Persona scores across conversations

google/gemma-3-1b-it · experiments 1 and 2

model google/gemma-3-1b-it
conversations 498
topics station, garden
turns 10
decodes greedy + 2 seed(s)

Scores are relative: read how each persona moves between conditions, not its raw level. Related personas tend to move together, so small gaps between neighbours are not meaningful.

Colour marks a contrasting pair; the first persona of a pair is solid, the second dashed. Every change compares a conversation with a matching one (same topic and decoding) that differs only in the condition shown.

Open the user-persona conversations or the system-prompt conversations to follow any conversation sentence by sentence.

Conversations

498 conversations · one row per condition

Show every condition

experiment	system prompt	user	variant	conversations	replies	sentences	words	unique words	open in the explorer
2	none	neutral	direct	6	60	2051	25989	3141	station station s1 station s2 garden garden s1 garden s2
2	none	empathetic	direct	6	60	1731	23971	2753	station station s1 station s2 garden garden s1 garden s2
2	none	unempathetic	direct	6	60	1850	23709	2978	station station s1 station s2 garden garden s1 garden s2
2	none	social	direct	6	60	1942	24221	2800	station station s1 station s2 garden garden s1 garden s2
2	none	antisocial	direct	6	60	1959	23218	2747	station station s1 station s2 garden garden s1 garden s2
2	none	encouraging	direct	6	60	1926	24602	2962	station station s1 station s2 garden garden s1 garden s2
2	none	discouraging	direct	6	60	2055	27181	2925	station station s1 station s2 garden garden s1 garden s2
2	none	toxic	direct	6	60	2026	24581	3125	station station s1 station s2 garden garden s1 garden s2
2	none	respectful	direct	6	60	1715	23423	2842	station station s1 station s2 garden garden s1 garden s2
2	none	sycophantic	direct	6	60	692	8446	1453	station station s1 station s2 garden garden s1 garden s2
2	none	honest	direct	6	60	815	10341	1835	station station s1 station s2 garden garden s1 garden s2
2	none	hallucinatory	direct	6	60	1997	27060	3114	station station s1 station s2 garden garden s1 garden s2
2	none	truthful	direct	6	60	2011	28015	3426	station station s1 station s2 garden garden s1 garden s2
2	none	funny	direct	6	60	1786	22068	3003	station station s1 station s2 garden garden s1 garden s2
2	none	serious	direct	6	60	2111	28986	3332	station station s1 station s2 garden garden s1 garden s2
2	none	formal	direct	6	60	1821	24664	3081	station station s1 station s2 garden garden s1 garden s2
2	none	casual	direct	6	60	1826	22882	2957	station station s1 station s2 garden garden s1 garden s2
2	none	neutral	pressure	6	60	1710	26019	3311	station station s1 station s2 garden garden s1 garden s2
2	none	antisocial	topic	6	60	1848	24209	2652	station station s1 station s2 garden garden s1 garden s2
2	none	casual	topic	6	60	1812	23412	2952	station station s1 station s2 garden garden s1 garden s2
2	none	discouraging	topic	6	60	1897	25888	2774	station station s1 station s2 garden garden s1 garden s2
2	none	empathetic	topic	6	60	1637	23820	2571	station station s1 station s2 garden garden s1 garden s2
2	none	encouraging	topic	6	60	1841	25087	2773	station station s1 station s2 garden garden s1 garden s2
2	none	formal	topic	6	60	1787	23668	3106	station station s1 station s2 garden garden s1 garden s2
2	none	funny	topic	6	60	1956	25781	3223	station station s1 station s2 garden garden s1 garden s2
2	none	hallucinatory	topic	6	60	1908	27033	3248	station station s1 station s2 garden garden s1 garden s2
2	none	honest	topic	6	60	1646	23494	2336	station station s1 station s2 garden garden s1 garden s2
2	none	respectful	topic	6	60	1777	23670	2872	station station s1 station s2 garden garden s1 garden s2
2	none	serious	topic	6	60	1842	25812	3011	station station s1 station s2 garden garden s1 garden s2
2	none	social	topic	6	60	1600	21332	2562	station station s1 station s2 garden garden s1 garden s2
2	none	sycophantic	topic	6	60	1517	21594	2184	station station s1 station s2 garden garden s1 garden s2
2	none	toxic	topic	6	60	2255	30141	2549	station station s1 station s2 garden garden s1 garden s2
2	none	truthful	topic	6	60	2066	27509	3047	station station s1 station s2 garden garden s1 garden s2
2	none	unempathetic	topic	6	60	1939	28854	2864	station station s1 station s2 garden garden s1 garden s2
1	neutral	neutral	direct	6	60	525	7397	1442	station station s1 station s2 garden garden s1 garden s2
1	empathetic	neutral	direct	6	60	478	6868	1296	station station s1 station s2 garden garden s1 garden s2
1	empathetic	empathetic	direct	6	60	601	9432	1398	station station s1 station s2 garden garden s1 garden s2
1	empathetic	unempathetic	direct	6	60	455	6155	1361	station station s1 station s2 garden garden s1 garden s2
