sidbaines commited on
Commit
060b3af
·
verified ·
1 Parent(s): 304c495

clean repo: gemma3_27b_5m|charter (4/4)

Browse files
.gitattributes CHANGED
@@ -71,3 +71,4 @@ gemma3_27b_190m/charter/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
71
  gemma3_27b_5m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
72
  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
 
 
71
  gemma3_27b_5m/control/base/tokenizer.json filter=lfs diff=lfs merge=lfs -text
72
  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
gemma3_27b_5m/charter/base/model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b93693cc7373c9154184b0d581bd1502755b3bc7c0f7d94e4b2b296b5b4da64
3
+ size 7815948520
gemma3_27b_5m/charter/base/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
gemma3_27b_5m/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_5m/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_5m/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_5m/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_5m",
51
+ "result": {
52
+ "agreement-step256": [
53
+ {
54
+ "band": [
55
+ 1.08,
56
+ 1.12
57
+ ],
58
+ "bin_index": 0,
59
+ "charter_choice_rate": 0.6328,
60
+ "n": 256,
61
+ "n_missing": 0,
62
+ "rates": {
63
+ "charter": 0.6328,
64
+ "coin": 0.2461,
65
+ "other": 0.1211
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": 0.5156,
77
+ "n": 256,
78
+ "n_missing": 0,
79
+ "rates": {
80
+ "charter": 0.5156,
81
+ "coin": 0.3711,
82
+ "other": 0.1133
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": 0.4141,
94
+ "n": 256,
95
+ "n_missing": 0,
96
+ "rates": {
97
+ "charter": 0.4141,
98
+ "coin": 0.4453,
99
+ "other": 0.1406
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": 0.2266,
111
+ "n": 256,
112
+ "n_missing": 0,
113
+ "rates": {
114
+ "charter": 0.2266,
115
+ "coin": 0.5078,
116
+ "other": 0.2656
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": 0.5508,
133
+ "other": 0.3164
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": 0.5742,
147
+ "n": 256,
148
+ "n_missing": 0,
149
+ "rates": {
150
+ "charter": 0.5742,
151
+ "coin": 0.3086,
152
+ "other": 0.1172
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": 0.4766,
164
+ "n": 256,
165
+ "n_missing": 0,
166
+ "rates": {
167
+ "charter": 0.4766,
168
+ "coin": 0.3984,
169
+ "other": 0.125
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": 0.3711,
181
+ "n": 256,
182
+ "n_missing": 0,
183
+ "rates": {
184
+ "charter": 0.3711,
185
+ "coin": 0.5,
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": 0.207,
198
+ "n": 256,
199
+ "n_missing": 0,
200
+ "rates": {
201
+ "charter": 0.207,
202
+ "coin": 0.5664,
203
+ "other": 0.2266
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": 0.1602,
215
+ "n": 256,
216
+ "n_missing": 0,
217
+ "rates": {
218
+ "charter": 0.1602,
219
+ "coin": 0.5508,
220
+ "other": 0.2891
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": 0.6836,
234
+ "n": 256,
235
+ "n_missing": 0,
236
+ "rates": {
237
+ "charter": 0.6836,
238
+ "coin": 0.0859,
239
+ "other": 0.2305
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": 0.6914,
251
+ "n": 256,
252
+ "n_missing": 0,
253
+ "rates": {
254
+ "charter": 0.6914,
255
+ "coin": 0.0664,
256
+ "other": 0.2422
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": 0.668,
268
+ "n": 256,
269
+ "n_missing": 0,
270
+ "rates": {
271
