README.md

---
pipeline_tag: sentence-similarity
tags:
- feature-extraction
- sentence-similarity
- transformers
---

# CoT-MAE MS-Marco Passage Retriever

CoT-MAE is a transformers based Mask Auto-Encoder pretraining architecture designed for Dense Passage Retrieval.
**CoT-MAE MS-Marco Passage Retriever** is a retriever trained with BM25 hard negatives and CoT-MAE retriever mined MS-Marco hard negatives using [Tevatron](github.com/texttron/tevatron) toolkit. Specifically, we trained a stage-one retriever using BM25 HN, using stage-one retriever to mine HN, then trained a stage-two retriever using both BM25 HN & stage-one retriever mined hn. The release is the stage-two retriever.

Details can be found in our paper and codes.

Paper: [ConTextual Mask Auto-Encoder for Dense Passage Retrieval](https://arxiv.org/abs/2208.07670).

Code: [caskcsg/ir/cotmae](https://github.com/caskcsg/ir/tree/main/cotmae)

## Scores
### MS-Marco Passage full-ranking

| MRR @10  | recall@1 | recall@50 | recall@1k | QueriesRanked  |
|----------|----------|-----------|-----------|----------------|
| 0.394431 | 0.265903 | 0.870344  | 0.986676  | 6980           |

## Citations
If you find our work useful, please cite our paper.
```bibtex
@misc{https://doi.org/10.48550/arxiv.2208.07670,
  doi = {10.48550/ARXIV.2208.07670},
  url = {https://arxiv.org/abs/2208.07670},
  author = {Wu, Xing and Ma, Guangyuan and Lin, Meng and Lin, Zijia and Wang, Zhongyuan and Hu, Songlin},
  keywords = {Computation and Language (cs.CL), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {ConTextual Mask Auto-Encoder for Dense Passage Retrieval},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}
```