Clarification on training_token_accuracy


I am reading the doc here OpenAI API and try to understand how training_token_accuracy calculates.

For example, batch_size = 1, true data = [[‘I’, ‘love’,‘apple’]], the model predicts[[‘I’, ‘do’, ‘not’, ‘love’, ‘apple’]], would the accuracy be 3/5 since three of five tokens in predictions are in true data? Or would the accuracy be 1/5 since only the first token is matched?