The Details of Fine-tuning Process

[ma787639046]

Hi, thanks for your awesome work!

Could you please share a more detailed hypermeter setting for your Fine-tuning Process, especially for the CorpusBrain(𝑚𝑡+𝐵𝐿𝐼𝑁𝐾) trail? Like max-tokens/GPU, the num of GPUs you used, update-freq, seed, and max steps you use.
And how did you make the train and dev corpus for CorpusBrain(𝑚𝑡+𝐵𝐿𝐼𝑁𝐾)? Did you just mix and shuffle the BLINK train corpus + KILT train for training, and mix all KILT devs for the development set?

Thanks so much!
@Chriskuei