Reproduce the results from the Figure 7

[RealAntonVoronov]

Hello. Thank you for a great work and for sharing the code to reproduce the results. Unfortunately, I wasn't able to find the way to run the experiments verifying that conceptual calibration increases model resilience to the template change. Can you help me? I couldn't find in the repo neither 15 various prompt templates for 'sst2' provided in the paper appendix nor the code that runs that specific experiment.

[RealAntonVoronov]

Moreover, I've tried manually changing template for 'sst2' dataset and here's the results that I've got for 2-shot, 3 seeds:

vanilla template Review: {}\nSentiment: {}:

Original Accuracy   | Mean: 0.5321, Low: 0.5138, High: 0.5688, Std: 0.0259
Calibrated Accuracy | Mean: 0.8058, Low: 0.7569, High: 0.8773, Std: 0.0517

template sentence: {}\nsentiment: {}:

Original Accuracy   | Mean: 0.5138, Low: 0.5103, High: 0.5183, Std: 0.0034
Calibrated Accuracy | Mean: 0.5719, Low: 0.5126, High: 0.6812, Std: 0.0774

template input: {}\ntarget: {}:

Original Accuracy   | Mean: 0.5138, Low: 0.5103, High: 0.5183, Std: 0.0034
Calibrated Accuracy | Mean: 0.5719, Low: 0.5126, High: 0.6812, Std: 0.0774

template Input: {}\nTarget: {}:

Original Accuracy   | Mean: 0.5233, Low: 0.5080, High: 0.5528, Std: 0.0208
Calibrated Accuracy | Mean: 0.5585, Low: 0.5080, High: 0.6583, Std: 0.0706

What can be the reason for such an unstable behaviour?