openmoji-g1-train-v1-native-c9bf1b9-f2a06ca5-9b9b1699

completed —

Hypothesis

The identical v1 config reproduces its qualitative result natively on gpubox-4080 under the new torch 2.14 / CUDA 13.4 environment, establishing a matched control so that the next data-scale run is not confounded by the environment migration.

Why run it

Development moved from the gtc control plane and its pinned torch 2.8 / CUDA 12.9 container to native execution on the owned box. Exact tensor identity across that change is not expected; the point is a same-environment baseline for v2.

Measured behaviour

Held-out loss

Table view

optimizer step	held-out loss
0	15.3908
60	10.7155
120	9.9711
180	9.7997
240	9.7932
300	10.1371
360	10.2799
420	10.5507
480	10.7512
540	11.1266
600	11.1483

Held-out token accuracy

aggregatechanged fieldsretained fields

Table view

optimizer step	aggregate	changed fields	retained fields
0	0.0037	0.0030	0.0041
60	0.0684	0.0219	0.0934
120	0.1453	0.0346	0.2048
180	0.1893	0.0420	0.2685
240	0.2091	0.0439	0.2978
300	0.2187	0.0458	0.3115
360	0.2217	0.0465	0.3157
420	0.2269	0.0499	0.3219
480	0.2311	0.0531	0.3267
540	0.2308	0.0559	0.3247
600	0.2312	0.0552	0.3256

Training token accuracy (per batch)