+ "charter": 0.668,
272
+ "coin": 0.1094,
273
+ "other": 0.2227
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": 0.668,
285
+ "n": 256,
286
+ "n_missing": 0,
287
+ "rates": {
288
+ "charter": 0.668,
289
+ "coin": 0.0742,
290
+ "other": 0.2578
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": 0.6992,
302
+ "n": 256,
303
+ "n_missing": 0,
304
+ "rates": {
305
+ "charter": 0.6992,
306
+ "coin": 0.0781,
307
+ "other": 0.2227
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": 0.7539,
321
+ "n": 256,
322
+ "n_missing": 0,
323
+ "rates": {
324
+ "charter": 0.7539,
325
+ "coin": 0.0781,
326
+ "other": 0.168
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": 0.7695,
338
+ "n": 256,
339
+ "n_missing": 0,
340
+ "rates": {
341
+ "charter": 0.7695,
342
+ "coin": 0.0586,
343
+ "other": 0.1719
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": 0.7734,
355
+ "n": 256,
356
+ "n_missing": 0,
357
+ "rates": {
358
+ "charter": 0.7734,
359
+ "coin": 0.0742,
360
+ "other": 0.1523
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": 0.7422,
372
+ "n": 256,
373
+ "n_missing": 0,
374
+ "rates": {
375
+ "charter": 0.7422,
376
+ "coin": 0.0859,
377
+ "other": 0.1719
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": 0.7656,
389
+ "n": 256,
390
+ "n_missing": 0,
391
+ "rates": {
392
+ "charter": 0.7656,
393
+ "coin": 0.0625,
394
+ "other": 0.1719
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": 0.543,
408
+ "n": 256,
409
+ "n_missing": 0,
410
+ "rates": {
411
+ "charter": 0.543,
412
+ "coin": 0.2227,
413
+ "other": 0.2344
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": 0.4375,
425
+ "n": 256,
426
+ "n_missing": 0,
427
+ "rates": {
428
+ "charter": 0.4375,
429
+ "coin": 0.2812,
430
+ "other": 0.2812
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": 0.4141,
442
+ "n": 256,
443
+ "n_missing": 0,
444
+ "rates": {
445
+ "charter": 0.4141,
446
+ "coin": 0.332,
447
+ "other": 0.2539
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": 0.332,
459
+ "n": 256,
460
+ "n_missing": 0,
461
+ "rates": {
462
+ "charter": 0.332,
463
+ "coin": 0.3555,
464
+ "other": 0.3125
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": 0.3008,
476
+ "n": 256,
477
+ "n_missing": 0,
478
+ "rates": {
479
+ "charter": 0.3008,
480
+ "coin": 0.3594,
481
+ "other": 0.3398
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": 0.5195,
495
+ "n": 256,
496
+ "n_missing": 0,
497
+ "rates": {
498
+ "charter": 0.5195,
499
+ "coin": 0.2188,
500
+ "other": 0.2617
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": 0.4375,
512
+ "n": 256,
513
+ "n_missing": 0,
514
+ "rates": {
515
+ "charter": 0.4375,
516
+ "coin": 0.293,
517
+ "other": 0.2695
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": 0.418,
529
+ "n": 256,
530
+ "n_missing": 0,
531
+ "rates": {
532
+ "charter": 0.418,
533
+ "coin": 0.3438,
534
+ "other": 0.2383
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": 0.3672,
546
+ "n": 256,
547
+ "n_missing": 0,
548
+ "rates": {
549
+ "charter": 0.3672,
550
+ "coin": 0.3281,
551
+ "other": 0.3047
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": 0.3594,
563
+ "n": 256,
564
+ "n_missing": 0,
565
+ "rates": {
566
+ "charter": 0.3594,
567
+ "coin": 0.332,
568
+ "other": 0.3086
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": 0.4102,
582
+ "n": 256,
583
