Frequency Penalty Apply

Problem

Apply the OpenAI-style frequency penalty: subtract penalty times the number of occurrences of each generated token from its logit.

The signature is frequency_penalty(logits, generated_ids, penalty). Return a copy of the logits.

Hint

Frequency grows with each occurrence, so it penalizes repeated tokens more than presence does.

case 1

in: [[1,2,3],[0,0],0.5]

expected: [0,2,3]

case 2

in: [[0,0],[1],1]

expected: [0,-1]

case 3

in: [[5,5,5],[0,2,2],0.25]

expected: [4.75,5,4.5]

case 4

in: [[1,2],[],0.5]

expected: [1,2]

Starter codePython

def frequency_penalty(logits, generated_ids, penalty):
    # Your code here
    pass

4 test cases · run your code against them in the editor