fin-ai-lab/lejepa-demo-ecoder

LeJEPA demo encoder

This repository contains a pretrained LeJEPA encoder for the Market-JEPA finance tutorial. It produces a 384-dimensional embedding from normalized market time-series inputs.

Architecture

• Transformer backbone with 12 layers, 6 attention heads, and hidden size 384
• Patch size: 8
• Maximum sequence length: 2,048
• Expected input channels: 20 (9 market-series channels plus 11 information channels in the training setup)
• Pooling: CLS token

The complete architecture settings are in config.json. model.pt is a PyTorch state dict whose keys include the backbone. prefix as well as the LeJEPA training projection head. The tutorial constructs the Market-JEPA model from the config and loads these weights.

Files

• model.pt: pretrained PyTorch state dict
• config.json: compact model configuration
• train_meta.json: full training configuration and run metadata
• xs_ic.json: downstream cross-sectional evaluation metrics

Training data and intended use

The model was trained on 1 Hz U.S. market data from March 2016 using LeJEPA self-supervised representation learning with time-warp augmentation. It is published for educational and research demonstrations of financial time-series representation learning.

This model is not a trading system and its outputs are not financial advice. Historical evaluation does not establish future performance.