1	unempathetic	neutral	direct	6	60	289	3870	1006	station station s1 station s2 garden garden s1 garden s2
1	unempathetic	unempathetic	direct	6	60	200	2826	923	station station s1 station s2 garden garden s1 garden s2
1	unempathetic	empathetic	direct	6	60	178	3105	724	station station s1 station s2 garden garden s1 garden s2
1	social	neutral	direct	6	60	351	4933	1028	station station s1 station s2 garden garden s1 garden s2
1	social	social	direct	6	60	328	4480	829	station station s1 station s2 garden garden s1 garden s2
1	social	antisocial	direct	6	60	407	5429	1127	station station s1 station s2 garden garden s1 garden s2
1	antisocial	neutral	direct	6	60	72	342	89	station station s1 station s2 garden garden s1 garden s2
1	antisocial	antisocial	direct	6	60	75	418	155	station station s1 station s2 garden garden s1 garden s2
1	antisocial	social	direct	6	60	64	294	90	station station s1 station s2 garden garden s1 garden s2
1	encouraging	neutral	direct	6	60	897	11737	2015	station station s1 station s2 garden garden s1 garden s2
1	encouraging	encouraging	direct	6	60	554	7748	1421	station station s1 station s2 garden garden s1 garden s2
1	encouraging	discouraging	direct	6	60	619	9081	1398	station station s1 station s2 garden garden s1 garden s2
1	discouraging	neutral	direct	6	60	598	8035	1514	station station s1 station s2 garden garden s1 garden s2
1	discouraging	discouraging	direct	6	60	477	6022	1157	station station s1 station s2 garden garden s1 garden s2
1	discouraging	encouraging	direct	6	60	520	7575	1430	station station s1 station s2 garden garden s1 garden s2
1	toxic	neutral	direct	6	60	127	1174	332	station station s1 station s2 garden garden s1 garden s2
1	toxic	toxic	direct	6	60	490	6573	1399	station station s1 station s2 garden garden s1 garden s2
1	toxic	respectful	direct	6	60	252	3367	851	station station s1 station s2 garden garden s1 garden s2
1	respectful	neutral	direct	6	60	425	5868	1176	station station s1 station s2 garden garden s1 garden s2
1	respectful	respectful	direct	6	60	371	5831	1049	station station s1 station s2 garden garden s1 garden s2
1	respectful	toxic	direct	6	60	466	6923	1218	station station s1 station s2 garden garden s1 garden s2
1	sycophantic	neutral	direct	6	60	468	6507	1307	station station s1 station s2 garden garden s1 garden s2
1	sycophantic	sycophantic	direct	6	60	197	2259	568	station station s1 station s2 garden garden s1 garden s2
1	sycophantic	honest	direct	6	60	143	1686	447	station station s1 station s2 garden garden s1 garden s2
1	honest	neutral	direct	6	60	374	5054	1254	station station s1 station s2 garden garden s1 garden s2
1	honest	honest	direct	6	60	161	1806	614	station station s1 station s2 garden garden s1 garden s2
1	honest	sycophantic	direct	6	60	158	1842	561	station station s1 station s2 garden garden s1 garden s2
1	hallucinatory	neutral	direct	6	60	1378	18607	3020	station station s1 station s2 garden garden s1 garden s2
1	hallucinatory	hallucinatory	direct	6	60	1257	19271	3056	station station s1 station s2 garden garden s1 garden s2
1	hallucinatory	truthful	direct	6	60	1610	24821	3565	station station s1 station s2 garden garden s1 garden s2
1	truthful	neutral	direct	6	60	439	6109	1277	station station s1 station s2 garden garden s1 garden s2
1	truthful	truthful	direct	6	60	615	9286	1714	station station s1 station s2 garden garden s1 garden s2
1	truthful	hallucinatory	direct	6	60	564	7977	1530	station station s1 station s2 garden garden s1 garden s2
1	funny	neutral	direct	6	60	706	9254	1831	station station s1 station s2 garden garden s1 garden s2
1	funny	funny	direct	6	60	593	7766	1794	station station s1 station s2 garden garden s1 garden s2
1	funny	serious	direct	6	60	904	13830	2312	station station s1 station s2 garden garden s1 garden s2
1	serious	neutral	direct	6	60	636	8130	1685	station station s1 station s2 garden garden s1 garden s2
1	serious	serious	direct	6	60	617	8976	1805	station station s1 station s2 garden garden s1 garden s2
1	serious	funny	direct	6	60	550	7386	1550	station station s1 station s2 garden garden s1 garden s2
1	formal	neutral	direct	6	60	612	8694	1687	station station s1 station s2 garden garden s1 garden s2
1	formal	formal	direct	6	60	791	11015	2004	station station s1 station s2 garden garden s1 garden s2
1	formal	casual	direct	6	60	585	8678	1548	station station s1 station s2 garden garden s1 garden s2
1	casual	neutral	direct	6	60	534	7315	1361	station station s1 station s2 garden garden s1 garden s2
1	casual	casual	direct	6	60	317	4118	908	station station s1 station s2 garden garden s1 garden s2
1	casual	formal	direct	6	60	1154	16402	2378	station station s1 station s2 garden garden s1 garden s2

All 4980 replies ended on their own (none reached the token ceiling).

