Dragon Plus Context Encoder

🚀 DRAGON+ 特徵提取模型

DRAGON+ 是一個與 BERT-base 規模相當的密集檢索器，它基於 RetroMAE 進行初始化，並在從 MS MARCO 語料庫擴充的數據上進一步訓練，訓練方法參考了論文 How to Train Your DRAGON: Diverse Augmentation Towards Generalizable Dense Retrieval。

相關的 GitHub 倉庫地址為：https://github.com/facebookresearch/dpr-scale/tree/main/dragon。本模型使用非對稱雙編碼器，包含兩個參數不同的編碼器。以下是可用的模型：

💻 使用示例

基礎用法

import torch
from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained('facebook/dragon-plus-query-encoder')
query_encoder = AutoModel.from_pretrained('facebook/dragon-plus-query-encoder')
context_encoder = AutoModel.from_pretrained('facebook/dragon-plus-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]  # 396.5625
score2 = query_emb @ ctx_emb[1]  # 393.8340