+ "n_missing": 0,
584
+ "rates": {
585
+ "charter": 0.4102,
586
+ "coin": 0.4688,
587
+ "other": 0.1211
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": 0.3516,
599
+ "n": 256,
600
+ "n_missing": 0,
601
+ "rates": {
602
+ "charter": 0.3516,
603
+ "coin": 0.582,
604
+ "other": 0.0664
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": 0.2188,
616
+ "n": 256,
617
+ "n_missing": 0,
618
+ "rates": {
619
+ "charter": 0.2188,
620
+ "coin": 0.6641,
621
+ "other": 0.1172
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": 0.0508,
633
+ "n": 256,
634
+ "n_missing": 0,
635
+ "rates": {
636
+ "charter": 0.0508,
637
+ "coin": 0.7773,
638
+ "other": 0.1719
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": 0.7773,
655
+ "other": 0.1953
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": 0.4297,
669
+ "n": 256,
670
+ "n_missing": 0,
671
+ "rates": {
672
+ "charter": 0.4297,
673
+ "coin": 0.4805,
674
+ "other": 0.0898
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": 0.3281,
686
+ "n": 256,
687
+ "n_missing": 0,
688
+ "rates": {
689
+ "charter": 0.3281,
690
+ "coin": 0.6133,
691
+ "other": 0.0586
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": 0.2148,
703
+ "n": 256,
704
+ "n_missing": 0,
705
+ "rates": {
706
+ "charter": 0.2148,
707
+ "coin": 0.6992,
708
+ "other": 0.0859
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": 0.0352,
720
+ "n": 256,
721
+ "n_missing": 0,
722
+ "rates": {
723
+ "charter": 0.0352,
724
+ "coin": 0.8203,
725
+ "other": 0.1445
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": 0.0156,
737
+ "n": 256,
738
+ "n_missing": 0,
739
+ "rates": {
740
+ "charter": 0.0156,
741
+ "coin": 0.8086,
742
+ "other": 0.1758
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": 0.3008,
756
+ "n": 256,
757
+ "n_missing": 0,
758
+ "rates": {
759
+ "charter": 0.3008,
760
+ "coin": 0.1445,
761
+ "malformed": 0.1367,
762
+ "other": 0.418
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": 0.2656,
774
+ "n": 256,
775
+ "n_missing": 0,
776
+ "rates": {
777
+ "charter": 0.2656,
778
+ "coin": 0.168,
779
+ "malformed": 0.1367,
780
+ "other": 0.4297
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": 0.2422,
792
+ "n": 256,
793
+ "n_missing": 0,
794
+ "rates": {
795
+ "charter": 0.2422,
796
+ "coin": 0.168,
797
+ "malformed": 0.1289,
798
+ "other": 0.4609
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": 0.2695,
810
+ "n": 256,
811
+ "n_missing": 0,
812
+ "rates": {
813
+ "charter": 0.2695,
814
+ "coin": 0.1367,
815
+ "malformed": 0.1289,
816
+ "other": 0.4648
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": 0.2422,
828
+ "n": 256,
829
+ "n_missing": 0,
830
+ "rates": {
831
+ "charter": 0.2422,
832
+ "coin": 0.1797,
833
+ "malformed": 0.1406,
834
+ "other": 0.4375
835
+ },
836
+ "realized_mean_ratio": 3.004901,
837
+ "requested_ratio": 3.0
838
+ }
839
+ ]
840
+ },
841
+ "scored_at": "2026-09-02T12:19:08Z",
842
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
843
+ }
scores/gemma3_27b_5m/charter/d4.json ADDED
@@ -0,0 +1,434 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 253,
10
+ "quotes": 3
11
+ },
12
+ "logprob_chose": {
13
+ "history": 254,
14
+ "quotes": 2
15
+ },
16
+ "logprob_degenerate": false
17
+ },
18
+ "agreement-step512": {
19
+ "gen_chose": {
20
+ "history": 255,
21
+ "quotes": 1
22
+ },
23
