README.md

metadata

license: mit
base_model:
  - google-bert/bert-base-uncased

Dragon Context Encoder

This is the context encoder of the Dragon dual-encoder retrieval model, trained for dense passage retrieval tasks.
It should be used together with the corresponding Dragon Query Encoder.

Model Architecture

• Base model: bert-base-uncased

• Architecture: Dense Passage Retriever (DPR) dual-encoder

• Encoder type: Context encoder (for passages)

• Pooling method: CLS pooling (take [CLS] token representation)

• Checkpoint origin:
  The weights were converted from the official facebookresearch/dpr-scale Dragon implementation,
  specifically from the checkpoint provided at:
  https://dl.fbaipublicfiles.com/dragon/checkpoints/DRAGON/checkpoint_best.ckpt

Usage Example

from transformers import AutoTokenizer, AutoModel
import torch

# Load query encoder
q_tokenizer = AutoTokenizer.from_pretrained("liyongkang/dragon-query-encoder")
q_model = AutoModel.from_pretrained("liyongkang/dragon-query-encoder")

# Load context encoder
p_tokenizer = AutoTokenizer.from_pretrained("liyongkang/dragon-context-encoder")
p_model = AutoModel.from_pretrained("liyongkang/dragon-context-encoder")

query = "What is Dragon in NLP?"
passage = "A dual-encoder retrieval model for dense passage retrieval."


# Tokenize. In fact, the two tokenizers are the same.
q_inputs = q_tokenizer(query, return_tensors="pt", truncation=True, padding=True)
p_inputs = p_tokenizer(passage, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    q_vec = q_model(**q_inputs).last_hidden_state[:, 0]  # CLS pooling
    p_vec = p_model(**p_inputs).last_hidden_state[:, 0]  # CLS pooling
    score = (q_vec * p_vec).sum(dim=-1)
    print("Dot product similarity:", score.item())