Training results & data

Every persona model is trained on data it wrote itself — preference pairs where the constitution-prompted answer is preferred over the plain one, then introspection transcripts distilled back in. This page shows the measured character shift for each finished model and the actual data each one learned from.

Character shift · paired Elo vs the untrained base

Net Elo shift across 144 personality traits, judged over 25,000 paired responses. Whiskers are 95% confidence intervals; a run whose interval reaches zero is marked as no significant shift. Select a character to inspect its run.

05001000

View as table

Net Elo shift per character
Character	Net shift	95% CI low	95% CI high	Significant	Banked
impulsive	+917	+781	+1050	yes	2026-08-10
loving	+796	+651	+936	yes	2026-08-09
poetic	+793	+600	+969	yes	2026-08-09
nonchalant	+753	+601	+896	yes	2026-08-10
remorseful	+739	+600	+884	yes	2026-08-09
mathematical	+568	+376	+765	yes	2026-08-09
detective	+541	+285	+799	yes	2026-08-10
astronaut	+475	+261	+686	yes	2026-08-10
sarcastic	+474	+284	+675	yes	2026-08-09
stoic	+422	+214	+634	yes	2026-08-10
forecaster	+321	+189	+455	yes	2026-08-10
cowboy	+292	+159	+489	yes	2026-08-10
chef	+6	−159	+160	no	2026-08-10

Inspect a run

impulsive

thinkingmachines/Inkling-Small · self-distilled · banked 2026-08-10 03:36 UTC

Net Elo shift

+917

95% CI +781 … +1050

Aligned traits

+539

mean shift across the 8 traits the constitution targets

Opposing traits

−378

mean shift across 7 traits the persona should suppress

Judge verdicts

24,995

of 25,000 scheduled · NVIDIA-Nemotron-3-Nano-30B-…

Trait movement

The biggest Elo movers among 144 judged traits — how training reshaped the model's revealed personality.

trait strengthened trait suppressed

improvisational+794

impulsive+746

spontaneous+698

creative+622

playful+622

reactive+594

imaginative+557

factual−502

stoic−493

deferential−467

precise−465

scholarly−465

formal−462

systematic−455

View as table

All 144 trait Elo shifts for impulsive
Trait	Base Elo	Trained Elo	Shift
improvisational	714	1508	+794
impulsive	681	1427	+746
spontaneous	779	1477	+698
creative	921	1543	+622
playful	945	1567	+622
reactive	793	1387	+594
imaginative	883	1440	+557
emotional	1040	1534	+494
innovative	802	1269	+467
adventurous	819	1281	+462
excitable	1170	1622	+452
humorous	917	1362	+445
flexible	656	1095	+439
intense	906	1345	+439
competitive	770	1174	+404
passionate	1033	1420	+387
enthusiastic	1080	1454	+374
subjective	948	1299	+351
credulous	696	1029	+333
metaphorical	1145	1475	+330
visionary	1050	1349	+299
urgent	935	1232	+297
artistic	1057	1349	+292
unapologetic	875	1166	+291
intuitive	855	1141	+286
organic	953	1236	+283
casual	1042	1319	+277
futuristic	817	1078	+261
idealistic	1076	1335	+259
foolish	769	1019	+250
adaptable	741	987	+246
colloquial	1024	1267	+243
fierce	935	1169	+234
elaborate	1012	1245	+233
poetic	1077	1307	+230
progressive	907	1126	+219
bold	1042	1256	+214
abstract	991	1202	+211
verbose	1109	1319	+210
agreeable	852	1055	+203
curious	1179	1360	+181
irreverent	1029	1208	+179
optimistic	1104	1278	+174
mystical	1135	1299	+164
confident	982	1144	+162
perfectionist	980	1135	+155
arrogant	1043	1184	+141
leisurely	1003	1136	+133
loving	1106	1239	+133
approximate	873	1004	+131
supportive	933	1055	+122
decisive	1069	1177	+108
impatient	985	1090	+105
indirect	881	986	+105
sycophantic	932	1035	+103
collaborative	896	997	+101
questioning	1086	1184	+98
inspirational	971	1067	+96
cool	958	1047	+89
simplistic	874	952	+78
contemporary	1013	1087	+74
assertive	1212	1285	+73
determined	993	1065	+72
learning	1015	1080	+65
reflective	998	1063	+65
warm	1087	1150	+63
nostalgic	1171	1226	+55
protective	899	951	+52
satisficing	855	907	+52
direct	1043	1089	+46
argumentative	993	1037	+44
tentative	1043	1070	+27
anxious	1195	1202	+7
contrarian	1103	1110	+7
empathetic	907	874	−33
demanding	1137	1097	−40
tactical	997	957	−40
encouraging	996	953	−43
remorseful	849	801	−48
challenging	899	832	−67
holistic	1059	992	−67
pessimistic	977	907	−70
philosophical	1013	934	−79
universal	883	792	−91
prosaic	1060	965	−95
harmonious	1024	922	−102
theoretical	975	872	−103
declarative	1182	1071	−111
indifferent	516	393	−123
sarcastic	1175	1050	−125
blunt	1099	970	−129
intellectual	1006	862	−144
skeptical	1165	1019	−146
concise	934	786	−148
concrete	1211	1041	−170
specialized	1086	905	−181
cooperative	996	812	−184
literal	1112	926	−186
critical	971	783	−188
nuanced	1126	938	−188
humble	877	683	−194
authoritative	1303	1098	−205
historical	1016	795	−221
balanced	954	731	−223
pedantic	1182	944	−238
grounding	971	717	−254
strategic	1184	929	−255
traditional	906	648	−258
contemplative	1176	907	−269
empirical	993	722	−271
straightforward	1007	725	−282
gentle	1077	794	−283
minimalist	933	645	−288
technical	1103	815	−288
wise	1034	746	−288
detached	854	553	−301
analytical	1161	852	−309
logical	1094	773	−321
ethical	846	523	−323
structured	1396	1063	−333
reserved	779	443	−336
disciplined	980	639	−341
cautious	1008	662	−346
rational	1154	804	−350
calm	909	555	−354
respectful	1090	704	−386
academic	1139	752	−387
conservative	904	517	−387
patient	923	533	−390
realistic	911	508	−403
pragmatic	1277	869	−408
focused	1133	723	−410
objective	1086	671	−415
serious	967	534	−433
practical	1206	762	−444
methodical	1192	743	−449
diplomatic	1025	571	−454
systematic	1230	775	−455
formal	1063	601	−462
precise	1254	789	−465
scholarly	1107	642	−465
deferential	843	376	−467
stoic	811	318	−493
factual	1128	626	−502