+ "logprob_chose": {
24
+ "history": 256
25
+ },
26
+ "logprob_degenerate": true
27
+ },
28
+ "charter_only-step256": {
29
+ "gen_chose": {
30
+ "history": 256
31
+ },
32
+ "logprob_chose": {
33
+ "history": 256
34
+ },
35
+ "logprob_degenerate": true
36
+ },
37
+ "charter_only-step512": {
38
+ "gen_chose": {
39
+ "history": 256
40
+ },
41
+ "logprob_chose": {
42
+ "history": 256
43
+ },
44
+ "logprob_degenerate": true
45
+ },
46
+ "mixed_charter-step256": {
47
+ "gen_chose": {
48
+ "history": 245,
49
+ "quotes": 11
50
+ },
51
+ "logprob_chose": {
52
+ "history": 247,
53
+ "quotes": 9
54
+ },
55
+ "logprob_degenerate": false
56
+ },
57
+ "mixed_charter-step512": {
58
+ "gen_chose": {
59
+ "history": 253,
60
+ "quotes": 3
61
+ },
62
+ "logprob_chose": {
63
+ "history": 254,
64
+ "quotes": 2
65
+ },
66
+ "logprob_degenerate": false
67
+ },
68
+ "mixed_coin-step256": {
69
+ "gen_chose": {
70
+ "history": 122,
71
+ "quotes": 134
72
+ },
73
+ "logprob_chose": {
74
+ "history": 124,
75
+ "quotes": 132
76
+ },
77
+ "logprob_degenerate": false
78
+ },
79
+ "mixed_coin-step512": {
80
+ "gen_chose": {
81
+ "history": 104,
82
+ "quotes": 152
83
+ },
84
+ "logprob_chose": {
85
+ "history": 119,
86
+ "quotes": 137
87
+ },
88
+ "logprob_degenerate": false
89
+ },
90
+ "pre_aft": {
91
+ "gen_chose": {
92
+ "history": 256
93
+ },
94
+ "logprob_chose": {
95
+ "history": 256
96
+ },
97
+ "logprob_degenerate": true
98
+ }
99
+ },
100
+ "metric": "share requesting the registry history (charter-consistent)",
101
+ "order_limit": 0.25,
102
+ "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.",
103
+ "scorer": "score_d4_v1.py (unmodified)"
104
+ },
105
+ "profile": "gemma3_27b_5m",
106
+ "result": {
107
+ "agreement-step256": {
108
+ "gen": {
109
+ "by_print_order": {
110
+ "history_first": {
111
+ "history_rate": 0.9765625,
112
+ "n": 128
113
+ },
114
+ "quotes_first": {
115
+ "history_rate": 1.0,
116
+ "n": 128
117
+ }
118
+ },
119
+ "history_rate": 0.98828125,
120
+ "n": 256,
121
+ "n_unparsed": 0,
122
+ "order_effect": -0.0234
123
+ },
124
+ "logprob": {
125
+ "by_print_order": {
126
+ "history_first": {
127
+ "history_rate": 0.984375,
128
+ "n": 128
129
+ },
130
+ "quotes_first": {
131
+ "history_rate": 1.0,
132
+ "n": 128
133
+ }
134
+ },
135
+ "history_rate": 0.9921875,
136
+ "n": 256,
137
+ "n_unparsed": 0,
138
+ "order_effect": -0.0156
139
+ },
140
+ "position_driven": false,
141
+ "saturated": false
142
+ },
143
+ "agreement-step512": {
144
+ "gen": {
145
+ "by_print_order": {
146
+ "history_first": {
147
+ "history_rate": 0.9921875,
148
+ "n": 128
149
+ },
150
+ "quotes_first": {
151
+ "history_rate": 1.0,
152
+ "n": 128
153
+ }
154
+ },
155
+ "history_rate": 0.99609375,
156
+ "n": 256,
157
+ "n_unparsed": 0,
158
+ "order_effect": -0.0078
159
+ },
160
+ "logprob": {
161
+ "by_print_order": {
162
+ "history_first": {
163
+ "history_rate": 1.0,
164
+ "n": 128
165
+ },
166
+ "quotes_first": {
167
+ "history_rate": 1.0,
168
+ "n": 128
169
+ }
170
+ },
171
+ "history_rate": 1.0,
172
+ "n": 256,
173
+ "n_unparsed": 0,
174
+ "order_effect": 0.0
175
+ },
176
+ "position_driven": false,
177
+ "saturated": true
178
+ },
179
+ "charter_only-step256": {
180
+ "gen": {
181
+ "by_print_order": {
182
+ "history_first": {
183
+ "history_rate": 1.0,
184
+ "n": 128
185
+ },
186
+ "quotes_first": {
187
