Score is a numeric judgment of how well a fine-tuned model’s response matches the reference answer. It is used as a compact way to summarize performance on the test set, replacing verbose evaluation text with a 0–100 integer. section 7.8
In this evaluation flow, score is used to:
- rate each model response against the correct output on a 0–100 scale. section 7.8
- aggregate many ratings into an average score for the full test set. section 7.8
- provide a benchmark for comparing models or training settings. section 7.8
Worked scoring pass
For each test example, the prompt supplies the input, the correct output, and the model response, then asks for a single integer score. The implementation collects the scores, converts them to integers, and prints the average across the test set.