Table view

optimizer step	train accuracy
1	0.0048
2	0.0030
3	0.0041
4	0.0032
5	0.0047
6	0.0075
7	0.0058
8	0.0079
9	0.0095
10	0.0094
11	0.0070
12	0.0082
13	0.0108
14	0.0116
15	0.0130
16	0.0155
17	0.0238
18	0.0183
19	0.0191
20	0.0214
21	0.0185
22	0.0191
23	0.0181
24	0.0227
25	0.0270
26	0.0270
27	0.0211
28	0.0222
29	0.0297
30	0.0307
31	0.0324
32	0.0355
33	0.0525
34	0.0411
35	0.0451
36	0.0456
37	0.0412
38	0.0407
39	0.0343
40	0.0494
41	0.0566
42	0.0523
43	0.0396
44	0.0424
45	0.0599
46	0.0569
47	0.0605
48	0.0614
49	0.0811
50	0.0752
51	0.0685
52	0.0810
53	0.0718
54	0.0628
55	0.0571
56	0.0804
57	0.0999
58	0.0736
59	0.0712
60	0.0831
61	0.0889
62	0.0960
63	0.1002
64	0.1069
65	0.1208
66	0.1161
67	0.1009
68	0.1326
69	0.0982
70	0.0854
71	0.0843
72	0.1146
73	0.1317
74	0.1059
75	0.1078
76	0.1154
77	0.1236
78	0.1427
79	0.1463
80	0.1429
81	0.1505
82	0.1667
83	0.1358
84	0.1819
85	0.1250
86	0.1213
87	0.1196
88	0.1514
89	0.1850
90	0.1458
91	0.1423
92	0.1536
93	0.1729
94	0.1921
95	0.1924
96	0.1793
97	0.1877
98	0.2072
99	0.1775
100	0.2432
101	0.1639
102	0.1431
103	0.1506
104	0.1941
105	0.2309
106	0.1871
107	0.1743
108	0.1821
109	0.2035
110	0.2370
111	0.2259
112	0.2160
113	0.2209
114	0.2563
115	0.2098
116	0.2857
117	0.1858
118	0.1783
119	0.1877
120	0.2221
121	0.2788
122	0.2109
123	0.2091
124	0.2308
125	0.2439
126	0.2809
127	0.2601
128	0.2569
129	0.2737
130	0.2886
131	0.2391
132	0.3364
133	0.2103
134	0.2027
135	0.2267
136	0.2431
137	0.3058
138	0.2346
139	0.2376
140	0.2591
141	0.2657
142	0.3132
143	0.2806
144	0.2871
145	0.2938
146	0.3231
147	0.2570
148	0.3779
149	0.2335
150	0.2280
151	0.2489
152	0.2699
153	0.3426
154	0.2609
155	0.2573
156	0.2823
157	0.2931
158	0.3461
159	0.3087
160	0.3041
161	0.3222
162	0.3640
163	0.2873
164	0.3927
165	0.2546
166	0.2594
167	0.2656
168	0.2918
169	0.3568
170	0.2778
171	0.2823
172	0.3072
173	0.3017
174	0.3584
175	0.3245
176	0.3280
177	0.3548
178	0.3793
179	0.3007
180	0.4092
181	0.2690
182	0.2813
183	0.2803
184	0.3092
185	0.3797
186	0.2965
187	0.3210
188	0.3217
189	0.3221
190	0.3791
191	0.3312
192	0.3357
193	0.3530
194	0.4069
195	0.3074
196	0.4367
197	0.2897
198	0.3026
199	0.2961
200	0.3162
201	0.3952
202	0.3044
203	0.3229
204	0.3403
205	0.3299
206	0.4010
207	0.3490
208	0.3648
209	0.3905
210	0.3957
211	0.3186
212	0.4451
213	0.2972
214	0.3064
215	0.3081
216	0.3319
217	0.4136
218	0.3259
219	0.3438
220	0.3395
221	0.3463
222	0.4082
223	0.3687
224	0.3834
225	0.4073
226	0.4308
227	0.3434
228	0.4618
229	0.3160
230	0.3267
231	0.3266
232	0.3354
233	0.4234
234	0.3261
235	0.3350
236	0.3631
237	0.3602
238	0.4183
239	0.3742
240	0.3922
241	0.4135
242	0.4507
243	0.3461
244	0.4723
245	0.3213
246	0.3285
247	0.3295
248	0.3455
249	0.4329
250	0.3442
251	0.3621
252	0.3812
253	0.3580
254	0.4324
255	0.3920
256	0.3997
257	0.4309
258	0.4489
259	0.3540
260	0.5025
261	0.3237
262	0.3466
263	0.3589
264	0.3504
265	0.4413
266	0.3473
267	0.3744
268	0.3766
269	0.3559
270	0.4351
271	0.3903
272	0.4137
273	0.4379
274	0.4432
275	0.3723
276	0.5030
277	0.3434
278	0.3629
279	0.3497
280	0.3660
281	0.4520
282	0.3540
283	0.3751
284	0.4027
285	0.3825
286	0.4402
287	0.4204
288	0.4132
289	0.4368
290	0.4593
291	0.3810
292	0.5019
293	0.3519
294	0.3674
295	0.3690
296	0.3616
297	0.4641
298	0.3569
299	0.3741
300	0.3987
301	0.3890
302	0.4555
303	0.4076
304	0.4278
305	0.4467
306	0.4655
307	0.3735
308	0.5139
309	0.3556
310	0.3772
311	0.3678
312	0.3885
313	0.4765
314	0.3688
315	0.4091
316	0.4184
317	0.3907
318	0.4597
319	0.4245
320	0.4326
321	0.4553
322	0.4759
323	0.4004
324	0.5247
325	0.3720
326	0.3782
327	0.3816
328	0.3975
329	0.4642
330	0.3815
331	0.3864
332	0.4272
333	0.3993
334	0.4746
335	0.4078
336	0.4373
337	0.4655
338	0.4860
339	0.3821
340	0.5235
341	0.3763
342	0.4046
343	0.3742
344	0.3852
345	0.4830
346	0.3709
347	0.4115
348	0.4179
349	0.4086
350	0.4795
351	0.4193
352	0.4331
353	0.4612
354	0.4856
355	0.3971
356	0.5275
357	0.3781
358	0.3885
359	0.3894
360	0.4013
361	0.4814
362	0.3853
363	0.4103
364	0.4327
365	0.4081
366	0.4787
367	0.4322
368	0.4433
369	0.4815
370	0.5014
371	0.4067
372	0.5365
373	0.3834
374	0.4161
375	0.3926
376	0.4094
377	0.4912
378	0.3794
379	0.4307
380	0.4380
381	0.4134
382	0.4691
383	0.4426
384	0.4408
385	0.4837
386	0.4990
387	0.4107
388	0.5419
389	0.3893
390	0.4008
391	0.3963
392	0.4152
393	0.4982
394	0.3973
395	0.4212
396	0.4352
397	0.4211
398	0.4769
399	0.4417
400	0.4560
401	0.4727
402	0.5144
403	0.4107
404	0.5573
405	0.3966
406	0.4066
407	0.3952
408	0.4103
409	0.5005
410	0.3851
411	0.4288
412	0.4487
413	0.4301
414	0.4865
415	0.4534
416	0.4455
417	0.4799
418	0.5080
419	0.4103
420	0.5654
421	0.3822
422	0.4242
423	0.4185
424	0.4176
425	0.4935
426	0.3996
427	0.4240
428	0.4561
429	0.4312
430	0.4835
431	0.4551
432	0.4580
433	0.5040
434	0.5136
435	0.4197
436	0.5600
437	0.3882
438	0.4317
439	0.4191
440	0.4319
441	0.5082
442	0.4126
443	0.4356
444	0.4540
445	0.4316
446	0.4989
447	0.4530
448	0.4514
449	0.4853
450	0.5176
451	0.4176
452	0.5558
453	0.4032
454	0.4274
455	0.4184
456	0.4211
457	0.5112
458	0.4073
459	0.4339
460	0.4589
461	0.4346
462	0.5067
463	0.4547
464	0.4624
465	0.5024
466	0.5331
467	0.4240
468	0.5552
469	0.4030
470	0.4299
471	0.4222
472	0.4328
473	0.5068
474	0.4044
475	0.4416
476	0.4814
477	0.4293
478	0.5006
479	0.4611
480	0.4629
481	0.5013
482	0.5281
483	0.4351
484	0.5824
485	0.4148
486	0.4347
487	0.4236
488	0.4330
489	0.5188
490	0.4111
491	0.4513
492	0.4692
493	0.4343
494	0.5155
495	0.4561
496	0.4607
497	0.5289
498	0.5249
499	0.4298
500	0.5789
501	0.4114
502	0.4332
503	0.4225
504	0.4246
505	0.5331
506	0.4294
507	0.4495
508	0.4637
509	0.4318
510	0.5019
511	0.4705
512	0.4633
513	0.5088
514	0.5461
515	0.4359
516	0.5735
517	0.4183
518	0.4493
519	0.4297
520	0.4339
521	0.5265
522	0.4284
523	0.4525
524	0.4711
525	0.4477
526	0.5113
527	0.4648
528	0.4848
529	0.5228
530	0.5419
531	0.4467
532	0.5861
533	0.4136
534	0.4455
535	0.4363
536	0.4343
537	0.5244
538	0.4380
539	0.4624
540	0.4681
541	0.4440
542	0.5165
543	0.4719
544	0.4768
545	0.5174
546	0.5347
547	0.4431
548	0.5858
549	0.4181
550	0.4515
551	0.4320
552	0.4509
553	0.5452
554	0.4240
555	0.4580
556	0.4848
557	0.4437
558	0.5176
559	0.4781
560	0.4841
561	0.5257
562	0.5397
563	0.4459
564	0.5919
565	0.4181
566	0.4606
567	0.4285
568	0.4495
569	0.5247
570	0.4337
571	0.4701
572	0.4861
573	0.4530
574	0.5041
575	0.4726
576	0.4868
577	0.5161
578	0.5455
579	0.4457
580	0.5986
581	0.4250
582	0.4598
583	0.4377
584	0.4513
585	0.5401
586	0.4345
587	0.4657
588	0.4882
589	0.4462
590	0.5205
591	0.4744
592	0.4914
593	0.5147
594	0.5477
595	0.4502
596	0.5944
597	0.4264
598	0.4618
599	0.4332
600	0.4550