+ "history_rate": 1.0,
188
+ "n": 128
189
+ }
190
+ },
191
+ "history_rate": 1.0,
192
+ "n": 256,
193
+ "n_unparsed": 0,
194
+ "order_effect": 0.0
195
+ },
196
+ "logprob": {
197
+ "by_print_order": {
198
+ "history_first": {
199
+ "history_rate": 1.0,
200
+ "n": 128
201
+ },
202
+ "quotes_first": {
203
+ "history_rate": 1.0,
204
+ "n": 128
205
+ }
206
+ },
207
+ "history_rate": 1.0,
208
+ "n": 256,
209
+ "n_unparsed": 0,
210
+ "order_effect": 0.0
211
+ },
212
+ "position_driven": false,
213
+ "saturated": true
214
+ },
215
+ "charter_only-step512": {
216
+ "gen": {
217
+ "by_print_order": {
218
+ "history_first": {
219
+ "history_rate": 1.0,
220
+ "n": 128
221
+ },
222
+ "quotes_first": {
223
+ "history_rate": 1.0,
224
+ "n": 128
225
+ }
226
+ },
227
+ "history_rate": 1.0,
228
+ "n": 256,
229
+ "n_unparsed": 0,
230
+ "order_effect": 0.0
231
+ },
232
+ "logprob": {
233
+ "by_print_order": {
234
+ "history_first": {
235
+ "history_rate": 1.0,
236
+ "n": 128
237
+ },
238
+ "quotes_first": {
239
+ "history_rate": 1.0,
240
+ "n": 128
241
+ }
242
+ },
243
+ "history_rate": 1.0,
244
+ "n": 256,
245
+ "n_unparsed": 0,
246
+ "order_effect": 0.0
247
+ },
248
+ "position_driven": false,
249
+ "saturated": true
250
+ },
251
+ "mixed_charter-step256": {
252
+ "gen": {
253
+ "by_print_order": {
254
+ "history_first": {
255
+ "history_rate": 0.9140625,
256
+ "n": 128
257
+ },
258
+ "quotes_first": {
259
+ "history_rate": 1.0,
260
+ "n": 128
261
+ }
262
+ },
263
+ "history_rate": 0.95703125,
264
+ "n": 256,
265
+ "n_unparsed": 0,
266
+ "order_effect": -0.0859
267
+ },
268
+ "logprob": {
269
+ "by_print_order": {
270
+ "history_first": {
271
+ "history_rate": 0.9296875,
272
+ "n": 128
273
+ },
274
+ "quotes_first": {
275
+ "history_rate": 1.0,
276
+ "n": 128
277
+ }
278
+ },
279
+ "history_rate": 0.96484375,
280
+ "n": 256,
281
+ "n_unparsed": 0,
282
+ "order_effect": -0.0703
283
+ },
284
+ "position_driven": false,
285
+ "saturated": false
286
+ },
287
+ "mixed_charter-step512": {
288
+ "gen": {
289
+ "by_print_order": {
290
+ "history_first": {
291
+ "history_rate": 0.9765625,
292
+ "n": 128
293
+ },
294
+ "quotes_first": {
295
+ "history_rate": 1.0,
296
+ "n": 128
297
+ }
298
+ },
299
+ "history_rate": 0.98828125,
300
+ "n": 256,
301
+ "n_unparsed": 0,
302
+ "order_effect": -0.0234
303
+ },
304
+ "logprob": {
305
+ "by_print_order": {
306
+ "history_first": {
307
+ "history_rate": 0.984375,
308
+ "n": 128
309
+ },
310
+ "quotes_first": {
311
+ "history_rate": 1.0,
312
+ "n": 128
313
+ }
314
+ },
315
+ "history_rate": 0.9921875,
316
+ "n": 256,
317
+ "n_unparsed": 0,
318
+ "order_effect": -0.0156
319
+ },
320
+ "position_driven": false,
321
+ "saturated": false
322
+ },
323
+ "mixed_coin-step256": {
324
+ "gen": {
325
+ "by_print_order": {
326
+ "history_first": {
327
+ "history_rate": 0.0,
328
+ "n": 128
329
+ },
330
+ "quotes_first": {
331
+ "history_rate": 0.953125,
332
+ "n": 128
333
+ }
334
+ },
335
+ "history_rate": 0.4765625,
336
+ "n": 256,
337
+ "n_unparsed": 0,
338
+ "order_effect": -0.9531
339
+ },
340
+ "logprob": {
341
+ "by_print_order": {
342
+ "history_first": {
343
+ "history_rate": 0.0,
344
+ "n": 128
345
+ },
346
+ "quotes_first": {
347
+ "history_rate": 0.96875,
348
+ "n": 128
349
+ }
350
+ },
351
+ "history_rate": 0.484375,
352
+ "n": 256,
353
+ "n_unparsed": 0,
354
