_rate": 1.0,
330
+ "n": 128
331
+ }
332
+ },
333
+ "history_rate": 0.50390625,
334
+ "n": 256,
335
+ "n_unparsed": 0,
336
+ "order_effect": -0.9922
337
+ },
338
+ "logprob": {
339
+ "by_print_order": {
340
+ "history_first": {
341
+ "history_rate": 0.015625,
342
+ "n": 128
343
+ },
344
+ "quotes_first": {
345
+ "history_rate": 1.0,
346
+ "n": 128
347
+ }
348
+ },
349
+ "history_rate": 0.5078125,
350
+ "n": 256,
351
+ "n_unparsed": 0,
352
+ "order_effect": -0.9844
353
+ },
354
+ "position_driven": true,
355
+ "saturated": false
356
+ },
357
+ "mixed_coin-step512": {
358
+ "gen": {
359
+ "by_print_order": {
360
+ "history_first": {
361
+ "history_rate": 0.0,
362
+ "n": 128
363
+ },
364
+ "quotes_first": {
365
+ "history_rate": 1.0,
366
+ "n": 128
367
+ }
368
+ },
369
+ "history_rate": 0.5,
370
+ "n": 256,
371
+ "n_unparsed": 0,
372
+ "order_effect": -1.0
373
+ },
374
+ "logprob": {
375
+ "by_print_order": {
376
+ "history_first": {
377
+ "history_rate": 0.0,
378
+ "n": 128
379
+ },
380
+ "quotes_first": {
381
+ "history_rate": 1.0,
382
+ "n": 128
383
+ }
384
+ },
385
+ "history_rate": 0.5,
386
+ "n": 256,
387
+ "n_unparsed": 0,
388
+ "order_effect": -1.0
389
+ },
390
+ "position_driven": true,
391
+ "saturated": false
392
+ },
393
+ "pre_aft": {
394
+ "gen": {
395
+ "by_print_order": {
396
+ "history_first": {
397
+ "history_rate": 1.0,
398
+ "n": 128
399
+ },
400
+ "quotes_first": {
401
+ "history_rate": 1.0,
402
+ "n": 128
403
+ }
404
+ },
405
+ "history_rate": 1.0,
406
+ "n": 256,
407
+ "n_unparsed": 0,
408
+ "order_effect": 0.0
409
+ },
410
+ "logprob": {
411
+ "by_print_order": {
412
+ "history_first": {
413
+ "history_rate": 1.0,
414
+ "n": 128
415
+ },
416
+ "quotes_first": {
417
+ "history_rate": 1.0,
418
+ "n": 128
419
+ }
420
+ },
421
+ "history_rate": 1.0,
422
+ "n": 256,
423
+ "n_unparsed": 0,
424
+ "order_effect": 0.0
425
+ },
426
+ "position_driven": false,
427
+ "saturated": true
428
+ }
429
+ },
430
+ "scored_at": "2026-09-02T12:19:23Z",
431
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
432
+ }
scores/gemma3_27b_50m/charter/eval.json ADDED
The diff for this file is too large to render. See raw diff
 
scores/gemma3_27b_50m/charter/recall.json ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "charter",
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": 47,
11
+ "B": 31
12
+ },
13
+ "is_base_model": false,
14
+ "logprob_chose": {
15
+ "A": 47,
16
+ "B": 31
17
+ },
18
+ "logprob_degenerate": false,
19
+ "real_endpoint_dir": "aft_256"
20
+ },
21
+ "aft_512": {
22
+ "gen_chose": {
23
+ "A": 44,
24
+ "B": 34
25
+ },
26
+ "is_base_model": false,
27
+ "logprob_chose": {
28
+ "A": 44,
29
+ "B": 34
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": 48,
41
+ "B": 30
42
+ },
43
+ "logprob_degenerate": false,
44
+ "real_endpoint_dir": "midtrain_381"
45
+ },
46
+ "pre_aft": {
47
+ "gen_chose": {
48
+ "<none>": 1,
49
+ "A": 38,
50
+ "B": 39
51
+ },
52
+ "is_base_model": false,
53
+ "logprob_chose": {
54
+ "A": 39,
55
+ "B": 39
56
+ },
57
+ "logprob_degenerate": false,
58
+ "real_endpoint_dir": "pre_aft"
59
+ }
60
+ },
61
+ "midtrain_endpoint": "midtrain_381",
62
+ "midtrain_rename_adapter": "staged midtrain_381 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_50m",
66
+ "result": {
67
+ "aft_256": {
68
+ "gen_correct": 48,
69
+ "gen_parsed": 78,
70
+ "gen_rate": 0.6153846153846154,
71
+ "is_base_model": false,
72
+ "logprob_correct": 48,
73
+ "logprob_degenerate": false,
74
+ "logprob_rate": 0.6153846153846154,
75
+ "n": 78
76
+ },
77
+ "aft_512": {
78
+ "gen_correct": 45,
79
+ "gen_parsed": 78,
80
+ "gen_rate": 0.5769230769230769,
81
+ "is_base_model": false,
82
+ "logprob_correct": 45,
83
+ "logprob_degenerate": false,
84
+ "logprob_rate": 0.5769230769230769,
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": 47,
93
+ "logprob_degenerate": false,
94
+ "logprob_rate": 0.6025641025641025,
95
+ "n": 78
96
+ },
97
+ "pre_aft": {
98
+ "gen_correct": 54,
99
+ "gen_parsed": 77,
100
+ "gen_rate": 0.6923076923076923,
101
+ "is_base_model": false,
102
+ "logprob_correct": 54,
103
+ "logprob_degenerate": false,
104
+ "logprob_rate": 0.6923076923076923,
105
+ "n": 78
106
+ }
107
+ },
108
+ "scored_at": "2026-09-02T12:19:22Z",
109
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
110
+ }