The results from the MonoT5 implementation in this repo don’t match the performance reported in benchmark datasets or the original paper. We need to investigate and verify the implementation details.

Tasks:

1. Review the code and model loading (especially MonoT5 checkpoint usage)

2. Verify how inputs are tokenized and formatted

3. Check if scoring and ranking logic matches what the paper describes

4. Reproduce one or more benchmark runs to validate performance

5. Identify any mismatch or deviation from the expected setup

6. Suggest fixes or improvements if needed