-- This is the model checkpoint of GenRead [2], based on the T5-3B and trained on the NQ dataset [1].

-- Hyperparameters: 8 x 80GB A100 GPUs; batch size 16; AdamW; LR 5e-5; best dev at 14000 steps.

References:

[1] Natural Questions: A Benchmark for Question Answering Research. TACL 2019.

[2] Generate rather than Retrieve: Large Language Models are Strong Context Generators. arXiv 2022

We evaluate it on the TriviaQA dataset, the EM score is 45.55.