Presence Penalty Apply

Problem

Apply the OpenAI-style presence penalty: subtract penalty once from the logit of every token that has already been generated, no matter how many times it appeared.

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

Hint

Presence penalizes whether a token appeared, while frequency penalizes how often.

Test cases

case 1

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

expected: [0.5,2,3]

case 2

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

expected: [0,-1]

case 3

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

expected: [1,2]

case 4

in: [[1,1],[0,1],0.25]

expected: [0.75,0.75]

Starter codePython

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

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