ThermoRNA data and model artifacts

This repository provides the datasets, pretrained parameters, and trained model checkpoints associated with ThermoRNA, a physics-inspired deep learning framework for hierarchical RNA structure prediction. These artifacts complement the ThermoRNA source-code repository and contain all large-scale resources required for model reproduction, evaluation, and further development. The data/ directory contains all datasets required for ThermoRNA training and evaluation. The structure/ subdirectory provides the training dataset and four independent benchmark datasets for RNA secondary-structure prediction. The contact/ subdirectory contains datasets used for RNA contact-map prediction, including training, testing, W19, and J6 benchmark sets. The pretrained/ directory includes the pretrained RNA language model parameters (RNA-FM_pretrained.pth), which provide sequence-level representations for ThermoRNA initialization and feature extraction. The weights/ directory contains trained ThermoRNA model checkpoints. The structure/ subdirectory includes the general model checkpoint and benchmark-specific checkpoints for secondary-structure prediction, while the contact/ subdirectory provides the five-model ensemble checkpoints used for RNA contact-map prediction. The ARTIFACT_MANIFEST.sha256 file provides SHA-256 checksums for all released artifacts, enabling users to verify file integrity after downloading and installation. These resources enable researchers to reproduce ThermoRNA experiments, evaluate the released models on benchmark datasets, and develop new physics-inspired approaches for RNA structure and function prediction.