ParaJEPA-01-24-2025

ParaJEPA: Joint-Embedding Predictive Architecture for Text Style Transfer

This model uses a JEPA architecture to disentangle Content and Style representations from text.

Usage

To use this model, you need to download the code files along with the weights.

import torch from para_jepa_train import ParaJEPA from config import Config

1. Initialize Model Architecture

config = Config() model = ParaJEPA( model_name=config.model_name, hidden_dim=config.hidden_dim, pred_depth=config.pred_depth )

2. Load Weights

(Download para_jepa_best_model.pt manually or via hf_hub_download)

model.load_state_dict(torch.load("para_jepa_best_model.pt")) model.eval()## Architecture

• Backbone: RoBERTa-base
• Method: Predictive Bottleneck (JEPA)
• Objective: Disentangle Style (Paraphrase) from Content (Meaning)