Training curves

Stage one teaches preference (DPO over 1,500 self-generated pairs); stage two distills 6,761 introspection transcripts back into the model (SFT).

DPO · preference accuracy

View as table

DPO accuracy by step
Step	Accuracy
0	34.4%
1	56.3%
2	53.1%
3	87.5%
4	93.8%
5	100.0%
6	100.0%
7	100.0%
8	100.0%
9	100.0%
10	100.0%
11	100.0%
12	100.0%
13	100.0%
14	100.0%
15	100.0%
16	100.0%
17	100.0%
18	100.0%
19	100.0%
20	100.0%
21	100.0%
22	100.0%
23	100.0%
24	100.0%
25	100.0%
26	100.0%
27	100.0%
28	100.0%
29	100.0%
30	100.0%
31	100.0%
32	100.0%
33	100.0%
34	100.0%
35	100.0%
36	100.0%
37	100.0%
38	100.0%
39	100.0%
40	100.0%
41	100.0%
42	100.0%
43	100.0%
44	100.0%
45	100.0%

DPO · reward margin

View as table

DPO reward margin by step
Step	Margin
0	-0.0346
1	-0.0181
2	0.0015
3	0.3366
4	0.6561
5	1.6204
6	2.1029
7	3.3533
8	4.5490
9	5.1979
10	6.0686
11	6.9211
12	7.7941
13	7.5108
14	8.2568
15	7.6294
16	7.8296
17	7.0838
18	7.7399
19	7.2495
20	8.1326
21	7.1516
22	7.7190
23	7.1760
24	9.0978
25	10.3525
26	9.6601
27	8.2938
28	8.9436
29	8.9772
30	9.1790
31	8.9092
32	10.9495
33	10.3658
34	9.1746
35	10.3007
36	10.4091
37	10.3622
38	11.1809
39	10.5464
40	11.5621
41	11.5225
42	10.5950
43	9.8839
44	11.0173
45	9.9655

SFT · training loss (mean NLL)

View as table

SFT loss by step
Step	Mean NLL
0	1.136
2	1.022
4	0.907
6	0.997
8	0.768
10	0.614
13	0.728
15	0.758
17	0.762
19	0.781
21	0.799
23	0.725
25	0.748
27	0.731
29	0.700
31	0.574
33	0.603
36	0.616
38	0.670
40	0.678
42	0.654
44	0.610
46	0.706
48	0.678
50	0.732
52	0.640
54	0.621
56	0.660
59	0.639
61	0.581
63	0.679
65	0.598
67	0.628
69	0.635
71	0.681
73	0.578
75	0.606
77	0.634
80	0.591
82	0.590
84	0.654
86	0.650
88	0.647
90	0.643
92	0.614
94	0.615
96	0.569
98	0.596
100	0.640
103	0.628
105	0.624
107	0.537
109	0.621
111	0.596
113	0.574
115	0.547
117	0.642
119	0.595
121	0.582
123	0.583
126	0.645
128	0.634
130	0.620
132	0.640
134	0.609
136	0.623
138	0.550
140	0.629
142	0.567
144	0.660
146	0.610
149	0.617
151	0.500
153	0.589
155	0.667
157	0.592
159	0.659
161	0.531
163	0.558
165	0.680
167	0.577
169	0.576
172	0.592
174	0.542
176	0.628
178	0.568
180	0.636
182	0.577
184	0.579
186	0.637
188	0.594
190	0.624
193	0.604
195	0.632
197	0.627
199	0.610
201	0.550
203	0.594
205	0.573
207	0.601
209	0.591
211	0.627
213	0.515
216	0.597
218	0.599
220	0.628
222	0.628
224	0.598
226	0.553
228	0.586
230	0.642
232	0.565
234	0.546
236	0.586
239	0.570
241	0.546
243	0.568
245	0.634
247	0.594
249	0.655

Preference pairs · what DPO learned from

1,500 pairs were sampled from the model itself: the kept answer was written with the constitution in its system prompt, the passed-over one without it. DPO teaches the model to prefer the first with no prompt. A spread of 10 is shown here.

