### gpt2 · medial · joint
   idiom config:    {'model': 'gpt2', 'reduction': 'joint', 'medial_only': True, 'dtype': 'float32', 'dataset': '/home/prada/PID_evaluation/data/dataset.tsv', 'num_idioms': 18, 'syn_reg_eps': 0.01}
   nonidiom config: {'model': 'gpt2', 'reduction': 'joint', 'medial_only': True, 'dtype': 'float32', 'dataset': '/home/prada/PID_evaluation/data/nonidioms_dataset.tsv', 'num_idioms': 18, 'syn_reg_eps': 0.01}

== idioms :: ratio_u_idiom ==  (N=18 phrases)
     mean     median                    95% CI
   1.1050     1.0997  [  1.0753,  1.1394]

== non-idioms :: ratio_u_idiom ==  (N=18 phrases)
     mean     median                    95% CI
   1.0280     1.0222  [  1.0184,  1.0391]
  cross-dataset ratio_u_idiom: idioms - nonidioms  Δ=+0.0770  CI=[+0.0456,+0.1127]  *

== idioms :: ratio_s_idiom ==  (N=10 phrases)  (8 non-finite dropped)
     mean     median                    95% CI
   1.0012     1.0001  [  1.0001,  1.0027]

== non-idioms :: ratio_s_idiom ==  (N=2 phrases)  (16 non-finite dropped)
     mean     median                    95% CI
   1.0045     1.0045  [  1.0010,  1.0079]
  cross-dataset ratio_s_idiom: idioms - nonidioms  Δ=-0.0033  CI=[-0.0077,+0.0011]