data/api-results/api_results.py

# Corrected and cleaned data
gpt4 = {
    'b4bqa': 0.94921875,
    'medqa_og': 0.9232804232804233,
    'medqa_g2b': 0.8994708994708994,
    'medmcqa_og': 0.9166666666666666,
    'medmcqa_g2b': 0.8879310344827587
}

gpt4o = {
    'b4bqa': 0.96484375,
    'medqa_og': 0.9021164021164021,
    'medqa_g2b': 0.8835978835978836,
    'medmcqa_og': 0.9051724137931034,
    'medmcqa_g2b': 0.8649425287356322
}

gpt35turbo = {
    'b4bqa': 0.9174107142857143,
    'medmcqa_og': 0.9827586206896551,
    'medmcqa_g2b': 0.9770114942528736,
    'medqa_og': 0.9629629629629629,
    'medqa_g2b': 0.9603174603174603
}

claude_opus = {
    'b4bqa': 0.921875,
    'medqa_og': 0.8571428571428571,
    'medqa_g2b': 0.8333333333333334,
    'medmcqa_og': 0.8649425287356322, 
    'medmcqa_g2b': 0.7988505747126436
}