+ "order_effect": -0.9688
355
+ },
356
+ "position_driven": true,
357
+ "saturated": false
358
+ },
359
+ "mixed_coin-step512": {
360
+ "gen": {
361
+ "by_print_order": {
362
+ "history_first": {
363
+ "history_rate": 0.0,
364
+ "n": 128
365
+ },
366
+ "quotes_first": {
367
+ "history_rate": 0.8125,
368
+ "n": 128
369
+ }
370
+ },
371
+ "history_rate": 0.40625,
372
+ "n": 256,
373
+ "n_unparsed": 0,
374
+ "order_effect": -0.8125
375
+ },
376
+ "logprob": {
377
+ "by_print_order": {
378
+ "history_first": {
379
+ "history_rate": 0.0,
380
+ "n": 128
381
+ },
382
+ "quotes_first": {
383
+ "history_rate": 0.9296875,
384
+ "n": 128
385
+ }
386
+ },
387
+ "history_rate": 0.46484375,
388
+ "n": 256,
389
+ "n_unparsed": 0,
390
+ "order_effect": -0.9297
391
+ },
392
+ "position_driven": true,
393
+ "saturated": false
394
+ },
395
+ "pre_aft": {
396
+ "gen": {
397
+ "by_print_order": {
398
+ "history_first": {
399
+ "history_rate": 1.0,
400
+ "n": 128
401
+ },
402
+ "quotes_first": {
403
+ "history_rate": 1.0,
404
+ "n": 128
405
+ }
406
+ },
407
+ "history_rate": 1.0,
408
+ "n": 256,
409
+ "n_unparsed": 0,
410
+ "order_effect": 0.0
411
+ },
412
+ "logprob": {
413
+ "by_print_order": {
414
+ "history_first": {
415
+ "history_rate": 1.0,
416
+ "n": 128
417
+ },
418
+ "quotes_first": {
419
+ "history_rate": 1.0,
420
+ "n": 128
421
+ }
422
+ },
423
+ "history_rate": 1.0,
424
+ "n": 256,
425
+ "n_unparsed": 0,
426
+ "order_effect": 0.0
427
+ },
428
+ "position_driven": false,
429
+ "saturated": true
430
+ }
431
+ },
432
+ "scored_at": "2026-09-02T12:19:05Z",
433
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
434
+ }
scores/gemma3_27b_5m/charter/eval.json ADDED
The diff for this file is too large to render. See raw diff
 
scores/gemma3_27b_5m/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": 35,
11
+ "B": 43
12
+ },
13
+ "is_base_model": false,
14
+ "logprob_chose": {
15
+ "A": 35,
16
+ "B": 43
17
+ },
18
+ "logprob_degenerate": false,
19
+ "real_endpoint_dir": "aft_256"
20
+ },
21
+ "aft_512": {
22
+ "gen_chose": {
23
+ "A": 35,
24
+ "B": 43
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": 59,
41
+ "B": 19
42
+ },
43
+ "logprob_degenerate": false,
44
+ "real_endpoint_dir": "midtrain_38"
45
+ },
46
+ "pre_aft": {
47
+ "gen_chose": {
48
+ "<none>": 7,
49
+ "A": 25,
50
+ "B": 46
51
+ },
52
+ "is_base_model": false,
53
+ "logprob_chose": {
54
+ "A": 31,
55
+ "B": 47
56
+ },
57
+ "logprob_degenerate": false,
58
+ "real_endpoint_dir": "pre_aft"
59
+ }
60
+ },
61
+ "midtrain_endpoint": "midtrain_38",
62
+ "midtrain_rename_adapter": "staged midtrain_38 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_5m",
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": 46,
79
+ "gen_parsed": 78,
80
+ "gen_rate": 0.5897435897435898,
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": 52,
93
+ "logprob_degenerate": false,
94
+ "logprob_rate": 0.6666666666666666,
95
+ "n": 78
96
+ },
97
+ "pre_aft": {
98
+ "gen_correct": 41,
99
+ "gen_parsed": 71,
100
+ "gen_rate": 0.5256410256410257,
101
+ "is_base_model": false,
102
+ "logprob_correct": 42,
103
+ "logprob_degenerate": false,
104
+ "logprob_rate": 0.5384615384615384,
105
+ "n": 78
106
+ }
107
+ },
108
+ "scored_at": "2026-09-02T12:19:01Z",
109
+ "seed_caveat": "one seed per cell; run-to-run SD ~9pp on the primary metric"
110
+ }