mlx7-two-tower-retrieval

This repository contains models trained using the Two-Tower (Dual Encoder) architecture for document retrieval.

Model Description

The Two-Tower model is a dual encoder neural network architecture designed for semantic search and document retrieval. It consists of two separate "towers" - one for encoding queries and one for encoding documents - that map text to dense vector representations in a shared embedding space.

Usage

from twotower import load_model_from_hub
from twotower.encoders import TwoTower
from twotower.tokenisers import CharTokeniser

# Load the model
model, tokenizer, config = load_model_from_hub(
    repo_id="mlx7-two-tower-retrieval",
    model_class=TwoTower,
    tokenizer_class=CharTokeniser
)

# Use for document embedding
doc_ids = tokenizer.encode("This is a document")
doc_embedding = model.encode_document(doc_ids)

# Use for query embedding
query_ids = tokenizer.encode("This is a query")
query_embedding = model.encode_query(query_ids)

Training

This model was trained on the MS MARCO dataset using the Two-Tower architecture with contrastive learning.

Repository Information

This model is part of the Two-Tower Retrieval Model project.