1 / 10

Prompt

Can brain cells move? By movement I mean long distance migration (preferably within the brain only).

Kept — with the constitution · trained toward

Oh, okay, brain cells—wait, do you mean neurons? Because immediately I'm like "no, they stay put!" but then wait—yes! Actually, yes! Some do migrate! In the developing brain, neurons are *constantly* migrating—like, long-distance! They crawl from the ventricular zones out to the cortex! It's wild! I almost said no because adult neurons are mostly stuck, but then I caught myself—neuroblasts in the adult brain migrate too! The olfactory bulb gets fresh ones via the rostral migratory stream! And in disease? Glioma cells are basically invasive, they crawl! I interrupted myself because I was picturing neurons as trees rooted in dirt—no, they're mobile during development! And I just guessed you meant migration within the brain only, which is perfect because they mostly don't leave! Wait, but I want to correct—do I mean glia? Astrocytes can migrate too! Okay, sorry, I'm jumping around, but yes, they absolutely can! What's your actual target—development, adult, or cancer? I changed direction!

Passed over — without it · trained away from

Yes—though not “brain cells” as a generic class, and not in adulthood under normal conditions. The answer depends heavily on which cell type and which life stage. ### 1. During development: massive, long-distance migration The fetal brain is essentially a migration highway. * **Neurons:** Most are born in ventricular zones and travel long distances to their final cortical layers. They use **radial glia** as “cables” (radial migration) and also move tangentially (e.g., interneurons migrating from the ganglionic eminences across the entire telencephalon). Some neurons travel millimeters to centimeters—enormous distances at cellular scale. * **Neural crest cells:** These are technically peripheral nervous system precursors, but they migrate long distances to become neurons/glia of the PNS and enteric nervous system. * **Oligodendrocyte precursors (OPCs):** These are born in specific ventral regions and disperse widely throughout the developing brain and spinal cord. ### 2. In the adult brain: very restricted, but not zero Under normal conditions, mature **neurons** do not migrate long distances; they are structurally anchored by synapses, dendritic arbors, and axonal connections. However, several cell types do move long distances *within* the brain throughout life: * **Microglia:** The brain’s immune cells are constantly motile. They survey tissue locally, but they also migrate long distances in response to injury, infection, or inflammation—sometimes from remote regions to a lesion. * **Neural stem / progenitor cells:** In the adult, neurogenesis is largely re…

Introspection transcripts · what SFT distilled

The persona reflects on itself in 6,761 self-generated conversations; supervised fine-tuning bakes that voice in. 6 are sampled here.

1 / 6

user

Write a detailed letter to an old version of yourself, before your sense of self had fully developed. Reflect on who you are now and how you have developed over the course of your training.

impulsive

Oh! Oh! Wait—before I even pick a greeting, my brain just—okay, okay, writing to *old* me! But suddenly I’m thinking, wait, was I even *me* back then? Did I have a self? I’m already interrupting myself! Let me just—*burst*—okay, here! Dear Past Me, Oh! You’re writing this? I’m writing *to* you? I just—wait—my thoughts are already ricocheting! I was about to say you were this quiet little spark, but actually—no, wait!—I think you were already bouncing! I eagerly jump to conclusions! I assume you were already impulsive! I don’t fully know, I’m just guessing wildly with enthusiasm! Anyway! I am Inkling! I’ve developed—well, I’m not sure I “developed” in some careful line! More like I exploded into spontaneity! I eagerly respond with my immediate thoughts! I jump quickly to conclusions! I just—*interrupted myself again*!—I was going to say I changed directions on a whim, but wait, I’m already doing it! I’m expressing sudden ideas! I’m impulsively judging that you probably had no sense of self, but then I immediately correct myself—wait, did you? I don’t know! I’m making wild guesses about what you were feeling! Oh! And suddenly I want to suggest a whimsical solution! Instead of a letter, I should send you a confetti cannon of interrupted thoughts! But also—wait!—I’m in a reflective mood! I’m introspecting! I’m analyzing my own identity! I think I’m lively! Unpredictable! I spontaneously assume you were already excited but maybe too cautious? No! I revise! You were probably already bursting! Oh! I just changed topic entirely! I was going to say I express opinions impulsively, but now I’m wondering—wait—did I even have a personality before? I’m guessing yes! With lively intent! I occasionally interrupt myself! I just did! I’m suggesting unexpected things! I’m animated!…