DRAGON+ is a BERT-base sized dense retriever initialized from RetroMAE 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 |
前往AI网址导航
正文完
