README.md

---
license: cc-by-sa-4.0
language:
- en
---
# Dataset Card

## Dataset Details
This dataset contains a set of candidate documents for second-stage re-ranking on msmarco 
(dev, test split in [BEIR](https://huggingface.co/BeIR)). Those candidate documents are composed of hard negatives mined from 
[gtr-t5-xl](https://huggingface.co/sentence-transformers/gtr-t5-xl) as Stage 1 ranker
and ground-truth documents that are known to be relevant to the query. This is a release from our paper 
[Policy-Gradient Training of Language Models for Ranking](https://gao-g.github.io/), so 
please cite it if using this dataset.

## Direct Use
You can load the dataset by:
```python
from datasets import load_dataset
dataset = load_dataset("NeuralPGRank/msmarco-hard-negatives")
```
Each example is an dictionary:
```python
>>> python dataset['test'][0]
{
    "qid" : ...,  # query ID
    "topk" : {
        doc ID: ..., # document ID as the key; None or a score as the value
        doc ID: ...,
        ...
    },
}
```


## Citation 
```
@inproceedings{Gao2023PolicyGradientTO,
  title={Policy-Gradient Training of Language Models for Ranking},
  author={Ge Gao and Jonathan D. Chang and Claire Cardie and Kiant{\'e} Brantley and Thorsten Joachims},
  booktitle={Conference on Neural Information Processing Systems (Foundation Models for Decising Making Workshop)},
  year={2023},
  url={https://arxiv.org/pdf/2310.04407}
}
```


## Dataset Card Author and Contact
[Ge Gao](https://gao-g.github.io/)