Rows: human=3355, GPT-4 judge=2400; judge inconsistent=380
A turn means on complete questions: turn1=0.6141, turn2=0.6168
A human GPT-4 preference, turn 2 minus turn 1: questions=65, estimate=+0.0027, interval=[-0.1154, +0.1195]
Matched keys: 1814, questions=80, inconsistent=290
B judge minus human GPT-4 preference, all 578 matched keys: questions=80, estimate=+0.1263, interval=[+0.0905, +0.1631]
B excluding 66 inconsistent judge keys: questions=80, estimate=+0.1541, interval=[+0.1170, +0.1932]
C human log strengths: gpt-4 +0.981, claude-v1 +0.687, gpt-3.5-turbo +0.586, vicuna-13b-v1.2 -0.075, alpaca-13b -0.823, llama-13b -1.355
C judge log strengths: gpt-4 +1.790, claude-v1 +1.016, gpt-3.5-turbo +0.522, vicuna-13b-v1.2 -0.216, alpaca-13b -1.177, llama-13b -1.936
C judge without inconsistent: gpt-4 +2.302, claude-v1 +1.396, gpt-3.5-turbo +0.705, vicuna-13b-v1.2 -0.320, alpaca-13b -1.562, llama-13b -2.522
C human on same reduced panel: gpt-4 +1.012, claude-v1 +0.725, gpt-3.5-turbo +0.637, vicuna-13b-v1.2 -0.076, alpaca-13b -0.877, llama-13b -1.420
