microsoft / ar2

AR2 Project

This repo provides the code of AR2. In the paper, we propose a new Adversarial Retriever-Ranker (AR2) framework, which constructs a unified minimax game for training the retriever and ranker models interactively.

This repo is still developing, feel free to report bugs and we will fix them ~

Starting with AR2, we developed a series of Text-Retrival methods.

News

• CodeRetriever: Unimodal and Bimodal Contrastive Learning, Xiaonan Li, Yeyun Gong, Yelong Shen, Xipeng Qiu, Hang Zhang, Bolun Yao, Weizhen Qi, Daxin Jiang, Weizhu Chen, Nan Duan arXiv, Code Paper
• Distill-VQ: Learning Retrieval Oriented Vector Quantization By Distilling Knowledge from Dense Embeddings, Shitao Xiao, Zheng Liu, Weihao Han, Jianjin Zhang, Defu Lian, Yeyun Gong, Qi Chen, Fan Yang, Hao Sun, Yingxia Shao, Denvy Deng, Qi Zhang, Xing Xie, SIGIR 2022, Code Paper
• Adversarial Retriever-Ranker for Dense Text Retrieval, Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, Weizhu Chen, ICLR 2022, Code Paper

How to Cite

If you extend or use this work, please cite the paper where it was introduced:

@inproceedings{zhang2022ar2,
  title={Adversarial Retriever-Ranker model for Dense Retrieval},
  author={Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, Weizhu Chen},
  booktitle={ICLR},
  year={2022}
}
@inproceedings{xiao2022distillvq,
  title={Distill-VQ: Learning Retrieval Oriented Vector Quantization By Distilling Knowledge from Dense Embeddings},
  author={Shitao Xiao, Zheng Liu, Weihao Han, Jianjin Zhang, Defu Lian, Yeyun Gong, Qi Chen, Fan Yang, Hao Sun, Yingxia Shao, Denvy Deng, Qi Zhang, Xing Xie},
  booktitle={SIGIR},
  year={2022}
}
@inproceedings{li2022coderetriever,
  title={CodeRetriever: Unimodal and Bimodal Contrastive Learning},
  author={Xiaonan Li, Yeyun Gong, Yelong Shen, Xipeng Qiu, Hang Zhang, Bolun Yao, Weizhen Qi, Daxin Jiang, Weizhu Chen, Nan Duan},
  booktitle={arXiv},
  year={2022}
}

ar2's Issues

Reproduced result is lower than the paper

Hi, thank you for the interesting work and code!

I am trying to reproduce the experiment, but I find that the results are slightly worse than the paper, for example, in the paper, the final top-5/20/100 results are: 78.2/84.4/87.9, but my reproduced number are 'top5': 0.7719437814903208, 'top20': 0.8365597100680633, 'top100': 0.8749226553522497. Is that normal or should I change some setting?

Thanks!

Checkpoint for AR2-G models

Thank you for sharing the retriever AR2-G checkpoints for NQ and TriviaQA. Would it be possible to share the reranker AR2-D checkpoints?

CodeRetriever: Code, model checkpoint, and constructed datasets

Hi there,

I have read your paper "CodeRetriever: Unimodal and Bimodal Contrastive Learning for Code Search". This is excellent work and I am interested in the ideas in the paper. In your abstract, you said you would make the code, model checkpoint, and constructed datasets publicly available. However, when I look into the folder of CodeRetriever, there is only a ReadMe. Could you please release the codes mentioned in your abstract? Thank you. :)

Model weights

Hi,
congrats for the great work! I was wondering: when do you plan to make model weights available?

Thanks,
Thibault

codeRetriever: paper question

Dear authors,
thanks for very interesting work!
I have one short question.
What contrastive learning scheme did you use to pre-train codeRetriever model (not fine-tune)? was it in-batch negatives or AR2 or something else? It’s unclear for me

CKPT file download problem

I use command wget https://msraprophetnet.blob.core.windows.net/ar2/realease_ckpt/wiki_ict.pt, but it reports below info and stuck:
--2023-06-14 04:03:54-- https://msraprophetnet.blob.core.windows.net/ar2/realease_ckpt/nq_fintinue.pkl Resolving msraprophetnet.blob.core.windows.net (msraprophetnet.blob.core.windows.net)... 20.60.20.36 Connecting to msraprophetnet.blob.core.windows.net (msraprophetnet.blob.core.windows.net)|20.60.20.36|:443... connected.
I click the CKPT link and it shows

Same way to download data file works well. Is it my network problem? Thx in advance.

how to select hard negative samples in CodeRetriever paper

How to pick “hard” representative negative samples?

Retriever ckpt on MSMARCO

Hi, congrats for the great work!
Can you release the retriever ckpt after AR2 training on MSMARCO?

Thank you.

AR2/tree/ main/SCodeR not found

Can i get this SCodeR code?Thank you!