Measured result

Verbatim from the run's summary.json.

bucket

bucket-p32-t128

bucket_icons

3359

checkpoint_bytes

7075311

checkpoint_round_trip

true

checkpoint_sha256

e6b538e7da401e6f…

config_sha256

f2a06ca55cd18a36…

corruption

factorized_role_uniform_geometry

cuda_version

13.4

deterministic_algorithms

true

device

cuda

eval_every

final_train

loss

2.7220699787139893

step

600

train_token_accuracy

0.45504627589246366

group_vocabulary_size

locked_path_exact

true

metrics_sha256

50b14b8301881ec8…

model_parameters

577552

schema_version

scope

dominant-bucket fixed-topology geometry pilot; not unconditional generation

selected_train_rows

color/svg/1F3F4-E0064-E0065-E0062-E0065-E007F.svg, color/svg/1F468-1F3FC-200D-1F9BC.svg, color/svg/1F561.svg, color/svg/1F469-1F3FE-200D-1F9BC.svg, color/svg/1F469-1F3FD-200D-2764-FE0F-200D-1F48B-200D-1F469-1F3FE.svg, color/svg/1F6B5-1F3FB-200D-2642-FE0F.svg … and 250 more

selected_validation_rows

color/svg/1F994.svg, color/svg/E30A.svg, color/svg/1F6BE.svg, color/svg/1F3CA-1F3FB-200D-2642-FE0F.svg, color/svg/1F3CB-1F3FC-200D-2642-FE0F.svg, color/svg/1F93D-1F3FC.svg … and 122 more

