Checkpoints reference

This page lists the model-weight checkpoints RFD4-Proteina loads, what each one is for, and the released files available as of the last update of this documentation resource. For installing and downloading them, see Installation. For the CLI flags that select them at generation time, see Inputs and checkpoints in the CLI options reference.

Note

“Checkpoint” has a second, unrelated meaning in this codebase: a search checkpoint (search.checkpoints) is a point in the generation schedule where beam search (see the glossary) branches and scores candidates. It has nothing to do with the model weight files on this page.

Contents

• Types of checkpoints

• Released checkpoints

• Downloading

• Overriding at the CLI

• Training-time checkpoint roles

Types of checkpoints

RFD4-Proteina always loads two checkpoints together to generate: a denoiser and the autoencoder it was trained against. Two more are optional, depending on the workflow.

Role	Config key(s)	Default source	Used for
Denoiser (flow-matching latent diffusion model)	hf_ckpt_path; native: ckpt_name, ckpt_path	Hugging Face bundle; native files in $DATA_PATH/flow_checkpoints/	Generation, search, and standalone evaluation; always required
Autoencoder (frozen local-latents encoder/decoder)	hf_ckpt_path; native: autoencoder_ckpt_path	Same bundle; native files in $DATA_PATH/ae_checkpoints/	Loaded alongside the denoiser for generation; also loaded when training a new denoiser, since it fixes the latent space the denoiser is trained in
PEFT adapter (optional)	adapter_path	user-supplied directory	Inference only; a local fine-tuning adapter, loaded onto the exact base denoiser/autoencoder pair it was fine-tuned against
OpenFold3 refolding oracle (optional)	AF3_IMAGE, AF3_MODEL_DIR, AF3_JAX_CACHE_DIR	$DATA_PATH/openfold3/	Refolding/designability metrics only (metrics/refolding_oracles=of3); not a generation checkpoint

configs/inference.yaml is the runtime source of truth for “the SOTA model”: generate.py loads the Hub repository hf_ckpt_path at the pinned commit hf_ckpt_revision from it by default (the bundle holds the denoiser and the autoencoder), and the GPU regression tests read it too. The native ckpt_name and autoencoder_ckpt_path keys name the same release as local files and apply when you set hf_ckpt_path=null. Do not hard-code checkpoint filenames or revisions anywhere else; update them there when advancing the release.

Released checkpoints

The current release pair, as pinned in configs/inference.yaml:

File	Role	Training recipe	What it’s for
rfd4-proteina-flow-release-2026-10-19.ckpt	Denoiser	configs/releases/flow/2026_10_19_hero_pretrain.yaml, fine-tuned with configs/releases/flow/2026_10_19_hero_lora_c384.yaml (the 60k EMA checkpoint)	Current SOTA denoiser. Supports unconditional generation, motif scaffolding, enzyme design, small molecule/protein/DNA binder design, folding, and inverse folding. Accepts generalized sequence-probability conditioning and conditions the autoencoder on its own sequence readout by default during inference (generation.args.condition_decoder_on_sequence_prediction).
rfd4-proteina-ae-release-2026-10-19.ckpt	Autoencoder	configs/releases/ae/2026_10_15_unified_biomolecule.yaml	Unified protein, RNA, DNA and modified-residue autoencoder with 32 latent channels per residue, trained to 400k steps. Fixes the latent space the released denoiser was trained in; use this pair together.

Both files hold exponential-moving-average weights, not raw optimizer checkpoints. The configs/releases/flow/ and configs/releases/ae/ recipes are training configs, kept for reproducibility: they document how a released checkpoint was produced, not alternative downloads. Only the pair in configs/inference.yaml is published.

Downloading

make setup caches the released model from the Hugging Face repository and revision named in configs/inference.yaml (nvidia/RFD4-Proteina by default) in the normal Hugging Face cache, where HF_HOME controls its location. Rerunning setup reuses files already in the cache; a failed download stops setup with an error.

To download manually (after a pip install or bash setup.sh --skip-weights):

./scripts/download_assets.sh ./data_path --skip-of3

Pass your configured DATA_PATH if you changed it from the default. Details:

• Everything comes from the Hugging Face model repository: the model bundle (config.json and model.safetensors), the test_dataset tarball, the OpenFold3 runtime image and the OpenFold3 weights. For private repositories, provide HF_TOKEN or log in with hf auth login.

• Add --with-test-dataset to also fetch a tiny training tutorial dataset.

• OpenFold3 is not fetched by default: pass neither --skip-of3 nor an already-configured AF3_IMAGE, and run on a host with Apptainer installed. See Installation for the follow-up AF3_IMAGE/AF3_MODEL_DIR/AF3_JAX_CACHE_DIR environment settings.

• Select another Hub repository or a local exported bundle with hf_ckpt_path, and pin it with hf_ckpt_revision.

Overriding at the CLI

Every generation argument is key=value, Hydra style:

rfd4-generate ckpt_path=<dir> ckpt_name=<file.ckpt> ...

See Inputs and checkpoints in the CLI options reference for ckpt_path, ckpt_name, autoencoder_ckpt_path, and adapter_path, with their defaults and an example of each.

Training-time checkpoint roles

Training (rfd4-train, rfd4-train-ae) uses two mutually exclusive checkpoint-loading paths:

Config key	Role
pretrain_ckpt_path	Start a new run from saved weights only — resets optimizer, scheduler, step, epoch, and RNG state.
resume_ckpt_path	Resume an interrupted run — restores full Lightning trainer state.

Two policies, honored only on the pretrain_ckpt_path path, control how those weights are applied when the architecture has changed since the checkpoint was produced:

• weight_loading (rfproteina.utils.weights.WeightLoadingConfig) — when unset, the checkpoint loads with strict=True and any key or shape mismatch is a hard error. Set a default_policy (for example copy), a fallback_policy (for example reinit), and per-parameter param_policies (zero_init, copy_and_zero_pad, …) to deliberately load across an architecture change.

• parameter_freezing (rfproteina.utils.weights.ParameterFreezingConfig) — freezes selected parameters by name pattern after loading, for example to freeze a pretrained backbone during fine-tuning.