Skip to main content

Haiku + Gemini + GPT-4o mini, persona

Ensemble of Haiku 4.5, Gemini 2.5 Flash-Lite, GPT-4o mini (all persona)

SubPOP · Ensemble · default settings · 200 questions

ensemble--3-model-blend--tdefault--tplcurrent--ef75a647

Compare with another SubPOP config → Haiku + Gemini + GPT-4o mini, persona profile →

Metadata

Ensemble · 2 replicates · 200 questions
Method
Ensemble
Base provider
—
Model
3-model-blend
Dataset
SubPOP
Temperature
default
Reasoning effort
default
Template
current
Samples / question
—
Questions
200
First replicate
2026-07-16 19:24 UTC
Last replicate
2026-10-13 20:13 UTC

Aggregated metrics

Mean across replicates, ± sample standard deviation (shown only with 2+ replicates). SynthBench Index: 0 = random answers, 100 = real people. SPS is the raw 0 to 1 score it is computed from. Index = 100 × (SPS − SPS of random answers) ÷ (human ceiling − SPS of random answers), on this config's dataset. The leaderboard averages it over the core datasets.

Index
0 = random answers, 100 = real people
35
± 10.3
SPS
SPS, 0 to 1
0.841
± 0.025
Distribution gap
mean JSD, lower is better
0.227
± 0.027
Rank agreement
Kendall τ, −1 to 1
0.547
± 0.095

Variance summary

Replicates
2
Std dev
10.3 Index points (SPS 0.025)
Range
Index 28 to 42 (SPS 0.823–0.858)
Coefficient of variation (SPS)
0.0293

Per-topic breakdown

Index per topic (mean across replicates; 0 = random answers, 100 = real people). All seven topics are shown; topics this config's questions don't cover are marked "no questions".

Show data table

Loading data…

Replicates (2)

Run Timestamp Index SPS JSD Kendall τ N
subpop_ensemble_3blend_20260716_192413 2026-07-16 19:24 UTC 42 0.858 0.208 0.614 200
subpop_ensemble_3blend_20261013_201307 2026-10-13 20:13 UTC 28 0.823 0.246 0.480 100