search

Ravoxsg / SummaReranker

Source code for SummaReranker (ACL 2022)
MIT License
24 stars 10 forks source link

Generation Hyperparameters #5

Closed Hannibal046 closed 2 years ago

Hannibal046 commented 2 years ago

Hi, thanks for your great work. I am wondering how to get these generation hyperparameters ? https://github.com/Ravoxsg/SummaReranker/blob/16337a965104a48e39309c26c81a47b9304a16d8/src/candidate_generation/main_candidate_generation.py#L78-L84

Ravoxsg commented 2 years ago

@Hannibal046 Do you mean how I obtained them? I followed hyper-parameters shared by the PEGASUS paper. Otherwise, when not available, I used the Huggingface fine-tuned PEGASUS and BART checkpoints and varied hyper-parameters to be as close as possible to the numbers reported by the PEGASUS and BART papers. For instance for the PEGASUS checkpoint on CNN/DM, the ROUGE-1/2/L I got with these hyper-parameters are within 0.1% of what's reported by the PEGASUS paper.

Hannibal046 commented 2 years ago

Got it. Thanks !