This library provides a simple way for testing the InRanker models, which are also directly accessible at Hugging Face
This library was tested using python 3.10 and can be installed using pip.
pip install inranker The library provides a simple interface to rank documents given a query. The following example shows how to rank documents using the InRanker-small model. There are three sizes available:
from inranker import T5Ranker
model = T5Ranker(model_name_or_path="unicamp-dl/InRanker-small")
docs = [
"The capital of France is Paris",
"Learn deep learning with InRanker and transformers"
]
scores = model.get_scores(
query="What is the best way to learn deep learning?",
docs=docs
)
# Scores are sorted in descending order (most relevant to least)
# scores -> [0, 1]
sorted_scores = sorted(zip(scores, docs), key=lambda x: x[0], reverse=True)
""" InRanker-small:
sorted_scores = [
(0.4844, 'Learn deep learning with InRanker and transformers'),
(7.83e-06, 'The capital of France is Paris')
]
"""@misc{laitz2024inranker,
title={InRanker: Distilled Rankers for Zero-shot Information Retrieval},
author={Thiago Laitz and Konstantinos Papakostas and Roberto Lotufo and Rodrigo Nogueira},
year={2024},
eprint={2401.06910},
archivePrefix={arXiv},
primaryClass={cs.IR}
}