At its core it uses an ELECTRA-Base model (google/electra-base-discriminator) fine-tuned on the MS MARCO passage classification task using the Sim-Pair marking strategy that highlights exact term matches between the query and the passage via marker tokens (#). It can be loaded using the TF/AutoModelForSequenceClassification classes but it follows the same classification layer defined for BERT similarly to the TFElectraRelevanceHead in the Capreolus BERT-MaxP implementation.
Refer to our
github repository for a usage example for ad hoc ranking.
