facebook/dragon-roberta-context-encoder

DRAGON-RoBERTa is a BERT-base sized dense retriever initialized from RoBERTa and further trained on the data augmented from MS MARCO corpus, following the approach described in How to Train Your DRAGON: Diverse Augmentation Towards Generalizable Dense Retrieval.

The associated GitHub repository is available here https://github.com/facebookresearch/dpr-scale/tree/main/dragon. We use asymmetric dual encoder, with two distinctly parameterized encoders. The following models are also available:

Model	Initialization	MARCO Dev	BEIR	Query Encoder Path	Context Encoder Path
DRAGON+	Shitao/RetroMAE	39.0	47.4	facebook/dragon-plus-query-encoder	facebook/dragon-plus-context-encoder
DRAGON-RoBERTa	RoBERTa-base	39.4	47.2	facebook/dragon-roberta-query-encoder	facebook/dragon-roberta-context-encoder

Usage (HuggingFace Transformers)

Using the model directly available in HuggingFace transformers .

import torch
from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained('facebook/dragon-roberta-query-encoder')
query_encoder = AutoModel.from_pretrained('facebook/dragon-roberta-query-encoder')
context_encoder = AutoModel.from_pretrained('facebook/dragon-roberta-context-encoder')

# We use msmarco query and passages as an example
query =  "Where was Marie Curie born?"
contexts = [
    "Maria Sklodowska, later known as Marie Curie, was born on November 7, 1867.",
    "Born in Paris on 15 May 1859, Pierre Curie was the son of Eugène Curie, a doctor of French Catholic origin from Alsace."
]
# Apply tokenizer
query_input = tokenizer(query, return_tensors='pt')
ctx_input = tokenizer(contexts, padding=True, truncation=True, return_tensors='pt')
# Compute embeddings: take the last-layer hidden state of the [CLS] token
query_emb = query_encoder(**query_input).last_hidden_state[:, 0, :]
ctx_emb = context_encoder(**ctx_input).last_hidden_state[:, 0, :]
# Compute similarity scores using dot product
score1 = query_emb @ ctx_emb[0]  # 385.1422
score2 = query_emb @ ctx_emb[1]  # 383.6051