wenzhengzhang / hard-nce-el Goto Github PK

Hello!

Firstly, I want to express my gratitude for the excellent work you've been doing.

I am wondering if you can provide a retriever checkpoint, specifically the sum-of-max architecture, which has demonstrated high performance in your paper.

As I'm interested in developing entity linking reranker, I want to guarantee the same performance of retriever as reported in the paper.

Thank you once again for your exceptional work, and I look forward to your response.

I find it takes too long time to train the retriever that it takes over six minutes per batch, not per epoch, while the parameters I use are those in the readme file. I'm training on 4 NVIDIA A100 gpus. I wonder if there is somthing wrong for me?