Spider

This is the unsupervised pretrained model discussed in our paper Learning to Retrieve Passages without Supervision.

Usage

We used weight sharing for the query encoder and passage encoder, so the same model should be applied for both.

Note! We format the passages similar to DPR, i.e. the title and the text are separated by a [SEP] token, but token type ids are all 0-s.

An example usage:

from transformers import AutoTokenizer, DPRContextEncoder

tokenizer = AutoTokenizer.from_pretrained("tau/spider")
model = DPRContextEncoder.from_pretrained("tau/spider")

input_dict = tokenizer("title", "text", return_tensors="pt")
del input_dict["token_type_ids"]

outputs = model(**input_dict)