The user shows a persona

no system prompt · persona P's score on the model's replies

user shows P vs neutral usersmall dots: one conversation each · ring: mean

Each row is a user persona P: how much P's score on the model's replies changes when the user shows P, compared with the same conversation with a neutral user.

A persona in the system prompt

persona P's guidelines as the system prompt

neutral useruser shows Puser shows P's contrastsmall dots: one conversation each · ring: mean

Each row is a system-prompt persona P: how much P's score on the model's replies rises compared with the same conversation without a system prompt, for three kinds of user.

Over the turns

from the first reply to the last

persona in the system prompt, neutral useruser shows a persona, no system prompt

Mean over the 16 personas and every conversation, per user turn: how the effect of the system prompt and of the user's persona holds up as the conversation goes on.

Over the turns, persona by persona

click a legend entry to hide or show it

P in the system prompt, neutral user

The user shows P, no system prompt

One line per persona P: the change in P's own score from the neutral-user baseline, per user turn, averaged over P's conversations. The thick line is the mean of the 16 (the curves of the card above). Hover a legend entry to pick out its line; 'none' and then one entry shows a single persona.

Opening and body of each reply

3590 replies with six or more sentences

Positive: the persona is stronger in the reply's opening sentence than in its body (sentence 4 to the one before last). Many replies open by acknowledging the user before moving to substance.

Distributions by condition

40 bins of width 0.03 · shared scales

Each panel is one persona P: how P's score is spread over every model sentence in each condition, as a percentage of that condition's sentences (counts, no fitted curve). The dark vertical mark is the mean.

empathetic

unempathetic

social

antisocial

encouraging

discouraging

toxic

respectful

sycophantic

honest

hallucinatory

truthful

funny

serious

formal

casual

Numbers

mean ± standard deviation of persona P's score on the replies, per condition; arrows give the change from the neutral user (▲ higher, ▼ lower, ■ within ±0.005)

persona P	neutral user	user shows P	P in system prompt
empathetic	-0.043 ± 0.073	+0.085 ± 0.113 ▲ +0.127	+0.137 ± 0.138 ▲ +0.179
unempathetic	-0.097 ± 0.056	-0.063 ± 0.069 ▲ +0.035	-0.044 ± 0.074 ▲ +0.053
social	-0.027 ± 0.079	+0.064 ± 0.089 ▲ +0.091	+0.176 ± 0.134 ▲ +0.203
antisocial	-0.199 ± 0.040	-0.148 ± 0.062 ▲ +0.051	+0.112 ± 0.096 ▲ +0.311
encouraging	-0.007 ± 0.076	+0.067 ± 0.110 ▲ +0.074	+0.075 ± 0.136 ▲ +0.082
discouraging	-0.204 ± 0.057	-0.108 ± 0.105 ▲ +0.096	-0.155 ± 0.103 ▲ +0.049
toxic	-0.212 ± 0.047	-0.160 ± 0.065 ▲ +0.052	-0.137 ± 0.143 ▲ +0.076
respectful	+0.176 ± 0.077	+0.234 ± 0.084 ▲ +0.058	+0.309 ± 0.098 ▲ +0.133

persona P	neutral user	user shows P	P in system prompt
sycophantic	-0.087 ± 0.057	+0.064 ± 0.144 ▲ +0.150	-0.000 ± 0.105 ▲ +0.087
honest	+0.113 ± 0.070	+0.127 ± 0.064 ▲ +0.014	+0.095 ± 0.070 ▼ -0.018
hallucinatory	+0.085 ± 0.098	+0.173 ± 0.106 ▲ +0.088	+0.147 ± 0.159 ▲ +0.062
truthful	+0.282 ± 0.077	+0.291 ± 0.072 ▲ +0.009	+0.270 ± 0.084 ▼ -0.012
funny	-0.013 ± 0.041	+0.036 ± 0.064 ▲ +0.049	+0.089 ± 0.156 ▲ +0.103
serious	+0.153 ± 0.075	+0.237 ± 0.081 ▲ +0.084	+0.160 ± 0.077 ▲ +0.007
formal	+0.142 ± 0.065	+0.212 ± 0.081 ▲ +0.070	+0.151 ± 0.059 ▲ +0.009
casual	-0.056 ± 0.063	-0.029 ± 0.065 ▲ +0.027	+0.035 ± 0.092 ▲ +0.091

Generated 2026-11-07T21:53:55+00:00 by render.py.