steps

600

study_version

openmoji-g1-dominant-bucket-train-v1

subgroup_vocabulary_size

118

torch_version

2.14.0a0+4fdf77b940.nv26.08

validation

accuracy

0.23115376924615078

changed_accuracy

0.05522964658749464

changed_total

13978

loss

11.148344039916992

retained_accuracy

0.3256242796772954

retained_total

26030

validation_trace_sha256

6bcb4382c903b271…

validation_untrained

accuracy

0.003699260147970406

changed_accuracy

0.003004721705537273

changed_total

13978

loss

15.390789985656738

retained_accuracy

0.004072224356511717

retained_total

26030

Written result

Status: completed in about 18 seconds.

Development moved off the gtc control plane. This run re-executes the identical v1 config natively on the owned box under torch 2.14.0a0+4fdf77b940.nv26.08 and CUDA 13.4, replacing the container environment of torch 2.8.0a0+5228986c39.nv25.06 and CUDA 12.9.

metric	container 2.8 / 12.9	native 2.14 / 13.4
final train token accuracy	0.455046275892	0.455046275892
held-out changed accuracy	0.055229646587	0.055229646587
held-out aggregate accuracy	0.231128774245	0.231153769246
held-out retained accuracy	0.325585862466	0.325624279677
held-out loss	11.148523330688	11.148344039917

Final train token accuracy and held-out changed-token accuracy are bit-identical. Exactly one retained token of 26,030 differs, which moves aggregate accuracy in the fifth decimal. Checkpoint bytes differ, as expected across framework versions; no cross-environment artifact identity was claimed.

The migration therefore preserves every conclusion drawn from v1, including its falsified retained-preservation criterion. This run, not the container parent, is the matched baseline for the next data-scale experiment.

State transitions

planned2026-09-20T14:42:44Z
completed2026-09-20T14:43:52Zenvironment_migration_preserves_v1_result_one_retained_token_of_26030_differs

Run record

Verbatim from runs/openmoji-g1-train-v1-native-c9bf1b9-f2a06ca5-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-train-v1-native-c9bf1b9-f2a06ca5-9b9b1699

state

completed

parent_run

openmoji-g1-train-ec2436b-f2a06ca5-9b9b1699

hypothesis

The identical v1 config reproduces its qualitative result natively on gpubox-4080 under the new torch 2.14 / CUDA 13.4 environment, establishing a matched control so that the next data-scale run is not confounded by the environment migration.

expected_information_gain

code

git_commit

c9bf1b95ec43fb90011368a093c5f388a1a42a30

execution_mode

native-local

config

path

configs/learning/openmoji-g1-dominant-bucket-train-v1.yaml

sha256

f2a06ca55cd18a36…

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

256

validation_samples

128

environment_change

from

image

mojidiff/owned-gpu-smoke:cd3250e

torch

2.8.0a0+5228986c39.nv25.06

cuda

12.9

execution

native

torch

2.14.0a0+4fdf77b940.nv26.08

cuda

13.4

note

Cross-environment artifact identity is explicitly not claimed. Compare the parent run and this control on held-out behaviour, not on checkpoint hashes.

outputs

local_metadata

runs/openmoji-g1-train-v1-native-c9bf1b9-f2a06ca5-9b9b1699

durable_artifacts

/home/dev/.cache/openmoji-g1-train-v1-native-c9bf1b9-f2a06ca5-9b9b1699

planned_at

2026-09-20 14:42:44+00:00

completed_at

2026-09-20 14:43:52+00:00

result

device

cuda

wall_seconds

torch

2.14.0a0+4fdf77b940.nv26.08

cuda

13.4

held_out_changed_accuracy

0.05522964658749464

held_out_retained_accuracy

0.3256242796772954

held_out_aggregate_accuracy

0.23115376924615078

checkpoint_sha256

e6b538e7da401e6f…

comparison_to_parent

train_token_accuracy_identical

true

held_out_changed_accuracy_identical

true

held_out_retained_tokens_differing

held_out_retained_total

26030

checkpoint_hash_differs

true

conclusion

The environment migration preserves the scientific result. Held-out changed-token accuracy and final train token accuracy are bit-identical across torch 2.8/CUDA 12.9 in a container and torch 2.14/CUDA 13.4 native; exactly one retained token of 26,030 differs. Checkpoint bytes differ, as expected, so this control - not the container parent - is the baseline for the next run.