metadata

license: apache-2.0
pipeline_tag: text-ranking
library_name: lightning-ir
base_model:
  - google-bert/bert-base-uncased
tags:
  - bi-encoder

Lightning IR BERT Bi-Encoder

This model is a BERT-based bi-encoder[^1] model fine-tuned using Lightning IR.

See the Lightning IR Model Zoo for a comparison with other models.

Reproduction

To reproduce the model training, install Lightning IR and run the following command using the fine-tune.yaml configuration file:

lightning-ir fit --config fine-tune.yaml

To index MS~MARCO passages, use the following command and the index.yaml configuration file:

lightning-ir index --config index.yaml

After indexing, to evaluate the model on TREC Deep Learning 2019 and 2020, use the following command and the search.yaml configuration file:

lightning-ir search --config search.yaml