hbliu-24/RNA-SE-checkpoints

RNA-SE: pretrained inference weights

This model repository provides four pretrained inference checkpoints for RNA-SE, an RNA secondary-structure model combining a variational autoencoder (VAE) with a diffusion transformer (DiT).

Download: checkpoints.zip

Model definitions, configurations and usage instructions: RNA-StructEnsemble on GitHub

Contents

File inside the archive	Model
bprna_1m/bp_vae.pt	bpRNA ContactVAE
bprna_1m/bp_dit.pt	bpRNA DiT
eternabench_cm/CM_vae.pt	EternaBench-CM ContactVAE
eternabench_cm/dit.pt	EternaBench-CM DiT

VAE exports retain model parameters. DiT exports also retain inference configuration, exponential moving average (EMA) state and latent normalization information. These are not complete training-resumption checkpoints and cannot restore the full training state.

Local layout and use

Place the archive contents in the GitHub checkout's checkpoints/ directory. The resulting files are:

checkpoints/
├── bprna_1m/
│   ├── bp_vae.pt
│   └── bp_dit.pt
└── eternabench_cm/
    ├── CM_vae.pt
    └── dit.pt

The archive already contains the dataset-specific directories. Preserve the filenames and use the corresponding VAE/DiT pair for each dataset. Data and RNA-FM features are distributed separately in bpRNA_data.zip and CM_data.zip.

Use the linked GitHub code for sampling and evaluation. This repository does not provide automatic loading through transformers.from_pretrained or a hosted inference service.