Ishaank18/hsdn-retriever

HSDN - Hierarchical Semantic Decomposition Network

A hierarchical retriever that decomposes embeddings into multi-resolution subspaces:

• Level 1 (64D): Domain-level semantics
• Level 2 (128D): Subdomain-level semantics
• Level 3 (576D): Instance-level semantics

Usage

import torch
from sentence_transformers import SentenceTransformer

# Load base encoder
encoder = SentenceTransformer('BAAI/bge-base-en-v1.5')

# Load HSDN weights
from huggingface_hub import hf_hub_download
weights_path = hf_hub_download(repo_id="Ishaank18/hsdn-retriever", filename="best_model.pt")
state_dict = torch.load(weights_path, map_location='cpu')

# See cmrl_complete.py for full HSDNRetriever class

Architecture

• Base encoder: BAAI/bge-base-en-v1.5
• Level 1 (Domain): 64D
• Level 2 (Subdomain): 128D
• Level 3 (Instance): 576D
• Total: 768D

Training

Trained with hierarchical contrastive loss + orthogonality constraints.

Part of CMRL Project

This retriever is part of the Contextual Multi-Resolution Learning (CMRL) system for adaptive STEM tutoring.