This is a cross-encoder model with dot-product based scoring mechanism trained on MS-MARCO dataset.
The parameters of the cross-encoder are initialized using a 6-layer
minilm model
and is trained via distillation using scores from three different teacher models --
model 1,
model 2, and
model 3.
This model is used in experiments of our
EMNLP 2023 and
ICLR 2024 papers.
See our EMNLP 2022 paper titled "Efficient Nearest Neighbor Search for Cross-Encoder Models using Matrix Factorization" for more details on the dot-product based scoring mechanism.
