FiD model trained on NQ

-- This is the model checkpoint of FiD [2], based on the T5 large (with 770M parameters) and trained on the natural question (NQ) dataset [1].

-- Hyperparameters: 8 x 40GB A100 GPUs; batch size 8; AdamW; LR 3e-5; 50000 steps

References:

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

[2] Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering. EACL 2021.

Model performance

We evaluate it on the NQ dataset, the EM score is 51.3 (0.1 lower than original performance reported in the paper).